Future Internet

37 pages, 4642 KB

Open AccessArticle

A Subsystem-Level Validation and Simulation Framework for a 12-DoF Biped Robot with Deep Reinforcement Learning Locomotion

by Michael Felipe Cifuentes-Molano, Kevin David Ortega-Quiñones, Byron Hernandez, Germán Andrés Holguín-Londoño and Mauricio Holguín-Londoño

Future Internet 2026, 18(9), 498; https://doi.org/10.3390/fi18090498 - 21 Sep 2026

Viewed by 338

Abstract

Simulation-based reinforcement-learning locomotion depends on the physical fidelity of the underlying model. This work presents a subsystem-level modelling, validation, and control framework for a 12-DoF biped robot, combining Denavit–Hartenberg kinematics, Euler–Lagrange dynamics, a Discrete Euler–Lagrange reference integrator, Hunt–Crossley contact, and Soft Actor-Critic training [...] Read more.

Simulation-based reinforcement-learning locomotion depends on the physical fidelity of the underlying model. This work presents a subsystem-level modelling, validation, and control framework for a 12-DoF biped robot, combining Denavit–Hartenberg kinematics, Euler–Lagrange dynamics, a Discrete Euler–Lagrange reference integrator, Hunt–Crossley contact, and Soft Actor-Critic training in PyBullet. Validation is scoped. For the fixed-hip leg, numerical damped-least-squares inverse kinematics achieved a round-trip error of 0.017 ± 0.022 mm, while a gravity path-integral test produced a residual of 0.006 J. On a one-DoF reference problem, DEL bounded energy error under a coarse-step stress test, whereas at the 1 ms training step, RK4 was more accurate; no RL-scale DEL advantage was established. Contact realism remained inconclusive because the available prescribed-penetration analysis was not a dynamically consistent whole-body impact test. The same nominal parameters were used in PyBullet for locomotion training, without establishing numerical equivalence between the two simulators. Across three asymmetric-reward runs, forward walking dominated final evaluations, but sustained velocity ranged from 0.62 to 1.24 m/s under unequal training budgets. An exploratory hybrid architecture reached 2.38 m/s in one run without controlled ablation. These results demonstrate subsystem-level diagnostics while identifying full-body validation, contact calibration, equal-budget replication, and architectural ablation as necessary future work. Full article

(This article belongs to the Topic Artificial Intelligence and Machine Learning in Cyber–Physical Systems)

► Show Figures

Graphical abstract

25 pages, 1963 KB

Open AccessArticle

GAT-MG-ECF: A Dependency Graph-Attention Extension of Multi-Granularity Entity–Context Fusion for Drug–Drug Interaction Extraction

by Hiba Chanaa, El Habib Nfaoui and Loqman Chakir

Future Internet 2026, 18(9), 497; https://doi.org/10.3390/fi18090497 - 21 Sep 2026

Viewed by 268

Abstract

Harmful drug–drug interactions (DDIs) are a common, preventable cause of hospitalizations, adverse drug events, and mistaken treatment decisions, and extracting them from the biomedical literature supports online pharmacovigilance and clinical decision support. Prior extractors fuse several views of a sentence from a pre-trained [...] Read more.

Harmful drug–drug interactions (DDIs) are a common, preventable cause of hospitalizations, adverse drug events, and mistaken treatment decisions, and extracting them from the biomedical literature supports online pharmacovigilance and clinical decision support. Prior extractors fuse several views of a sentence from a pre-trained encoder, but they pool every view from contiguous token spans and use syntactic dependency structure only implicitly. We propose GAT-MG-ECF(Graph-Attention Multi-Granularity Entity–Context Fusion): it adds a fourth, dependency-graph view—a graph attention network (GAT) over the sentence’s dependency parse—to the entity, context-bridge, and global views of its three-view predecessor, MG-ECF. The graph is built from the sentence alone, so the model needs no external knowledge base. On DDI-2013, under the official four-class protocol (three seeds), GAT-MG-ECF with BiomedBERT reaches micro-F1

0.9092 \pm 0.0039

—to the best of our knowledge, the highest reported under this protocol; it surpasses the three-view MG-ECF (

0.905

) and outperforms the strongest knowledge-graph-enhanced system, HKG-DDIE (

0.854

), by 5.5 percentage points. A second backbone, BioLinkBERT, also exceeds the prior best (

0.9076

), and the architecture transfers to ChemProt without re-tuning (micro-F1

0.9064

). The graph view’s own contribution is modest but consistent across both backbones, adding complementary rather than redundant signal. Explicit syntactic structure is a useful, low-cost addition to strong contextual encoders for biomedical relation classification. Full article

(This article belongs to the Special Issue Artificial Intelligence-Enabled Smart Healthcare—2nd Edition)

► Show Figures

Figure 1

44 pages, 25272 KB

Open AccessReview

Serverless Functions in Cloud–Edge Environments: A Comprehensive Critical Review and Taxonomy

by Abdullah Abbasi, Dil Nawaz Hakro, Asad Ullah, Suhail S. M. Alqrinawi, Akhtar Hussain, Osama Al Rahbi, Mohammed Izaan Kari, Suad Mohammed Al Qassabi and Muhammad Hafidz Fazli Bin Md Fauadi

Future Internet 2026, 18(9), 496; https://doi.org/10.3390/fi18090496 - 20 Sep 2026

Viewed by 781

Abstract

Cloud–edge continuums are driving the shift of cloud-native applications from centralized data centers to latency-, mobility-, privacy-, and energy-saving applications. The serverless architecture provides an attractive “Function-as-a-Service” (FaaS) model in this transition, as it is driven by events, elastic and fine-grained, and controlled [...] Read more.

Cloud–edge continuums are driving the shift of cloud-native applications from centralized data centers to latency-, mobility-, privacy-, and energy-saving applications. The serverless architecture provides an attractive “Function-as-a-Service” (FaaS) model in this transition, as it is driven by events, elastic and fine-grained, and controlled by the platform. But introducing a mix of heterogeneous edge nodes, fog/MEC resources, regional clouds, and hyperscale data centers creates a seemingly simple FaaS deployment problem to solve with a set of multi-objective orchestration challenges: runtime selection, autoscaling, cold start mitigation, placement, migration, workflow coordination, state management, trust, cost, energy, and carbon. In this article, we provide an extensive critical review of serverless functions in cloud–edge environments. While some surveys are narrowly focused on aspects of autoscaling, offloading, IoT, or security, the review brings together architectural evolution, runtime mechanisms, platform ecosystems, governance issues, sustainability issues, and emerging applications using AI. It builds a multidimensional taxonomy ranging from runtime systems, autoscaling, cold start mitigation, function placement, and offloading/migration, to workflow orchestration, state and data management, intelligent scheduling, security, sustainability, and industrial serverless platforms. It also presents a built-in conceptual model that connects application needs, runtime environment, orchestration intelligence, governance policies, and system-level results. The synthesis reveals that cloud–edge serverless systems need accountable placement, state-aware workflows, reproducible benchmarking, trustworthy orchestration, and carbon-aware lifecycle control, which can be achieved only by going beyond latency and elasticity. The paper ends with research directions on adaptive, interoperable, explainable, and sustainable serverless systems on the cloud–edge continuum. Full article

(This article belongs to the Special Issue Distributed, Cloud, Edge, and Fog Computing in Next-Generation Wireless Networks)

► Show Figures

Figure 1

38 pages, 8131 KB

Open AccessArticle

HFS-SVE: A Hybrid Feature Selection and Soft Voting Ensemble for Android Malware Detection

by Hany F. Atlam and Samyak M. Jeevane

Future Internet 2026, 18(9), 495; https://doi.org/10.3390/fi18090495 - 20 Sep 2026

Viewed by 293

Abstract

Android malware continues to evolve in complexity, creating challenges for detection systems that must distinguish malicious applications from increasingly heterogeneous benign applications. Although machine learning provides effective mechanisms for learning malware characteristics, the high dimensionality of Android malware datasets can introduce redundant and [...] Read more.

Android malware continues to evolve in complexity, creating challenges for detection systems that must distinguish malicious applications from increasingly heterogeneous benign applications. Although machine learning provides effective mechanisms for learning malware characteristics, the high dimensionality of Android malware datasets can introduce redundant and weakly informative features and increase computational requirements. To address this problem, this paper proposes a Hybrid Feature Selection and Soft-Voting Ensemble (HFS-SVE) framework that integrates complementary feature-selection and ensemble-learning strategies. The proposed framework sequentially applies Random Forest (RF) feature importance, Chi-square-based SelectKBest, correlation filtering, and L1 regularisation, reducing the original 489-feature representation to 13 selected features. These features are subsequently classified using RF, XGBoost, and LightGBM, whose probability outputs are combined through soft voting. Experimental results on the KronoDroid dataset demonstrate that the proposed HFS-SVE achieves 99.41% accuracy, 99.52% precision, 99.30% recall, 99.40% F1-score, and 99.41% ROC-AUC. The proposed framework also records the lowest measured detection time among the evaluated feature-selection strategies. Cross-dataset evaluation on Malgenome, TUANDROMD, and Drebin achieves accuracy above 98% on each dataset. The findings demonstrate that the proposed HFS-SVE can combine substantial feature-space reduction with strong Android malware detection performance, while the cross-dataset results highlight the importance of dataset variation and feature provenance when assessing generalisation. Full article

(This article belongs to the Special Issue Cybersecurity and Resilience in IoT and Distributed Networks (Including AI‑Enabled Approaches))

► Show Figures

Graphical abstract

27 pages, 5089 KB

Open AccessArticle

Comparative Analysis of Transfer Learning Architectures for Human Posture Classification Using Silhouette Images

by Ajit Kumar Pradhan, Sarita Tripathy, Kalaiarasi Sonai Muthu Anbananthen, Prasant Kumar Pattnaik and Amirthaa Anbananthen

Future Internet 2026, 18(9), 494; https://doi.org/10.3390/fi18090494 - 20 Sep 2026

Viewed by 196

Abstract

Human posture classification is an important computer vision task for intelligent monitoring, rehabilitation, assistive systems, and human–machine interaction. This study presents a comparative evaluation of four transfer learning configurations, TL-ResNet-18, TL-ResNet-50, TL-MobileNet, and TL-Xception, under a common experimental protocol—for classifying four human postures [...] Read more.

Human posture classification is an important computer vision task for intelligent monitoring, rehabilitation, assistive systems, and human–machine interaction. This study presents a comparative evaluation of four transfer learning configurations, TL-ResNet-18, TL-ResNet-50, TL-MobileNet, and TL-Xception, under a common experimental protocol—for classifying four human postures (standing, sitting, lying, and bending) from silhouette images. The architectures were evaluated under a common experimental protocol using the same dataset partitioning, preprocessing, augmentation procedures, and principal training settings, while retaining architecture-specific fine-tuning configurations. Performance was assessed using accuracy, F1-score, and ROC-AUC across the training, validation, and test partitions. Among the evaluated architectures, TL-MobileNet achieved the strongest overall held-out classification performance, with a test accuracy of 93.47%, an F1-score of 93.00%, and a ROC-AUC of 99.25%. TL-Xception achieved the same test ROC-AUC of 99.25%, but a lower test accuracy of 92.08%, while TL-ResNet-18 and TL-ResNet-50 achieved test accuracies of 91.81% and 90.00%, respectively. The results demonstrate performance differences among the evaluated transfer learning configurations. Within the evaluated ResNet configurations, greater network depth did not correspond to improved classification performance under the evaluated conditions. Silhouette-based posture classification shows potential for supporting future human-centered smart applications while reducing reliance on appearance-rich imagery. Further evaluation is required to establish cross-dataset generalizability, computational efficiency, and real-time deployment performance. Full article

► Show Figures

Figure 1

39 pages, 2494 KB

Open AccessArticle

Toward a Physical Operating System for Agentic Commerce: The Emerging Role of Amazon Supply Chain Services

by Chihiro Watanabe, Shanyu Lei, Akira Nagamatsu and Yuji Tou

Future Internet 2026, 18(9), 493; https://doi.org/10.3390/fi18090493 - 19 Sep 2026

Viewed by 383

Abstract

This study examines the Physical Execution Layer, in which AI agents extend digital decision-making into real-world tasks. It asks why Amazon Supply Chain Services (ASCS), launched in May 2026, may be structurally favored as infrastructure for this transition and evaluates it as an [...] Read more.

This study examines the Physical Execution Layer, in which AI agents extend digital decision-making into real-world tasks. It asks why Amazon Supply Chain Services (ASCS), launched in May 2026, may be structurally favored as infrastructure for this transition and evaluates it as an emerging Physical Operating System (Physical OS) candidate. Based on Amazon’s 10-K filings for 2013–2025 and an exploratory proxy series for ASCS revenue, the study applies logistic curve fitting and regression analysis with sensitivity analysis to compare the growth trajectories of Amazon Web Services (AWS) and ASCS from 2013 to 2026, including an exploratory 2026 reference year. AWS revenue is closely fitted in-sample by a logistic S-curve, whereas the ASCS proxy reflects a more physically constrained trajectory; their growth rates nevertheless exhibit substantial co-movement. The pattern is consistent with a proposed Dual-OS architecture in which AWS supplies compute, inference, and optimization while ASCS supports real-world execution within a shared technology and infrastructure investment base. CX physicalization and system-level supply-chain decarbonization may further increase the strategic importance of logistics performance. The findings are descriptive and correlational, not causal, and the ASCS series is a proxy. The Dual-OS architecture is presented as an analytical framework for future digital-physical execution networks and manufacturing reconfiguration. Full article

(This article belongs to the Special Issue Advances in Smart Environments and Digital Twin Technologies—2nd Edition)

► Show Figures

Figure 1

28 pages, 389 KB

Open AccessArticle

Assessing Generative Artificial Intelligence Tools Usage Guidelines at the Top 200 Universities Worldwide

by Midrar Ullah and Maged N. Kamel Boulos

Future Internet 2026, 18(9), 492; https://doi.org/10.3390/fi18090492 - 19 Sep 2026

Viewed by 869

Abstract

This study aimed to evaluate the guidelines for the use of GenAI tools available on the websites of the world’s top 200 universities. We collected all publicly available guidelines from the websites of the top 200 universities listed in the 2026 QS World [...] Read more.

This study aimed to evaluate the guidelines for the use of GenAI tools available on the websites of the world’s top 200 universities. We collected all publicly available guidelines from the websites of the top 200 universities listed in the 2026 QS World University Rankings. These guidelines were evaluated using the same 24-item checklist developed for our 2024 survey of the top 50 universities worldwide. The assessment revealed that 180 institutions have publicly available guidelines. All universities permit the use of GenAI tools under specific conditions. These policies are primarily established by academic bodies that focus on teaching, learning, and innovation. Most guidelines emphasize the use of university-licensed options to enhance data security. The most cited GenAI tools included ChatGPT, Microsoft Copilot, and Google Gemini (formerly Bard). More than 80% of the guidelines covered academic integrity, the limitations of GenAI tools, data privacy, the integration of GenAI in teaching and assessment, and the need to acknowledge its use. However, the operation of AI algorithms, reporting mechanisms for non-permitted use, documentation of prompts, and adoption of outputs were addressed in fewer than 40% of the guidelines. A comparison with our 2024 data for the subset of 50 universities re-evaluated in 2026 shows progress in guidelines coverage, reflecting institutions’ rapid responses to the evolving AI landscape. Despite the QS top 200’s concentration in wealthy nations, our checklist and findings are actionable for universities across the Global South. Full article

(This article belongs to the Special Issue 2026 and 2027 Feature Papers from Future Internet’s Editorial Board Members)

33 pages, 14730 KB

Open AccessArticle

Performance Evaluation of ADS-B Receivers Implemented Using Software-Defined Radio Platforms and GNU Radio

by Vlad-Stefan Hociung, Alexandru-Gabriel Gherghina, Cezar-Petrut Onu, Calin Vladeanu and Alexandru Martian

Future Internet 2026, 18(9), 491; https://doi.org/10.3390/fi18090491 - 19 Sep 2026

Viewed by 555

Abstract

Automatic Dependent Surveillance-Broadcast (ADS-B) is one of the most critical technologies utilized in contemporary air traffic control, providing an automated and broadcast means of periodically sending aircraft identification and location, as well as velocity and other state-related information. This paper assesses the receiving [...] Read more.

Automatic Dependent Surveillance-Broadcast (ADS-B) is one of the most critical technologies utilized in contemporary air traffic control, providing an automated and broadcast means of periodically sending aircraft identification and location, as well as velocity and other state-related information. This paper assesses the receiving performance of ADS-B signals utilizing multiple software-defined radio (SDR) platforms. Four different SDR platforms (DX Patrol MK4, Adalm-Pluto, USRP B200mini and USRP B210) were considered for evaluation, using a single antenna feed distribution via an active RF splitter. Each receiver’s performance was evaluated by measuring the rate at which each platform was able to decode messages from aircraft, the number of aircraft that were detected, the number of valid position reports received from each aircraft, the distance from the receiver to the aircraft at which each platform could receive valid position reports and each platform’s susceptibility to various forms of interference. The results indicate that the best cost-performance in case of interference-free ADS-B reception is obtained for the Adalm-Pluto platform (882 ADS-B messages received in the analyzed period, 2.84 cost/message), whereas the USRP B210 SDR exhibits the best performance in the presence of strong interference (109 ADS-B messages received). These findings provide insight into the relative trade-offs between low-cost SDR platforms and higher-performance SDR platforms, specifically related to analog-to-digital converter (ADC) resolution, RF front-end architecture, host interface, sensitivity and decoding reliability. Full article

(This article belongs to the Special Issue Efficient and Secure Wireless Communications and Networking: Architecture and Applications)

► Show Figures

Figure 1

33 pages, 16287 KB

Open AccessArticle

Explainable and Analyst-Driven Random Forest for Intrusion Detection

by Saloua Bellouch, Mostapha Zbakh, Siham Aouad and An Braeken

Future Internet 2026, 18(9), 490; https://doi.org/10.3390/fi18090490 - 18 Sep 2026

Viewed by 276

Abstract

Random Forest and other tree-ensemble classifiers achieve high accuracy in network intrusion detection; however, their aggregate decision logic prevents analysts from auditing or deploying individual predictions as operational rules. Post hoc explanation methods introduce latencies incompatible with security operation center (SOC) requirements and [...] Read more.

Random Forest and other tree-ensemble classifiers achieve high accuracy in network intrusion detection; however, their aggregate decision logic prevents analysts from auditing or deploying individual predictions as operational rules. Post hoc explanation methods introduce latencies incompatible with security operation center (SOC) requirements and produce conditions unsuitable for firewall configuration. Among the systems reviewed in this study, none unifies intrinsic explanation, rule deployment, ATT&CK attribution, cross-dataset validation, and adaptive feedback in one pipeline. This work presents a depth-limited Random Forest with deterministic, per-instance explanations at a fraction of gradient-based attribution latency. Complementary mechanisms generate analyst-deployable rule specifications, technique-level adversary attribution, and a feedback protocol that models label noise, missed reviews, and bounded correction budget. Evaluated on a large, multi-category network-traffic benchmark, the system attains high detection accuracy (macro recall

0.86

, driven substantially by the majority normal-traffic class at

68 %

of flows) while sustaining throughput beyond SOC requirements; a stealthy reconnaissance-and-exploitation category remains markedly harder to detect under this class imbalance. Cross-dataset evaluation on a more recent benchmark attains strong performance after limited target-domain retraining. The adaptive feedback protocol yields a statistically significant false-positive reduction over repeated simulated reviews, requiring only modest weekly analyst effort. Together, these capabilities enable auditable and SOC-integrable detection pipelines. Full article

(This article belongs to the Section Cybersecurity)

► Show Figures

Figure 1

29 pages, 5228 KB

Open AccessArticle

An Effective Cooperative Coevolution–Differential Evolution Algorithm for Improving the Performance and Robustness of Ridesharing Systems with Trust Requirements

by Fu-Shiung Hsieh

Future Internet 2026, 18(9), 489; https://doi.org/10.3390/fi18090489 - 18 Sep 2026

Viewed by 238

Abstract

Despite the growth of ridesharing services worldwide, the adoption of ridesharing remains low compared with other modes of transportation. Five important factors influencing the willingness to consider ridesharing include time/cost, service experience, traffic/environment, privacy, and safety. Information from social networks can provide potential [...] Read more.

Despite the growth of ridesharing services worldwide, the adoption of ridesharing remains low compared with other modes of transportation. Five important factors influencing the willingness to consider ridesharing include time/cost, service experience, traffic/environment, privacy, and safety. Information from social networks can provide potential value for improving service experience and trustworthiness of ridesharing services. However, the problem of optimizing ridesharing decisions based on consideration of trust requirements of participating drivers and riders and other constraints in ridesharing systems poses a challenge in the development of a solution algorithm due to high computational complexity. In addition, a ridesharing optimization problem considering trust requirements is typically non-convex and non-linear with discrete decision variables, making exact methods not applicable. Metaheuristic approaches can be applied to find solutions for non-convex and non-linear discrete constrained optimization problems. The goal of this paper is to develop an effective solution algorithm to improve the performance of ridesharing systems with trust requirements. To achieve the goal of this study, we develop a variant of the Differential Evolution (DE) algorithm by combining the Cooperative Coevolution approach with the DE approach. To verify the effectiveness of the new algorithm for solving the ridesharing optimization problem with trust requirements, we conducted experiments and compared the results obtained by the new algorithm with those obtained by sixteen other competitive algorithms. Comparison with other competitive algorithms based on the experimental results shows that the proposed algorithm significantly outperforms other competitive algorithms in terms of performance and robustness. The CC–DE algorithm achieves the highest average fitness values for all test cases, whereas the other algorithms achieve the highest average fitness values for at most 80% of the test cases. The CC–DE algorithm achieves a zero standard deviation of fitness function values for all test cases, whereas the other algorithms achieve a zero standard deviation for at most 80% of the test cases. For the two largest test cases, the CC–DE algorithm outperforms the other algorithms by at least 9.246% and 12.9867%, respectively, in terms of performance. Full article

(This article belongs to the Special Issue Intelligent Computational Methods for Urban Traffic Management and Control)

► Show Figures

Figure 1

23 pages, 901 KB

Open AccessArticle

COAu-IoD: A Cloud and Offline Computing-Assisted Authentication Framework for Lightweight UAV Communication in IoD

by Pingyuan Zhang, Chen Fu and Qikun Zhang

Future Internet 2026, 18(9), 488; https://doi.org/10.3390/fi18090488 - 18 Sep 2026

Viewed by 179

Abstract

The Internet of Drones (IoD) enables efficient communications among unmanned aerial vehicles (UAV) via wireless links. Restricted by limited onboard computing resources and complicated operating environments, such a network is vulnerable to external or internal security threats. It is therefore urgent to deploy [...] Read more.

The Internet of Drones (IoD) enables efficient communications among unmanned aerial vehicles (UAV) via wireless links. Restricted by limited onboard computing resources and complicated operating environments, such a network is vulnerable to external or internal security threats. It is therefore urgent to deploy an effective authentication mechanism to isolate malicious attackers prior to UAV communications. Nevertheless, the conventional identity-based IoD authentication schemes must account for total computation and communication overhead, regardless of the effective instant of authentication messages. In this work, we consider the offline precomputation in the authentication phase before the message is available and add cloud servers as an extra component in IoD to assist limited onboard resources to share partial authentication calculation. As a result, this work proposes a new cloud- and offline-computing-assisted authentication framework, known as COAu-IoD, to provide lightweight communication between a UAV and ground station by transferring heavy computation overhead to the offline phase. We give a security model for this authentication framework and provide an efficient construction based on this framework and a known identity-based signature scheme. What is more, we give its security proof and evaluate its performance using the standard and recent state-of-the-art IoD authentication schemes. The results demonstrate that our COAu-IoD scheme achieves lower online computation and communication overhead for UAVs, at the cost of acceptable offline precomputation overhead. Full article

(This article belongs to the Section Internet of Things)

► Show Figures

Figure 1

33 pages, 419 KB

Open AccessArticle

GDPR Dark Patterns in Cookie Consent: An Automated Study of High-Traffic Websites Accessed from Denmark

by Christos Ntemkas, Laura Vieira Teixeira, Lejla Islami, Gaurav Choudhary and Nicola Dragoni

Future Internet 2026, 18(9), 487; https://doi.org/10.3390/fi18090487 - 17 Sep 2026

Viewed by 240

Abstract

This paper examines how often GDPR-relevant dark patterns appear in cookie-consent banners and how often users receive a choice that is free and informed. We use an adapted version of CCrawler to crawl 99 high-traffic websites accessed from Denmark, detect consent interfaces, and [...] Read more.

This paper examines how often GDPR-relevant dark patterns appear in cookie-consent banners and how often users receive a choice that is free and informed. We use an adapted version of CCrawler to crawl 99 high-traffic websites accessed from Denmark, detect consent interfaces, and extract first-layer choice pathways and settings-layer elements for compliance-risk coding. We then assess detected banners against four GDPR-informed interface-level criteria: first-layer refusal availability, absence of pre-ticked or default-enabled options, relative accept/reject prominence, and non-obstructive presentation. These operational criteria are observable interface proxies rather than final legal determinations of GDPR compliance. Cookie banners were detected on 90 of 99 sites; the nine non-detections are reported separately and are not treated as evidence of non-compliance. The results indicate a structural imbalance: acceptance is usually available on the first layer, while refusal is often placed behind additional steps or settings dialogues. A subset of reachable settings interfaces also contained pre-ticked options. To place these observations in the European context, we compare the two directly comparable first-layer choice indicators with a recent 31-country study. Acceptance availability is close to that study’s Denmark-specific estimate, whereas first-layer rejection in our high-traffic Denmark-accessed sample lies between its Denmark-specific estimate and its 31-country aggregate. These cross-study comparisons are descriptive and are not treated as matched statistical estimates because the sampling frames and measurement procedures differ. The findings provide Denmark-specific, time-bounded evidence that known consent-interface asymmetries remain visible on high-traffic websites, together with a reproducible workflow for future audits. Full article

► Show Figures

Figure 1

19 pages, 480 KB

Open AccessArticle

Transformer-Based Generation of Route Opening Patterns for Crowd Evacuation

by Akihiro Morita, Koichi Kobayashi and Yuh Yamashita

Future Internet 2026, 18(9), 486; https://doi.org/10.3390/fi18090486 - 17 Sep 2026

Viewed by 223

Abstract

Developing methods for achieving safe and efficient crowd evacuation is an important research issue. In particular, controlling pedestrian flows based on predictions from a mathematical model is important. In this paper, we propose a new method for generating route opening patterns. The target [...] Read more.

Developing methods for achieving safe and efficient crowd evacuation is an important research issue. In particular, controlling pedestrian flows based on predictions from a mathematical model is important. In this paper, we propose a new method for generating route opening patterns. The target area is modeled by an undirected graph, and pedestrian flows are represented by the temporal change in density at each vertex. We assume that the density at each vertex can be observed using equipment such as IoT devices. Based on the model predictive control framework, we consider the problem of finding a route opening pattern that minimizes a safety-related penalty. To solve this problem, we propose a Transformer-based solution method that combines offline and online computations. The effectiveness of the proposed method is demonstrated through a numerical simulation. In the numerical example, a route opening pattern satisfying the Signal Temporal Logic (STL) formula was obtained using the proposed method. The evacuation rate, which represents the proportion of pedestrians whose movement is complete, was

0.8526

, the penalty was 2193, and the maximum online computation time was

8.5199

s. In the offline computation, the training dataset for the Transformer model was generated using a genetic algorithm (GA). By combining offline and online computations, we confirmed that an appropriate route opening pattern can be generated during online computation. Full article

(This article belongs to the Special Issue Smart Technology: Artificial Intelligence, Robotics and Algorithms)

► Show Figures

Figure 1

48 pages, 27955 KB

Open AccessArticle

An Explainable AI Framework for Identity Document Authentication in AML/KYC Verification

by Eldeena Huey Yinn Lim and Tee Connie

Future Internet 2026, 18(9), 485; https://doi.org/10.3390/fi18090485 - 16 Sep 2026

Viewed by 221

Abstract

This study investigates the development of an AI-driven document authentication framework for Anti-Money Laundering (AML) and Know Your Customer (KYC) verification environments. Conventional manual inspection and rule-based verification techniques often fail to detect sophisticated forged identity documents containing subtle visual or semantic manipulations. [...] Read more.

This study investigates the development of an AI-driven document authentication framework for Anti-Money Laundering (AML) and Know Your Customer (KYC) verification environments. Conventional manual inspection and rule-based verification techniques often fail to detect sophisticated forged identity documents containing subtle visual or semantic manipulations. To address this limitation, the proposed framework combines handcrafted forensic feature extraction, OCR-driven semantic analysis, rule-based semantic field extraction and Random Forest classification to identify inconsistencies within identity documents captured under realistic mobile imaging conditions. Experimental evaluation was conducted using selected MIDV-2020 identity document subsets consisting of Albanian identity cards, Latvian passports, and Slovakian identity cards. The proposed framework achieved a recall rate of 92.31% and an overall accuracy of 84.85% on the held-out test set, while maintaining interpretable forensic feature analysis suitable for regulated AML/KYC environments. The results demonstrate that lightweight and explainable machine learning approaches can provide effective forged-document detection without requiring computationally intensive deep learning architectures. Full article

(This article belongs to the Special Issue Securing Artificial Intelligence Against Attacks)

► Show Figures

Figure 1

27 pages, 10423 KB

Open AccessArticle

Understanding Differential Privacy in Decentralized Federated Learning: A Controlled Privacy–Utility Comparison

by Alsharif Hasan Mohamad Aburbeian, Manuel Fernández-Veiga, Ana Fernández-Vilas, Majdi Owda and Amani Yousef Owda

Future Internet 2026, 18(9), 484; https://doi.org/10.3390/fi18090484 - 16 Sep 2026

Viewed by 871

Abstract

Centralized Federated Learning (FL) enables collaborative model training without sharing raw data. Differential privacy (DP) is widely used to protect sensitive information in FL; however, its behavior in decentralized environments remains poorly understood. This study empirically compares centralized FL and sequential Decentralized Federated [...] Read more.

Centralized Federated Learning (FL) enables collaborative model training without sharing raw data. Differential privacy (DP) is widely used to protect sensitive information in FL; however, its behavior in decentralized environments remains poorly understood. This study empirically compares centralized FL and sequential Decentralized Federated Learning (DFL) under matched clipping and perturbation settings to examine model utility and privacy leakage. Both frameworks were evaluated under a common experimental setup with non-IID data, and each configuration was evaluated across five independent seeds. Utility was evaluated under matched experimental perturbation parameters, whereas formal client-level privacy accounting was applied to the perturbed round-end model releases, with each client’s complete dataset treated as the protected unit. Empirical leakage was evaluated separately using membership inference and gradient inversion attacks. Within the evaluated MNIST configuration, the results show that clipping, perturbation, and the learning procedure jointly influence the observed privacy–utility behavior. At

C = 1

and

ϵ_{cal} = 2

, the mean final accuracy was

73.83 %

for FL and

97.74 %

for sequential DFL. At

ϵ_{cal} = 4

, InvGrad reconstruction for FL produced an MSE of

0.048

, PSNR of

13.71

, and SSIM of

0.280

, compared with

0.130

,

10.10

, and

0.184

for sequential DFL, respectively. Future research will investigate topology-aware privacy mechanisms and adaptive noise allocation for decentralized systems. Full article

(This article belongs to the Special Issue Privacy-Preserving and Secure Machine Learning)

► Show Figures

Graphical abstract

20 pages, 1210 KB

Open AccessArticle

On the Role of Data Normalization in Adversarial Robustness Evaluation: A Comparative Analysis of XGBoost and Random Forest for IoT Network Intrusion Detection

by Inas Mostefai, Sofiane Boukli-Hacene and Abdelhafid Abouaissa

Future Internet 2026, 18(9), 483; https://doi.org/10.3390/fi18090483 - 16 Sep 2026

Viewed by 182

Abstract

Intrusion detection systems (IDS) based on machine learning (ML) have demonstrated remarkable effectiveness in detecting cyber-attacks in Internet of Things (IoT) environments. Despite their high predictive performance, the robustness of these systems against adversarial attacks remains insufficiently understood, particularly with regard to the [...] Read more.

Intrusion detection systems (IDS) based on machine learning (ML) have demonstrated remarkable effectiveness in detecting cyber-attacks in Internet of Things (IoT) environments. Despite their high predictive performance, the robustness of these systems against adversarial attacks remains insufficiently understood, particularly with regard to the influence of data pre-processing techniques such as feature normalisation. This study examines the impact of Min-Max normalisation on the robustness of two widely used tree-based classifiers, XGBoost and Random Forest, against adversarial attacks of the Fast Gradient Method (FGM) and Projected Gradient Descent (PGD) types. Using the ACI-IoT-2023 dataset, we conduct comprehensive experiments across three scenarios: models trained on normalized data and evaluated using a fixed perturbation magnitude (

ϵ

), models trained on raw data and evaluated using the same fixed

ϵ

, and models trained on raw data and evaluated using an adaptive

ϵ

calibrated to the scale of each feature. Our results demonstrate that Min-Max normalization alone does not inherently make these models robust against adversarial attacks. The apparent robustness observed on non-normalised data with a fixed perturbation budget is an artefact caused by a misalignment in the scaling of the perturbation. When using an adaptive epsilon strategy calibrated according to feature scales, the accuracies of XGBoost and Random Forest decreased to 2.30% and 2.42% respectively under the

L_{\infty}

norm with the PGD attack. These results highlight the crucial importance of taking feature scaling into account when assessing robustness against adversarial attacks and provide practical recommendations for the design of reliable evaluation protocols for intrusion detection systems in the Internet of Things. Full article

(This article belongs to the Section Internet of Things)

► Show Figures

Graphical abstract

3 pages, 129 KB

Open AccessEditorial

Software-Defined Networking (SDN) and Network Function Virtualization (NFV) for a Hyperconnected World

by Faycal Bouhafs and Sachin Sharma

Future Internet 2026, 18(9), 482; https://doi.org/10.3390/fi18090482 - 15 Sep 2026

Viewed by 214

Abstract

Data networks are a fundamental infrastructure of modern societies, supporting industries, transport, communication, government services, finance, etc [...] Full article

(This article belongs to the Special Issue Software-Defined Networking (SDN) and Network Function Virtualization (NFV) for a Hyperconnected World)

27 pages, 472 KB

Open AccessArticle

From Semantic Retrieval to Conversational Agent: A Web-Based RAG Architecture for Interactive System Dynamics Modeling

by Pavel Kyurkchiev and Anton Iliev

Future Internet 2026, 18(9), 481; https://doi.org/10.3390/fi18090481 - 15 Sep 2026

Viewed by 378

Abstract

Keyword-based semantic search performs poorly on complex knowledge repositories such as System Dynamics model databases, where users know the behavior they want to simulate but not the structural vocabulary needed to retrieve it. We replace the static search field with an interactive web-based [...] Read more.

Keyword-based semantic search performs poorly on complex knowledge repositories such as System Dynamics model databases, where users know the behavior they want to simulate but not the structural vocabulary needed to retrieve it. We replace the static search field with an interactive web-based conversational agent, built as a Retrieval-Augmented Generation architecture in which the system asks context-aware clarifying questions to narrow the search scope across multiple turns. The architecture was evaluated in an ablation study of 37 benchmark scenarios over a curated corpus of 63 models, comparing six retrieval strategies against an expert semantic baseline using Precision@5, Recall@5, MRR@5, nDCG@5 and Hit@5. Conversational refinement raised mean nDCG@5 from 0.1066 to 0.4422 and Hit@5 from 0.1892 to 0.5946. The improvement over the broad-intent baseline is significant on nDCG@5 and MRR@5 under Holm-corrected Wilcoxon signed-rank tests. This comparison aggregates the clarification exchange with the additional user input it elicits. A separate condition that bypasses the generative rewriting step bounds the contribution of that step. A residual gap to the expert semantic baseline (nDCG@5 = 0.5750) remains and is significant on MRR@5. Lexical BM25 applied to expert queries outperformed dense retrieval on every metric (nDCG@5 = 0.8053), showing that sparse matching retains a decisive advantage where the structural vocabulary is exact. We conclude that conversational elicitation is an effective mechanism for narrowing the expertise gap in this domain, and that the lexical results motivate pairing it with hybrid sparse–dense retrieval. Full article

(This article belongs to the Special Issue Human-Centered Artificial Intelligence—2nd Edition)

► Show Figures

Graphical abstract

16 pages, 791 KB

Open AccessArticle

Privacy-Preserving Information Fusion of Heterogeneous Cross-Jurisdictional Sources for Traffic Accident Severity Prediction

by Ashik Shah Jahangeer and Shanmugavadivu Pichai

Future Internet 2026, 18(9), 480; https://doi.org/10.3390/fi18090480 - 14 Sep 2026

Viewed by 272

Abstract

Road safety authorities each hold accident records that, when combined, could train stronger severity prediction models, yet these records can be neither centralized, for privacy and governance reasons, nor naively merged, because jurisdictions encode severity under incompatible ontologies. This paper recasts that impasse [...] Read more.

Road safety authorities each hold accident records that, when combined, could train stronger severity prediction models, yet these records can be neither centralized, for privacy and governance reasons, nor naively merged, because jurisdictions encode severity under incompatible ontologies. This paper recasts that impasse as an information fusion problem and fuses model updates from multiple road safety data silos into a single severity model while every raw record stays at its source. Three components act together: model-level fusion under differential privacy, a reliability-weighted aggregation rule that trusts each source based on its measured quality rather than its size, and a per-source centered logit adjustment layer that reconciles mismatched label priors without double-correcting the shared class imbalance. The primary evaluation is a clean cross-silo setting: five United States state datasets (US Accidents) that share one severity ontology but are held by distinct custodians. Here, over five seeds with 95% confidence intervals, effect sizes, and Holm–Bonferroni correction, private fusion recovers most of a centralized upper bound while keeping data local (

0.599

balanced accuracy versus

0.624

when centralized and

0.544

when local-only), and reliability-weighted fusion attains the highest macro F1 of all methods (

0.567

). Reliability weighting yields a small but consistent robustness advantage under privacy noise; in leave-one-state-out transfer, its improvement over uniform averaging is large on every held-out state but, after Holm correction, survives in two out of five cases. Crucially, we also report a boundary honestly; a United Kingdom source that encodes injury severity—an ontologically different target from the US traffic impact scale—is used as a deliberate out-of-ontology transfer stress test, and a transfer to it collapses to chance (

0.50

,

p = 0.62

). Two further honest results are reported: alignment raises accuracy everywhere but does not close the across-source gap, and a membership inference attack reveals no measurable leakage for differential privacy to remove. Full article

(This article belongs to the Section Big Data and Augmented Intelligence)

► Show Figures

Figure 1

35 pages, 9748 KB

Open AccessReview

Digital Twin in Vehicular Communications: Challenges and Opportunities

by Junliang Ye, Yuna Jiang, Ziwei Chen, Zijing He, Deqiao Gan, Xiaomeng Ai, Ling Ma and Xiaohu Ge

Future Internet 2026, 18(9), 479; https://doi.org/10.3390/fi18090479 - 14 Sep 2026

Viewed by 413

Abstract

Digital twins (DTs) are increasingly studied together with vehicular communications, but the literature spans different twinned entities, synchronization assumptions, computing placements, and levels of experimental evidence, which makes results difficult to compare. This survey provides a communication-centric review of DT-enabled vehicular systems. We [...] Read more.

Digital twins (DTs) are increasingly studied together with vehicular communications, but the literature spans different twinned entities, synchronization assumptions, computing placements, and levels of experimental evidence, which makes results difficult to compare. This survey provides a communication-centric review of DT-enabled vehicular systems. We first introduce an operational definition for a vehicular DT and distinguish a synchronized DT from a static simulator or digital model. We then organize the field according to twin granularity, physical–virtual synchronization, vehicle–edge–cloud placement, enabling V2X technologies, and measurable evaluation dimensions. The standardization discussion is updated from DSRC and IEEE 802.11p through 5G-Advanced and ongoing 3GPP work toward 6G, together with recent digital-twin and network-digital-twin standards. Rather than treating reported benefits as established outcomes, we compare the assumptions, metrics, evidence, limitations, deployment constraints, and safety and security implications of representative studies. Finally, we identify research priorities involving synchronization staleness, twin placement and migration, communication–computation–fidelity tradeoffs, validation and uncertainty, interoperable data models, lifecycle management, and safety-aware closed-loop operation. The survey is intended to provide researchers and practitioners with a structured basis for deciding when DT techniques are appropriate for vehicular communication systems and how such systems should be evaluated. Full article

(This article belongs to the Special Issue Progress and Challenges in Wireless Communication)

► Show Figures

Figure 1

21 pages, 9045 KB

Open AccessReview

The Impact of Intelligent Transport Systems on Safety, Emissions Reduction, and Travel Time: A Review

by Nadica Stojanovic, Ivan Grujic, Suzana Petrovic Savic, Miladin Stefanovic and Aleksandar Djordjevic

Future Internet 2026, 18(9), 478; https://doi.org/10.3390/fi18090478 - 14 Sep 2026

Viewed by 463

Abstract

The intensive development of road transportation and the increasing number of vehicles have led to significant challenges related to road safety, traffic congestion, travel time, energy consumption, and negative environmental impacts. In this context, intelligent transport systems (ITS) represent a significant approach to [...] Read more.

The intensive development of road transportation and the increasing number of vehicles have led to significant challenges related to road safety, traffic congestion, travel time, energy consumption, and negative environmental impacts. In this context, intelligent transport systems (ITS) represent a significant approach to improving the efficiency and sustainability of modern transportation systems. The aim of this paper is to present and systematize the application of modern ITS technologies for improving road safety, reducing emissions, and shortening travel time. Based on an analysis of the relevant literature, the fundamental components and architecture of ITS are presented, including sensor systems, V2X communication, IoT, cloud and edge computing, as well as the application of artificial intelligence in traffic data processing and prediction. The analyzed studies demonstrate that ITS enables dynamic traffic flow management, route optimization, reduction in congestion and emissions, and more efficient responses to emergency situations. Particular attention is devoted to the possibility of simultaneously considering travel time, energy consumption, emissions, noise, and road safety. As a synthesis of the analyzed findings, an integrated algorithm for intelligent traffic management is proposed, operating as a closed feedback loop encompassing data collection, state assessment, prediction, optimization, and control. Future ITS development is expected to focus on the integration of AI, IoT, 6G, and edge computing technologies and their validation using real-world traffic data. Full article

(This article belongs to the Special Issue Next-Generation Intelligent Transportation Systems)

► Show Figures

Figure 1

27 pages, 2868 KB

Open AccessArticle

FinDS-Agent: A Cloud–Edge Collaborative Data Science Agent for Financial Analytics

by Xiaozheng Du, Ruijun Deng, Cheng Wang, Feng Zhou, Shijing Hu, Zhihui Lu and Simon Fong

Future Internet 2026, 18(9), 477; https://doi.org/10.3390/fi18090477 - 13 Sep 2026

Viewed by 313

Abstract

Large language model agents can automate data science workflows, but cloud-centric deployment exposes sensitive context and edge-only deployment limits analytical capability. We present FinDS-Agent, a cloud–edge framework that keeps raw records and program execution at the trusted edge while providing a policy-screened, sanitized [...] Read more.

Large language model agents can automate data science workflows, but cloud-centric deployment exposes sensitive context and edge-only deployment limits analytical capability. We present FinDS-Agent, a cloud–edge framework that keeps raw records and program execution at the trusted edge while providing a policy-screened, sanitized context to support cloud planning. FinDS-Agent integrates a Three-Stage Cascaded Privacy Gate (TCPG), a Multi-Dimensional Joint Router (MJR), contract-guided ToolGraph planning, edge-side verification, and bounded repair. On 222 DataSciBench tasks over three runs, FinDS-Agent achieved a 69.93% completion rate and 57.06% success rate, improving over Edge-Only by 9.50 and 5.71 percentage points while invoking the cloud for 32.27% of eligible task-runs. On FinDS-Privacy-Bench, TCPG increased sensitive-field recall from 58.20% to 98.10%; no payload-leakage event was observed in the full set (0/200; Wilson 95% CI: 0–1.8845%) or blind split (0/100; 0–3.6993%) under the specified audit and threat model. External evaluation gave pass rates of 33.2%, 53.1%, and 62.3% for Edge-Only, FinDS-Agent, and Cloud-Only on DS-1000. These empirical results support selective cloud planning while delimiting statistical, privacy, and transfer claims. Full article

(This article belongs to the Special Issue LLM-Driven Agentic AI in Edge-Cloud Computing)

► Show Figures

Graphical abstract

21 pages, 4446 KB

Open AccessArticle

AI-Enabled Smart Monitoring of Bovine Embryo Development Using Time-Lapse Imaging and Transfer Learning

by Manickavasagan Shivaani, Meenakshi P.L and Pavneesh Madan

Future Internet 2026, 18(9), 476; https://doi.org/10.3390/fi18090476 - 12 Sep 2026

Viewed by 276

Abstract

Time-lapse incubation systems enable the continuous, non-invasive monitoring of embryonic development; however, identifying developmental stages still requires substantial manual assessment, making the process time-consuming, labor-intensive, and potentially subjective. This study evaluated the developmental kinetics of bovine embryos cultured in synthetic oviductal fluid (SOF) [...] Read more.

Time-lapse incubation systems enable the continuous, non-invasive monitoring of embryonic development; however, identifying developmental stages still requires substantial manual assessment, making the process time-consuming, labor-intensive, and potentially subjective. This study evaluated the developmental kinetics of bovine embryos cultured in synthetic oviductal fluid (SOF) and Gx-TL™ media using MIRI time-lapse imaging, and investigated the effectiveness of transfer learning-based convolutional neural networks (CNNs) for automated embryo-stage classification. A total of 311 zygotes were individually cultured under standard in vitro fertilization conditions, including 152 embryos in SOF medium and 159 embryos in Gx-TL™ medium. In the SOF group, 81 embryos reached the two-cell stage, 10 developed to the morula stage, and 6 reached the blastocyst stage. In the Gx-TL™ group, 77 embryos reached the two-cell stage, 18 developed to the morula stage, and 10 reached the blastocyst stage. Time-lapse images were manually annotated according to key developmental stages, including the two-cell through eight-cell stages, morula, and blastocyst. The annotated image dataset was augmented to 5000 images and used to train three pretrained CNN architectures: ResNet18, DenseNet121, and EfficientNet-B0. All three models achieved 100% accuracy in the two-class classification task across both culture media. Classification accuracy ranged from 95% to 100% for the nine-class model and from 98% to 100% for the ten-class model. These findings provide preliminary evidence that transfer learning-based convolutional neural networks (CNNs) can support the automated classification of bovine embryonic developmental stages using time-lapse images. Full article

(This article belongs to the Special Issue Machine Learning, Big Data, and Artificial Intelligence in Smart Systems)

► Show Figures

Figure 1

50 pages, 5536 KB

Open AccessArticle

Technical and Regulatory Prerequisites for Blockchain-Enabled Point-of-Sale Systems: A Tanzanian Case Study

by Julius Massawe, Bonny Mgawe, Cleverence Kombe and Anael Sam

Future Internet 2026, 18(9), 475; https://doi.org/10.3390/fi18090475 - 12 Sep 2026

Viewed by 378

Abstract

The continued expansion of digital payment technologies has encouraged Tanzanian district councils to use Point-of-Sale (POS) systems to collect service fees. In current POS systems, authorized POS terminals capture payment details, and through the centralized server, transactions are recorded on the POS database, [...] Read more.

The continued expansion of digital payment technologies has encouraged Tanzanian district councils to use Point-of-Sale (POS) systems to collect service fees. In current POS systems, authorized POS terminals capture payment details, and through the centralized server, transactions are recorded on the POS database, with receipts printed as confirmation. Although this architecture supports recording and monitoring transaction revenue, it provides limited support for verifying the identity of the actor authorizing the transaction and for independent confirmation of transaction integrity during auditing. To address these limitations, this study investigated the essential requirements, standards, and protocols for integrating a Self-Sovereign Identity (SSI) as a blockchain-based identity solution with existing POS systems and for using a permissioned blockchain platform to verify integrity. To achieve the study objectives, an exploratory qualitative approach was used, involving 32 semi-structured interviews with POS operators, revenue accountants, internal auditors, Information and Communication Technology (ICT) administrators, a regulator, and blockchain experts. A hybrid deductive–inductive thematic analysis was used to establish three requirement themes, namely, security and identity management, legal and regulatory compliance, and data management and integrity assurance; two standard themes, namely, security and cryptographic standards, and identity and decentralized identification; and two protocol themes, namely, security and user authentication protocols, and data management and identity portability protocols. The findings were mapped to applicable legal obligations, compliance standards, technical specifications, and implementation controls, indicating how these requirements were translated into the conceptual SSI-POS integration for district-council POS systems. The proposed solution separates credential issuance, credential holder, and device POS management; verifier and POS transaction processing; blockchain and integrity evidence; and the assurance domain across defined trust boundaries. After a signed transaction is approved by the verifier, a complete transaction receipt remains in the existing POS database, while receipt hashes or the corresponding Merkle roots are anchored on the permissioned blockchain to provide tamper-evident verification. Hyperledger Besu with Quorum Byzantine Fault Tolerance (QBFT) was selected for its fit with permissioned, multi-organizational governance and for independent replication of receipt-hash evidence. The study provides a stakeholder-derived, regulatory-aligned conceptual foundation for SSI-POS integration without replacing the existing POS system workflow. Full article

(This article belongs to the Section Cybersecurity)

► Show Figures

Graphical abstract

18 pages, 1515 KB

Open AccessArticle

Intelligent Synchronization of Machine Learning Models Using Graph Neural Networks: Application to Flood Prediction

by Boban Temelkovski, Rexhep Mustafovski, Jugoslav Achkoski, Georgi Dimirovski and Mile Stankovski

Future Internet 2026, 18(9), 474; https://doi.org/10.3390/fi18090474 - 11 Sep 2026

Viewed by 269

Abstract

Flood prediction remains a critical challenge in environmental risk management and disaster preparedness. Accurate river-level forecasting is essential for the development of reliable early warning systems and the mitigation of flood-related risks. However, conventional ensemble approaches, such as averaging and majority voting, often [...] Read more.

Flood prediction remains a critical challenge in environmental risk management and disaster preparedness. Accurate river-level forecasting is essential for the development of reliable early warning systems and the mitigation of flood-related risks. However, conventional ensemble approaches, such as averaging and majority voting, often exhibit limited adaptability when individual models respond differently to anomalies or incomplete data. To address this limitation, this study proposes a graph-based synchronization framework that integrates XGBoost and Random Forest models using a Graph Convolutional Network (GCN). The proposed framework represents the outputs of the base prediction models as graph nodes and employs graph message passing to learn context-dependent relationships between their predictions. The framework is evaluated using real-world hydrological observations from the Lepenec River Basin in North Macedonia together with meteorological data obtained from the OpenWeatherMap API. Experimental results demonstrate that the proposed GCN-based synchronization framework outperforms both the standalone prediction models and the previously proposed linear synchronization method, achieving an R² value of 0.91 and a Mean Absolute Error (MAE) of 0.21. The obtained results indicate that graph-based synchronization provides an adaptive approach for integrating heterogeneous machine-learning models and has the potential to support future flood early-warning systems and intelligent environmental monitoring applications. Full article

(This article belongs to the Section Smart System Infrastructure and Applications)

► Show Figures

Graphical abstract

22 pages, 11832 KB

Open AccessArticle

FQDA-ML: A Hierarchical Machine Learning-Powered Data Quality Framework for Evaluating Agile Sprint Performance: A Case Study in Community Engagement Projects

by Mario Pérez-Cargua, Elizabeth Salazar-Jácome, Javier De la Torre-Guzmán, Félix Chávez-Jácome and Wilson Sánchez-Ocaña

Future Internet 2026, 18(9), 473; https://doi.org/10.3390/fi18090473 - 11 Sep 2026

Viewed by 267

Abstract

Agile software development continuously generates operational data through sprint execution, task completion, and effort estimation. However, the quality of these data is rarely assessed before they are used for analytics and predictive modeling, particularly in non-industrial settings. Existing data quality approaches primarily focus [...] Read more.

Agile software development continuously generates operational data through sprint execution, task completion, and effort estimation. However, the quality of these data is rarely assessed before they are used for analytics and predictive modeling, particularly in non-industrial settings. Existing data quality approaches primarily focus on industrial big data contexts and provide limited guidance for evaluating agile sprint repositories. This study proposes FQDA-ML, a hierarchical machine learning-powered framework for assessing sprint data quality and analyzing its relationship with delivery performance. The framework adapts Cai and Zhu’s five-dimensional data quality model and operationalizes it through 13 measurable indicators aggregated into the Sprint Data Quality Index (SDQI). The framework was validated using 129 real sprints from 13 software development teams during one academic year, organized into two consecutive academic semesters (2024-S1 and 2024-S2), corresponding to two student cohorts involved in community engagement software projects. The results showed a mean SDQI of 0.603 ± 0.217 and a strong association between SDQI and sprint completion rate

(ρ = 0.719 . p < 0.001)

. The predictive evaluation achieved an AUC-ROC of 0.968 under standard five-fold cross-validation and 0.726 under Leave-One-Group-Out validation, highlighting the influence of team-level dependency on model generalization. The findings provide an empirically validated framework for data quality assessment and predictive software analytics in agile environments. Full article

(This article belongs to the Topic Data Intelligence and Computational Analytics)

► Show Figures

Figure 1

38 pages, 19248 KB

Open AccessSystematic Review

Sustainability–Resilience Trade-Offs in Edge-Enabled Systems: A Comprehensive Survey

by Nithya Nedungadi and Sriram Sankaran

Future Internet 2026, 18(9), 472; https://doi.org/10.3390/fi18090472 - 8 Sep 2026

Viewed by 386

Abstract

Edge-enabled Internet of Things (IoT) systems are rapidly becoming the operational substrate of mission-critical infrastructure spanning industrial automation, smart healthcare, vehicular ecosystems, and cyber–physical environments. The distributed, resource-constrained, and physically exposed nature of these systems makes them persistent targets for a diverse and [...] Read more.

Edge-enabled Internet of Things (IoT) systems are rapidly becoming the operational substrate of mission-critical infrastructure spanning industrial automation, smart healthcare, vehicular ecosystems, and cyber–physical environments. The distributed, resource-constrained, and physically exposed nature of these systems makes them persistent targets for a diverse and evolving spectrum of cyber attacks. Critically, cyber attacks on edge-enabled IoT systems do not merely threaten data confidentiality; they simultaneously erode two interdependent operational objectives: sustainability, the ability of the system to maintain continuous, energy-efficient operation within its resource envelope and resilience, the ability to absorb adversarial disruptions, recover operational continuity, and adapt to prevent recurrence. The structural conflict between defending sustainability and maintaining resilience under active cyberattack conditions constitutes a research gap that prior surveys have not systematically addressed. This survey introduces a cyber attack-driven Sustainability–Resilience (S-R) framework that positions cyber threats as the primary stressor forcing a bilateral trade-off between operational efficiency and continuity in edge-enabled IoT systems. A five-layer, attack-centric taxonomy is developed spanning: network-layer attacks (DDoS, MitM, routing manipulation, jamming); device and firmware attacks (malware injection, firmware compromise, sensor spoofing); data and AI/ML attacks (adversarial inputs, data poisoning, model inversion); federated and Byzantine attacks (gradient poisoning, backdoor injection, free-riding); and advanced persistent threats (APT-class intrusions, ransomware, LLM prompt injection, zero-day exploitation). For each attack class, the survey systematically analyses the impact on sustainability and resilience objectives, the resulting S-R conflict, and the state-of-the-art defensive strategies. The framework is formalised as a maximin optimisation over the joint S-R objective surface, incorporating the adaptive, goal-directed nature of the adversary through a game-theoretic formulation. Cross-domain analysis spanning Industrial IoT, smart healthcare, Internet of Vehicles, UAV-assisted IoT, smart grids, and tactical edge networks establishes domain-specific S-R operating constraints under representative attack scenarios. The survey concludes with a structured characterisation of open research challenges and forward-looking directions, providing a prioritised research agenda for advancing simultaneously sustainable and adversarially resilient edge-enabled IoT ecosystems. Full article

(This article belongs to the Special Issue Security and Privacy Issues in the Internet of Cloud—2nd Edition)

► Show Figures

Figure 1

26 pages, 455 KB

Open AccessArticle

LLM-Assisted Porting of Security-Critical C Libraries to Idiomatic Rust: A Multi-Model Empirical Study

by Marco Parrillo, Marco Grassi and Luigi Laura

Future Internet 2026, 18(9), 471; https://doi.org/10.3390/fi18090471 - 7 Sep 2026

Viewed by 522

Abstract

Memory-safety vulnerabilities remain the dominant class of security defects in C/C++ software underpinning Internet infrastructure. Rust offers a structural solution through its ownership system, yet migrating existing codebases remains costly. This paper defines a structured methodology for LLM-assisted porting of security-critical C libraries [...] Read more.

Memory-safety vulnerabilities remain the dominant class of security defects in C/C++ software underpinning Internet infrastructure. Rust offers a structural solution through its ownership system, yet migrating existing codebases remains costly. This paper defines a structured methodology for LLM-assisted porting of security-critical C libraries to idiomatic Rust and applies it to cJSON (∼3200 LOC, 14 CVEs). A manual expert porting serves as the baseline; five LLMs (Claude Opus 4.6, Gemini 3 Pro, GPT-5.4, Kimi K2.7-Code, and Qwen3.5-27B) produce independent portings in agentic mode. Verification uses an end-to-end pipeline: CVE-specific tests, coverage-guided and differential fuzzing (>1.7 billion executions), Miri analysis, and comparative benchmarking. All six portings eliminate all in-scope CVE classes by construction, with zero unsafe blocks and zero memory-safety crashes. In this case study, structural safety holds consistently across all five evaluated models and across all five Kimi repetitions, whereas code quality varies widely (0–9 residual bugs). Because only Kimi was repeated (

N = 5

), its variance bounds run-to-run noise at the 95% confidence level, against which some but not all between-model differences are distinguishable from chance; a single porting attempt costs approximately $3 in API usage. Differential fuzzing reveals complementary bugs in the manual and LLM portings, supporting a hybrid workflow, and we translate these findings into concrete practical guidance for teams planning a similar migration. These results are scoped to one compact, single-threaded C library. The entire codebase and evaluation pipeline are publicly released. Full article

(This article belongs to the Special Issue Advanced Artificial Intelligence and Machine Learning for Cybersecurity)

► Show Figures

Figure 1

36 pages, 639 KB

Open AccessSystematic Review

A Systematic Literature Review on Machine Learning for Intrusion Detection Systems

by Ali Ahmed, Ramy Mostafa, Mahmoud H. Qutqut and Noha Ragab

Future Internet 2026, 18(9), 470; https://doi.org/10.3390/fi18090470 - 7 Sep 2026

Viewed by 584

Abstract

The use of Artificial Intelligence (AI) and Machine Learning (ML) in cybersecurity, especially for creating Intrusion Detection Systems (IDSs), has become increasingly important. These systems are essential for detecting malicious behaviour, identifying network issues, and stopping cyberattacks in real time. Despite extensive research [...] Read more.

The use of Artificial Intelligence (AI) and Machine Learning (ML) in cybersecurity, especially for creating Intrusion Detection Systems (IDSs), has become increasingly important. These systems are essential for detecting malicious behaviour, identifying network issues, and stopping cyberattacks in real time. Despite extensive research on various ML and Deep Learning (DL) models for IDS, the current literature remains incomplete. It has many different datasets, methods, and evaluation standards. As cyber threats become more advanced, it is crucial to conduct a thorough analysis of ML techniques for intrusion detection. The goal of this Systematic Literature Review (SLR) is to provide a full picture of the most recent academic articles on ML-based IDS. The study addresses important research questions about the most widely used algorithms, the types of attacks and network environments covered, the methodological problems that remain unsolved, and the new trends that should shape future research. Following the PRISMA framework, we conducted a systematic review of peer-reviewed articles published between January 2022 and May 2025. We searched IEEE Xplore, ACM Digital Library, and SpringerLink, yielding 22,558 initial records. After carefully applying strict inclusion criteria, 125 papers were selected for the final analysis. We created a standardised data extraction form (i.e., using MS Excel) to gather bibliographic details, research emphasis, methodological strategies, datasets, evaluation criteria, and recognised constraints. We employed thematic analysis to develop a clear taxonomy. We identified five main research themes in our analysis: (1) ensemble and hybrid learning pipelines focused on performance optimisation (30 papers), (2) context-specific IDS designs for Internet of Things (IoT), cloud, and Software-Defined Networking (SDN) environments (34 papers), (3) data-centric engineering that deals with class imbalance and feature selection (20 papers), (4) deep neural architectures for representation learning (31 papers), and (5) trustworthiness concerns like adversarial robustness, zero-day detection, and Explainable AI (XAI) (10 papers). Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM), and Random Forests are the most commonly used algorithms, often combined. Nonetheless, significant deficiencies remain: about 2% of papers incorporate XAI, only 4% focus on adversarial robustness, and none validate their models in real-world production settings. Denial-of-Service (DoS) and Distributed DoS (DDoS) attacks are the most common types in the literature, whereas Web attacks, ransomware, and advanced persistent threats remain poorly studied. The number of publications grows at an average of 30.2% annually, but the field still relies on legacy benchmark datasets rather than operational validation. Full article

(This article belongs to the Special Issue Privacy-Preserving and Secure Machine Learning)

► Show Figures

Figure 1

62 pages, 14001 KB

Open AccessArticle

Evaluation of Vanilla and RAG-Supported LLM Comprehension of the EU AI Act

by Eleni Tsalera, Gerasimos Pagiatakis and Andreas Papadakis

Future Internet 2026, 18(9), 469; https://doi.org/10.3390/fi18090469 - 4 Sep 2026

Viewed by 430

Abstract

This paper presents a comparative benchmarking study evaluating the regulatory comprehension of four open-source large language models, TinyLlama-1.1B-Chat, Gemma-2B-Instruct, Llama-3.1-8B-Instruct, and Mistral-7B-Instruct-v0.3, on the EU Artificial Intelligence Act (Regulation EU 2024/1689). A custom benchmark of 100 multiple-choice questions was constructed and classified across [...] Read more.

This paper presents a comparative benchmarking study evaluating the regulatory comprehension of four open-source large language models, TinyLlama-1.1B-Chat, Gemma-2B-Instruct, Llama-3.1-8B-Instruct, and Mistral-7B-Instruct-v0.3, on the EU Artificial Intelligence Act (Regulation EU 2024/1689). A custom benchmark of 100 multiple-choice questions was constructed and classified across three dimensions, namely cognitive difficulty, knowledge type, and regulatory mechanism, with a balanced answer key distribution. Each model was evaluated under vanilla inference, relying exclusively on parametric knowledge, and retrieval-augmented generation (RAG), in which relevant Act passages are retrieved using a FAISS-indexed sentence embedding pipeline with standardized top three chunk retrieval applied uniformly across all models. Vanilla accuracy ranges from 22.0% for TinyLlama-1.1B to 80.0% for Llama-3.1-8B, indicating that larger models perform better. Under the retrieval configuration employed, RAG improves performance of the evaluated models, with gains from 3.0 percentage points for TinyLlama-1.1B to 17.0 for Gemma-2B and 10.0 for both 7B–8B models. Questions related to procedural knowledge, governance and enforcement emerge as weaknesses in the baseline, vanilla setting, partially mitigated by retrieval. Qualitative analysis identifies two RAG failure modes: retrieval failures, where the embedding mechanism returns informationally insufficient passages, and integration failures, where the correct passage is retrieved but not correctly exploited. The study contributes an empirical characterization of small and medium language model regulatory comprehension and a reusable 100-question benchmark. Full article

(This article belongs to the Section Cybersecurity)

► Show Figures

Figure 1

Journal Menu

Journal Browser

Future Internet, Volume 18, Issue 9 (September 2026) – 54 articles

Further Information

Guidelines

MDPI Initiatives

Follow MDPI