Future Internet doi: 10.3390/fi18060309

Authors: Ivona Plamenova Velkova Valentin Stefanov Kisimov

Reliable public-sector labour-market forecasting requires models that can be updated as data sources, AI tools, and labour-market signals evolve. This paper proposes a provider-independent multi-agent framework for dynamic predictive evaluation of national and regional labour markets in Bulgaria. Implemented as a Model Context Protocol (MCP) server, the system coordinates specialised agents for data ingestion, preprocessing, semantic extraction, AI-adjusted transformation modelling, automated model evaluation, and reporting through stable input–output contracts. The empirical application integrates Bulgarian Employment Agency administrative registered-unemployment indicators, Eurostat labour-market data, World Bank macroeconomic data, and textual, audio, and video evidence on AI, skills, and employment change. The analysis covers the period 2015–2030. Observed official data are used up to 2025 for model construction and validation, while the 2026–2030 values are reported only as forecast and scenario projections. For youth unemployment among persons aged 24 years or younger, the semantic-enhanced model achieves the best predictive accuracy (RMSE = 0.2033; MAE = 0.1457), representing a small improvement over the structured baseline (RMSE = 0.2057; MAE = 0.1462) and a substantial RMSE reduction relative to the persistence benchmark (RMSE = 0.4750; MAE = 0.2891). The AI-adjusted coefficient does not reduce holdout error relative to the semantic-enhanced model, but provides an explicit and sensitivity-tested mechanism for regional scenario interpretation. Regional forecasts indicate persistent spatial inequality, with the Northwest remaining the highest-risk region and the Southwest the lowest-risk region.

Future Internet doi: 10.3390/fi18060308

Authors: Khaled Chahine Hassan N. Noura

Industrial control systems (ICS) are increasingly targeted by cyberattacks that manipulate physical processes while evading data-driven detectors trained on raw time-series data. This paper extracts 34–41 control-theoretic features, including tracking error, valve mismatch, sensor liveness, and their temporal derivatives, from Proportional–Integral–Derivative (PID) control loops and evaluates them using an Isolation Forest combined with a maximum z-score. On HAI 21.03, Stage 1 achieves a PA-F1 score of 0.8945, detecting 48 out of 50 attacks. On HAI 23.05, Stage 1 attains a PA-F1 score of 0.9210, surpassing seven deep-learning baselines by at least 23 PA-F1 points; the closest baseline, a learned Graph Neural Network (GNN), achieves 0.6890. Re-implementations of ConvBiLSTM-AE (PA-F1 = 0.6689) and TranAD (PA-F1 = 0.6838) on the same evaluation split confirm this performance gap. A controlled USAD experiment, with PA-F1 = 0.7343 for physics features versus 0.6687 for raw Supervisory Control and Data Acquisition (SCADA), demonstrates that the extracted features provide the detection signal independently of the model architecture. Adding a bidirectional Gated Recurrent Unit (GRU) refinement stage improves PA-F1 by 8.1 percentage points on HAI 21.03, but the same stage reduces it by 6.8 percentage points on HAI 23.05, where attacks manifest as brief perturbations; four alternative Stage 2 designs reproduce this degradation. We therefore characterize temporal refinement as beneficial only for sustained-deviation attacks and identify Stage 1 as the primary deployable detector. This study is the first to apply physics-informed features, report both PA-F1 and eTaPR on HAI 23.05, and perform per-window error diagnosis on this dataset. Results show that 10 of 15 detected windows are covered by fewer than 10% of their timesteps, revealing a structural tension between PA-F1 and eTaPR.

Future Internet doi: 10.3390/fi18060307

Authors: Francesco Pilotti Aurora Pavone Lia Di Sabatino Farinelli Simone Tinelli Gaetanino Paolone

The rapid evolution of Smart Homes and Smart Buildings is driven by the transition from the Internet of Things (IoT) to the Artificial Intelligence of Things (AIoT). Within this scenario, Security Systems are particularly critical and data-intensive systems. Despite extensive research, a high-level synthesis focusing exclusively on the synergy between AIoT and Security Systems in Smart Home and Smart Building application domains is still lacking. To bridge this gap, this paper presents a systematic Tertiary Study (TS) following a well-known research protocol. 13 Secondary Studies (SSs) were synthesized and discussed from an initial pool of 139 publications (years 2024–2025). Findings reveal that monitoring is the most addressed system, followed by security and alarm, while surveillance and access control remain comparatively underexplored. Moreover, results highlight a definitive shift toward Edge and Fog computing to meet latency and privacy requirements, whereas Deep Learning and Ensemble Learning techniques predominate for anomaly detection and predictive maintenance. This study identifies open challenges and future research directions, providing a foundational roadmap for resilient, cognitive-driven security infrastructures in smart environments.

Future Internet doi: 10.3390/fi18060306

Authors: Yerlan Tursynbek Nurtay Albanbay Djamel Djenouri Shahid Latif Ainur Akhmediyarova Zhibek Alibiyeva Janna Alimkulova Dina Oralbekova

Federated learning (FL) enables distributed model training in IoT environments while keeping raw data on local devices. However, protecting model-update exchange is difficult on microcontroller-class devices due to strict latency, memory, and energy constraints. Existing studies often evaluate lightweight cryptography outside complete FL pipelines or on more powerful hardware, leaving its practical overhead on MCU-class devices insufficiently explored. This paper presents an end-to-end, hardware-validated secure framework for exchanging model updates in federated learning on resource-constrained IoT microcontrollers. Implemented on ESP32-based edge devices, the framework combines lightweight block ciphers (SPECK, SIMON, and PRESENT), HMAC-SHA256 for integrity verification, and ECDH-HKDF for session-key establishment. The evaluation assessed latency, throughput, RAM/ROM footprint, and energy consumption. Results show that SPECK provides the lowest overhead (0.13 µs/byte, 8.68 MB/s, 138.3 mJ), SIMON offers intermediate performance (0.41 µs/byte, 1.96 MB/s, 184.9 mJ), and PRESENT incurs the highest computational cost (89.37 µs/byte, 0.011 MB/s, 446.2 mJ). In the CICIoT2023 federated intrusion-detection evaluation, the secure model maintained stable convergence and achieved 85.43% accuracy after 20 rounds, remaining close to the centralized baseline. These findings demonstrate the practical feasibility of secure model-update exchange in FL on real IoT microcontrollers and provide hardware-grounded guidance for cipher selection under tight resource budgets.

Future Internet doi: 10.3390/fi18060305

Authors: Abdalilah Alhalangy Saleh Abdulrahman Alkhamis Eman Abouelkheir

Effective intrusion detection for Internet of Things (IoT) environments requires balancing predictive performance, resource efficiency, and interpretability—particularly in real-world deployments where traffic distributions and attack scenarios vary. While many studies report near-perfect detection on benchmark datasets, this often overlooks model stability under distribution shifts. This paper addresses this gap by introducing a stability-focused evaluation of lightweight, explainable intrusion detection models using compact IoT-23 scenarios and a constrained set of 14 connection-level features for interpretability. Four lightweight models—logistic regression, random forest, XGBoost, and LightGBM—are assessed within a unified pipeline. Beyond standard internal validation, we implement a strict cross-scenario evaluation framework featuring a fully unseen malware capture. Our proposed Internal–External Stability Gap (IESG) framework, enhanced with normalized and multi-metric measures, highlights the degradation in consistency between internal and external metrics. Surprisingly, even models with high internal F1 scores (up to 0.9994) may experience considerable drops in external macro-F1 and specificity, exposing weaknesses in conventional evaluation. Experimentally, LightGBM provides the best trade-off between performance and compactness (606 KB) and shows the smallest stability gap for malicious detection. Nevertheless, all models show reduced balanced performance under scenario shift, underscoring that deployment readiness hinges on stability under changing conditions. Feature ablation reveals that leveraging high-impact features, such as port information, can boost internal accuracy at the expense of generalization. In summary, we demonstrate that while lightweight models deliver strong detection, only those proven stable across scenarios are viable for real-world IoT intrusion detection. Our evaluation framework offers a practical, interpretable tool for assessing model robustness.

Future Internet doi: 10.3390/fi18060304

Authors: Lassaad Ben Ammar Altahir Saad Ahod Alghuried

Knee Osteoarthritis (KOA) is a leading contributor to global physical disability, where delayed diagnosis often results in irreversible joint damage and socio-economic cost. Early diagnosis remains challenging due to subtle radiographic biomarkers and limited access to specialized expertise, particularly in distributed healthcare settings. Within the evolving landscape of the Future Internet, characterized by Internet of Medical Things (IoMT), edge–cloud computing, and intelligent digital health infrastructures, there is an increasing demand for scalable, low-latency, and explainable AI-driven diagnostic solutions. In this work, we propose a Dual-Stream Wavelet Fusion Network (DS-WFN) alongside a distributed edge-cloud architectural roadmap tailored for deployment in distributed and edge-enabled healthcare ecosystems. The framework integrates a spatial morphological stream with a spectral wavelet stream, augmented by an Adaptive Wavelet Selection Mechanism (AWSM). The AWSM dynamically selects optimal frequency bases (Haar, Symlet, Daubechies) to preserve fine-grained diagnostic features typically lost in conventional CNN architectures. An Adaptive Spatial Alignment (ASA) module further ensures efficient fusion of heterogeneous representations, enabling robust feature integration across computational nodes. Experimental results across a five-fold patient-isolated cross-validation protocol demonstrate that the DS-WFN achieves a mean classification accuracy of 76.3% (95% CI: 71.6–80.8%) and a macro-averaged F1-score of 0.747 (95% CI: 0.697–0.795), consistently outperforming single-stream baselines while preventing patient-level data leakage. Furthermore, Grad-CAM visualizations provide interpretable outputs aligned with clinical diagnostic criteria, supporting trustworthy AI integration into digital healthcare workflows. Furthermore, we disclose a methodological framework for edge-based implementation, highlighting how localized inference ensures data sovereignty and real-time clinical support. By combining multiscale signal processing with deep learning under a Future Internet paradigm, this work contributes a scalable, explainable, and edge-ready diagnostic framework for early KOA detection, enabling intelligent, connected, and resource-efficient healthcare services.

Future Internet doi: 10.3390/fi18060303

Authors: Adolfo A. Jurado Rosas Marina Fernández Miranda Gladys L. Peña Pazos Elberth E. García Panta Carlos A. Ramos Reyes Milagros P. Córdova de Chang José H. Chang Valdiviezo Olga P. Gamarra Chirinos Carlos E. Esquerre Aguirre

This study analyzes the transition of the Artificial Intelligence of Things (AIoT) toward a Human-Centered Artificial Intelligence (HCAI) approach. Following PRISMA 2020 guidelines, a Systematic Literature Review was conducted on 1 April 2026, retrieving literature from Scopus, Web of Science, SciELO, and Springer Nature Link. The inclusion criteria prioritized open-access, peer-reviewed English articles published between 2020 and 2025 that addressed AIoT architectures and explainability mechanisms. The screening procedure involved a dual independent review process, followed by a rigorous methodological quality assessment to minimize the risk of bias, culminating in a final sample of 40 studies from an initial pool of 971 records. The findings reveal a structural paradox: while intelligent systems achieve greater operational autonomy, legal and moral accountability remains inexorably bound to the human operator. Furthermore, 77.5% of the evaluated implementations employ superficial explainability, functioning merely as a psychological buffer to manage automation anxiety rather than providing a genuine interactive control mechanism. It is concluded that programming based on HCAI principles must shift from a post hoc feature to an inherent architectural requirement. Establishing explainability by design is imperative to guarantee an interactive audit capability that comprehensively safeguards operational integrity and preserves human agency, although the exclusive reliance on open-access literature limits visibility into proprietary commercial models.

Future Internet doi: 10.3390/fi18060302

Authors: Amir Anees Ding Ming Gnana Bharathy Lois Holloway

With the growing need for collaborative machine learning across institutions holding sensitive data, ensuring data privacy without compromising model performance has become an important challenge. This work introduces secure federated learning algorithms that use encryption and masking techniques to protect the privacy of data during collaborative model training. Three federated learning algorithms were developed: one for vertical federated learning and two combining horizontal and vertical data partitioning. The proposed algorithms are designed such that participating clients communicate only with the server, even when data exchange between clients is required. This exchange occurs through the server with the help of encryption and masking. The performance of the algorithms, evaluated in terms of accuracy and loss, shows competitive results. The accuracy remains unchanged compared to the centralised scenario for the vertical federated learning algorithm and one of the combined federated learning algorithms, and it remains highly competitive with the other combined federated learning algorithm. The privacy analyses conducted as part of this work demonstrate no risk of data leakage ensuring that no party involved can infer sensitive information.

Future Internet doi: 10.3390/fi18060301

Authors: Hari Hara Babu Saripalli Jyothsna Laxmi Saripalli Leonel Lagos Himanshu Upadhyay

Unauthorized commercial drone activity around critical infrastructure motivates the development of trajectory-level anomaly detection. We present a rigorous benchmarking study of variational autoencoder methods for drone trajectory anomaly detection in a simulated nuclear facility protection scenario, evaluating six methods (bidirectional LSTM-VAE, unidirectional LSTM-VAE, fully connected VAE, standard autoencoder, One-Class SVM, Isolation Forest) on 2500 trajectories using identical raw features and training pipelines. Across five random seeds, all VAE variants achieve AUC-ROC of approximately 0.92 versus 0.73 to 0.80 for the non-VAE baselines, isolating variational regularization rather than bidirectionality or temporal encoding alone as the dominant performance driver in this domain. Building on this benchmark, we propose a domain-aware LSTM-VAE incorporating two facility-specific architectural elements: a polar coordinate input representation expressing trajectories relative to the protected facility and a distance-weighted reconstruction loss that allocates model capacity toward near-facility timesteps. The domain-aware variant achieves AUC-ROC of 0.962 ± 0.007 on the original test set and 0.973 ± 0.005 on an augmented hard anomalies test set, a 3 to 4 percentage-point improvement over generic VAE methods at no additional parameter cost. A bootstrap evaluation under 99:1 class imbalance confirms that the domain-aware variant maintains its precision advantage at low false positive rate operating points.

Future Internet doi: 10.3390/fi18060300

Authors: Md Manirul Islam Umme Salsabil Mekhriddin Nurmamatov Sazzad Hossain

Future internet services are expected to increasingly depend on IoT–edge deployments, in which intrusion detection must operate close to constrained, heterogeneous devices rather than only in cloud or data-center environments. Although the literature focuses on many “lightweight” intrusion detection systems (IDSs), the evidence supporting deployability is uneven and often limited to accuracy-oriented benchmark results. This PRISMA-ScR review, which was cross-checked against the PRISMA 2020 reporting items, synthesizes 78 peer-reviewed studies published between January 2017 and March 2026 and evaluates how they report model compactness, data and preprocessing burden, system placement, hardware measurements, operational robustness, and reproducibility. The reviewers independently screened 1162 deduplicated records and charted the included studies. This review found that architectural compactness is commonly reported, whereas target device latency, runtime memory, measured power or energy, zero-day evaluation, time-aware splitting, and device shift validation remain inconsistent. To make these gaps auditable, this study introduces a five-dimensional deployability framework using log-scale normalization, bounded benefit coding, completeness penalties, scorer agreement checks, and scenario-based sensitivity analysis. The results show that no IDS family dominates across all deployment scenarios: rankings change when hardware constraints or operational robustness receive priority. This review concludes with a benchmark blueprint, reporting protocol, completed PRISMA checklist, and research agenda for deployment-grade IoT–edge IDS studies.

Future Internet doi: 10.3390/fi18060299

Authors: Sergiu Weisz Dragoș Petre Andreea-Cătălina Mazilu Virgile Robles Maria-Elena Mihăilescu Vlad-Iulius Năstase Mihai Carabaș Jacek Andrzejewski Dawid Lazaj Andrzej Bartoszek

Recent improvements in hardware and software have enabled a paradigm shift in satellite computing, moving from purpose-built satellites running a single application to platforms capable of executing and even receiving new workloads on orbit. This evolution has allowed image processing to migrate from ground stations to single- or multi-node satellite clusters, with only processed results transmitted, significantly reducing end-to-end latency. This paper proposes ORCHIDE, an orchestration solution built on cloud-native technologies such as Kubernetes and Argo, purpose built for space edge computing. A key capability of ORCHIDE is its support for unikernels—minimal, single-application virtual machines—alongside containers. Compared to traditional containerized deployments, unikernels substantially reduce CPU and memory footprint, achieve short boot times, and produce smaller binary images. ORCHIDE further enables unikernel workloads to leverage heterogeneous accelerator hardware, including FPGAs, through a dedicated accelerator management library. We describe the system architecture, the scheduling model, and the minimum target hardware required for deployment. Three clusters of varying topology were used to evaluate ORCHIDE, demonstrating that it operates effectively on both single- and multi-node heterogeneous configurations. Preliminary results show the ORCHIDE platform being able to run in heterogeneous and single-node environments with as low as 4 cores and 8 GB of memory, offering potential users the flexibility to compose satellite hardware to best match their mission requirements.

Future Internet doi: 10.3390/fi18060298

Authors: M. A. Elsabagh Rana Albelaihi Esraa Hassan

The integration of heterogeneous healthcare data sources remains a major challenge in developing reliable and personalized predictive systems for digital healthcare applications. Traditional machine learning methods perform well on structured clinical data but often fail to effectively exploit multimodal information, while deep learning approaches may suffer from instability, weak generalization, and poor calibration when dealing with limited modalities. To address these limitations, this study proposes HA-PI-MADT, a hybrid adaptive healthcare-informed multimodal digital twin-inspired framework that combines deep multimodal representation learning with ensemble-based predictive modeling for robust and trustworthy healthcare prediction. The proposed framework integrates wearable sensor signals, electronic health records (EHRs), CT/MRI imaging representations, and population-level risk prototypes derived from the UCI diabetes dataset within a unified multimodal healthcare representation architecture. In addition, a modality-aware adaptive fusion mechanism dynamically adjusts the contribution of each modality according to its relevance and data quality, while a hybrid stacking strategy combines deep multimodal embeddings with classical ensemble learners to improve predictive robustness and ranking performance. To enhance clinical trustworthiness, calibration-aware optimization is incorporated to improve probabilistic reliability and uncertainty estimation. Extensive experiments conducted on a multimodal healthcare dataset demonstrate that HA-PI-MADT achieves a balanced performance profile across discrimination, ranking, and calibration-oriented evaluation metrics compared with several unimodal, multimodal, and ensemble baselines. The proposed framework achieves strong ranking-oriented and classification performance, including the highest AUPRC (0.6388) and F1-score (0.6327), while also demonstrating competitive calibration-oriented reliability through lower Brier score and negative log-likelihood values. The results demonstrate the effectiveness of the proposed hybrid adaptive multimodal digital twin-inspired framework for reliable, robust, and clinically trustworthy healthcare prediction.

Future Internet doi: 10.3390/fi18060297

Authors: Yi Ren Feng Luo Yingpeng Tong Yanhua Yu Zeqi Liao Yuezhen Xiao

Driven by the rapid advancement of intelligence and connectivity, traditional distributed and signal-oriented automotive architectures are gradually being replaced by centralized, service-oriented architectures. In response to this transition, In-Vehicle Networks (IVNs) are expected to deliver high bandwidth, hard real-time performance, high reliability, and service-oriented capabilities. Data Distribution Service (DDS) and Time-Sensitive Networking (TSN) provide key technical support from the perspectives of service orientation and quality of service, respectively. Consequently, the integration of DDS and TSN has become a focal point in the field of IVNs. However, existing DDS message scheduling mechanisms cannot eliminate publishing time jitter, which prevents effective integration with deterministic scheduling mechanisms at the TSN layer, particularly the Time-Aware Shaper (TAS). To enable deterministic DDS communication in the DDS over TSN Architecture (DoTA), a Time-Triggered (TT) communication strategy based on message preemption and guard band mechanisms is proposed. This strategy is integrated into the flow controller of the DDS middleware. By scheduling a timed-event table, the publishing time of Time-Sensitive (TS) DDS messages is precisely controlled to align with the TAS mechanism. In addition, a schedulability analysis method is proposed to estimate the Worst-Case End-to-end Delay (WCED) of TS messages in DoTA. Experimental results from a physical testbed demonstrate that the proposed TT strategy can constrain the publishing time deviation of TS messages within 3 μs. When the TT strategy is jointly deployed with the TAS mechanism, both the end-to-end delay and jitter satisfy the requirements of safety-critical in-vehicle applications. Furthermore, the maximum deviation between the experimental results and the WCED estimated from the schedulability analysis is 15.4%. This indicates that the proposed method can effectively validate the feasibility of network designs and provide sufficient safety margins.

Future Internet doi: 10.3390/fi18060296

Authors: Leandro Pazmiño-Ortiz Alan Cuenca-Sánchez Byron Loarte-Cajamarca

Artificial Intelligence of Things (AIoT) is transforming Internet of Things (IoT) systems from cloud-centric data processing into distributed intelligence across device, edge, fog, and cloud tiers. However, existing reviews often emphasize specific computational layers, learning paradigms, or application domains rather than the cross-domain problem of Artificial Intelligence (AI) workload placement under real deployment constraints. This paper presents a structured integrative review of AI workload placement in AIoT, based on a multi-stage literature search, two-stage screening process, and thematic synthesis of 132 sources. The review does not propose a new physical architecture; instead, it develops a terminology-harmonized and AI-centric perspective for assessing where AI functions should reside according to latency, privacy, bandwidth, power, scalability, resilience, and model complexity. Evidence is synthesized across Industrial Internet of Things (IIoT), smart cities, Internet of Medical Things (IoMT), and smart agriculture. The findings show that placement drivers are domain-dependent: deterministic response and reliability dominate IIoT, interoperability and scale shape smart cities, privacy and human oversight constrain IoMT, and energy scarcity and intermittent connectivity define agriculture. The review concludes that robust AIoT requires hybrid multi-layer architectures combining Tiny Machine Learning (TinyML), edge/fog coordination, cloud-scale optimization, and Federated Learning (FL) where appropriate.

Future Internet doi: 10.3390/fi18060295

Authors: Georgi Tsochev Ivo Gergov

Critical infrastructures are increasingly Internet-connected cyber–physical systems whose recovery after cyber incidents must satisfy safety, timing, regulatory, and interdependency constraints. Yet, the use of large language models (LLMs) for generating recovery plans remains fragmented across cybersecurity, industrial control, digital twins, and AI assurance research. This review synthesizes that emerging field through a structured critical survey of studies on LLMs in incident response, OT/ICS resilience, and cyber–physical recovery, with a focused perspective on grounding, trust, and assurance mechanisms relevant to recovery-plan generation. It develops an architecture-centric taxonomy spanning prompt-only assistants, retrieval-augmented copilots, graph-aware planners, multi-agent systems, and hybrid verification/simulation pipelines; maps realistic applications across energy, water, manufacturing, transportation, healthcare, and telecommunications; and organizes limitations into technical, security, governance, and human-factor categories. Based on this synthesis, the paper proposes the Grounded Recovery Planning Stack as a reference architecture and outlines a staged roadmap from human-in-the-loop copilots to bounded orchestration. The main conclusion is that near-term value lies in grounded, auditable, compliance-aware copilots, whereas autonomous recovery execution remains premature without stronger validation, state-aware grounding, sector-specific benchmarks, and formal safeguards.

Future Internet doi: 10.3390/fi18060294

Authors: Mona Scheerer Yogachandran Rahulamathavan

In this work, the newly developed functional bootstrapping (FBT) for the Cheon–Kim–Kim–Song (CKKS) scheme is used for the first time to homomorphically evaluate an encrypted neural network. The advantage of FBT over previous approaches for the homomorphic evaluation of non-linear activation functions is that it combines bootstrapping and homomorphic function evaluation. For this purpose, FBT for CKKS is extended to be applied to real input values by evaluating the first order Hermite interpolation function not only on its interpolation points but on the entire domain [0,1]. For the sigmoid function, to respect the internal representation of negative values in CKKS and the convergence behaviour of trigonometric interpolation, a glueing of shifted and reflected sigmoid functions that is periodic and continuous is used as an input function for FBT. The experimental results yield an accuracy of 97.33% with a relative loss of 0% compared to the Hermite plaintext counterpart that were obtained with a fully connected neural network with 100 hidden neurons on the MNIST test set at a security level of 128 bits. The current implementation required approximately 1.66 s per image (amortised time) and about 201 GB RAM.

Future Internet doi: 10.3390/fi18060293

Authors: Basil Alshehri Nayef Aljhani Ahmed Albalawi Waleed Abdulghani Talal Alfawzan Ahmad J. Alkhodair

This paper presents the design, implementation, and controlled evaluation of a proof-of-concept ear-level wearable system that integrates local artificial intelligence for real-time physiological monitoring in construction safety applications. The proposed architecture combines photoplethysmography (PPG), non-contact infrared thermometry, and nine-axis inertial sensing on a Raspberry Pi Pico microcontroller, enabling local inference that reduces dependence on cloud processing. A lightweight logistic regression model with three binary outputs, trained on a subset of the publicly available WESAD dataset (subjects S2–S4), classifies three physiological states relevant to worker safety—elevated PPG variability, drowsiness, and fatigue—directly from the device’s 2 MB flash memory. The principal contribution is demonstrating that ear-level multi-sensor fusion combined with on-device machine learning achieves high agreement with clustering-derived proxy labels under controlled conditions (average F1-score: 97.80% on an unseen test subject) while sustaining sub-second inference latency (<0.5 s). These results support timely supervisor alerting and motivate subsequent field validation in operational construction environments.

Future Internet doi: 10.3390/fi18060292

Authors: Zhengze Liu Nianmin Yao Shengyuan Bai Qibin Li

In highly dynamic vehicular ad hoc networks (VANETs), vehicles frequently move across the coverage areas of multiple roadside units (RSUs), making secure and efficient continuous vehicle-to-infrastructure access essential. However, repeated full authentication and key agreement for each new RSU access impose considerable computational and communication overhead. This paper proposes a state-assisted privacy-preserving mutual authentication and key agreement scheme for lightweight multi-RSU access in VANETs. The proposed scheme consists of initial and subsequent authentication phases. In the initial phase, elliptic curve cryptography (ECC) is used to achieve anonymous mutual authentication and session key establishment between vehicles and RSUs. In the subsequent authentication phase, a vehicle leverages follow-up authentication state securely forwarded by the previous RSU to complete fast authentication with a neighboring RSU using only hash and XOR operations. In addition, physically unclonable functions (PUFs) are deployed on both vehicles and RSUs to protect critical secrets. Security analysis shows that the proposed scheme achieves mutual authentication, anonymity preservation, and resistance to common attacks. Performance evaluation shows that it reduces the computational cost of subsequent authentication by more than 90% while maintaining low communication overhead.

Future Internet doi: 10.3390/fi18060291

Authors: Florin Dragomir Otilia Elena Dragomir

The rapid global expansion of photovoltaic capacity, now exceeding 1 TW, has transformed solar power forecasting from an engineering problem into a financially critical investment decision. Yet virtually all published forecasting studies optimise statistical accuracy metrics without translating improvements into monetised operational value. This paper introduces a unified cost–benefit framework that maps forecast errors across three operationally distinct time horizons onto imbalance costs, arbitrage revenues, and AI deployment costs. The economic conclusions are grounded in Romanian Balancing Market conditions (mean up-regulation price λ+ ≈ 85 €/MWh, mean down-regulation price λ− ≈ 42 €/MWh; 15 min settlement interval), a five-year dataset (2018–2022) from a 10 MW utility-scale PV installation in Romania, and an annual AI system cost of 36,000 €/MW decomposed into data infrastructure, cloud GPU compute, and model-monitoring personnel. A Temporal Fusion Transformer ensemble, benchmarked against CNN-LSTM, Informer, and smart-persistence baselines, achieves a 0.38 Skill Score at the day-ahead horizon and a 0.28 Value Score, translating to a net economic benefit of €142,000 per installed MW per annum after full AI system cost deduction. While the framework is designed to be reusable across markets, all reported economic values are specific to the stated Romanian market parameters and should be recalibrated for other regulatory jurisdictions.

Future Internet doi: 10.3390/fi18060290

Authors: Akzhibek Amirova Aliya Abdiraman Laura Aldasheva Ibraheem Shayea Didar Yedilkhan Akhmet Tussupov

Handover management is a fundamental process in modern mobile networks, ensuring service continuity under user mobility. However, the relationship between network conditions and handover behavior remains insufficiently understood under real-world measurement conditions. This study presents a data-driven analysis of handover behavior based on drive-test measurements collected in an urban environment. A formal definition of handover events is proposed and implemented for automatic detection using changes in the serving cell identifier. The dataset is further analyzed to assess the influence of radio signal indicators, QoS metrics, and mobility-related variables on handover occurrence. Logistic Regression is used as an interpretable baseline, while Random Forest is applied to capture nonlinear feature interactions. The results show that individual QoS indicators demonstrate limited direct explanatory capability when considered independently. Random Forest achieved higher predictive performance than Logistic Regression, with AUC = 0.902 compared to 0.787, indicating the importance of nonlinear relationships in handover behavior. Degradation events are additionally identified using a threshold-based proxy, showing that latency is a more sensitive indicator of degraded conditions than throughput. Overall, the findings suggest that handover behavior depends on multiple interacting network conditions rather than a single dominant predictor, highlighting the importance of QoS-aware and data-driven mobility analysis in 5G networks and beyond.

Future Internet doi: 10.3390/fi18060289

Authors: Hiba Chanaa Loqman Chakir El Habib Nfaoui

Adverse drug–drug interactions (DDIs) are a leading cause of preventable medication-related harm, and their automatic detection from the biomedical literature is critical for pharmacovigilance and clinical decision support. Most existing systems derive relation representations from a single source, under-using the rich contextual structure of DDI sentences. We propose MG-ECF (Multi-Granularity Entity-Context Fusion), a relation-classification architecture that extracts three complementary views from a shared biomedical encoder—entity-level, inter-entity contextual, and global sentence representations—which are adaptively combined through a temperature-scaled gating mechanism regularized by view-dropout. MG-ECF was evaluated on the DDI-2013 benchmark under the official shared-task protocol, with multi-seed experiments on BioBERT and BiomedBERT backbones and a focal-loss objective to address severe class imbalance. MG-ECF achieves a mean micro-F1 of 90.55% with BiomedBERT and 88.8% with BioBERT, an absolute improvement of 2.7 F1 points over the strongest previously reported PLM-based system (BioMCL-DDI, 87.8%). Systematic component analyses confirm the contribution of each representational view, demonstrating the effectiveness of multi-granularity fusion for DDI classification and its potential as a research-stage building block for Internet-based pharmacovigilance platforms and networked clinical decision support systems, pending real-world clinical validation.

Future Internet doi: 10.3390/fi18060288

Authors: Yiyu Wang Zhufang Kuang Mingxiao Lei

As 6G mobile communications evolve, Low Earth Orbit (LEO) satellite mobile edge computing (MEC) enables globally seamless computing. However, the high mobility of LEO satellites disrupts service continuity and resource stability. Existing approaches often use oversimplified models that ignore multi-beam interference and dynamic task queueing. To address this, we establish a hierarchical Geostationary Earth Orbit (GEO)–LEO synergistic architecture, where the integration is implemented by utilizing GEO satellites as stability anchors and remote cloud relays, while LEO satellites provide low-latency edge processing. We formulate fine-grained models for two-level beam-centric communication and preemptive dynamic queueing. The resulting joint task offloading and resource allocation problem is a complex mixed-integer nonlinear program (MINLP). To effectively solve this MINLP, we decouple it hierarchically: first determine discrete offloading decisions, then optimize continuous resource allocations based on them, proposing a novel framework termed G2DRL (GNN-enhanced Game-theoretic and deep reinforcement learning). Simulation results demonstrate that G2DRL significantly reduces the weighted sum of system delay and energy, showing superior convergence stability and performance over state-of-the-art DRL baselines.

Future Internet doi: 10.3390/fi18060287

Authors: Chekwas Ifeanyi Chikezie Abraham Usman Usman Michael David Sulieman Zubair Henry Ohiani Ohize Joseph Ojeniyi

Machine-learning-based intrusion detection for Internet of Things systems has often been evaluated through model-centered pipelines that use weakly governed partitioning, limited leakage auditing, and closed-set assumptions. Consequently, reported performance could reflect data-handling artifacts rather than reliable security intelligence. This paper introduces a scientific integrity framework that treats preprocessing as a primary research object for open-set Internet of Things intrusion detection. The framework integrated device-disjoint split governance, feasibility-aware zero-day isolation, quantified leakage control, train-only preprocessing, shared-safe feature selection, diagnostic-harness verification, baseline split comparison, and auditable artifact generation. Applied to the CICIoT-DIAD 2024 corpus with Institute of Electrical and Electronics Engineers Organizationally Unique Identifier-based vendor enrichment, the protocol locked 28 canonical classes, eight semantic attack families, and five policy labels before constructing a device-disjoint, vendor-aware grouped split. When strict device-level zero-day holdout was infeasible, the framework activated an audited row-level fallback that preserved contamination-free holdout isolation without claiming strict device-novel zero-day evaluation. On 35,672,407 flows from 180 files, the accepted run achieved zero device overlap, zero flow-signature Jaccard leakage risk, 100 percent zero-day purity, a Feature Distribution Stability Score of 0.00518, a Device-Feature Dependency Index of 0.00000, an Attack Invariance Score of 0.92964, and an Attack Semantic Consistency Score of 0.90714. The diagnostic harness produced zero hard failures and zero warnings, while baseline comparison showed stronger preprocessing integrity than random stratified and simple device-disjoint splitting. This study did not claim downstream classifier superiority; rather, it established an auditable preprocessing substrate for later classifier-level experiments.

Future Internet doi: 10.3390/fi18060286

Authors: Krishnaprasath Suryanarayanan Ramanamurthy Gajula

Background: Continuously monitoring an ECG signal is vital for patients who cannot articulate their symptoms in a healthcare setting. To identify abnormalities, we introduce a compact one-dimensional convolutional autoencoder (CNN-AE) that can operate at low power consumption on TinyML-enabled edge devices. Methods: Comprising two convolutional layers and a bottleneck fully connected layer, the CNN-AE was trained on normal ECG signals from the Kaggle heartbeat Categorization dataset, derived from the PTB Diagnostic ECG Database. Anomalies were detected by comparing the reconstruction errors with the thresholds. To deploy the model at the edge, symmetric INT8 post-training quantization was applied, and the threshold was fine-tuned with calibration data. Results: The FP32 CNN-AE registered an accuracy of 83.12%, an F1-score of 88.67%, a precision of 85.88%, and a recall of 91.47% on 4046 normal and 10,506 abnormal ECG beats. The INT8 quantized model reduced the memory size by 75% (5.44 KB to 1.36 KB) while maintaining its performance (82.19% accuracy, 83.66% precision, 93.61% recall, and F1-score 88.36%). Inspection of the reconstruction errors and QRS morphology verified that quantization did not compromise the integrity of the ECG signals. Conclusions: The CNN-AE achieves a balance between accuracy, resource efficiency, and memory usage, allowing for real-time abnormality detection in assisted living and low-resource healthcare environments. Its capability on TinyML platforms indicates its role in integration with edge–cloud architectures, cloud-assisted optimization, and secure offloading strategies for scalable healthcare informatics.

Future Internet doi: 10.3390/fi18060285

Authors: Fangfang Shan Lulu Fan Yuhang Liu Zhuo Chen Yifan Mao

Federated Learning (FL) enables privacy-preserving model training in edge and IoT environments. However, in adversarial settings, FL suffers from two key challenges: robustness degradation due to data heterogeneity and poisoning attacks, and runtime instability on resource-constrained devices. Existing work mainly focuses on robustness while overlooking system-level stability. To address this, we propose FedCASKD, a robustness- and stability-aware FL framework. It employs a score-based soft aggregation mechanism to suppress unreliable client updates without requiring a trusted dataset, and introduces a selection-aware bidirectional knowledge distillation protocol to mitigate model drift under Non-IID data. The novelty lies in integrating aggregation and distillation into a unified feedback framework that enhances robustness and stability. Experiments on AGNews and SogouNews show that FedCASKD outperforms baselines under label-flipping attacks and heterogeneous settings. Memory and Out-of-Memory (OOM) tests further demonstrate its superior runtime stability in edge environments.

Future Internet doi: 10.3390/fi18060284

Authors: Cristian Cosentino Simone Squillace Fabrizio Marozzo

Financial analysis increasingly depends on the ability to transform heterogeneous textual evidence into reliable, verifiable, and actionable knowledge. However, adoption in finance requires generated outputs to be not only accurate, but also traceable and auditable. This work presents an audit-oriented LLM-RAG architecture for financial document intelligence. Rather than proposing a new foundation model, the contribution is a reproducible pipeline that integrates financial document processing, hybrid retrieval, evidence-grounded generation, structured validation, and persistent audit artifacts within a state-machine-based workflow. Designed for analyst-facing use, the system produces structured answers linked to explicit evidence while preserving the intermediate artifacts needed to inspect, reproduce, and validate each result. Experiments on AI-FinanceQA, a benchmark of heterogeneous financial documents and analyst-style questions, show that hybrid retrieval with reranking improves evidence selection over single-signal baselines and that the selected LLM backend achieves a compliance-oriented score of Scomp=0.9527. Additional experiments on FinQA confirm that targeted evidence selection improves numerical robustness and semantic alignment compared with uncontrolled context expansion. Overall, the proposed architecture provides an evidence-grounded and audit-oriented framework that supports human review rather than replacing expert financial judgment.

Future Internet doi: 10.3390/fi18060283

Authors: Iulian Tiță Luca-Ionuț Corățu Mihai Cătălin Cujbă Nicolae Țăpuș

Contemporary vulnerability management relies on the Common Vulnerability Scoring System (CVSS) and the Exploit Prediction Scoring System (EPSS), both of which evaluate Common Vulnerabilities and Exposures (CVE) entry in isolation, disregarding the network topology in which vulnerable components operate. We present the Dynamic Security Resistance Distance (DSRD) framework, which parses Docker Compose, GNS3, and Containerlab configuration files into weighted attack graphs where edge conductance reflects EPSS exploitability. A version-aware filtering stage matches discovered CVEs against the software versions declared in container image tags, reducing version-irrelevant CVE matches by up to 97%. Kirchhoff effective resistance, computed via the Moore-Penrose pseudoinverse of the graph Laplacian, yields a structural compromise affinity—a monotone score guaranteed not to increase upon patching. Four algorithms—Ant Colony Optimization, Physarum, Fungal Network Growth, and Greedy Kirchhoff-rank vulnerabilities by their structural impact on network-wide risk. Evaluation on nine representative topologies derived from public IaC artifacts, spanning six Docker Compose and three GNS3 deployments, with 895 version-relevant vulnerability nodes from cvelistV5 shows that graph-aware prioritization reduces structural risk by up to 5.62×10−2 after ten patches, whereas EPSS-only ordering achieves at most 1.28×10−2 on the same topology. EPSS-only targets high-probability CVEs on entry points that do not lie on critical paths; graph-aware methods instead prioritize CVEs on high-resistance paths toward critical assets. The advantage depends on infrastructure heterogeneity and topology structure: topologies with diverse vendors and well-defined structural bottlenecks benefit most, while densely connected or homogeneous environments show marginal improvement. We release the full pipeline as open-source software.

Future Internet doi: 10.3390/fi18060282

Authors: Murtadha D. Hssayeni Imadeldin Mahgoub

The possibility of zero-day attacks on Internet of Things (IoT) networks is high, particularly in dynamic and heterogeneous IoT environments, including emerging battlefield scenarios (IoBT). Detecting these attacks requires adaptive and generalizable security mechanisms. Due to the unique and unknown signatures of these attacks, they go undetected using signature-based Intrusion Detection Systems (IDSs) on the one side. On the other side, current anomaly-based IDSs that employ traditional machine learning on statistical features struggle to adapt and generalize to unknown networks, which is the case in IoBT. Transformer-based deep learning models have shown the capability of learning complex sequential patterns. This ability can be leveraged to analyze packet payloads that encompass opcodes capable of executing malicious patterns within an IoT network. In this work, we propose a dual-stage Transformer IDS that operates on the raw payload of network packets to detect zero-day attacks. Due to the lack of IoBT datasets, we evaluate the algorithm on three comprehensive IoT traffic benchmarks—MQTT-IoT, IoT-23, and CIC-IoT-2022—which have a high number of IoT devices and various attacks. Importantly, model evaluation is performed in two cross-validation settings to address the key operational challenges associated with unseen scenarios and networks. The evaluation settings are split-at-scenario to evaluate the detection ability of zero-day attacks and split-at-dataset to evaluate the model’s generalizability to new environments. In the former, the average increase in the F1-score of the proposed algorithm over the baseline model is 44% in detecting four zero-day attacks presented in the MQTT-IoT dataset. In the latter, the average increase in the F1-score is 16% in detecting malicious attacks across the three datasets. These results show the benefit of advanced AI in securing the next generation of IoT systems in future Internet applications.

Future Internet doi: 10.3390/fi18060281

Authors: Peiying Zhang Lijuan Chen Jian Wang Yujie Yuan Chun Sing Lai Lizhuang Tan

Dynamic multicast routing in software-defined networks is challenging due to continuously changing network states, multicast branch coupling, and the dependency between local forwarding decisions and global multicast tree construction. Existing multicast routing approaches mainly rely on static heuristics or snapshot-based optimization, which makes them difficult to maintain routing adaptability and decision stability under dynamic network conditions. To address these limitations, this paper proposes a topology-aware deep reinforcement learning multicast routing algorithm, named Graph-structured Hierarchical Actor–Critic for Multicast Routing (GHAC-MR). Specifically, the multicast routing process is formulated as a sequential tree construction problem, where each forwarding action incrementally affects the subsequent multicast tree evolution. A graph-structured state representation mechanism is designed to encode network topology information, link resource states, and multicast branch dependencies, enabling the routing agent to capture structural correlations among multicast forwarding nodes. Furthermore, a hierarchical actor–critic learning architecture is introduced to jointly optimize multicast forwarding policies and long-term routing rewards, thereby improving routing adaptability and convergence stability in dynamic network environments. Experimental results on multiple representative network topologies demonstrate that the proposed GHAC-MR algorithm achieves superior performance in multicast acceptance ratio, resource utilization efficiency, and routing adaptability compared with representative heuristic, evolutionary, and reinforcement learning-based multicast routing schemes.

Future Internet doi: 10.3390/fi18060280

Authors: Ming-Hseng Tseng Yu-Chuan Chen Wei-Ting Chen

Large language models (LLMs) have demonstrated strong potential in medical knowledge applications; however, their reliability in knowledge-intensive medical reasoning—remains limited due to hallucination, inadequate domain grounding, and unstable inference behavior. These limitations are particularly pronounced in tasks of professional medical reasoning that require strict logical consistency and authoritative knowledge support. This study proposes a generative AI architecture that integrates RAG (Retrieval-Augmented Generation) with parameter-efficient supervised fine-tuning based on Low-Rank Adaptation (LoRA) to improve reasoning stability and diagnostic accuracy in complex medical domains. The architecture combines internalized domain reasoning learned through LoRA-based fine-tuning with external knowledge grounding enabled by a dynamic RAG mechanism, allowing the model to selectively retrieve domain-specific knowledge only when it is semantically relevant and evidence supported. To validate the proposed architecture, a large-scale real-world dataset comprising 11,476 multiple-choice questions from Taiwan’s national Traditional Chinese Medicine (TCM) licensing examinations (2005–2025) is constructed as a representative case study of knowledge-intensive medical reasoning. The experimental results show that the baseline LLM achieves an accuracy of 61.0%. Incorporating RAG improves accuracy to 89.0%, while combined LoRA-based fine-tuning and RAG architecture further increases accuracy to 90.1%, with reduced variance across repeated evaluations. Statistical analysis using McNemar’s test confirms that the performance improvements introduced by the retrieval mechanism are highly significant. The results demonstrate that integrating parameter-efficient fine-tuning with dynamically controlled retrieval is critical to balancing reasoning stability and knowledge enhancement in generative AI systems. Beyond the specific medical case study examined in this work, the proposed architecture offers a reproducible and extensible framework for developing reliable generative AI systems in other knowledge-intensive professional reasoning and educational domains.

Future Internet doi: 10.3390/fi18060279

Authors: Lachlan Jones Benjamin Turnbull Nour Moustafa

Metric-based software security plays a crucial role in allowing software developers to make informed decisions about their development practices, while also allowing software users to evaluate the security risks associated with the software they use. Metrics are increasingly used to ensure code security, but there has been little formal evaluation of their broader applicability to date, and interpretation of their results remains a qualitative task. To address this gap, we introduce DevSecTrust, a standardised evaluation framework for measuring and comparing software development security metrics. DevSecTrust provides: (i) a unified control-mapped metric schema, (ii) outcome-based calibration and validation against real vulnerability and maintenance data, and (iii) robustness and manipulability testing to assess metric reliability. This paper analyses two software development security tools, MITRE’s Hipcheck and OpenSSF’s Scorecard, to evaluate and contrast the metrics they produce against widely used open-source software projects. Our quantitative comparison identified low correlation and inconsistent distributions between the tools’ outputs, and our qualitative analysis of feature weighting and scoring logic revealed foundational differences in how each tool conceptualises “secure development”. These inconsistencies complicate trust in development security metrics and hinder their interpretability and operational value. This contributes to a path toward standardised measurement of software development security.

Future Internet doi: 10.3390/fi18060278

Authors: Aeshah Mohammed Alshehri Samer H. Atawneh Hussein Al Bazar Roxane Elias Mallouhy

Cybersecurity has become a critical concern for individuals, organizations, and governments, especially with the rise of sophisticated cyberattacks and remote work environments. Traditional security approaches are no longer sufficient, leading to the adoption of advanced frameworks such as the zero-trust model, which operates on the principle “never trust, always verify.” This model enforces strict access controls and continuous monitoring across all network activities. Designing an intelligent zero-trust system is challenging due to the complexity of network environments and the evolving nature of malicious threats. This project proposes an advanced zero-trust architecture that integrates machine learning and multi-factor authentication (MFA) to strengthen security. Specifically, it employs Multilayer Perceptron models and k-Nearest Neighbors algorithms to analyze system logs and user behavior, enabling real-time anomaly detection and adaptive authentication mechanisms. The proposed framework is experimentally evaluated using the H-MOG behavioral–contextual authentication dataset, which captures multimodal user interaction patterns and supports continuous authentication analysis within Zero Trust environments. The integration of machine learning enhances the system’s ability to identify suspicious activities quickly and accurately, while MFA provides an additional layer of protection against unauthorized access. Moreover, the proposed framework emphasizes usability, ensuring that enhanced security does not impose excessive burden on users or IT teams. This allows the framework to respond more effectively to potential threats while maintaining usability. Overall, the proposed approach offers a practical and scalable solution that improves detection performance and strengthens continuous authentication and adaptive access control within Zero Trust environments.

Future Internet doi: 10.3390/fi18060277

Authors: Özlem Kaya Lorenzo Bacchiani Andrea Melis Roberta Presta Chan-Tong Lam Giovanni Pau Roberto Girau

Digital Twin (DT) technologies are increasingly adopted in the automotive domain to support real-time monitoring, predictive analytics, and connected decision-making across vehicles, drivers, and road infrastructure. However, research on Vehicle, Driver, and Road Digital Twins (VDTs, DrDTs, and RDTs) remains fragmented, with heterogeneous definitions, architectural assumptions, and integration strategies. This paper presents a critical review of seventy-six studies published between 2008 and 2025, examining how these three DT domains are modeled, evaluated, and connected within intelligent mobility scenarios. The review synthesizes recurring architectural patterns, communication and computing choices, and the role of interoperability and standardization in multi-twin systems. It also highlights open challenges involving distributed coordination, semantic alignment, real-time operation, and driver-aware adaptation. Based on this analysis, the paper presents a unified conceptual framework for connected automotive digital twins and discusses key directions for building scalable and safety-aware mobility services.

Future Internet doi: 10.3390/fi18060276

Authors: Mohammed Al Qurashi Ibtihaj Al Qarni

Sybil attacks remain a primary challenge for Proof-of-Stake (PoS) blockchain systems, as low-cost identity creation can distort validator participation and limit consensus reliability. This study proposes a hybrid participation–governance framework that integrates Attribute-Based Access Control (ABAC) and Reputation-Based Access Control (RpBAC) with a trust-based PoS workflow to reduce the influence of suspicious identities during validator selection and block validation. The proposed framework also incorporates graylisting and dynamic reward–penalty updates to support adaptive participation control. The strategy was evaluated in a simulation environment informed by Ethereum-derived block metadata, using network sizes ranging from 100 to 1000 nodes and Sybil attack ratios of 30%, 40%, and 50%. Its performance was compared with PoS-only and PoS + ABAC baselines using both security and performance indicators. The results show that the full ABAC + RpBAC configuration achieved the strongest and most stable security performance across the evaluated settings while introducing additional overhead at larger network sizes. These findings suggest that combining policy-based eligibility control with behavior-based reputation control strengthens the resilience against Sybil in PoS-like blockchain environments. However, this improvement requires a measurable trade-off between security and performance.

Future Internet doi: 10.3390/fi18060274

Authors: Marco Romano Teresa Onorati

In the last few years, extended reality (XR) technologies, including virtual reality (VR), augmented reality (AR), and mixed reality (MR), have gradually expanded beyond their traditional entertainment domain and are now used in many other contexts [...]

Future Internet doi: 10.3390/fi18060275

Authors: Peixin Zhao Chengqun Wang

Modulation and signal classification are two highly correlated core tasks in wireless communications and are the core foundation of intelligent spectrum management in Future Internet and 6G networks. Although their objectives differ, the two tasks often share a substantial amount of underlying information in the feature space. However, focusing solely on their commonalities while neglecting their intrinsic differences may lead to suboptimal model performance. Therefore, by taking into account both the correlation and inherent differences between the two tasks, we propose TAMTNet, a task-adaptive multi-task network for edge deployment in Future Internet. TAMTNet introduces Extremely Efficient Spatial Pyramid (EESP) into the shared layer to efficiently extract multi-scale temporal information. In addition, a multi-gate mixture-of-experts (MMoE) mechanism is employed after the shared layer to enhance the modeling capability of task-specific features. Furthermore, to address the difficulty of deploying deep models on resource-constrained edge devices, a joint lightweight framework combining quantization-aware training and knowledge distillation is proposed, which significantly reduces model complexity while maintaining performance. Extensive experiments conducted on the simulation and real-world over-the-air transmission datasets demonstrate that the TAMTNet model achieves excellent performance on both modulation and signal classification tasks across a wide range of signal-to-noise ratios and radio transmit gain conditions. Meanwhile, the low-bitwidth lightweight models are able to maintain classification performance comparable to the full-precision model while significantly reducing model storage and computational complexity.

Future Internet doi: 10.3390/fi18050273

Authors: Isaac Kofi Nti

Behavioral ransomware detection often achieves high accuracy under standard evaluation settings, but such results may not generalize under distribution shift or when previously unseen ransomware families are encountered. This study evaluates detection performance on the MLRan dataset, which contains 4880 samples from 64 ransomware families, using four evaluation protocols: stratified, temporal, family-disjoint, and open-set. The family-disjoint and open-set protocols were constructed at the family level to limit overlap between learned and held-out ransomware families. The study proposes the Hierarchical Sparse Neural Network (HSNN), a taxonomy-aligned model that uses group-level and branch-level gating to support structured interpretability and modality-level analysis. Compared with the FlatMLP baseline, HSNN achieved a slightly lower average macro-F1 score (0.9839 vs. 0.9860) but showed better calibration and lower model complexity. Specifically, HSNN reduced expected calibration error by 34.1% and parameter count by 42%. HSNN also showed slightly lower variability across random seeds and stable gate patterns. Under the open-set family protocol, HSNN achieved one of the strongest macro-F1 scores (0.9930 vs. 0.9913 for FlatMLP) using a maximum-softmax novelty baseline. Feature analysis indicates that string-based artifacts remain strong predictors, while the hierarchical structure distributes importance across multiple behavioral modalities. These results position HSNN as a competitive alternative to dense neural baselines when calibration, compactness, and structured interpretability are considered alongside macro-F1 performance.

Future Internet doi: 10.3390/fi18050272

Authors: Evelio Astaiza Hoyos Héctor Fabio Bermúdez-Orozco Nasly Cristina Rodriguez-Idrobo

The sixth generation of mobile networks (6G) represents a paradigmatic shift in the conception of wireless communication systems, where Artificial Intelligence (AI) is not integrated as an additional feature but is conceived as a native and fundamental component of the physical layer (PHY). This paper presents a comprehensive survey of the state of the art in AI-native physical layer for 6G, synthesizing approximately 100 references from the period 1948–2025. The survey systematically covers 5 main PHY components (channel coding, channel estimation, signal detection, beamforming, and semantic communications) and analyzes 8 AI architectural families (autoencoders, CNN, RNN/LSTM, Transformers, GNN, GAN, Diffusion Models, and Foundation Models), addressing theoretical foundations, proposed architectures, learning algorithms, implementation challenges, and future research directions. A rigorous mathematical framework underpinning these developments is presented, including optimization formulations, convergence analysis, and theoretical performance characterization. Published results from the literature demonstrate that AI-native physical layer can improve conventional performance metrics and enable emerging capabilities essential to 6G, such as semantic communications, predictive environmental adaptation, and operation in previously inaccessible computational complexity regimes. However, such gains are conditional on adequate training resources, robust channel-matched data, and careful consideration of known limitations including generalization across channel distributions, sample inefficiency, model interpretability, and hardware implementation constraints—all of which are critically analyzed in this survey. A reproducible proof-of-concept benchmark further confirms that, under severe resource constraints, autoencoder-based codes currently underperform conventional schemes, highlighting the gap between theoretical potential and practical deployment readiness.

Future Internet doi: 10.3390/fi18050271

Authors: Oussama Nabil Cherkaoui Leghris

Tomato (Solanum lycopersicum) cultivation is highly vulnerable to fungal, bacterial, and viral leaf diseases that can significantly reduce crop yield and fruit quality when not detected at early stages. Although recent deep learning approaches have achieved remarkable performance in plant disease classification, many state-of-the-art architectures remain computationally expensive and therefore difficult to deploy on resource-constrained edge devices commonly used in smart agriculture environments. To address this challenge, this paper introduces EcoTomHybridNet, an adaptive resource-aware CNN–Transformer framework designed for efficient tomato leaf disease classification under edge-computing constraints. The proposed architecture combines a lightweight convolutional backbone with a dual-branch inference mechanism composed of a fast convolutional branch for computationally efficient prediction and a Transformer-enhanced branch with local self-attention for richer contextual feature extraction. Unlike conventional lightweight hybrid models relying on static inference pipelines, EcoTomHybridNet integrates a lightweight policy-guided routing mechanism that dynamically allocates inputs between the fast convolutional branch and the Transformer-enhanced branch according to input complexity. This adaptive inference strategy dynamically reduces unnecessary Transformer computations for simpler samples while preserving strong predictive performance on more challenging inputs through policy-guided branch allocation. To further improve representation capability without significantly increasing computational complexity, the proposed student network is trained using knowledge distillation from a ViT-Tiny teacher model. Experimental results on the PlantVillage tomato dataset demonstrate that EcoTomHybridNet achieves 99.42% test accuracy and 99.0% validation accuracy under the full hybrid inference configuration. Additional validation strategies, including 5-fold cross-validation and robustness evaluation under Gaussian noise and motion blur perturbations, indicate stable performance across different data splits and moderate image degradations, suggesting improved generalization capability beyond simple dataset memorization. Furthermore, adaptive routing experiments using a lightweight threshold-based policy mechanism achieved 99.20% test accuracy while reducing computational complexity from 0.36 GFLOPs to 0.25 GFLOPs per image, corresponding to approximately 30% computational savings. These results demonstrate the effectiveness of policy-guided adaptive inference for balancing predictive performance and computational efficiency in edge-oriented plant disease classification. Overall, EcoTomHybridNet provides an efficient and adaptive framework for intelligent plant disease monitoring in IoT-enabled smart agriculture systems.

Future Internet doi: 10.3390/fi18050270

Authors: Ruize Gu Xiaoying Wang Fangfang Cui Guoqing Yang Shuai Liu Panpan Qi

Existing anomalous traffic detection methods based on feature fusion in Software-Defined Networking (SDN) lack adaptability in weight allocation mechanisms. Consequently, their detection accuracy and model generalization capabilities fail to meet practical security requirements. To solve these limitations, this paper proposes a refined detection method based on hybrid feature selection and gated fusion. First, the framework employs XGBoost combined with the Recursive Feature Elimination (RFE) algorithm. This process identifies shallow statistical features with high discriminative power. Simultaneously, the method utilizes a 1D Convolutional Neural Network (1D-CNN) integrated with a Squeeze-and-Excitation (SE) block to extract deep temporal semantic features. Subsequently, a tailored gated fusion mechanism incorporating linear projection layers for feature alignment adaptively integrates these two categories of features. The fused features are then input into a Multilayer Perceptron (MLP) to execute anomalous traffic detection. Experimental results demonstrate that the proposed method achieves superior performance. Specifically, on the InSDN Dataset, the binary and multi-classification accuracy rates reach 99.91% and 99.88%. Similarly, the accuracy rates on the NSL-KDD dataset are 99.78% and 99.76%. Finally, we established a local simulation environment. Experimental results demonstrate that our method attains an average precision exceeding 93% for anomalous traffic detection in simulated real scenarios.

Future Internet doi: 10.3390/fi18050269

Authors: Abdelrahman Radwan Mohammad Hamdan Zhuldyz Ismagulova Mohammad Ma’aitah Ala’a Alshubbak Mohammad Nasir

Wireless sensor networks (WSNs) are fundamental to the Internet of Things (IoT) and are widely used in environmental, industrial, and healthcare applications. However, their operational lifetime is constrained by the limited energy resources of sensor nodes. The Low-Energy Adaptive Clustering Hierarchy (LEACH) protocol reduces energy consumption through clustering but suffers from random cluster head (CH) selection, leading to uneven energy usage and reduced stability. This study introduces a hybrid optimization approach, LEACH-CSA, which integrates the Crow Search Algorithm (CSA) with LEACH to enhance CH selection and positioning. The proposed method employs CSA’s intelligent search behavior to minimize intra-cluster distances and balance energy consumption across nodes. MATLAB simulations with 100 sensor nodes in a 100 × 100 m2 area demonstrate that LEACH-CSA significantly reduces energy consumption and extends network lifetime compared with LEACH and its variants. Furthermore, CSA parameters were optimized using a progressive randomized tuning strategy with 1000, 2000, and 4000 candidate configurations. A comparative evaluation against LEACH-based GA, PSO, GWO, and WOA demonstrated that LEACH-CSA consistently improved the FND metric under different node density and area-scaling scenarios.

Future Internet doi: 10.3390/fi18050268

Authors: Geno Stefanov Valentin Kisimov

The growing complexity of enterprise resource planning (ERP) systems necessitates intelligent approaches for dynamically identifying and evaluating key performance indicators (KPIs) that accurately reflect organizational performance. This paper proposes a multi-agent architecture for dynamic KPI management over Oracle E-Business Suite (EBS). The core design combines a dynamic multi-agent analytics layer, an extendable dedicated EBS KPI Model Context Protocol (MCP) server layer, and a data layer. The dynamic multi-agent analytics layer defines a set of independent large language model (LLM) agents, each responsible for a specific task determined by the business requirements of a particular company. The EBS KPI MCP server layer defines the tools required to access and transform Oracle EBS data and exposes them to the AI agents in the upper layer. Above these layers is the user layer, where the user actively participates in the process through a human-in-the-loop approach. Based on this general architecture, we proposed and implemented, as a proof of concept (PoC), a multi-agent system for dynamic business KPI selection, evaluation, and quantification, in which three distinct agents for KPI selection, KPI quantification, and KPI forecasting were instantiated within the multi-agent analytics layer. This demonstrates the practical applicability of the proposed general architecture. The study contributes to intelligent business analytics by showing how coordinated LLM agents can automate KPI lifecycle activities within ERP ecosystems, enabling adaptive, data-driven performance management aligned with evolving organizational needs.

Future Internet doi: 10.3390/fi18050267

Authors: Zuzana Soltysova Julia Nazarejova

The sharing economy has become a significant phenomenon of the modern economy in recent years, enabling more efficient use of resources through digital platforms. At the same time, the development of generative artificial intelligence (AI) has begun to reshape the functioning of these platforms. This article explores the intersection of sustainability and AI within sharing economy platforms, focusing on environmental, economic and social dimensions. The study adopts a conceptual and exploratory research design combining a literature review with a comparative case study analysis of selected sharing economy platforms, namely Airbnb and BlaBlaCar, complemented by an industrial platform example (Xometry). The analysis examines how generative AI can support sustainable consumption, operational efficiency, and user engagement while raising important ethical concerns related to data usage, trust, bias, and algorithmic governance. The findings suggest that AI integration can improve resource utilization, accessibility, and platform efficiency, but simultaneously introduces new ethical challenges related to transparency, data governance, and algorithmic decision-making. These results highlight the dual role of AI as both a driver of sustainability and a source of emerging ethical risks in digital platform ecosystems.

Future Internet doi: 10.3390/fi18050266

Authors: Mehdi Moucharraf Mohammed Ridouani Fatima Salahdine Naima Kaabouch

By allowing continuous connectivity, automation, and data-driven decision-making across these areas, Internet of Things (IoT) has transformed certain facets of daily life, including home automation and healthcare, as well as business operations like supply chain management and smart manufacturing. IoT systems are susceptible to different cyberattacks, though, because of different designs, lack of funds, and inadequate security policies, which creates major security issues given their fast growth. Covering important topics including protocols, architectures, attack classification, detection methods, countermeasures, and research issues, this paper offers a thorough study of IoT security. Emphasizing their relevance in enhancing the security of IoTs, the article offers a thorough analysis of machine and deep learning-based detection techniques. It also offers recommendations for future paths to handle changing risks by means of particular proposals and provides tools and datasets required for IoT security studies. When considering recent progress, however, there are still some major limitations in scaling, real-time detection, dataset availability, and versatility of current solutions. We identified these issues and provided guidance on future research; we also offered a selected set of tools and datasets for further research. Additionally, this paper provides an overview of the most important issues related to IoT security as documented in the current literature, providing a framework for developing resilient and adaptable IoT security solutions in the future.

Future Internet doi: 10.3390/fi18050265

Authors: Sicheng Liang Chunpu Huang Yexuan Li Zhao Wang Benhao Zhu Jiawei Ye Jie Wu

Multi-agent workflows have emerged as an important execution pattern for intelligent applications, where specialized agents collaborate through dependent stages such as planning, retrieval, execution, and verification. When such workflows are deployed over edge-cloud infrastructures, scheduling becomes challenging because task dependencies, heterogeneous resource conditions, and topology-dependent communication overhead must be considered jointly. We study the problem of scheduling multi-agent workflows in edge-cloud environments and propose a graph-aware scheduling method that models workflow execution as a task graph and the underlying infrastructure as a resource graph. The method combines structure-aware task and resource representations, communication-sensitive assignment scoring, and online resource-state updates to improve task placement quality. Experiments across different workflow complexities, system scales, and dynamic operating conditions show that the proposed method achieves a favorable balance among latency, communication overhead, and execution cost, particularly in communication-sensitive and medium-to-large-scale settings. These results suggest that jointly modeling workflow structure and resource topology can improve scheduling quality for multi-agent workflow execution in heterogeneous edge-cloud environments.

Future Internet doi: 10.3390/fi18050264

Authors: Larysa S. Globa Serhii M. Ushakov

Industry 4.0 requires IoT ontologies that are interoperable, scalable, and adaptive in non-stationary industrial environments. This study combines methodological ontology optimization with a hybrid elastic framework for dynamic semantic updates and feedback-driven refinement. The methodological component systematizes literature and industrial practices to identify structural gaps and derive practical requirements. The engineering component integrates truth-table-based data structuring, vector–matrix automata for real-time classification and clustering, and in-memory event processing for low-latency operation. Experimental evaluation across no-drift, abrupt-drift, gradual-drift, and cyclic-drift scenarios shows a trade-off between semantic proximity and operational robustness: the rule-based approach reaches lower semantic distance in drift regimes, while the hybrid approach delivers higher stability and fewer false alarms in cyclic dynamics. All tested configurations preserve sub-millisecond processing latency, supporting edge/fog deployment. The results indicate that combining methodological analysis with elastic architecture is a practical pathway from static to adaptive IoT ontologies and a relevant step toward human-centric Industry 5.0 systems.

Future Internet doi: 10.3390/fi18050263

Authors: Ding Pan Yawen Chen Yan Li Yunpeng Ma

Predictive business process monitoring (PBPM) plays an important role in intelligent workflow management by enabling organizations to anticipate future process behavior and support operational decisions. However, many existing approaches represent execution traces primarily as linear prefixes, thereby limiting their capacity to explicitly capture the control-flow semantics of non-sequential processes. To address this limitation, this paper proposes KG-MTPM, a knowledge-graph-enhanced multi-task framework that integrates process-model-level structural knowledge with prefix-level runtime dynamics in a unified predictive architecture. In particular, control-flow relations are organized as a process knowledge graph so that non-linear execution dependencies can be explicitly represented during prediction. Based on the integrated representation, the model jointly predicts the next-activity, next-activity time, and remaining-time of an ongoing case. Experiments on three real-world event log datasets demonstrate that KG-MTPM achieves the best overall performance among the evaluated baselines, with a marked advantage in time-related prediction tasks. Relative to the best-performing baseline, KG-MTPM improves next-activity prediction accuracy from 0.84 to 0.85, while reducing the mean absolute error (MAE) of next-activity time prediction from 0.81 to 0.25 and that of remaining-time prediction from 0.98 to 0.47. Ablation results confirm the contributions of both the structure-aware representation and the multi-task learning scheme. Overall, the findings suggest that explicit modeling of process structure is beneficial for predictive monitoring in business processes with complex execution behavior.

Future Internet doi: 10.3390/fi18050262

Authors: El Hariri Ayyoub Mouiti Mohammed Lazaar Mohamed

Securing IoT networks presents fundamental challenges rooted in hardware constraints: firmware is often non-upgradeable and every security boundary is fixed at manufacture. Machine learning-based intrusion detection offers a scalable response, yet nearly all published systems assume clean training data and clean inference conditions. Production IoT environments satisfy neither assumption. Sensors degrade, packets drop, and adversaries deliberately corrupt telemetry streams to evade detection. The framework described here is built around that reality. The proposed framework is distinguished from prior work by four design decisions. First, three encoding branches, a residual DNN, a 1D-CNN, and a BiLSTM, are run in parallel and are fused by concatenation, each capturing structural patterns in tabular traffic data that the others miss. Second, a dual-view consistency loss trains the model under simultaneous feature masking and Gaussian noise, penalizing prediction divergence between two independently corrupted views of the same sample. Third, we introduce entropy-weighted attention: rather than fixed learned weights, per-feature importance is adjusted dynamically from information entropy measured across training batches, giving higher-entropy features stronger influence because they carry more discriminative variation. Fourth, branch-dropout regularization randomly silences entire branches during training, forcing each to develop independently useful representations instead of co-adapting. Class imbalance is handled through severity-aware loss weighting which scales contributions by the operational cost of missing each attack category, not purely by inverse frequency. On UNSW-NB15, the full model achieves 99.99% accuracy, 100% precision, 99.97% recall, and a false-negative rate of 2.65 × 10−4—the lowest across all compared architectures.

Future Internet doi: 10.3390/fi18050261

Authors: Ina Asenova Naydenova Zlatinka Svetoslavova Kovacheva Iliya Krasimirov Georgiev

User interface evaluation remains fragmented across performance metrics, subjective assessments, and user-dependent factors, limiting the comparability and interpretability of results across methodological traditions. This paper proposes a multidimensional evaluation framework that integrates these perspectives into a coherent analytical structure. The framework consists of three dimensions—Functional–Objective, Cognitive–Perceptual, and Contextual–Individual—each capturing a distinct facet of interface quality. A key feature of the proposed approach is the use of profile-dependent weighting, which enables evaluation results to reflect the specific priorities of different user groups. The framework’s operational logic is demonstrated through structured illustrative scenarios, showing how the model can be applied in practice to support more informed design and evaluation decisions. By aligning heterogeneous evaluation logics within a unified structure, the proposed approach provides a systematic basis for more consistent, transparent, and context-sensitive assessment of user interfaces.

Future Internet doi: 10.3390/fi18050260

Authors: Ernesto Cadena Muñoz Diego Armando Giral César Hernández Suárez

Efficient spectrum sensing remains a main challenge for Cognitive Radio Networks (CRNs), especially in a wireless environment where methods like energy detection have high uncertainty. This work proposes an entropy-based spectrum-sensing system enhanced with machine-learning algorithms and implemented on a Software-Defined Radio (SDR) platform for real scenario testing. Entropy measures, such as Shannon and Rényi entropies, are used as discriminative features to distinguish occupied and idle frequency bands and release the channel if needed. Machine learning classifiers have achieved good results. In this research, Support Vector Machines (SVMs), K-Nearest Neighbors (KNNs), and Random Forests (RFs) are used with data captured via a GNU Radio and the Universal Software Radio Peripheral (USRP)-based SDR testbed. The experimental results demonstrate a probability of detection (Pd) above 0.9 and a false alarm rate (Pfa) below 0.1, indicating a substantial improvement over the classical energy detector of more than 20% for some signal-to-noise ratio (SNR) values. The integration of entropy metrics with machine learning (ML) models enables a dynamic detection in variable spectral environments, providing a practical framework for CRNs.

Future Internet doi: 10.3390/fi18050258

Authors: Anna Kushchazli Kseniia Leonteva Darina Shiyapova Alexandr Priscepov Irina Kochetkova

Resource-constrained Multi-Access Edge Computing (MEC) nodes cannot fully replace cloud infrastructure, yet existing service placement models treat edge hosting as an all-or-nothing decision. This paper proposes a queueing-theoretic framework for split-user offloading in hybrid MEC–cloud environments. The system is modeled as a Continuous-Time Markov Chain (CTMC) over a load-vector state space that admits a product-form stationary distribution. A delay-aware greedy orchestration policy determines, at every arrival and departure event, which service occupies the MEC node and how many of its users are offloaded from the cloud. Closed-form expressions are derived for average end-to-end (E2E) delay, MEC occupancy and saturation probabilities, per-service hosting probabilities, and delay-saving indicators. Numerical analysis of a five-service industrial scenario shows that the proposed split-user mechanism keeps the MEC node occupied for most of the observation time (around 97% at the baseline load), naturally prioritizes services with the largest aggregate latency benefit, and substantially reduces the average delay compared with a cloud-only configuration. The analytical results are validated by discrete-event simulation, which matches the CTMC values with relative discrepancy below 1% under the Poisson/exponential assumptions; additional simulations quantify the sensitivity to alternative arrival and service-time distributions. The framework provides analytically tractable, interpretable decision logic with negligible runtime overhead, making it a suitable analytical foundation for cloud service orchestration platforms that must meet strict QoS targets in next-generation edge networks.

Future Internet doi: 10.3390/fi18050259

Authors: Mohammad AbdulJawad Mohammad Z. Masoud Álvaro Álesanco José García

Industrial control systems (ICSs) increasingly rely on legacy communication protocols such as ModbusTCP, which lack built-in security mechanisms and remain widely exposed to network-based attacks. This paper investigates the security limitations of authentication mechanisms in ModbusTCP-enabled programmable logic controllers (PLCs) and demonstrates how plaintext credential transmission and limited connection handling capabilities can be exploited to perform brute-force and denial-of-service (DoS) attacks. An experimental testbed based on two industrial Delta PLC families (DVP-13SE and DVP-311SV3) was developed to systematically evaluate these vulnerabilities under realistic conditions. The results show that authentication credentials can be easily captured through network sniffing, while the PLC communication stack supports a maximum of 16 concurrent connections and can process up to approximately 8600 Modbus operations per second, making it susceptible to resource exhaustion and performance degradation under distributed attack scenarios. To address these limitations, this paper proposes a lightweight deception-based protection mechanism, termed the PLC misleading algorithm (PMA), which is implemented directly within the PLC ladder logic. Unlike traditional network-level defenses, PMA operates at the device level and dynamically misleads attackers by generating controlled randomized responses while preserving consistent behavior for legitimate clients. Experimental results demonstrate that PMA significantly mitigates brute-force effectiveness by preventing reliable password extraction while introducing minimal overhead (2.2% memory usage) and maintaining acceptable communication latency. Additionally, the proposed approach significantly reduces observable attack traffic, with only 0.246 Modbus operations per second observed during the attack phase, thereby limiting the effectiveness of automated exploitation tools. These findings highlight the potential of in-device deception mechanisms as a practical and deployable security layer for legacy industrial systems, and provide new insights into the resilience of PLC-based infrastructures against network-level attacks. This work bridges the gap between lightweight PLC-level protections and the growing need for robust cybersecurity mechanisms in industrial IoT environments.

Future Internet doi: 10.3390/fi18050257

Authors: Zuocheng Yu Ming Xu Xiaogang Xing Yuanhao Lin Yuwen Shu Xiaohan Qi

With the rapid development of the Internet, open-source graph datasets are increasingly shared and reused in intelligent networked services, making robust infringement verification increasingly important. Backdoor-based watermarking for graph neural networks (GNNs) can be used to check whether a suspicious model has been trained on protected data without authorization. However, existing dataset infringement verification methods have limited applicability and are mainly designed for private datasets. Directly applying them to open-source datasets would cause models trained by legitimate users to learn backdoor behavior, which would expose them to security risks. In this paper, we propose a new infringement verification method for open-source graph datasets, which reduces backdoor-related security risks in models trained by legitimate users. The core idea is to introduce an additional expansion-class and re-label watermarked samples as belonging to this class. This design completely separates the learning of watermark patterns from the original feature-label mappings during training. As a result, only trigger-bearing samples are directly involved in infringement verification, which helps prevent watermark patterns from being associated with existing classes in the original task. The proposed method provides a practical solution for trustworthy graph data sharing and infringement verification in Internet environments. Extensive experiments on benchmark datasets demonstrate that the proposed method achieves a high verification success rate while largely preserving the model’s clean accuracy.

Future Internet doi: 10.3390/fi18050256

Authors: Muhammed Fatih Koc Nouman Ashraf Pramod Pathak Sachin Sharma

In recent years, Intelligent Transportation Systems (ITSs) have emerged as a solution to mitigate the problem of traffic congestion. Understanding human driving styles such as aggressive, normal, and cautious is crucial for safe driving. In particular, predicting lane-change manoeuvres may be further supported by combining vehicle state information with driving style information. However, existing vehicle trajectory datasets lack driving style information, making classification challenging. To address this limitation, this paper proposes a fuzzy logic-based driving style classification framework in a Vehicle-to-Everything (V2X) environment. The model uses vehicle state information, including speed, longitudinal acceleration, lateral acceleration, and distance headway to classify style as cautious, normal, or aggressive. The proposed system is interpretable, aligns with human reasoning, and remains computationally efficient for real-time applications. The performance of the proposed work has been evaluated through comprehensive experiments on highway data. Results show a separation of driving styles, achieving 77% accuracy on a balanced dataset, showing moderate agreement with deterministic labelling while maintaining interpretability. In V2X-enabled lane-change prediction scenarios, computational latency is essential, as Roadside Units (RSUs) must understand driving style and update prediction models. Since lane-change intentions should be predicted around 3 s before manoeuvre, delays in inference reduce reaction time. The proposed classifier achieves an inference latency of approximately 8 ms, ensuring that it does not become a bottleneck in real-time systems. Furthermore, the usefulness of driving style information is tested by integrating it into a lane-change prediction task. Experimental results demonstrate that incorporating driving style enhances prediction accuracy from 75% to 84%. Lastly, the proposed method provides a balanced result between interpretability, computational efficiency, and predictive performance, supporting RSUs to issue timely warnings and support safer decision-making in highway environments.

Future Internet doi: 10.3390/fi18050255

Authors: Vasilios A. Orfanos Stavros D. Kaminaris Panagiotis Papageorgas Dimitrios Piromalis Dionisis Kandris

The Internet of Things (IoT) has a decentralized smart home ecosystem, as each protocol has its own gateway infrastructure needs. This study advances gateway convergence by proposing and rigorously evaluating a scalable architectural framework for future smart-home infrastructure. Specifically, this paper provides a detailed analysis of a proposed integrated multi-protocol gateway design that supports 18 of the most widely used IoT communication protocols simultaneously. It is a one-device implementation combining wireless technologies, including short-range radios (Sub-1 GHz, 2.4 GHz), LPWANs (Long Power Wide Area Networks), cellular (LTE, Long-Term Evolution), and wired (Ethernet, KNX). Using the ns-3 network simulator, this paper shows that this architecture is practical in a simulated smart-home environment with a large number of interconnected devices distributed across various zones. The results demonstrate substantial reductions in energy consumption and operational complexity, without compromising quality of service across heterogeneous communication technologies.

Future Internet doi: 10.3390/fi18050254

Authors: Juan Manuel Díaz-Gómez Enrique Quiceno-Rua Cristian David Correa-Álvarez

This review examines the technologies shaping real-time web application development, with particular attention to bidirectional communication protocols, distributed reactive architectures, and computing platforms designed for low-latency, high-concurrency environments. Based on a systematic analysis of 62 studies published from 2020 through September 2025, the review identifies clear areas of convergence around WebSockets, hybrid edge–cloud architectures, and JavaScript-based ecosystems built on Node.js and React. The findings show a broader shift toward decoupled, event-driven systems that rely on asynchronous communication, while multi-user synchronization and horizontal scalability continue to pose major challenges. Bibliometric analysis also reveals a sharp increase in publications since 2023, with most studies appearing in IEEE conference proceedings and journals focused on software and systems architecture. The evidence suggests a growing preference for microservice-based architectures over monolithic designs because of their scalability, fault isolation, and support for asynchronous workflows, although the most effective architectural choice still depends on the application context. Current research is limited by the frequent use of controlled experimental settings, the lack of standardized benchmarks, and the relatively limited attention paid to interoperability. Overall, this review brings together the current evidence and outlines directions for designing efficient, scalable, and secure real-time web systems.

Future Internet doi: 10.3390/fi18050253

Authors: Leonilde Varela Goran D. Putnik Luis Ferreira Vijaya Kumar Manupati Pedro Pinheiro Catia Alves Paulo Avila Helio Castro

The integration of artificial intelligence (AI) is a cornerstone of Industry 4.0, driving significant gains in automation, efficiency, and adaptability. In parallel, manufacturing environments are evolving into cyber–physical systems (CPS), where physical processes are deeply integrated with computational intelligence. While machine learning and deep learning techniques have become standard practice in manufacturing CPS, the emergence of advanced and foundation AI models—such as reinforcement learning, agent-based AI systems, large language models, and neuro-symbolic approaches—brings fresh opportunities and challenges that are not fully understandable. This paper offers a comprehensive systematic literature review (SLR) on AI applications in manufacturing cyber–physical systems, with a particular focus on the role, maturity, and industrial readiness of emergent AI models. Following the PRISMA 2020 guidelines, a structured search was carried out in Scopus and Web of Science, producing over 4200 publications, out of which a final set of 172 publications were retained following a rigorous multi-stage screening and eligibility process. We analysed the selected literature through complementary descriptive, longitudinal, and mapping syntheses to identify publication trends, paradigm evolution, and relationships between AI paradigms and manufacturing functions. Our findings show a clear transition from rule-based and conventional machine learning approaches toward more adaptive, decentralized, and learning-driven AI paradigms. However, despite their conceptual suitability for complex and dynamic manufacturing environments, emergent AI models are mostly limited to experimental, hybrid, or decision-support contexts, with limited integration into core manufacturing operations. Critical research gaps regarding the industrial readiness of these models—specifically concerning integration frameworks, empirical validation, safety, and trust—are identified. Furthermore, the study outlines future research directions for advancing the next generation of intelligent and autonomous manufacturing CPS. Overall, this review underscores the rapid growth and current fragmentation of the field, highlighting the need for more integrative and production-ready AI frameworks in the evolution of manufacturing CPS.

Future Internet doi: 10.3390/fi18050252

Authors: Antônio João Azambuja Marcos Guilherme João Victor Fernandes de Castro Jean Phelipe de Oliveira Lima Leonardo B. Oliveira Anderson da Silva Soares

The rapid integration of Large Language Models (LLMs) into critical personal and professional environments has exacerbated security risks, particularly adversarial attacks such as prompt injection and jailbreaking, which aim to bypass safety alignment. This study evaluates the efficacy of NVIDIA’s Llama-3.1-nemoguard-8b-content-safety model acting as a semantic firewall to mitigate these threats. To ensure a robust assessment, we utilized the ‘Do Not Answer’ dataset, augmented with 939 synthetically generated benign prompts to create a balanced corpus of 1878 samples. The evaluation methodology encompasses a risk-category analysis, standard binary classification metrics, and a novel metric, the Compensation Rate, which measures the firewall’s ability to block responses when the underlying LLM fails. Results indicate a high Precision (94.57%) but a moderate Sensitivity (51.97%), uncovering a critical performance trade-off: the model exhibits a conservative bias, prioritizing high precision to minimize false positives at the expense of recall for nuanced adversarial prompts, particularly in categories involving sensitive data leakage and misinformation. Furthermore, the proposed Compensation Rate achieved 34.8%, suggesting that the semantic firewall successfully mitigated 34.8% of instances where the foundational LLM’s internal safety alignment failed. These findings indicate that while the system effectively blocks explicit threats, its efficacy as a secondary defense diminishes against context-dependent vulnerabilities, notably data exfiltration and misinformation.

Future Internet doi: 10.3390/fi18050251

Authors: Sufian A. Badawi Maen Takruri Khouloud Salameh Mohammad Al-Badawi Nowar Alani Isam ElBadawi Aws Al-Qaisi Ghaleb Aldoboni

This work presents an innovative approach to test the Arabic language proficiency assessment via Automatic Speech Recognition (ASR) by enhancing the proficiency of the Whisper model in transcribing Arabic speech. The core of our research involved fine-tuning the Whisper model using a substantial, large-scale Arabic speech corpus, with a specific focus on Modern Standard Arabic. This process used a 2000-h Arabic-labeled speech corpus, the QASR dataset, and improved the model’s Word Error Rate (WER). After optimization, the fine-tuned Whisper model’s WER was reduced from 35% to 7% on the QASR dataset, corresponding to an absolute reduction of 28 percentage points (approximately 80% relative reduction). These results demonstrate the strong generalization ability of the fine-tuned model across multiple Arabic ASR benchmarks. A key component of our methodology was the development of a sophisticated scoring system. This system integrates various similarity metrics, such as cosine similarity, the Jaccard index, and the Levenshtein distance, with a machine learning regression model. This multifaceted system provides a comprehensive assessment of reading proficiency, proposing a practical automated assessment method that contributes to the field of AI language transcription and to its application in the assessment of students’ reading. Our research also introduces the ICONET dataset, an augmented Arabic speech corpus comprising 3160 h of diverse and tailored audio–text pairs designed for fine-tuning ASR models. This study demonstrates the potential of fine-tuning pretrained models for specific linguistic contexts (Arabic), establishing a foundation for future research in ASR and language technology.

Future Internet doi: 10.3390/fi18050250

Authors: Xiaojiang Ding Zhaowei Hu Han Diao Mengting Wu

To address the dual challenges of static privacy budgets and declining data utility in trajectory sharing, this paper proposes a threat-aware differential privacy protection method with dual-factor dynamic adjustment. A lightweight LSTM-Lite model is employed to perceive and quantify environmental threats in real time, thereby dynamically linking threat scores with privacy budget adjustment and reducing the disconnection between threat detection and parameter control. Based on threat scores and semantic sensitivity, a dual-factor controller is designed to enable adaptive regulation of the privacy budget ϵ, thereby improving the adaptability of privacy protection in adversarial environments. Furthermore, a Bayesian-inference-based trajectory consistency reconstruction algorithm is developed to mitigate trajectory jitter and drift caused by Laplace noise. The proposed algorithm calibrates the filtering gain through posterior estimation and incorporates kinematic constraints to improve reconstruction quality. Experimental results show that the proposed method can accurately identify abnormal threats, achieving an F1-score of 0.885, and adaptively regulate privacy intensity under dynamically changing risks. Compared with fixed-budget and single-stage recovery methods, the proposed approach achieves a better balance between privacy protection and trajectory utility, effectively reducing spatial deviation and speed jitter in perturbed trajectories.

Future Internet doi: 10.3390/fi18050249

Authors: Ary Setijadi Prihatmanto Agus Sukoco Rahadian Yusuf Dewi Tresnawati Azizah Zakiah

The rapid evolution of Internet-enabled smart systems has accelerated the adoption of the Internet of Things (IoT), Cyber–Physical Systems (CPS), Big Data, Artificial Intelligence (AI), and Human–Computer Interaction (HCI/AR–VR) across distributed digital ecosystems. Despite these advances, the architectural integration of sensing, information processing, and system-level reasoning remains fragmented, limiting system coherence and accountability. This study derives an architectural foundation through a systematic synthesis of smart system architectures. An activity-mediated socio-technical perspective is employed to analyze diverse paradigms—including IoT-centric frameworks, AI-driven infrastructures, digital twins, Big Data pipelines, and cyber–physical systems—as well as reference architectures such as RAMI 4.0, IIRA, and other representative smart system architectures. Here, activity-mediated denotes an architectural mediation mechanism that coordinates sensing, data-driven reasoning, and human–AI interaction. The synthesis reveals a lack of explicit mechanisms for vertical integration and alignment between bottom-up data flows and top-down goal propagation. In response, this study derives three architectural pillars that integrate interaction, governance, and smart technologies. Their operationalization reveals a structured transformation process in which activity-derived signals are translated into actionable intelligence and adaptive interventions, enabling feedback-driven behavior and cross-layer traceability.

Future Internet doi: 10.3390/fi18050248

Authors: Yong Zhuang Yiheng Jing Wenzhe Yi Xiaoyang Xu Juan Wang

Text-to-image diffusion models such as Stable Diffusion enable high-quality image synthesis from text and are widely deployed due to their open-source nature and low computational requirements. However, this accessibility also makes them attractive targets for misuse, including the generation of not-safe-for-work and otherwise restricted content. In this paper, we propose EvilPrompt, a jailbreak attack that exploits the cross-attention mechanism in Stable Diffusion. The attack operates purely at inference time using plain-text prompts and does not require fine-tuning or modification of model parameters. By selectively reweighting cross-attention for specific tokens, EvilPrompt preserves the overall semantic structure of the prompt while steering the generation toward prohibited content. This enables fine-grained control over malicious semantics without introducing explicit unsafe keywords. We evaluate EvilPrompt on two real-world prompt sets, 4chan and Lexica, each containing 500 prompts. The attack achieves an Attack Success Rate (ASR) of 97.4% on 4chan and 98.0% on Lexica, yielding an overall average ASR of 97.7%. The attack maintains high semantic alignment between prompts and generated images. Bootstrapping Language-Image Pre-training (BLIP) similarity consistently exceeds 0.75 across all categories on both datasets. Human evaluation further confirms high visual realism, with mean scores above 7.0 on a 10-point scale, and strong semantic consistency, with mean scores above 7.3. These results demonstrate that cross-attention manipulation provides an effective and practical jailbreak pathway. We further analyze how commonly used text-level moderation affects the success of such attacks. Although the strongest defense configuration (HateCoT with GPT-4) reduces the ASR to 5.9%, it introduces 21.5 s of additional latency and a cost of $0.01182 per query. Lighter-weight alternatives such as Perspective API leave nearly half (45.0%) of attacks successful. These observations indicate that safeguards acting only on the input or final output are insufficient to capture attention-level manipulations. Overall, our results reveal a fundamental limitation of post-generation safety pipelines when confronted with inference-time control of cross-attention.

Future Internet doi: 10.3390/fi18050247

Authors: Xiang Sun Bin Wang Shaoying Shi Luda Zhao Jun Sun

To address the critical requirements for electromagnetic spectrum orchestration in complex ultra-dense communication environments, this paper proposes an Expert Knowledge-Guided improved NSGA-III framework to solve large-scale frequency assignment problems efficiently. which is built upon the standard NSGA-III architecture as the algorithmic backbone. Traditional multi-objective evolutionary algorithms often struggle with slow convergence and insufficient local search capabilities when navigating high-dimensional, strongly constrained search spaces. In this study, we first introduce a Conflict Graph-based Intelligent Initialization strategy to generate high-quality initial populations by constructing an interference conflict graph based on network topology. Second, a Knowledge-Guided Mutation operator is designed to precisely identify and reconfigure conflicting communication nodes using physical layer indicators. Furthermore, a Best-individual Guided Double-Scale Mutation mechanism is incorporated to dynamically balance global exploration and local exploitation. Experimental results on complex multi-node datasets demonstrate that EKG-NSGA-III significantly outperforms the standard NSGA-III and other baseline algorithms in terms of Hypervolume and Inverted Generational Distance. Specifically, for the 200 nodes scenario, the proposed method achieves a 25.7% improvement in IGD and a 4.2% increase in HV compared to the standard NSGA-III. The proposed algorithm provides a robust and efficient solution for spectrum management in complex urban electromagnetic environments, such as future smart city infrastructures.

Future Internet doi: 10.3390/fi18050246

Authors: Lijun Dong Jingjing Wu Yitong Yang

Significant challenges will be encountered in next-generation cellular networks to achieve both high spectral efficiency (SE) and diverse quality of service (QoS) requirements simultaneously, particularly under stringent bandwidth and power budgets within highly dynamic and dense topologies. To address these challenges, we formulate an optimization problem in a multi-slice non-orthogonal multiple access (NOMA) system with underlay device-to-device (D2D) communications. This problem aims to maximize SE and satisfy user QoS demands by jointly optimizing power allocation and resource block (RB) assignment. To solve this non-convex and NP-hard problem, we propose a resource allocation mechanism based on joint optimization and cooperative multi-agent deep reinforcement learning (MADRL). Specifically, we construct an optimization framework based on successive convex approximation (SCA) and the Lagrange duality method to derive an analytical iterative solution for the optimal power allocation under a given RB assignment, thereby avoiding the inherent discretization error of the action space in pure learning methods. Furthermore, we propose a cooperative multi-agent algorithm based on dueling double deep Q-Network (CMAD3QN) to address the discrete RB assignment problem. Simulation results demonstrate that, compared with benchmark schemes, the proposed scheme exhibits faster convergence speed and significantly enhances system spectral efficiency while ensuring slice isolation and resource constraints.

Future Internet doi: 10.3390/fi18050245

Authors: Lordwin Cecil Prabhaker Micheal Xavier Fernando Mathan Kumar Arumugasamy Neelamegam Devarasu Daisy Merina Rathinarajan

Autonomous vehicle (AV) networks require secure and efficient data processing under strict latency and resource constraints. This paper proposes a secure, lightweight edge-centric framework, SLEVA-AV, for Internet of Things (IoT)-enabled autonomous vehicle communication. The framework integrates multi-modal sensor data processing, lightweight key management, multi-stage encryption, and integrity verification within a unified pipeline. A key derivation function (KDF) is employed to generate session keys using contextual parameters, enabling efficient re-keying during vehicular mobility without repeated handshake overhead. The encryption process combines PRESENT, SPECK, and lightweight encryption algorithm (LEA) ciphers to enhance cryptographic strength, while SHA-256 ensures data integrity. The proposed system is implemented using a CARLA-based simulation environment and validated through CrypTool 2-based cryptographic analysis. Performance evaluation over 10,000 samples demonstrates low latency (0.039–0.794 s), reduced energy consumption (0.0196–0.0589 J), and negligible key management overhead. Comparative analysis with recent state-of-the-art approaches shows improved scalability and efficiency. Security validation through attack simulations demonstrates resistance against brute-force (2336 key space), differential (2−185), replay, and tampering attacks, achieving 100% detection accuracy. The results indicate that the proposed framework strikes a balanced trade-off among security strength, computational efficiency, and real-time performance, and it is suitable for deployment in IoT environments with high mobility and dynamic edge connectivity.

Future Internet doi: 10.3390/fi18050244

Authors: Mutsam A. Jarajreh

This paper presents a novel integration of quadrature amplitude modulation (QAM)-based fast optical orthogonal frequency-division multiplexing (F-OFDM) with machine learning (ML)-based equalization in spatial division multiplexing (SDM) applications, using few-mode fibers (FMFs). The FMFs support four LP modes, resulting in a total of 12 orthogonal modes, each accommodating two polarizations. A digital multiple-input multiple-output channel equalizer is employed at the receiver’s digital signal processing (DSP) unit to effectively mitigate channel crosstalk. The study harnesses supervised ML-DSP techniques, in particular recurrent neural networks (RNNs) and deep neural networks (DNNs), achieving substantial reductions in bit error rates (BERs). In addition, higher-complexity architectures, namely convolutional neural networks (CNNs) and long short-term memory (LSTM) networks, are evaluated to assess the impact of advanced spatial and temporal feature extraction. It is shown that F-OFDM demonstrates superior performance over conventional optical OFDM, particularly when supported by ML techniques. Simulation results reveal that RNNs achieve a BER of 0.0019 over 15 km at 12 Gbaud (worst-case selected channel), showcasing a remarkable 52.5% improvement compared to linear equalization. DNNs achieve a BER of 0.0025, reflecting a 37.5% enhancement. While RNNs perform better, their computational demands pose challenges for real-time applications, and the more complex models (CNN and LSTM) do not provide additional performance gains. The paper also explores cyclic prefix management and subcarrier number strategies in F-OFDM to optimize performance, paving the way for future advancements in SDM networks.

Future Internet doi: 10.3390/fi18050243

Authors: Lluis Mas Jordi Vilaplana Josep Rius Radu Spaimoc Jordi Mateo

The Model Context Protocol (MCP) enables Large Language Models (LLMs) to act as autonomous agents that orchestrate complex workflows over distributed systems, while MCP resolves integration bottlenecks by standardizing agent-to-resource communication. Its current registry relies on an unverified pointer architecture, exposing agentic workflows to supply chain poisoning and dynamic capability mutation (“Rug Pull”) attacks. This paper identifies this gap and proposes a three-layer architectural framework for a Trustworthy MCP Registry. The novelty of our contribution lies not in the individual standards employed (RFC 8615, Sigstore, and JCS/JWS are established technologies), but in their specific composition to address MCP’s unique runtime security requirements: (1) RFC 8615 Well-Known URIs for decentralized server discovery and domain-bound identity; (2) Sigstore Keyless signing to bind server artifacts to audited CI/CD environments without managing long-lived keys; and (3) JSON Canonicalization Scheme (RFC 8785) combined with JWS to provide deterministic, per-message integrity verification of live capability updates. We present a prototype implementation and an experimental evaluation conducted in a controlled, synthetic environment. Results indicate that the cryptographic overhead averages 0.61 ms per signing operation and that the Layer 3 mechanism correctly rejects all 100 simulated Rug Pull attempts, as expected by construction, since an attacker without the server’s private key cannot produce a valid signature. These findings suggest that the proposed approach is feasible and warrants further evaluation in real-world deployment settings.

Future Internet doi: 10.3390/fi18050242

Authors: Tuo Nie Dajiang Yang Xin Guo Wenxuan Zhu Bochao Su

Smart care applications impose increasingly stringent requirements on low-latency execution, privacy preservation, and continuous monitoring. These requirements are driving intelligent services from cloud-centric architectures toward edge-side deployment. When multiple care-related workloads are deployed on resource-constrained edge devices, performance bottlenecks arise not only from model inference itself, but also from process scheduling, inter-process communication, and resource coordination overhead. To address this issue, this paper presents a task scheduling and management platform for multi-workload smart elderly care on a single pure-edge CPU–TPU heterogeneous node. The platform adopts a shared-memory and event-driven synchronization mechanism together with fine-grained process partitioning, thereby establishing a data-sharing and runtime-coordination framework for concurrent multi-workload execution. To evaluate the effectiveness of the proposed platform, experiments were conducted under single-workload, multi-workload, multi-resolution, and long-term runtime settings. The results show that, compared with two baseline schemes, the proposed platform improves the average frame rate by 66.7% and 71.1%, reduces net memory usage by 96.3% and 45.3%, and lowers net power consumption by 46.8% and 37.7%, respectively, under the single-workload setting. Under 10 concurrent workload instances, the system still maintains a stable frame rate of 42.03 ± 0.73 fps, demonstrating strong concurrency scalability. Multi-resolution experiments further indicate that the performance degradation at higher resolutions is mainly constrained by the front-end data supply stage. A continuous 10-day runtime experiment additionally verifies the sustained operating capability and resource stability of the platform under pure-edge deployment. These results demonstrate that node-level shared-memory and event-driven coordination can effectively improve the execution efficiency, scalability, and stability of real-time multi-workload analytics on such pure-edge heterogeneous nodes, providing a useful basis for future extensions to multi-node edge environments and edge–cloud collaborative task scheduling.

Future Internet doi: 10.3390/fi18050241

Authors: Christos Pergamalis Eleftherios Tsampasis Panagiotis K. Gkonis Charalambos N. Elias

The increasing adoption of electric vehicles in residential buildings creates challenges for charging infrastructure management, particularly in pricing services to balance revenue, user satisfaction, and grid stability. Traditional pricing methods, such as fixed rates and time-of-use tariffs, cannot adapt to the dynamic nature of charging demand. We propose a reinforcement learning framework for dynamic pricing of residential EV charging stations. The framework formulates the pricing problem as a Markov decision process and employs proximal policy optimization to learn a pricing policy based on real-time conditions. The state representation includes ten features covering temporal indicators, charging loads, grid status, traffic, and weather. A multi-objective reward function balances revenue, station utilization, grid stability, and user satisfaction. The system is trained on 6878 charging sessions from a residential complex in Trondheim, Norway. Compared with fixed pricing and time-of-use pricing, the proposed method achieves an overall score of 0.569, representing improvements of 32.9% and 48.9%, respectively. Sensitivity analysis confirms that the model remains robust across different demand response assumptions. The main contributions include a custom reinforcement learning environment for residential EV charging and empirical evidence that learned policies outperform traditional pricing approaches.

Future Internet doi: 10.3390/fi18050240

Authors: Dimitrios Dechouniotis Ioannis Dimolitsas

With the advent of the Internet of Things (IoT), the centralized cloud computing service delivery paradigm has been gradually transformed into a cloud continuum that includes edge and fog computing and heterogeneous IoT devices with varying computing and power capabilities [...]

Future Internet doi: 10.3390/fi18050239

Authors: Mohammed A. Makarem Muneef A. Razaz Zead Saleh

The evolving landscape of cybersecurity threats, characterized by increasingly sophisticated and adaptive attackers, poses major challenges to traditional static network defense mechanisms. To address these limitations, this paper proposes an adaptive cyber defense framework that integrates Reinforcement Learning (RL) with Attack Graph (AG) modeling. The interaction between attacker and defender is formulated as a repeated zero-sum stochastic game over a partially observable Attack Graph-guided environment, allowing both agents to adapt their strategies through repeated interaction. Two value-based learning approaches are investigated, namely tabular Q-learning and Deep Q-Networks (DQN), under a unified attacker–defender setting. Experimental results across multiple training scenarios show that defender performance improves substantially as the training budget increases. Under limited training, Q-learning provides a computationally efficient and stable baseline, while DQN requires more training and careful tuning to achieve strong performance. However, with extended training, the DQN-based defender attains the highest win rate, albeit at a significantly greater computational cost. In addition, multi-run statistical comparisons highlight a clear trade-off between defensive effectiveness and runtime efficiency: Q-learning remains far more lightweight, whereas DQN offers stronger asymptotic performance when sufficient resources are available. These findings demonstrate the promise of learning-based adaptive defense over attack graphs while also emphasizing the importance of training budget, computational constraints, and model selection in practical cyber defense deployment.

Future Internet doi: 10.3390/fi18050238

Authors: Ionuț Petre Ella Magdalena Ciupercă Ion Alexandru Marinescu Dragoș Daniel Iordache Alin Zamfiroiu

The growing integration of immersive technologies into education is opening new possibilities for teaching and learning, while also raising concerns about the reliability and potential distortion of knowledge in artificial intelligence-mediated environments. Understanding how users perceive and accept artificial intelligence-generated content in immersive learning systems is therefore essential. This study explores the factors that influence user acceptance of artificial intelligence-driven virtual reality educational applications and explains it through a multidimensional framework that extends the Technology Acceptance Model, the Theory of Reasoned Action, and the Theory of Planned Behavior—a new ED-SCALE model. We innovated the previous models by adding an ergonomic dimension, often overlooked in virtual reality-based education. To test the model, we developed an artificial intelligence-driven virtual reality educational escape room designed to simulate adaptive and interactive learning experiences. Data were collected from 213 participants using a questionnaire measuring subjective norms, perceived behavioral control, attitudes toward artificial intelligence-mediated instruction, perceived informational efficacy, and ergonomic quality. The findings show that ergonomic quality, intuitive interfaces, physical comfort, and social influence play an important role in shaping user trust and long-term adoption intentions. The results suggest that the success of artificial intelligence-driven immersive learning systems depends not only on technological performance but also on user experience and social context, confirming our first hypothesis regarding new variables that are conditional for virtual technology acceptance.

Future Internet doi: 10.3390/fi18050237

Authors: Daniel Schneider Ali Imran Jehangiri Daniel Müller Hannes Frey Maria Anna Wimmer

Radio network planning is critical for 5G deployments, particularly for temporary installations in rural areas where terrain and vegetation significantly impact signal propagation. While empirical path loss (PL) models characterize propagation environments through scenario-specific parameters—leading to inherently noisy predictions at individual sites—machine learning (ML) approaches can predict site-specific path loss from multiple features simultaneously. This study conducts a systematic literature review of rural path loss prediction methods and introduces a novel dataset collected via a 5G nomadic measurement platform in a vineyard environment, capturing real-world propagation characteristics. We present a comprehensive comparison of machine learning and interpretable machine learning techniques, demonstrating that vegetation dynamics (quantified through the Normalized Difference Vegetation Index, NDVI) is an important driver of path loss variability when combining data across seasonal campaigns—though not within individual campaigns, where distance dominates. Cross-campaign NDVI transfer, however, is sensitive to satellite resolution, which appears to conflate vine canopy with seasonally managed inter-row ground cover. In cross-campaign transfer, XGBoost proves substantially less susceptible to NDVI-induced degradation than Explainable Boosting Machines (EBM), and a hybrid Log-Normal Shadowing (LNS) and XGBoost model confirms that NDVI captures seasonal variability more effectively than empirical path loss parameters alone. Still, the data captured the expected seasonal trend between April and June 2025, from which our interpretable models derived useful propagation insights. Tree-based models like Random Forest and XGBoost achieved the highest prediction accuracy (R2 up to 0.924 on individual campaigns, 0.891 on combined data, and up to 0.945 (individual) and 0.907 (combined) with antenna pattern-corrected path loss), while explainable boosting machines achieved near-parity (R2 up to 0.919; 0.876 on combined data) with the advantage of interpretability. Among individual campaigns, June—with densest canopy cover—yielded the highest R2 values. These findings provide actionable insights for optimizing temporary 5G networks in precision agriculture and other rural applications.

Future Internet doi: 10.3390/fi18050236

Authors: Zhaohui Liu Haocheng Yang Wenjie Xie

Deep-learning-based vulnerability detection has made substantial progress, but two limitations remain prominent. Sequence-based methods linearize source code and thus weaken the explicit modeling of control-flow and data-flow dependencies. Graph-based methods preserve program structure, yet conventional graph neural networks still have difficulty capturing long-range interactions in large code property graphs (CPGs). In addition, standard CPGs usually lack explicit variable semantics and security-critical node roles, which limits their ability to represent vulnerability-relevant program behavior. To address these issues, we propose SV-GEN, a vulnerability detection framework that combines large-language-model-driven semantic enhancement with hybrid sequence-graph learning. The novelty of SV-GEN lies in introducing a semantically enriched code property graph, termed Sem-CPG, which augments conventional CPGs with variable semantic roles and security-oriented node labels, and in coupling this representation with an adaptive fusion mechanism over structural and sequential views. Specifically, we use a large language model as an external semantic annotator to assign variable roles and identify source, sink, and sanitizer nodes, and then encode the resulting Sem-CPG with a Graph Transformer while modeling the code sequence with GraphCodeBERT. A learnable gating module is further used to adaptively fuse the graph-level and sequence-level representations for final prediction. Experiments on Devign, ReVeal, and DiverseVul show that SV-GEN achieves competitive or superior overall performance across benchmarks, with particularly strong improvements on the large and highly imbalanced DiverseVul dataset.

Future Internet doi: 10.3390/fi18050235

Authors: Sindhu Nair Y. S. Rao

Text summarization involves analyzing large amounts of text, selecting the salient text features, and arranging them coherently. The graph-based TextRank and statistical topic modeling are unsupervised approaches for generating an extractive synopsis. Deep learning models are supervised, data-driven, and pre-trained on a huge corpus of data, making a significant contribution to automatic text summarization systems. Despite grammatical correctness and coherence, deep learning-based summarization systems are prone to factual inconsistency. This has hindered the applicability of transformer-based summarizers, particularly in critical domains where misleading summarization systems can lead to severe consequences due to their significant social impact. This work proposes an ingenious hybrid hierarchical approach that combines unsupervised approaches, such as the TextRank algorithm and Latent Dirichlet Allocation (LDA)-based summaries, with contemporary transformer-based language models. When validated on three benchmark summarization datasets, empirical results prove that our hybrid hierarchical transformer-based approach mitigates the factual inconsistency problem inherent in abstractive summarization. The improved summary consistency score of the abstractive summaries generated with our multilevel hybrid approach, in comparison to the fine-tuned baseline transformer-based language models, increases trust in transformer-based summarizers.

Future Internet doi: 10.3390/fi18050234

Authors: Antanios Kaissar Ali Bou Nassif Ahmed Bouridane

Quantum computing introduces new computational capabilities that can support advanced cybersecurity solutions when combined with machine learning. In recent years, quantum machine learning (QML) has emerged as a promising approach for enhancing network intrusion detection systems (IDS), particularly for analyzing complex and high-dimensional network traffic. This paper presents a systematic survey of QML techniques applied to network intrusion detection. The survey reviews peer-reviewed studies published up to January 2026 that employ quantum, hybrid quantum–classical, and quantum-inspired learning models for IDS. The selected studies are analyzed with respect to the algorithms used, intrusion detection datasets, and evaluation metrics reported. The analysis shows that most current approaches rely on simulated quantum environments and legacy datasets, while evaluation practices remain inconsistent across studies. These findings highlight the early developmental stage of QML-based IDS and the need for standardized evaluation protocols and more realistic experimental settings. Finally, open challenges and future research directions are identified to support the development of reliable, scalable, and practically deployable QML-based intrusion detection systems.

Future Internet doi: 10.3390/fi18050233

Authors: Rodrigo Eduardo Arevalo-Ancona Manuel Cedillo-Hernandez Antonio Cedillo-Hernandez Francisco Javier Garcia-Ugalde

Deepfake content forgery compromises the integrity of digital media and the protection of personal identity, making its detection essential for preserving trust and enabling effective forensic analysis. Most deepfake detection approaches focus on global classification with a binary decision, which is inadequate for precise localization of manipulated regions. This limitation becomes particularly evident under image processing distortions. This paper proposes a dual-decoder architecture for the detection and segmentation of original and deepfake facial manipulations. Unlike conventional single-decoder segmentation models, the proposed approach introduces two decoding branches that learn complementary feature representations of authentic and forgery facial textures. In addition, attention mechanism modules are incorporated to refine encoder features based on decoder context, introducing adaptive feature selection during reconstruction. This architectural design reduces feature interference during reconstruction and enhances the localization of subtle inconsistencies introduced by deepfake manipulations. This approach generates complementary masks for real and forged regions, providing more precise boundary delineation. Experimental results highlight the robustness of the proposed method under image processing distortions, achieving intersection over union (IoU) scores of 0.9387 for real faces and 0.9254 for deepfake segmentation. These results underscore the effectiveness of the dual-decoder architecture in accurately detecting and localizing deepfake facial manipulations.

Future Internet doi: 10.3390/fi18050232

Authors: Nikolay Lipskiy Stephen V. Flowerday

Healthcare artificial intelligence (AI) is moving from the laboratory into the infrastructure of care. As these systems become embedded in imaging, electronic health records, triage, and clinical decision support, their failures can affect not only individual encounters but also institutions and patient populations. Yet governance still centers on model development, local validation, and one-time compliance, with limited attention to cross-site failure after deployment. This article examines how public health preparedness can help close that gap. It presents a conceptual analysis grounded in two cases: a pneumonia-screening convolutional neural network that learned institutional confounders rather than portable clinical signals, and a widely deployed sepsis prediction model whose external performance and alert burden fell short of developer claims. Together, these cases reveal five governance features of systemic healthcare AI risk: population-level exposure, cascade effects across shared infrastructures, unequal vulnerability, delayed recognition, and coordination needs beyond any single institution. In response, we propose a tripartite framework combining stronger pre-deployment assurance, post-deployment surveillance with escalation thresholds, and tertiary response through investigation, rollback, remediation, and cross-site learning. The argument is not that AI failures are epidemics, but that high-impact clinical AI systems now function as critical digital health infrastructure requiring preparedness alongside lifecycle oversight.

Future Internet doi: 10.3390/fi18050231

Authors: Qi Wang Longjuan Wang Weiwei Zhao

With the development of artificial intelligence technologies, various models have become mainstream methods in malicious code detection. The application of these models brings significant advantages in automation, intelligence, and proactivity. However, as malicious code continuously evolves and updates, discrepancies emerge between the distribution of malicious code characteristics and those in the model’s training dataset. This leads to a decline in the model’s detection performance, a phenomenon known as concept drift. Existing research still lacks a systematic review that comprehensively explains how concept drift impacts malicious software detection models and how to effectively address this issue. Therefore, this paper reviews and analyzes the current research on this topic in five aspects: enhanced machine learning methods, deep neural network models, graph neural network models, continual learning strategies, and meta-learning strategies. By analyzing, comparing, summarizing, and discussing the various methods, this paper aims to provide insights into future improvements for reducing concept drift in malicious code detection models. This paper helps researchers understand the basic principles behind concept drift, current mitigation techniques, existing challenges, and future development directions, providing support for further research and improvement of existing methods.

Future Internet doi: 10.3390/fi18050230

Authors: Muhammad Nawaz Khan Inam Ullah Sokjoon Lee Mohsin Shah

The Internet of Things (IoT) and quantum computing revolutionized the era of conventional and classical computing into a new paradigm of Quantum-IoT where qubits and entanglement make IoT more interactive, powerful, and secure. They facilitate numerous tasks by increasing productivity and efficiency, paving the path for a smarter and more connected future. In this article, we propose a novel authentication scheme, “Securing the Internet of Things, Lightweight Mutual Authentication Based on Quantum Key Distribution (LMA-QIoT)”. LMA-QIoT enables mutual authentication using various parameters including quantum key distribution, symmetric keys and timestamps, as well as additional quantum random numbers. All these parameters play a crucial role in thwarting man-in-the-middle, backtracking and nonce reuse attacks. The evaluation of LMA-QIoT demonstrates that quantum key distribution and quantum numbers enhance system performance by reducing CPU usage by 25% and memory requirements 30% compared to an IoT edge-based system and without a server, respectively. In the reconfiguration ratio, the efficiency metric grows exponentially and remains constant on the initial line in edge-server-based systems. In comparison, LMA-QIoT confirms a much reduced overall computational complexity by 16.64%, with the lowest computational cost of O(n2). At 1024 Bytes, the original data length and increased data length (normalized) sizes stay constant with 2logn(klogn). Comparing the total overhead, LMA-QIoT demonstrates a reduction of 33 ms, which corresponds to approximately 16.63% less than the baseline mechanisms.

Future Internet doi: 10.3390/fi18050229

Authors: Mehdi Moucharraf Mohammed Ridouani Fatima Salahdine Naima Kaabouch

The rapid growth of IoT ecosystem has significantly increased the potential threats and attack vectors in the recent times, thereby requiring intrusion detection mechanisms that are highly accurate and scalable in nature. This paper presents a hybrid intrusion detection system that involves the usage of both supervised and unsupervised machine learning methods to detect different kinds of attacks present in the IoT network. In the first step, Random Forest-based feature extraction is adopted to determine the most important features from the highly dimensional network traffic data. After this, the extracted features are compressed using the Deep AutoEncoder model into latent features that are fed into multiple classifiers to classify the traffic into various IoT attack classes and normal traffic class. Specifically, the classifiers used in the process include XGBoost, SVM, Logistic Regression, Naive Bayes and Multilayer Perceptron models. Multiple IoT benchmark datasets, such as N-BaIoT and CICIoT2023, are used to evaluate the performance of the proposed hybrid intrusion detection system. It was found that the XGBoost classifier performed better than others, obtaining an accuracy rate of 99.63% and 98.94% on the N-BaIoT and CICIoT2023 datasets, respectively. The above-discussed results show the high potential of the proposed architecture for generalization in various IoT environments. From the results, one can see that it is highly effective to integrate deep learning for extracting features from data and using boosting techniques for classification to develop an efficient IDS system.

Future Internet doi: 10.3390/fi18050228

Authors: José-Amelio Medina-Merodio Mikel Ferrer-Oliva Alejandro Ruiz-Zambrano José Fernández-López Luis De-Marcos

This study proposes a methodological framework for extending Malware Information Sharing Platform (MISP) taxonomies in the domain of Dark Web drug forums through the integration of large language models (LLMs) and Human-in-the-Loop (HITL) validation. The research addresses the existing ontological gap between traditional MISP taxonomies, focused on technical or chemical indicators, and the linguistic and morphological complexity of illicit digital markets. By modelling the primary physical form as an ontological predicate with mutually exclusive values (for example, powder, pill–tablet–capsule, liquid, and plant-matter), the proposed approach captures the material dimension of the discourse, enhancing semantic disambiguation and forensic traceability. The Mistral 7B model was used in the morphology-classification stage conducted on a stratified analytical subset of 2904 drug-related Dark Web posts, extracted from a final corpus of 6456 posts after data cleaning and relevance filtering. In the first pass, 76.48% of posts were directly assigned to one of the base morphological categories, while 23.52% were labelled as unclear and subsequently reviewed through the HITL stage. Following HITL refinement and full reclassification, the proportion of posts labelled as unclear decreased from 23.52% to 11.29%, corresponding to a 51.99% relative reduction in ambiguity. Network visualisation with VOSviewer revealed three major discursive axes—recreational–commercial, pharmaceutical–opioid, and transnational–logistical—reflecting the hybrid semantic structure of digital drug markets. The results show that combining LLM-based inference with expert oversight improves the interpretability, reproducibility and ontological robustness of cyberintelligence models, offering a replicable framework for other sensitive domains such as terrorism or child exploitation.

Future Internet doi: 10.3390/fi18050227

Authors: Kuljaree Tantayakul Adisak Intana Aung Aung Riadh Dhaou

One of the challenges in managing mobility in a heterogeneous network domain remains a significant challenge in Software-Defined Networking (SDN). While SDN has effectively facilitated intra-domain mobility, inter-domain mobility has been a major issue, leading to service interruptions, packet loss, and unstable communication sessions. This article presents a new concept in mobility management: a hybrid SDN-LISP network that facilitates inter-domain communication by integrating SDN with the Locator/Identifier Separation Protocol (LISP). The main idea is to introduce a new event-based orchestration model that uses OpenFlow Packet-In messages to provide instantaneous updates to Endpoint Identifiers-to-Routing Locators (EID-to-RLOC) mappings, unlike traditional LISP, which relies on timers for updates. The proposed framework has been implemented and evaluated on a Mininet-WiFi testbed under various mobility conditions. The results obtained from the experimental evaluation reveal that packet loss is reduced by 92.32% when using the proposed framework over the conventional SDN Mobility approach. Although there is a slight increase in jitter overhead due to LISP encapsulation of 0.628 ms, the framework does not compromise Transmission Control Protocol (TCP) session continuity. In addition, the control plane synchronization time is also minimized to 277.5 ms. This reveals that the proposed framework is a stable mobility solution that does not depend on any conventional IP mobility solutions and can be used in future network environments requiring seamless inter-domain connectivity.

Future Internet doi: 10.3390/fi18050226

Authors: Zhiling Zeng Yuxuan Jiang Na Niu

The increase in delay-sensitive application tasks requires heterogeneous edge clusters to maintain low online latency and energy efficiency without relying on rigid scheduling policies. To address this, we propose HERO (Hybrid Energy-aware Ranking and Optimization), a lightweight collaborative scheduling framework. HERO utilizes a perturbation-based communication-aware multi-layer perceptron (MLP) predictor to quantify global time sensitivity and discover latent time slack in non-critical paths. A hybrid budget mechanism then converts this slack into customized DVFS decisions. These decisions are based on the inherent computational load and topological criticality to optimize energy consumption. A communication-aware hole-filling strategy dynamically recovers sporadic idle times fragmented by heterogeneous communication overhead. Extensive simulations were conducted across varying DAG depths, parallelism levels, and system utilizations. Compared to state-of-the-art algorithms (NSGA-II, SSA, TOM, and DPMC), HERO reduced the completion time by an average of 10.89% under high-density topologies, and achieved up to 4.04% energy savings across varying task depths.

Future Internet doi: 10.3390/fi18050225

Authors: Haiyang You Zhenhua Wei Wenpeng Wu Chenxi Li Jianwei Zhan Zhaoguang Zhang

Addressing the high-dimensional, strongly constrained multi-objective optimization problem of air–ground collaborative jamming scheduling in complex terrain, existing methods are often limited by incomplete modeling and low optimization efficiency in discrete feasible regions. This paper proposes a Terrain-Aware Multi-Scale Discrete Operator (TA-MSDO). A joint optimization model integrating discrete terrain characteristics and practical combat constraints is first constructed. Then, by leveraging the topological adjacency of terrain units, TA-MSDO employs a block-level crossover and a multi-scale mutation mechanism, replacing traditional continuous genetic operations to enable efficient and directional exploration of the discrete feasible region. Integrating TA-MSDO into the NSGA-III framework yields the enhanced ENSGA3 algorithm. Experimental results in a typical hilly terrain scenario demonstrate that ENSGA3 achieves a statistically significant performance improvement over the decomposition-based MOEA/D algorithm in terms of maximum achievable suppression effectiveness and hypervolume. As a comprehensive metric integrating convergence and Pareto frontier coverage, hypervolume further verifies the superior comprehensive optimization capability of the proposed algorithm. Meanwhile, compared with other classic mainstream multi-objective optimization algorithms including NSGA-II, standard NSGA-III and SPEA2, the proposed algorithm exhibits clear positive advantages in the upper bound of suppression effectiveness for elite solutions and operational stability across random initializations, with a favorable trend in Pareto frontier coverage for multi-objective collaborative optimization. This work provides an effective solution for jamming resource scheduling in complex battlefield environments.

Future Internet doi: 10.3390/fi18050224

Authors: Fahad Alanazi Osama Rabie

The aviation industry operates in a security-sensitive environment where customer feedback may contain not only expressions of satisfaction or dissatisfaction but also threatening or violent language with potential security implications. While conventional sentiment analysis effectively captures customer opinions, it remains insufficient for identifying security-relevant linguistic cues that could signal risks requiring proactive intervention. This study addresses this gap by introducing a security-aware ambient intelligence framework for detecting violent language in airline customer reviews. This framework supports intelligent internet-based monitoring systems and real-time threat detection. We present the first annotated dataset of airline reviews specifically labeled for violent and threatening content, derived from 3629 reviews and balanced through manual resampling to achieve equal representation across positive, neutral, negative, and violent classes. The proposed framework employs VADER-based sentiment analysis for initial polarity estimation, combined with a validated annotation process to identify violent or threat-related content, followed by comprehensive feature engineering combining TF-IDF (2000 features) with text statistics and sentiment scores. We systematically evaluate individual classifiers (Random Forest, Decision Tree, SVM, Naive Bayes) against ensemble methods (Voting, Stacking, Boosting) using accuracy, precision, recall, F1-score, and ROC AUC metrics. Results demonstrate that Stacking achieves the highest raw performance (98.57% accuracy, F1-macro 0.9856), while Naive Bayes offers an optimal balance between effectiveness and computational efficiency (81.79% accuracy, F1-macro 0.8172, training time 0.03 s). This is the first dataset and framework designed for security-aware analysis of airline reviews. The selected Naive Bayes model achieves per-class F1-scores of 0.9978 for neutral, 0.7814 for negative, 0.7482 for violent, and 0.7415 for positive reviews, with a macro-average ROC AUC of 0.7123. The framework is deployed with serialized components enabling real-time prediction, supporting both single-review analysis and batch processing for integration into airline security monitoring systems. This work establishes a foundation for security-aware natural language processing in critical infrastructure contexts, bridging the gap between conventional sentiment analysis and proactive threat detection.

Future Internet doi: 10.3390/fi18050223

Authors: Asif Yasin Seyed Ebrahim Hosseini Muhammad Nadeem Shahbaz Pervez

As 5G Standalone Core networks grow, Application Programming Interface (APIs) have become a key part of how network systems talk to each other. They allow different functions to share data and complete tasks quickly. However, this also makes them targets for attacks. 5G Standalone Core networks rely on Service-Based Architecture (SBA), where network functions communicate through exposed APIs. These APIs are attractive targets for cyberattacks because they are externally accessible, handle sensitive control-plane operations, and exchange structured data using Hypertext Transfer Protocol version 2 (HTTP/2) and JavaScript Object Notation (JSON) protocols. Most older security tools work using fixed rules, which cannot always detect new or changing threats. This study aimed to fix that gap by using Artificial Intelligence to make API security smarter. Two AI models were tested: Long Short-Term Memory (LSTM), which learns from past traffic and Reinforcement Learning (RL), which learns by adapting to network behavior. Both were used to assess API traffic and assign a real-time risk score. Synthetic traffic was created using Python, including both normal API calls and different types of attacks like Distributed Denial-of-Service (DDoS), brute force, and Structured Query Language (SQL) injection. The results show that both LSTM and RL models were better than traditional rule-based systems. They found more threats, gave fewer false alerts, and responded faster. RL was especially strong at handling unknown or changing attacks. Experimental results show that the proposed LSTM and RL models achieved approximately 95% detection accuracy, significantly outperforming the static rule-based baseline model, which achieved 58% accuracy. The results demonstrate the effectiveness of adaptive AI-based security mechanisms for detecting evolving API threats. This research shows that AI can help protect 5G APIs in a smarter and more flexible way. It can support telecom networks by making threat detection faster, more accurate, and ready for future challenges.

Future Internet doi: 10.3390/fi18040222

Authors: Antón Román-Portabales Martín López-Nores

Short-term load forecasting is a key capability for smart-grid operation, but real smart-meter streams are affected by missing values, communication noise, and non-stationary consumption patterns. This paper studies forecasting using raw smart-meter data collected from domestic consumers in a medium-sized city in southern Spain. In particular, we assess Hierarchical Temporal Memory (HTM), a biologically inspired online sequence learner, against a family of Long Short-Term Memory (LSTM)-based recurrent baselines. HTM offers continual adaptation and avoids a separate training phase, whereas LSTM relies on offline supervised training and may require retraining or fine-tuning under distribution shift. For five-step-ahead forecasting, HTM achieved a test RMSE of 251 kWh (about 15% of average consumption). After hyperparameter optimization, the best tested LSTM configuration achieved a test RMSE of approximately 250 kWh under clean conditions, indicating nearly identical point accuracy between the two approaches. Under synthetic Gaussian-noise injection, however, HTM remained comparatively stable, whereas the optimized LSTM configuration degraded markedly under the tested perturbation protocol. In addition, HTM exhibited a lower runtime in the tested CPU-based implementation. These findings suggest that HTM is a viable online alternative for aggregated smart-meter forecasting, offering competitive accuracy together with a favorable operational profile under the specific evaluation setup considered here.

Future Internet doi: 10.3390/fi18040221

Authors: Jianzhuo Li Ye Yu Zhao Dai Panyu Dai

Driver emotion recognition is a crucial method for reducing traffic accidents. Most existing research focuses on experienced drivers as the primary research subjects, overlooking novice drivers, who are inexperienced in driving. However, novice drivers can easily lose control of their emotions due to the high mental load during driving, which can lead to serious traffic accidents. Therefore, to recognize the emotions of novice drivers for timely warnings, we propose an emotion recognition model based on multimodal information. The model consists of a facial feature extraction module, an eye movement feature extraction module and a classifier. The facial feature extraction module uses the ViT-B/16 to extract the facial features of novice drivers. The eye movement feature extraction module is a hybrid network containing Bi-LSTM and Transformer. It extracts eye movement features of novice drivers. Facial features and eye movement features are fused and fed to the classifier. The classifier can output the five major emotion categories of surprise, anger, calm, happy, and other for novice drivers. The experimental results demonstrate that our model accurately recognizes the emotions of novice drivers with an accuracy of 98.72%, surpassing that of other models.

Future Internet doi: 10.3390/fi18040220

Authors: Hassan N. Noura Zaid Allal Ola Salman Khaled Chahine

The increasing reliance on smart grid communication systems has significantly raised the demand for robust cybersecurity measures to defend against advanced threats. This paper proposes a two-tier classification framework to enhance the detection and diagnosis of man-in-the-middle attacks within smart grid communication protocols. Initially, the model detects the presence of an attack and then identifies the specific type of man-in-the-middle attack through subsequent inferences. To achieve this, the “Man-in-the-Middle Attacks Targeting Modbus TCP/IP and MMS Protocols in the Smart Grid” dataset was carefully preprocessed and analyzed to better understand the underlying hidden characteristics. This understanding, coupled with existing works on fault detection and diagnosis, facilitated the engineering of new features from the original dataset. Four classifiers were employed in each tier: Random Forest, XGBoost, LightGBM, and CatBoost. The first tier exhibited exceptional performance, with the CatBoost framework achieving 99.6% accuracy. The second tier also demonstrated strong results, with the same model achieving 99.1% accuracy. Systematic model explainability was conducted using SHapley Additive exPlanations for both tiers and revealed that the highest accuracy was achieved using five features for the first and six for the second. The average inference time was approximately 4.76 milliseconds. The proposed framework is accurate, fast, interpretable, lightweight, and well-optimized for direct implementation in smart grid systems to detect and diagnose man-in-the-middle attacks.

Future Internet doi: 10.3390/fi18040219

Authors: Diego Vergara Pablo Fernández-Arias

Artificial intelligence (AI) is driving profound changes in the digital transformation of society [...]

Future Internet doi: 10.3390/fi18040218

Authors: Aymé Escobar Díaz Ricardo Rivadeneira Walter Fuertes Washington Loza

Hate speech on social media reproduces norms of inequality and gender stereotypes, disproportionately affecting women. This study proposes a hybrid approach that integrates emotional tone classification with explicit hostility detection to strengthen preventive moderation. We constructed a corpus from three open data sets (1,236,371 records; 1,003,991 after ETL) and represented the text using TF-IDF and contextual RoBERTa embeddings. We trained individual models (RoBERTa fine-tuned, Random Forest, and XGBoost) and a stacking metamodel (Gradient Boosting) that combines their probabilities. On the test set, the ensemble outperformed the base classifiers, achieving accuracy of 0.93 in hate detection and 0.90 in emotion classification, with an AUC of 0.98 for emotion classification. We implemented a RESTful API and a web client to validate the moderation flow before publication, along with an administration panel for auditing. Performance tests in a prototype deployment (Google Colab exposed through an Ngrok tunnel) provided proof-of-concept validation, revealing concurrency limitations from around 300 users due to infrastructure constraints. In general, the results indicate that incorporating emotional tone analysis improves the model’s ability to identify implicit hostility and offers a practical way to promote safer digital environments. The probabilistic outputs produced by the ensemble model were subsequently analyzed using the Bayesian Calibration and Optimal Design under Asymmetric Risk (BACON-AR) framework, which serves as a mathematical post hoc decision layer for evaluating classification behaviour under unequal error costs. Rather than modifying the trained architecture or improving its predictive performance, the framework identifies a cost-sensitive operating threshold that minimizes the total expected risk under the selected asymmetric cost configuration. The experiments were conducted using an English-language data set; therefore, the findings of this study are limited to hate speech detection in English.

Future Internet doi: 10.3390/fi18040217

Authors: Gianluigi Ferrari

As indicated on the journal’s website, Future Internet fosters contributions to the future Internet ecosystem, which, in turn, is expected to lead to significant improvement in well-being in all spheres of human life (private, public, professional) [...]

Future Internet doi: 10.3390/fi18040216

Authors: Askhat Bolatbek Ömer Faruk Beyca Batyrbek Zholamanov Madiyar Nurgaliyev Gulbakhar Dosymbetova Dinara Almen Ahmet Saymbetov Botakoz Yertaikyzy Sayat Orynbassar Ainur Kapparova

The rapid growth of large-scale Internet of Things (IoT) deployments in urban environments requires accurate and energy-efficient localization methods for low-power wireless devices. In long-range wide-area networks (LoRaWAN), traditional GPS-based positioning is often impractical due to energy consumption constraints and signal propagation challenges in urban areas. This study proposes a hybrid localization system that integrates weighted centroid localization (WCL) with a machine learning (ML) regression model to improve outdoor positioning accuracy. The proposed approach first estimates approximate transmitter coordinates using a physically grounded WCL method based on received signal strength indicator (RSSI) measurements. These initial estimates are subsequently refined by ML models trained to learn nonlinear residual corrections. In addition to random partitioning, a spatial data splitting strategy is proposed and evaluated using a publicly available LoRaWAN dataset. The experimental results demonstrate that the hybrid WCL framework combined with a multilayer perceptron (MLP) significantly outperforms other ML models. The proposed method achieves a mean localization error of 160.47 m and a median error of 73.78 m. Compared to the baseline model, the integration of WCL reduces the mean localization error by approximately 29%, highlighting the effectiveness of incorporating physically interpretable priors into localization models.

Future Internet doi: 10.3390/fi18040215

Authors: Youjian Yu Zhaowei Song Sifeng Zhu Qinghua Zhang

To improve vehicle user service quality and address data privacy and security issues in intelligent transportation vehicle networking systems, a three-tier communication architecture with cloud-edge-end collaboration was designed in this paper. A Bayesian decision criterion was utilized to divide user data segments into fine-grained slices based on their privacy levels, and differential privacy techniques were applied to protect the offloaded data. To achieve multi-objective optimization between user service quality and data privacy and security, the problem was formulated as a constrained Markov decision process. A communication model, a caching model, a latency model, an energy consumption model, and a data-fragment privacy protection model were designed. Additionally, a deep reinforcement learning algorithm based on the actor–critic approach was proposed for the collaborative and centralized training of multiple intelligent agents (CTMA-AC), enabling multi-objective optimization decision-making for the protection of offloaded private user data. Simulation experiments demonstrate that the proposed multi-agent collaborative privacy data offloading protection strategy can effectively safeguard private user data while ensuring high service quality.

Future Internet doi: 10.3390/fi18040214

Authors: Xuxun Ye Anliang Cai

The millisecond-level volatile fluctuations in workloads in large-scale intelligent computing clusters pose significant challenges to traditional electricity markets. Constrained by optical–electrical–optical conversion bottlenecks, these markets struggle to achieve real-time response and risk substantial social welfare loss. Leveraging existing fiber-optic infrastructure to build All-Optical Networks (AONs) presents a cost-effective evolutionary path. This paper develops a distributed energy trading strategy based on all-optical multicast. By utilizing the physical multicast properties of the underlying light-tree architecture instead of traditional protocols, the proposed strategy bypasses end-to-end latency constraints. This enables rapid transaction synchronization and dynamic tracking of social welfare optima within millisecond-level time-slots. Simulation results demonstrate that the proposed scheme elevates the transaction saturation threshold by two orders of magnitude compared with traditional strategies, effectively breaking the physical locking effect of latency on system throughput. Across various topologies, the social welfare gains exceed those of conventional schemes by more than 20 times. This study validates the engineering value of all-optical architectures for high-frequency trading and provides critical technical support for ultra-dynamic power trading algorithms.

Future Internet doi: 10.3390/fi18040213

Authors: Norah Alrusayni Asma A. Al-Shargabi

In resource-constrained environments, distributed split learning allows for collaborative training; however, the system suffers from high communication overhead and is sensitive to system heterogeneity. Despite advances in IoT data reduction and distributed learning, existing approaches treat heterogeneity, adaptability, and communication efficiency as separate problems. As a result, the Heterogeneity-Aware Dynamic Federated Split Learning with Adaptive Compression (HADFL-AC) framework is proposed, enabling adaptive adjustment of communication payloads to instantaneous bandwidth conditions during training. This approach distinguishes itself by focusing on feature-representation-level adaptation, offering seamless transitions between linear PCA, nonlinear Tiny Autoencoder (TinyAE), and hybrid PCA–AE compression methods without requiring changes to architecture or retraining. Experiments were conducted using the CIFAR10 and CI=NIC datasets with a lightweight ResNet-18 backbone under Dirichlet-based non-IID data partitioning and fluctuating network scenarios. HADFL-AC achieves significant communication reductions of 80.86% on CIFAR-10 and 77.2% on CINIC-10, as well as significant reductions in training time and energy consumption. In addition, the framework achieved these gains while maintaining competitive performance, reaching 79.58% on CIFAR-10 and exhibiting stable convergence on CINIC-10. Consequently, the results demonstrate that leveraging network heterogeneity as an adaptive signal facilitates efficient and scalable distributed learning while effectively balancing communication efficiency and model accuracy.

Future Internet doi: 10.3390/fi18040212

Authors: Jan Sawicki Maria Ganzha Marcin Paprzycki Jihui Han Subhajit Sahu

Network dismantling—the targeted removal of nodes to degrade large-scale connectivity—plays a central role in resilience analysis, epidemic containment, and systemic-risk mitigation. Recent work shows that dismantling performance depends strongly on mesoscale modular structure, suggesting that community-aware strategies may offer advantages over classical centrality-based heuristics. In this work, we perform a large-scale, systematic evaluation of dismantling strategies and introduce gateways as a new mesoscale dismantling concept. While similar experiments exist using degree- and betweenness-based dismantling strategies, we check a new strategy based on gateways, which capture asymmetric entry points into communities and generalize the notion of inter-community connectors. Furthermore, we process a massive dataset of 568,584 LFR benchmark graphs, covering a wide range of degree distributions, community sizes, and mixing parameters. For evaluation, we use both extrinsic (ARI, NMI, FMI, VI) and intrinsic (Modularity, Coverage, Performance, Average Conductance, Average Internal Density) metrics. We find that across parameter regimes and evaluation metrics, classical strategies (degree, betweenness, community connections) and gateway-based dismantling exhibit broadly similar performance. Our results also corroborate recent findings that dismantling effectiveness is robust to the specific partitioning algorithm and that inter-community connectivity plays a dominant role in global fragmentation. The evaluation provides large-scale evidence that gateway-aware dismantling captures an operationally relevant mesoscale mechanism as good as previous approaches and motivates further empirical studies on real networks and cost-aware settings.

Future Internet doi: 10.3390/fi18040211

Authors: Timothy Malche Govind Murari Upadhyay Sumegh Tharewal Vipin Balyan Vikash Kumar Mishra Gunjan Gupta Pramod Kumar Soni

In this study, we propose a wearable fall detection system that combines wearable sensors, TinyML model, and IoT-based communication for real-time monitoring and detection of falls. The system is designed for resource-constrained IoT devices where memory, power, and processing capacity are limited. The system works by collecting body motion data using accelerometer sensors placed on the human body. The data is then processed using a feedforward neural network trained on preprocessed signals. The trained model is quantized so that it can run on low-power embedded hardware with small memory size. The model performs inference directly on the device. This reduces latency and avoids sending raw sensor data to the cloud. When a fall is detected, the result is sent through Bluetooth to a gateway. The gateway forwards the data to a cloud server using the MQTT protocol. The cloud stores the data and supports monitoring and analysis. The experimental results show that the quantized TinyML model achieves 98.40% accuracy with more than 80% F1-score and more than 99% recall. The deployed model uses only ∼5 KB of RAM and ∼40 KB of flash memory. The inference time is 7 ms per class. These results show that wearable sensing with quantized TinyML models and IoT communication can provide fast and reliable fall detection for real-world safety monitoring systems.

Future Internet doi: 10.3390/fi18040210

Authors: Otuekong Ekpo Valentina Casola Alessandra De Benedictis Philip Asuquo Bright Agbor

Smart cities are increasingly dependent on interconnected transportation systems; however, this connectivity exposes smart mobility networks to significant cybersecurity risks. Traditional Intrusion Detection Systems are ill-equipped for this environment, as they are designed for isolated systems or fixed network boundaries. Thus, they struggle to secure the complex and heterogeneous smart mobility networks, where various protocols and resource-constrained edge devices require more adaptive solutions. To address this limitation, we propose a novel hybrid deep learning framework that combines convolutional neural networks for spatial feature extraction, long short-term memory networks for temporal pattern recognition, and an attention mechanism for adaptive feature weighting, together forming a context-aware Intrusion Detection System. Our approach is evaluated across six benchmark datasets spanning vehicular networks, IoT ecosystems, cloud computing, and 5G environments—VeReMi Extension, CICIoV2024, Edge-IIoTset, UNSW-NB15, Car Hacking, and 5G-NIDD—a deliberately diverse selection that represents the heterogeneous nature of real-world smart mobility networks. Empirical evaluation using three different random seeds reveals the proposed framework achieves detection accuracy exceeding 98% on each dataset, a mean F1 score of 98.94%, and an inference latency of just 4.96 ms per sample. Our results show that the proposed model achieves consistently high detection performance across six heterogeneous benchmark datasets, making it a potentially robust candidate for real-time intrusion detection in smart mobility systems.