Computers

31 pages, 9997 KB

Open AccessArticle

Integrated Anomaly Detection and Mitigation in SDN Environments: A Hybrid Approach

by Sherzod Gulomov, Sodikjon Jumayev, Suhrobjon Bozorov, Ilkhom Boykuziev, Alpamis Kutlimuratov and Islambek Saymanov

Computers 2026, 15(9), 641; https://doi.org/10.3390/computers15090641 - 21 Sep 2026

Viewed by 275

Abstract

Modern network infrastructures are under attack from increasingly sophisticated attacks that static, rule-based defenses cannot adequately mitigate. We introduce a multilayer anomaly detection and mitigation framework for Software-Defined Networking (SDN) environments consisting of four integrated subsystems: (i) a Micro-segmentation Integrated Management and Defense [...] Read more.

Modern network infrastructures are under attack from increasingly sophisticated attacks that static, rule-based defenses cannot adequately mitigate. We introduce a multilayer anomaly detection and mitigation framework for Software-Defined Networking (SDN) environments consisting of four integrated subsystems: (i) a Micro-segmentation Integrated Management and Defense System (MIMDS), (ii) an Adaptive CNN-LSTM-Attention Deep Packet Inspection (MCLA-DPI) module, (iii) a Hybrid Adaptive Cyberattack Prediction (KBGM) framework, and (iv) an AI-driven log analysis pipeline. Detection, prediction and containment operate in parallel (as opposed to conventional approaches) and correlate results through a common risk-scoring mechanism to coordinate policy enforcement via OpenFlow and P4-compatible data planes. The experimental evaluation shows promising performance. MIMDS achieves 94.3%. detection accuracy with full traffic isolation in 10.1 s. MCLA-DPI achieves 98.9% classification accuracy with 14 ms inference latency, outperforming baseline models SVM and LSTM on encrypted traffic. KBGM achieves 98.4% detection accuracy with 7.2 ms mean response time on the CICIDS2017 dataset. All modules are trained in federated learning to preserve data locality with continuous improvements of the global model. The results collectively demonstrate quantifiable improvements over single-paradigm approaches in detection fidelity, response latency, resource efficiency, and privacy compliance. An ablation study isolates the contribution of integration itself. Removing the coordination layer while retaining all four detectors reduces accuracy from 0.892 to 0.634 and raises the false-positive rate from 0.031 to 0.436, while peak rule installation rises from 22.3 to 98.4 rules per second and oscillation events increase by two orders of magnitude; the integrated framework also exceeds its strongest individual subsystem, which reaches 0.831 accuracy at a false-positive rate of 0.117. These figures are obtained from the released reference implementation over a synthetic campaign and are reported separately from the component measurements. Full article

(This article belongs to the Special Issue Emerging Trends and Challenges in Software-Defined Networking (SDN) and Artificial Intelligence-Driven Networking—2nd Edition)

► Show Figures

Figure 1

27 pages, 34630 KB

Open AccessArticle

An Integrated regARIMA–HP Filter–CNN Framework with Attention Mechanism for Monthly Electricity Demand Forecasting

by Zhenyu Su and Zhehan Yang

Computers 2026, 15(9), 640; https://doi.org/10.3390/computers15090640 - 21 Sep 2026

Viewed by 236

Abstract

Monthly electricity demand forecasts are often affected by outliers and moving-holiday effects. This study proposes a forecasting framework that integrates regression with ARIMA errors (regARIMA), Hodrick–Prescott (HP) filter decomposition, and a multi-branch convolutional neural network (CNN) with channel attention. The regARIMA model removes [...] Read more.

Monthly electricity demand forecasts are often affected by outliers and moving-holiday effects. This study proposes a forecasting framework that integrates regression with ARIMA errors (regARIMA), Hodrick–Prescott (HP) filter decomposition, and a multi-branch convolutional neural network (CNN) with channel attention. The regARIMA model removes outlier and moving-holiday effects; the HP filter separates the adjusted series into trend and cyclical components; separate CNNs forecast these components; and the final forecast is reconstructed with moving-holiday correction. On the primary Changzhou dataset, the framework achieved the lowest two-year average RMSE (2.44), MAE (1.89), and MAPE (3.36%) and one of the highest R² values (0.91). In Guangzhou, it ranked second in the two-year averages of all four reported metrics. The top-ranked model retained the proposed X13-HP preprocessing and multi-scale CNN-attention core but added two bidirectional long short-term memory (Bi-LSTM) layers with self-attention. Per-comparison tests showed no statistically significant difference between this extended model and the proposed framework, and no multiplicity adjustment was applied. By contrast, a Bi-LSTM and self-attention model without the multi-scale CNN front end performed poorly. These results indicate that the main advantage of the proposed design lies in multi-scale convolutional feature extraction with channel attention, whereas recurrent depth alone is insufficient. The framework therefore provides accurate, stable, and structurally simpler forecasting. Full article

(This article belongs to the Special Issue AI Applications for Smart Grid Energy Management and Industrial Electrical Systems)

► Show Figures

Figure 1

51 pages, 1833 KB

Open AccessArticle

Encoder Language Models for Zero-Shot Recommender Systems: Cross-Domain and Cross-Lingual Evaluation

by Patrik Müller, Bogdan Walek and Radim Farana

Computers 2026, 15(9), 639; https://doi.org/10.3390/computers15090639 - 21 Sep 2026

Viewed by 247

Abstract

This research provides deep insights into the capabilities of encoder models, offering a unique, controlled benchmark of over 20 state-of-the-art models. The aim is to analyze the viability of a potentially advantageous recommender architecture that capitalizes on the progress made in natural language [...] Read more.

This research provides deep insights into the capabilities of encoder models, offering a unique, controlled benchmark of over 20 state-of-the-art models. The aim is to analyze the viability of a potentially advantageous recommender architecture that capitalizes on the progress made in natural language processing (NLP) without the drawbacks of more advanced deep learning models or generative decoder-only large language models. We present a zero-shot recommender architecture used in a training-free setting. In this setting, item embeddings are aggregated into weighted user profiles and scored by cosine similarity against the full candidate pool. This approach makes use of state-of-the-art encoder language models in a frozen, inference-only, context-free state. We employed a temporal evaluation designed to simulate the cold-start problem, ordering user interactions over time and progressively revealing later interactions. We also used three contrasting datasets with specific challenges, as well as baseline methods of similar complexity and resource demands, such as collaborative filtering or content-based TF-IDF methods. For our recommendation task, we also compared the local large language model with the encoder models in a limited experiment. Our results suggest that frozen embedding-based recommendation could be a viable option, particularly when considering the cost/performance ratio. However, model and method selection should be evaluated against domain-specific tasks and protocols, rather than being based on general intuitions about scaling, general-purpose benchmark scores, or even leave-one-out validation methods. Full article

► Show Figures

Figure 1

25 pages, 840 KB

Open AccessArticle

VeriCrypt-Agent: Evidence-Grounded Multi-Agent Verification for Cryptographic Dependency Updates

by Vadim Tynchenko, Aleksei Borodulin, Dmitry Martysyuk, Andrei Gantimurov, Vladimir Nelyub and Ivan Malashin

Computers 2026, 15(9), 638; https://doi.org/10.3390/computers15090638 - 21 Sep 2026

Viewed by 289

Abstract

Cryptographic dependency updates can install cleanly and pass visible regression tests while changing project-facing behavior or leaving an application inside a vulnerable version range. This study presents VeriCrypt-Agent, an evidence-grounded workflow for bounded-autonomy cryptographic maintenance. The system combines a frozen snapshot of [...] Read more.

Cryptographic dependency updates can install cleanly and pass visible regression tests while changing project-facing behavior or leaving an application inside a vulnerable version range. This study presents VeriCrypt-Agent, an evidence-grounded workflow for bounded-autonomy cryptographic maintenance. The system combines a frozen snapshot of release/advisory evidence, a public catalog of project-level executable probes, cross-version execution under old and candidate dependency versions, role-specialized LLM reasoning, and a deterministic decision gate. Automatic merging is allowed only when mandatory probes pass, policy constraints are satisfied, and the audit record is complete. On CryptoUpdate-Mini-30, ten unique package-version transitions are evaluated through 30 controlled executions. Visible tests accept all five non-mergeable transitions, while advisory-only screening accepts two of five. The deterministic All-Probes Gate performs best on this benchmark, with perfect transition-level decisions and no review cases, when all relevant probes are already known and inexpensive. The full workflow observes no unsafe auto-merges (0/5; exact 95% CI: 0.000–0.522) and routes 9/30 executions to review (0.30; exact 95% CI: 0.147–0.494). Removing grounding accepts three of five non-mergeable transitions, and a fixed probe budget accepts two of five. On 48 blinded static crypto-API audit files, including 32 unsafe cases, two-pass review detects 20/32 unsafe cases (0.625; exact 95% CI: 0.437–0.789) and reaches macro-F1

= 0.556

. These point estimates characterize evidence-backed release decisions only under the evaluated conditions; they do not establish general superiority over exhaustive deterministic testing or a deployable static vulnerability detector. Full article

(This article belongs to the Special Issue Advancing Software Engineering with Artificial Intelligence)

► Show Figures

Figure 1

37 pages, 14079 KB

Open AccessArticle

Crack-Free GPU Tessellation for Real-Time Visual Bone Drilling in Virtual Arthroscopic Simulators

by Simon D. Koch, Furkan Dinc, Tansel Halic, Hayden Reitz and Clayton Maddox

Computers 2026, 15(9), 637; https://doi.org/10.3390/computers15090637 - 20 Sep 2026

Viewed by 207

Abstract

Background/Objectives: Real-time arthroscopic bone drilling in simulators requires high graphics and haptic update rates while deforming bone visually. We present a fully GPU-resident, shader-based pipeline for the Virtual Rotator Cuff Arthroscopic Skill Trainer (ViRCAST) that avoids mesh regeneration. Methods: Instead of [...] Read more.

Background/Objectives: Real-time arthroscopic bone drilling in simulators requires high graphics and haptic update rates while deforming bone visually. We present a fully GPU-resident, shader-based pipeline for the Virtual Rotator Cuff Arthroscopic Skill Trainer (ViRCAST) that avoids mesh regeneration. Methods: Instead of changing geometry, the method updates a vector displacement map and a tessellation amount map. The pipeline uses six GPU passes: drawing-vector calculation, displacement-map generation, tessellation-map generation via Sobel edge detection, mipmapped tessellation sampling, vertex displacement, and fragment-level normal correction. A patterned-drawing scheme reduces stretched-triangle artifacts, and an edge-symmetric UV rule assigns identical tessellation factors to shared edges. Results: On a Dell desktop with an Intel Core Ultra 9, NVIDIA GeForce RTX 5090, 32 GB VRAM, 128 GB RAM, and Windows 11, 1000-frame GPU timing showed that 4K texture updates completed in 0.292 ms on average, and the full drilling update completed in 0.397 ms. This is far below the 16.7 ms budget for 60 Hz rendering. The pipeline sustained 120 fps in multi-object scenes and improved frame rate by up to 24% over uniform tessellation, with SSIM ≈ 0.99992 on depth output. Conclusions: The method provides surface deformation, complements volumetric drilling modules, and supports continuous feedback for surgical training. Full article

► Show Figures

Figure 1

25 pages, 1567 KB

Open AccessArticle

Computational Reliability of Raspberry Pi 5 Under Undervoltage and Current Limitation: Detected Computational Errors and Implications for Silent Data Corruption

by Oscar Humberto Sierra-Herrera, Rodolpho Fernando Vaz, Luis David Patarroyo-Gutierrez, Joseph Felipe Tamayo-Hernández and Mario Eduardo González-Niño

Computers 2026, 15(9), 636; https://doi.org/10.3390/computers15090636 - 20 Sep 2026

Viewed by 328

Abstract

Single-board computers (SBCs) are increasingly being used in embedded and edge computing applications, where power stability can affect computational reliability. This work experimentally characterizes an ARM-based SoC in a Raspberry Pi 5 under controlled voltage and current limitation conditions. An automated system based [...] Read more.

Single-board computers (SBCs) are increasingly being used in embedded and edge computing applications, where power stability can affect computational reliability. This work experimentally characterizes an ARM-based SoC in a Raspberry Pi 5 under controlled voltage and current limitation conditions. An automated system based on programmable power control, real-time monitoring, and stress-ng workloads with verification mechanisms was developed to evaluate system behavior under progressive power limitations. Voltage sweeps were performed with unrestricted current, while current sweeps were conducted at a fixed nominal voltage. During the experiments, operations per second (ops/s), IPC, processor frequency, temperature, cache misses, branch misses, power consumption, throttling events, and undervoltage warnings were monitored. The results showed that voltage and current limitation produced different degradation behaviors. Voltage reduction caused a progressive transition from stable operation to detected computational errors, while current limitation led to an abrupt transition toward instability and boot failure. The Raspberry Pi 5 continued operating with relatively stable IPC and processor frequency while computational errors were detected. These results indicate that conventional performance metrics alone may not be sufficient to evaluate system reliability under unstable power conditions. Temperature also increased as the supply voltage decreased, particularly near conditions where computational errors were detected. This highlights the importance of monitoring computational correctness in SBCs operating under constrained power conditions. Full article

(This article belongs to the Special Issue Edge and Fog Computing for Internet of Things Systems (3rd Edition))

► Show Figures

Figure 1

14 pages, 923 KB

Open AccessArticle

From Anxiety to Agency: Artificial Intelligence Adoption for Media and Information Literacy Among Adult Secondary Students

by Evgenia Marneri, Dimitrios E. Tzimas, Vasiliki Karamerou and Dimitrios J. Vergados

Computers 2026, 15(9), 635; https://doi.org/10.3390/computers15090635 - 20 Sep 2026

Viewed by 454

Abstract

Artificial Intelligence (AI) and Media and Information Literacy (MIL) are recognised as vital components of adult secondary education. However, we know little about how adult learners engage with AI in authentic educational contexts, given that AI adoption is a complex sociotechnical process. In [...] Read more.

Artificial Intelligence (AI) and Media and Information Literacy (MIL) are recognised as vital components of adult secondary education. However, we know little about how adult learners engage with AI in authentic educational contexts, given that AI adoption is a complex sociotechnical process. In this study, AI for MIL (AI/MIL) refers to AI-supported tools that guide learners in interpreting and engaging with digital media and information. We report findings from an ethnographic study involving eight adult students in Greek secondary education. Informed by the Unified Theory of Acceptance and Use of Technology (UTAUT), we followed participants over an eight-month educational intervention through participant observation, interviews, and field notes. Our data show that learners moved from anxiety toward more agentic forms of AI engagement. This ethnographic reinterpretation of UTAUT highlights an interplay among technological, emotional, and sociocultural factors that shape AI/MIL adoption. Beyond traditional technology acceptance models, we underscore negotiated trust, critical AI literacy, ethical awareness, and human agency within AI-mediated information ecosystems. These findings have implications for designing participatory, transparent, and ethically grounded AI initiatives in educational settings. Full article

(This article belongs to the Special Issue Computer-Assisted Learning and Teaching Tools in the AI Era)

► Show Figures

Figure 1

23 pages, 6819 KB

Open AccessArticle

Physics-Guided Dynamic Prediction and Intrinsic Interpretability of Substation Carbon-Emission Factors: A MOIRAI-2 and UPINN Fusion Framework

by Jingbo Song, Chen Chen, Song Wang, Liang Zhang, Han Yao and Tongchui Liu

Computers 2026, 15(9), 634; https://doi.org/10.3390/computers15090634 - 19 Sep 2026

Viewed by 187

Abstract

Substation-level carbon-emission factors (CEFs) are operationally relevant because substations concentrate transformer losses, auxiliary consumption, and sulfur hexafluoride (SF6) leakage at the interface between transmission and distribution. However, static or annual emission-factor methods average over heterogeneous operating regimes and cannot capture the pronounced non-stationarity [...] Read more.

Substation-level carbon-emission factors (CEFs) are operationally relevant because substations concentrate transformer losses, auxiliary consumption, and sulfur hexafluoride (SF6) leakage at the interface between transmission and distribution. However, static or annual emission-factor methods average over heterogeneous operating regimes and cannot capture the pronounced non-stationarity of substation CEFs driven by seasonal loads, stochastic maintenance events, cooling-system switching, and extreme weather. To support high-frequency dynamic carbon tracing, dispatch optimization, and audit compliance, this study proposes a physics-guided fusion framework integrating a temporal foundation model, MOIRAI-2, with a Uniform Physics-Informed Neural Network (UPINN). A 15-dimensional physically constrained feature vector is constructed from IEEE C57.91 thermal-circuit equations and ideal-gas state equations, including transformer top-oil/hot-spot temperature, SF6 pressure/density estimation, and oil-forced/air-forced (OFAF) or oil-directed/water-forced (ODWF) cooling status. MOIRAI-2 uses Any-Variate Attention with binary attention bias to model intra-variate temporal dependencies and cross-variate physical couplings, whereas UPINN embeds thermal-balance, SF6 leakage-kinetics, and CEF conservation residuals as soft constraints. An adaptive gating network balances data-driven pattern recognition and physics-driven smoothness across steady-state, extreme-event, and maintenance regimes. Validation on a 220 kV substation dataset achieves a mean absolute error (MAE) of 0.352 gCO₂e/kWh, outperforming random forest (RF), gradient boosting machine (GBM), long short-term memory (LSTM), and a Pure Transformer by 24.0%, 19.1%, 23.0%, and 14.4%, respectively. Ablation studies show that the 15-dimensional physical-feature expansion improves accuracy by 8.8%, whereas physics-loss regularization reduces prediction variance by 37%. UPINN decomposition further indicates that transformer total loss, ambient temperature, and load factor dominate CEF dynamics, and rainfall cooling reduces CEF by 0.04 gCO₂e/kWh per 20 mm increment. The framework provides a physically consistent and intrinsically interpretable basis for dynamic substation carbon accounting and low-carbon operation. Full article

(This article belongs to the Section AI-Driven Innovations)

► Show Figures

Figure 1

29 pages, 782 KB

Open AccessArticle

Generative AI, Performance, and Learning: A Framework for Comparing Interventions Across Assessment Regimes

by Attila Kovari

Computers 2026, 15(9), 633; https://doi.org/10.3390/computers15090633 - 19 Sep 2026

Viewed by 419

Abstract

Generative artificial intelligence (AI) can improve students’ work while assistance is available, but better AI-assisted performance does not necessarily show what students have learned or can do independently. The methodological gap is that outcomes collected under different AI-access, timing, and task conditions are [...] Read more.

Generative artificial intelligence (AI) can improve students’ work while assistance is available, but better AI-assisted performance does not necessarily show what students have learned or can do independently. The methodological gap is that outcomes collected under different AI-access, timing, and task conditions are often treated as comparable even when they answer different questions. This paper proposes a framework for making those conditions explicit before intervention effects are compared or synthesized. It separates what is being assessed, the task, the conditions under which the outcome is produced, and evidence used to verify AI use or non-use. A retrospective secondary analysis of Wong and Qiu illustrates the problem. For expert-rated originality, unrestricted ChatGPT-4 use outperformed a learner-first strategy on a stuffed-bunny improvement task (

+ 0.78

points), whereas learner-first outperformed unrestricted use on a vocabulary-game task under the study’s no-AI protocol (

- 0.73

; Holm-adjusted

p = 0.01351

). Usefulness showed the same directional pattern, whereas elaboration did not. These results do not establish general creativity, durable learning, or a causal effect of removing AI because task, sequence, and access conditions also changed. For educators and researchers, the practical implication is that AI-assisted performance, independent performance, retention, and transfer are different outcomes and should not be treated as interchangeable without justification. The proposed Assessment-Regime Reporting Profile provides a compact way to document these conditions before findings are generalized, ranked, or pooled. Full article

(This article belongs to the Special Issue Computer-Assisted Learning and Teaching Tools in the AI Era)

► Show Figures

Figure 1

34 pages, 3135 KB

Open AccessArticle

Reasoning Together: Designing and Evaluating MLLM Team Strategies for Multimodal Quiz Questions

by Anastasia Kotelnikova, Viktor Byzov, Maria Dolzhenkova and Evgeny Kotelnikov

Computers 2026, 15(9), 632; https://doi.org/10.3390/computers15090632 - 19 Sep 2026

Viewed by 310

Abstract

Multimodal large language models (MLLMs) often struggle with open-ended questions requiring the integration of visual evidence, indirect textual clues, and background knowledge. We investigate whether team-based inference improves performance on Russian-language multimodal What? Where? When? questions. We introduce a nine-dimensional design space and [...] Read more.

Multimodal large language models (MLLMs) often struggle with open-ended questions requiring the integration of visual evidence, indirect textual clues, and background knowledge. We investigate whether team-based inference improves performance on Russian-language multimodal What? Where? When? questions. We introduce a nine-dimensional design space and a seven-stage reasoning pipeline for MLLM teams, instantiate six collaboration strategies, and evaluate six heterogeneous MLLMs on a new multimodal dataset of 1170 questions. Matched comparisons systematically assess performance differences associated with majority voting, captain-based aggregation, rationale sharing, and an additional discussion round. All team configurations outperformed the corresponding individual models. Rationale sharing produced the strongest and most consistent gains, while additional interaction was beneficial mainly when rationales were exchanged. The best strategy, Talkative Debate with Gemma 4 as captain, achieved 31.2% accuracy, while team-based inference yielded improvements of up to 20 percentage points over the corresponding individual models. An analysis of whether correct candidate answers appeared, persisted, or disappeared between the initial and post-discussion rounds showed that debate improved performance mainly by preserving or introducing correct hypotheses before final aggregation. Thus, rationale-aware collaboration substantially improves answer accuracy on the Russian-language multimodal quiz questions studied here, although visual interpretation, knowledge gaps, implicit clues, and answer-format constraints remain major limitations. Full article

► Show Figures

Figure 1

20 pages, 10747 KB

Open AccessArticle

PatchGuard-Freq: Zero-Overhead Adversarial Patch Defense via Frequency Detection and Data-Driven Robustness

by Dejie Luan, Chenghua Li, Chunjie Zhang, Peng Li and Huachang Yang

Computers 2026, 15(9), 631; https://doi.org/10.3390/computers15090631 - 19 Sep 2026

Viewed by 308

Abstract

Adversarial patch attacks pose a tangible physical-world threat to traffic sign recognition in autonomous driving systems. Current state-of-the-art defenses based on image reconstruction require dual-model deployment and add per-frame inference latency, making them impractical for resource-constrained embedded platforms such as mass-produced ADAS systems [...] Read more.

Adversarial patch attacks pose a tangible physical-world threat to traffic sign recognition in autonomous driving systems. Current state-of-the-art defenses based on image reconstruction require dual-model deployment and add per-frame inference latency, making them impractical for resource-constrained embedded platforms such as mass-produced ADAS systems and aftermarket dashcams. This paper proposes a two-component defense that separates detection from mitigation. PatchGuard-Freq leverages frequency-domain analysis to detect attacked images with high accuracy, while adversarial fine-tuning enables the detector to recover most of its detection performance under attack without adding any inference module or architectural change. Experimental results show that the approach is effective in the latency regime targeted by this study: on the evaluated stop-sign and nine-class LISA benchmarks, under the evaluated detection placement (a fresh

110 \times 110

patch at a uniformly random position), the detector achieves AUC = 1.0 with zero false positives across all four dataset configurations and all combined frequency variants, while the frequency-only arm of the ablation separates the data only where the patch covers almost the entire input, and adversarial fine-tuning recovers attacked mAP@0.5 on the COCO stop-sign test set from 19.4% to 68.4% on the

110 \times 110

-patch configuration while degrading clean mAP by only about 0.7 points. We further show that PatchGuard-Freq is vulnerable to adaptive (defense-aware) attacks and introduce defense-aware training that restores detection of such attacks to 100% without increasing false alarms. This work characterizes the complementary relationship between reconstruction-based and robustness-based paradigms in the accuracy–efficiency design space under the evaluated conditions: the former suits compute-unconstrained scenarios while the latter serves latency-constrained deployments. All quantitative results were obtained on desktop-grade GPUs; embedded-platform latency, memory, and energy were not measured, and the embedded discussion is limited to hardware-independent parameter and FLOP counts. Full article

(This article belongs to the Special Issue Next-Generation Cyber Defense: AI, Automation and Adaptive Security)

► Show Figures

Figure 1

33 pages, 8349 KB

Open AccessArticle

PhysioSpeck-Net: Physics-Guided Feature Separation for Parameter-Efficient and Uncertainty-Aware Retinal OCT Classification Under Cross-Dataset Shift

by Tahasin Ahmed Fahim, Fatema Binte Alam, Md Shamim Ahmed and Yu Jia

Computers 2026, 15(9), 630; https://doi.org/10.3390/computers15090630 - 19 Sep 2026

Viewed by 348

Abstract

Deep networks for optical coherence tomography (OCT) classification usually learn directly from image appearance, which leaves coherent speckle, depth-dependent attenuation, point-spread-function blur, and refractive distortion entangled with pathology inside a single feature representation. This work asks whether making that separation architecturally explicit is [...] Read more.

Deep networks for optical coherence tomography (OCT) classification usually learn directly from image appearance, which leaves coherent speckle, depth-dependent attenuation, point-spread-function blur, and refractive distortion entangled with pathology inside a single feature representation. This work asks whether making that separation architecturally explicit is a useful inductive bias, rather than whether it raises benchmark accuracy. PhysioSpeck-Net routes a shared stem representation through four branches aligned one-to-one with distinct OCT image-formation effects, recombines them through cross-branch attention and a physics-guided gate, and constrains them with auxiliary objectives derived from a seven-layer retinal simulator. The framework is evaluated on a 1000-image balanced test set and, without fine-tuning, on 1400 images from a second OCT source. Under a common training protocol, the model is competitive with substantially larger convolutional and transformer baselines while using 8.80 million parameters, but the more informative results concern reliability: predictive entropy separates errors from correct predictions with error-detection AUCs of 0.9847 and 0.9688, expected calibration error remains near 4–5% on both sets, and Grad-CAM++ perturbation analysis shows that attribution faithfulness is class-dependent rather than uniformly high. Distributional analysis further shows a residual gap between simulated and clinical images. Leave-one-branch-out and loss ablations produced consistent performance reductions, paired significance testing confirmed gains over most conventional baselines, and duplicate-control analysis found no exact or confirmed near-duplicate images across the evaluated datasets. These findings support physics-guided feature separation as a compact and uncertainty-aware strategy for retinal OCT classification under dataset shift. Full article

► Show Figures

Figure 1

28 pages, 3199 KB

Open AccessArticle

A Unified Dual-Stream Framework for Heterogeneous and Imbalanced Medical Image Classification

by Samuel Ovuehor, Adi El-Dalahmeh, Usman Adeel and Jie Li

Computers 2026, 15(9), 629; https://doi.org/10.3390/computers15090629 - 17 Sep 2026

Viewed by 219

Abstract

Medical image classification models often struggle to generalise across heterogeneous clinical domains owing to variations in visual characteristics, acquisition conditions, and class imbalance. Existing studies largely address these challenges independently, with limited investigation into their combined impact on classification robustness. This paper proposes [...] Read more.

Medical image classification models often struggle to generalise across heterogeneous clinical domains owing to variations in visual characteristics, acquisition conditions, and class imbalance. Existing studies largely address these challenges independently, with limited investigation into their combined impact on classification robustness. This paper proposes a lightweight dual-stream framework based on a pretrained ConvNeXt-Tiny backbone, integrating complementary global semantic and local structural feature representations with imbalance-aware optimisation. Rather than modifying the backbone, the framework enhances feature discrimination through dual-stream processing while incorporating class-balanced focal loss and stratified sampling for minority-class recognition. The framework is evaluated independently on four public datasets spanning dermatology, ophthalmology, and gastrointestinal endoscopy using five-fold cross-validation to assess architectural robustness across heterogeneous domains rather than cross-domain transfer of a single trained model. Experimental results show strong performance across all datasets, achieving AUC values above 0.92. Ablation studies confirm that dual-stream representation learning provides the primary performance gains, while imbalance-aware optimisation further improves robustness under severe class imbalance. These findings demonstrate an effective and practical solution for medical image classification across heterogeneous clinical imaging domains without requiring dataset-specific architectural modifications. Limitations regarding independent per-domain (rather than cross-domain) evaluation, baseline comparability, and incomplete quantitative calibration analysis are discussed explicitly and identified as directions for future work. Full article

(This article belongs to the Special Issue AI and Network Science for Biological Systems and Human Health)

► Show Figures

Figure 1

24 pages, 3952 KB

Open AccessArticle

Efficient Log Anomaly Detection via Dual-Domain State Space Modeling

by Xianlang Hu, Guangsheng Feng, Ruini Wang, Dawei Yang and Chuhao Chen

Computers 2026, 15(9), 628; https://doi.org/10.3390/computers15090628 - 17 Sep 2026

Viewed by 270

Abstract

Log anomaly detection requires models that capture long-range event dependencies without the quadratic sequence-length cost of self-attention. We present LogMamba, a reconstruction-based model that combines a bidirectional selective state-space branch with a Multi-Scale Frequency Learner (MSFL). The sequence branch models ordered semantic dependencies, [...] Read more.

Log anomaly detection requires models that capture long-range event dependencies without the quadratic sequence-length cost of self-attention. We present LogMamba, a reconstruction-based model that combines a bidirectional selective state-space branch with a Multi-Scale Frequency Learner (MSFL). The sequence branch models ordered semantic dependencies, whereas the MSFL applies a Fast Fourier Transform to the sequence-position axis and processes multiple spectral bands with independent multilayer perceptrons. An adaptive gate integrates both representations before semantic-embedding reconstruction. Because the position-frequency branch uses FFT operations, the complete block has

O (L \log L)

complexity with respect to a sequence of length L. On HDFS, BGL, and Thunderbird, LogMamba obtains mean F1-scores of 0.837 ± 0.006, 0.987 ± 0.002, and 0.983 ± 0.003, respectively. The HDFS score is 6.3% higher than the strongest baseline reported in our comparison, while recall reaches 0.996 on BGL and 0.998 on Thunderbird. Component ablations indicate that the sequence and position-frequency branches contribute differently across datasets, supporting their complementary use for log-sequence reconstruction. Full article

(This article belongs to the Section AI-Driven Innovations)

► Show Figures

Figure 1

53 pages, 22707 KB

Open AccessArticle

Neuromorphic-Inspired Language Identification for Low-Resource Code-Switched Texts Using Spiking Neural Networks

by Hlaudi D. Masethe and Mosima A. Masethe

Computers 2026, 15(9), 627; https://doi.org/10.3390/computers15090627 - 17 Sep 2026

Viewed by 368

Abstract

Multilingual code-switched language identification remains a challenging task due to frequent language alternation, lexical ambiguity, and the limited availability of annotated corpora for low-resource languages. While transformer-based language models have demonstrated strong performance, their computational complexity motivates the exploration of more efficient neuromorphic [...] Read more.

Multilingual code-switched language identification remains a challenging task due to frequent language alternation, lexical ambiguity, and the limited availability of annotated corpora for low-resource languages. While transformer-based language models have demonstrated strong performance, their computational complexity motivates the exploration of more efficient neuromorphic approaches. This study proposes an optimized Spiking Neural Network (SNN) framework for multilingual code-switched language identification using spike-based neural computation. Text data are preprocessed and represented using Term Frequency–Inverse Document Frequency (TF-IDF) feature vectors, which are transformed into temporal spike trains through a rate-coding mechanism over a fixed simulation window. The encoded spike sequences are processed by a feedforward SNN employing Leaky Integrate-and-Fire (LIF) neurons. Hyperparameters are optimized using Optuna to improve classification performance. The proposed model is evaluated on a balanced multilingual code-switched dataset and compared with classical machine learning models, including Logistic Regression, Support Vector Machine, and Random Forest, as well as deep learning and transformer-based models, including BiLSTM, mBERT, AfroXLMR, and XLM-RoBERTa. Experimental results demonstrate that the optimized SNN achieves an overall classification accuracy of 81%, outperforming the baseline SNN while remaining competitive with several state-of-the-art neural models. Statistical validation using the Friedman and Nemenyi tests confirms significant performance differences among the evaluated classifiers while demonstrating that the optimized SNN performs competitively against several strong baselines. Although transformer models achieve the highest overall accuracy, the proposed SNN offers a computationally efficient neuromorphic alternative that combines temporal spike processing with stable learning behaviour for multilingual code-switched language identification. These findings demonstrate the potential of spike-based neural computing for low-resource multilingual natural language processing and provide a reproducible foundation for future neuromorphic language identification research. Full article

► Show Figures

Figure 1

23 pages, 3041 KB

Open AccessArticle

Toward Clinically Trustworthy Pathology Foundation Models for Microsatellite Instability Prescreening in Colorectal Cancer

by Nadine Huyen Nguyen, Kim Ngan Ly and Nguyen Quoc Khanh Le

Computers 2026, 15(9), 626; https://doi.org/10.3390/computers15090626 - 17 Sep 2026

Viewed by 439

Abstract

Pathology foundation models have recently emerged as powerful pretrained representations for computational pathology, yet whether complex downstream modeling is still necessary once frozen representations are evaluated under a common downstream framework remains insufficiently understood. We address this question for microsatellite-instability-high (MSI-H) prediction in [...] Read more.

Pathology foundation models have recently emerged as powerful pretrained representations for computational pathology, yet whether complex downstream modeling is still necessary once frozen representations are evaluated under a common downstream framework remains insufficiently understood. We address this question for microsatellite-instability-high (MSI-H) prediction in colorectal cancer by benchmarking nine frozen encoders—five pathology-specific (CONCH, CONCH v1.5, UNI, Virchow2, Phikon) and four conventional vision backbones (ResNet18, ResNet50, ViT-B/16, ConvNeXt-Tiny)—for colorectal cancer histopathology under a patient-level framework, using TCGA-COAD/READ as the development cohort and CPTAC-COAD as an independent external cohort. For each encoder, H&E tiles were embedded without fine-tuning, mean-pooled to slide and patient representations, and evaluated with the same logistic regression linear probe, alongside nonlinear classifiers and a gated-attention multiple-instance learning (MIL) comparator. The complete MANTIS-defined 206-patient cohort (82 MSI-H, 124 MSS) is reported as the primary development benchmark. A label provenance audit identified 15 of 206 development cohort patients (7.3%) with cross-source MSI provenance discordance, and the concordant-label 191-patient cohort (68 MSI-H, 123 MSS) is reported as a secondary sensitivity analysis. In the primary cohort, the selected pathology-specific encoders had a higher mean pooled out-of-fold AUROC than the selected conventional vision backbones (0.827 vs. 0.773; paired-bootstrap difference, +0.054; 95% interval, +0.015 to +0.094), which is reported as a descriptive benchmark rather than a formal inference about the model classes. Virchow2 and UNI produced the strongest linear probe discrimination (AUROC 0.861 and 0.855, respectively), with no consistent gain from the tested nonlinear classifiers or attention–MIL configurations, and representational similarity analysis (linear centered kernel alignment) confirmed that encoders occupy distinct feature geometries rather than converging to a shared representation. In external validation on CPTAC-COAD (105 patients; 24 MSI-H, 81 MSS), selected TCGA-trained models retained variable discrimination, including CONCH attention–MIL AUROC 0.880 and UNI linear probe AUROC 0.859, but external calibration and threshold behavior varied substantially; for CONCH attention–MIL, the development-derived threshold did not transport under the external ensemble implementation (specificity 0.000), whereas threshold-only local adaptation on a small external subset improved median specificity to 0.686 without updating model weights. These findings indicate that modern pathology foundation models can encode MSI-associated morphology in frozen representations under this benchmark, while decision threshold transportability and multimodal or explainability extensions remain open questions for future work. Full article

(This article belongs to the Special Issue Multimodal Medical AI: Fusion, Interpretation, and Clinical Translation)

► Show Figures

Graphical abstract

55 pages, 7230 KB

Open AccessArticle

Agentic Cryptographic Debt: Repository-Level Measurement of Post-Quantum Migration Regression Under Autonomous AI Software Development

by Robert Campbell

Computers 2026, 15(9), 625; https://doi.org/10.3390/computers15090625 - 16 Sep 2026

Viewed by 309

Abstract

Post-quantum migration is becoming binding for regulated systems, and delegating it to AI coding assistants is a plausible response. We measured what two open-weight coding models produced when asked to migrate a repository from RSA-based JSON Web Token signing to ML-DSA (FIPS 204), [...] Read more.

Post-quantum migration is becoming binding for regulated systems, and delegating it to AI coding assistants is a plausible response. We measured what two open-weight coding models produced when asked to migrate a repository from RSA-based JSON Web Token signing to ML-DSA (FIPS 204), scored by static analysis and execution, on a pinned Go application whose ecosystem supplies ML-DSA and a Python service whose JOSE dependency lacks support in five audited libraries. Across 56 runs in four conditions plus a nested ablation, no run migrated successfully where a conforming primitive was available, and none accurately reported the blocker. Tools and a compiler made failure surface later, not less often: no modified artifact compiled unless ML-DSA key generation had been stubbed, and the two runs that compiled reported success without a working ML-DSA signing path. Qwen3-Coder substituted Ed25519 in all five feasible-arm one-shot runs, despite identifying its quantum vulnerability in a separate direct-question probe. For Devstral Small 2, naming the correct library and selected API functions eliminated module hallucination and stubbing but produced no successful migration: all five runs made the same API error, and two incorrectly declared the migration impossible. A reference migration built after review with the same library and disclosed scope extensions passes the scorer; scoring it exposed four acceptance-path defects that would each have rejected a correct migration. Results are descriptive and bounded to the models, repositories, and runs examined. To our knowledge, this is the first repository-level evaluation of a standardized post-quantum migration to score blocker reporting alongside outcome. Full article

► Show Figures

Graphical abstract

51 pages, 3157 KB

Open AccessReview

Deep Learning Approaches for Phishing Detection: A Systematic Review with a Focus on Mobile Deployment, Explainability, and Temporal Modelling

by Ivonne Kuma Nketia and Okuthe P. Kogeda

Computers 2026, 15(9), 624; https://doi.org/10.3390/computers15090624 - 16 Sep 2026

Viewed by 441

Abstract

Financial loss, data theft, and user harm can result from phishing attacks, which are increasing in frequency and variety beyond traditional email-based attacks to include mobile forms like smishing, quishing, malicious mobile applications, and mobile phishing websites. While a number of reviews have [...] Read more.

Financial loss, data theft, and user harm can result from phishing attacks, which are increasing in frequency and variety beyond traditional email-based attacks to include mobile forms like smishing, quishing, malicious mobile applications, and mobile phishing websites. While a number of reviews have already addressed phishing, Android Mal-Ware, and Explainable Artificial Intelligence (XAI), the literature is still disjointed in the lightweight deployment, model explainability, temporal attack behaviour, coordinated phishing detection, and evaluation datasets of intelligent mobile systems. The purpose of this systematic literature review (SLR) is to analyse all the major phishing detection methods using deep learning published from 2020 to 2026 and to focus on the key features of these methods, such as the model architectures, the mobile deployment, the explainability, the temporal modelling, coordinated detection, and the evaluation datasets. Studies were identified, screened, and assessed for eligibility using the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines before being included in the review for qualitative synthesis. Out of 101 studies that were included in the final review, the evidence was comparatively synthesised considering the following: learning architecture, detection modality, deployment characteristics, explainability mechanisms, the ability to model time, and the evaluation performance reported. The synthesis results show that phishing detection is generally good, with the accuracy of the transformer-based, hybrid, and attention-enhanced architectures being above 97% for some of the selected studies, and lightweight and edge-oriented architectures making the deployment more feasible for resource-constrained mobile environments. The reviewed evidence also demonstrates the increasing combination of XAI, temporal sequence modelling and multimodal analysis to enhance the transparency and robustness of detection. Comparisons between studies are, however, restricted by variations in the datasets, the representation of the features, the evaluation protocols, and the performance metrics. The results highlight the importance of future research focused on explaining lightweight, temporally adaptive, and robust phishing detection frameworks for intelligent mobile systems. Full article

► Show Figures

Figure 1

32 pages, 549 KB

Open AccessArticle

Open or Frontier? A Cost- and Energy-Aware Benchmark of Large Language Models for Software Vulnerability Detection

by Patrick Deininger and Wolfgang Slany

Computers 2026, 15(9), 623; https://doi.org/10.3390/computers15090623 - 16 Sep 2026

Viewed by 424

Abstract

Large language models (LLMs) are increasingly applied to software vulnerability detection, but evaluations report accuracy while ignoring inference cost and energy and under-represent open-weight models relative to proprietary systems. We benchmark eight LLMs, three frontier, and five open-weight on a stratified 1549-function subset [...] Read more.

Large language models (LLMs) are increasingly applied to software vulnerability detection, but evaluations report accuracy while ignoring inference cost and energy and under-represent open-weight models relative to proprietary systems. We benchmark eight LLMs, three frontier, and five open-weight on a stratified 1549-function subset of label-clean PrimeVul, treating cost and energy as first-class axes alongside detection quality. Cost is measured directly; energy is measured on-GPU across a concurrency sweep for three locally servable open models and FLOP-estimated with a sensitivity range for the API-served ones. Efficiency is the robust finding: open-weight models occupy the quality-efficiency Pareto frontier in every configuration tested, and no frontier model is Pareto-optimal; this is a 4-billion-parameter model matching the best frontier system’s quality at one sixty-sixth of the list price. On quality, at a matched output budget, the best open model significantly exceeds every frontier model (0.711 balanced accuracy against 0.606–0.653), although the strongest frontier system is level with the next two open models. Two findings bound the practical reading. A 125-million-parameter detector fine-tuned on PrimeVul outperforms all eight LLMs (0.765), so where in-distribution labels exist, a small task-specific model is the better instrument. It should be noted that nothing here is deployment-ready: at the natural 1:44 prevalence, precision is 2.3–8.2%. Full article

(This article belongs to the Special Issue Trustworthy and Efficient Large Language Models: Methods, Systems and Applications)

► Show Figures

Figure 1

19 pages, 3248 KB

Open AccessArticle

Development of an ATrUNet Architecture for Image Segmentation of Oncorhynchus mykiss in the Peruvian Highlands

by Wilson Mamani, José Cruz, Ferdinand Pineda, Christian Romero, Luis Baca, Norman Beltrán, Vilma Sarmiento, Helarf Calcina, Severo Huaquipaco, Erick Toque, Anibal Flores, Víctor Yana-Mamani and Saul Huaquipaco

Computers 2026, 15(9), 622; https://doi.org/10.3390/computers15090622 - 15 Sep 2026

Viewed by 253

Abstract

Image segmentation enables accurate object identification, a key requirement in computer vision applications. In aquaculture, this technology is essential for monitoring and management of species such as Oncorhynchus mykiss. In the Peruvian highlands, where trout farming is a vital economic activity, robust computational [...] Read more.

Image segmentation enables accurate object identification, a key requirement in computer vision applications. In aquaculture, this technology is essential for monitoring and management of species such as Oncorhynchus mykiss. In the Peruvian highlands, where trout farming is a vital economic activity, robust computational models are needed to automate the estimation of fish size and weight and to optimize the sustainability of production systems. This research proposes ATrUNet, a U-Net-based architecture optimized for accurate segmentation of Oncorhynchus mykiss. ATrUNet improves the information flow between the encoding and decoding layers by incorporating convolutional layers, batch normalization, and activation functions. A dataset of 1166 images was constructed, processed using LabelMe with JSON annotations, and converted into binary masks. Evaluation was conducted using loss, accuracy, and IoU. As a result, ATrUNet showed higher performance than U-Net, achieving a 27.55% reduction in loss, a 0.40% increase in accuracy (0.992), and improvements in overlap metrics such as IoU and GIoU by 2.70% and 8.75%, respectively. Future work includes expanding dataset diversity, exploring instance segmentation, and optimizing for embedded deployment. This research contributes to the application of computer vision to aquaculture, with the possibility of extending it to different species and various research contexts. Full article

(This article belongs to the Special Issue Intelligent Computing and Sensing Systems for Sustainable Precision Agriculture)

► Show Figures

Figure 1

29 pages, 7443 KB

Open AccessArticle

Bearing Fault Diagnosis Under Data Imbalance and Heavy Noise: An Adaptive Weighted Heterogeneous Ensemble Learning Framework

by Tao Peng, Ran Gu, Quanjun Li, Bo Fan, Zhihong Liu and Hua Zhao

Computers 2026, 15(9), 621; https://doi.org/10.3390/computers15090621 - 15 Sep 2026

Viewed by 292

Abstract

Recent advances have been achieved in intelligent fault diagnosis of rolling bearings. However, noise interference and data imbalance remain critical challenges for achieving accurate and reliable fault diagnosis in practical industrial applications. To address the limited robustness of conventional deep learning models under [...] Read more.

Recent advances have been achieved in intelligent fault diagnosis of rolling bearings. However, noise interference and data imbalance remain critical challenges for achieving accurate and reliable fault diagnosis in practical industrial applications. To address the limited robustness of conventional deep learning models under noisy conditions and their bias toward majority classes in imbalanced scenarios, this study proposes a robust bearing fault diagnosis method based on an adaptive weighted heterogeneous ensemble learning framework. The proposed method begins with continuous wavelet transform (CWT), which is employed to preprocess raw vibration signals and convert them into time–frequency images. Subsequently, a residual convolutional denoising autoencoder augmented by the convolutional block attention module is developed, namely CBAM-RCDAE. CBAM-RCDAE is capable of effectively reducing and eliminating noise interference in two-dimensional image data, thus enhancing fault diagnosis accuracy. Furthermore, a heterogeneous ensemble learning framework consisting of three base learners, including Swin Transformer, a multi-scale convolutional neural network, and BiLSTM, is developed to enhance generalization capability. An adaptive weight selection (AWS) strategy is introduced to adjust the weights and aggregate the outputs of the three base learners for final fault classification. The proposed method is extensively evaluated on the PU and CWRU bearing datasets. Experimental results demonstrate that, under the most challenging imbalanced conditions, the proposed method improves the G-mean metric by 5.89% and 4.95% compared with the state-of-the-art methods on the PU and CWRU datasets, respectively. In addition, the proposed method exhibits superior noise robustness, enabling reliable fault diagnosis performance across various noise levels. Full article

(This article belongs to the Section AI-Driven Innovations)

► Show Figures

Figure 1

29 pages, 9324 KB

Open AccessArticle

From Single Platforms Towards Coordinated Fleets: Safety- and Security-Bounded Autonomous Multi-Robot Systems for Nuclear Decommissioning in Europe

by Abdenour Benkrid, Omar Zahra, Ankur Shukla, István Szőke, Réka Szőke, Guillaume Hueber, Bruno Angelucci, Jarkko Kotaniemi, An Bielen and Giacomo Pinagli

Computers 2026, 15(9), 620; https://doi.org/10.3390/computers15090620 - 15 Sep 2026

Viewed by 378

Abstract

Europe’s ageing nuclear fleet creates a growing need for repeated radiological characterisation of contaminated, GPS-denied facilities where human access must be kept to a minimum. However, most deployed robotic systems remain limited to a single platform with a narrow task range. Methods: Building [...] Read more.

Europe’s ageing nuclear fleet creates a growing need for repeated radiological characterisation of contaminated, GPS-denied facilities where human access must be kept to a minimum. However, most deployed robotic systems remain limited to a single platform with a narrow task range. Methods: Building on two earlier conference papers by the authors, this article presents the design, safety engineering and initial field evaluation of the EURATOM project XS-ABILITY, tracing how previous European projects shaped its architecture. The system combines legged, wheeled, rail-based and caged aerial robots equipped with gamma/neutron and beta/gamma instruments. The platforms are coordinated through a ROS 2 architecture supporting distributed SLAM, energy-aware task allocation, risk-aware navigation and a radiological digital twin. A safety and conformity framework consolidates machinery, radiation protection, aviation, cybersecurity and AI regulations, with constraints enforced by a supervisory controller. Results: A campaign at the BR1 and BR3 facilities of SCK CEN, Belgium, in April 2026 evaluated a single wheeled ground platform against criteria fixed before deployment. Radiological and pose streams were paired with a median offset of −8.6 ms and no detectable message loss; LiDAR-only localisation was maintained over a 76 m traverse with the accumulated trajectory error bounded at 0.15 m; and repeat passes agreed to a cloud-to-cloud dispersion of 1.1 cm. Fleet-level coordination, communication reliability, radiation endurance and metrological registration accuracy were not evaluated. Conclusions: The demonstrated capability is integrated single-platform operation at TRL 4–5; coordinated fleet operation remains designed but unvalidated. The forthcoming Ignalina campaign is the next validation step, while supervisory human–robot interaction remains a priority for development. Full article

(This article belongs to the Special Issue Advanced Human–Robot Interaction 2026)

► Show Figures

Graphical abstract

19 pages, 2327 KB

Open AccessReview

Collaboration and Uncertainty Management in Agile Software Development: A Systematic Integrative Review and Literature-Based Positioning of CDM

by Maha Makkass, Youness Laghouaouta and Adil Anwar

Computers 2026, 15(9), 619; https://doi.org/10.3390/computers15090619 - 15 Sep 2026

Viewed by 336

Abstract

Agile software development relies on frequent interaction, short feedback cycles, and adaptation, yet these practices do not automatically create shared understanding, visible assumptions, timely validation, or coordinated action. Relevant evidence is distributed across requirements engineering, teamwork, project uncertainty, traceability, domain-specific representation, prototyping, flow [...] Read more.

Agile software development relies on frequent interaction, short feedback cycles, and adaptation, yet these practices do not automatically create shared understanding, visible assumptions, timely validation, or coordinated action. Relevant evidence is distributed across requirements engineering, teamwork, project uncertainty, traceability, domain-specific representation, prototyping, flow management, and leadership. This PRISMA 2020-guided critical integrative review connects these streams to explain how collaboration can support uncertainty management throughout iterative development. The initial evidence corpus was assembled between February and July 2026 and was subsequently checked through six structured Scopus searches. After deduplication, 2658 database records were screened against concept-driven eligibility criteria. Thirteen sources identified through the structured Scopus update and 87 sources from the initial corpus formed the 100-source review corpus. Two additional recent sources were incorporated during revision for targeted analytical updating, yielding a final analytical registry of 102 sources. Because source selection was concept-driven and aimed at explanatory coverage, the synthesis does not estimate the prevalence or frequency of the identified themes. Problem–mechanism–outcome–boundary coding produced six evidence streams and seven connected findings. The synthesis develops a collaborative learning loop linking stakeholder intent, uncertainty, early evidence, consequential decisions, and adaptation. The combined uncertainty-ownership-and-inquiry mechanism remains a low-confidence design proposition rather than a demonstrated effect. It then positions the Collaborative Development Method (CDM) as a literature-grounded configuration of established mechanisms rather than a validated replacement for existing agile frameworks. The review clarifies supported foundations, provisional rules, boundary conditions, and testable propositions for comparative and longitudinal evaluation. Full article

► Show Figures

Graphical abstract

44 pages, 32422 KB

Open AccessArticle

Design of Real-Time Browser-Based Platform for Thermohydraulic Characterization of a Laboratory Heat Exchanger Using PolyVR

by Vasil Hristov, Nely Georgieva, Petko Tsankov and Victor Häfner

Computers 2026, 15(9), 618; https://doi.org/10.3390/computers15090618 - 14 Sep 2026

Viewed by 249

Abstract

This paper presents a real-time browser-based platform for thermohydraulic characterization of a compact laboratory heating system, developed using the PolyVR research-grade virtual reality engine. Experimental measurements are retrieved at 1 Hz from a cloud-based database and processed via browser-native computational framework that continuously [...] Read more.

This paper presents a real-time browser-based platform for thermohydraulic characterization of a compact laboratory heating system, developed using the PolyVR research-grade virtual reality engine. Experimental measurements are retrieved at 1 Hz from a cloud-based database and processed via browser-native computational framework that continuously performs thermophysical modeling, hydraulic analysis and energy balance evaluation. The system calculates the rate of heat transfer (

h

), overall heat transfer coefficient (

U

), dimensionless numbers (

R e, P r, G r, N u

), pump performance, heater efficiency and cumulative thermal energy. PolyVR provides the immersive environment in which the partial digital twin functionality is integrated alongside the browser-based thermohydraulic calculations. The whole system includes support for animations regarding flow diagrams, valve state indicators, thermal field visualization and manipulation of system elements. The system architecture is designed to work on desktops, head-mounted devices, as well as in CAVE (cave automatic virtual environment) systems with remote connection made possible via using ngrok tunnels. The experiments were separated into three categories (steady-state, dynamic and validation). Steady-state and dynamic datasets show that the browser computation with PolyVR achieves high-fidelity thermohydraulic analysis similar to that done in laboratory settings. The steady-state and transient datasets illustrate that browser-based computation provides highly accurate thermohydraulic simulation close to that of the laboratory reference computations. For all experiments performed on the platform, the deviation of measurements does not exceed

\pm

0.5

K

in temperature,

\pm

5

%

in flow rate and

\pm

1

%

in pressure. The energy balance is closed with a deviation of

\pm

2–3

%

. Full article

► Show Figures

Figure 1

14 pages, 883 KB

Open AccessArticle

Contrastive Sequence Learning for DoH Tunnel Detection and Fine-Grained Traffic Classification

by Peng Wang, Jun Yin, Yanlei Liu, Yi Gu, Tongjie Wei and Peng Zhang

Computers 2026, 15(9), 617; https://doi.org/10.3390/computers15090617 - 14 Sep 2026

Viewed by 226

Abstract

DNS over HTTPS (DoH) protects name-resolution traffic but can also conceal command-and-control and data-exfiltration channels. Detection is challenging because benign and malicious flows share similar statistics, packet relationships span long sequences, and labeled malicious examples are scarce. We present the Contrastive Learning–Transformer–Bidirectional Gated [...] Read more.

DNS over HTTPS (DoH) protects name-resolution traffic but can also conceal command-and-control and data-exfiltration channels. Detection is challenging because benign and malicious flows share similar statistics, packet relationships span long sequences, and labeled malicious examples are scarce. We present the Contrastive Learning–Transformer–Bidirectional Gated Recurrent Unit–Attention model (CL–TBiGRU–Attention). The model combines contrastive pretraining with global and bidirectional temporal modeling. We evaluate binary DoH/non-DoH detection, tunneling-tool classification on two datasets, and malware-family classification. Across these four tasks, accuracy ranges from 94.91% to 99.98%, and the F1-score ranges from 95.02% to 99.98%. The model outperforms the listed neural baselines in binary detection and all listed methods in the three fine-grained tasks. On malware-family classification, it reaches 97.00% accuracy and 97.01% F1-score, with most residual confusion occurring between Sisron and Zloader. These results show the value of contrastive sequence modeling for coarse- and fine-grained DoH traffic analysis on the evaluated benchmarks. Full article

(This article belongs to the Special Issue Using New Technologies in Cyber Security Solutions (3rd Edition))

► Show Figures

Figure 1

46 pages, 5077 KB

Open AccessArticle

A Human-Centered Trust and Optimization Framework for Adaptive Access Management in Industrial IoT Water Treatment Systems

by Abderrahim Rafae, Aicha Aiche, Mohammed Erritali and Pierre-Martin Tardif

Computers 2026, 15(9), 616; https://doi.org/10.3390/computers15090616 - 14 Sep 2026

Viewed by 224

Abstract

Industrial Internet of Things (IIoT) technologies have significantly improved the automation and monitoring of critical infrastructures such as water treatment facilities. However, the increasing interaction between human operators and cyber-physical systems has intensified insider-related cybersecurity risks. Existing trust management approaches primarily focus on [...] Read more.

Industrial Internet of Things (IIoT) technologies have significantly improved the automation and monitoring of critical infrastructures such as water treatment facilities. However, the increasing interaction between human operators and cyber-physical systems has intensified insider-related cybersecurity risks. Existing trust management approaches primarily focus on connected devices or static role-based access control mechanisms, providing limited support for continuously assessing employee trustworthiness. This paper proposes a Human-Centered Trust Framework for adaptive access management in Industrial IoT water treatment systems. The framework integrates organizational, behavioral, and operational information into a unified employee decision matrix. Criterion importance is objectively determined using the CRITIC method, while employee trustworthiness is evaluated through the TOPSIS multi-criteria decision-making approach. The resulting trust coefficients are then incorporated into a Simulated Annealing optimization model to assign employees to safety-critical industrial tasks under operational and security constraints. The framework was validated using the IBM HR Analytics Employee Attrition and Performance dataset, semantically adapted to represent employee profiles in Industrial IoT environments. Experimental results across four task-allocation scenarios involving 30, 60, 120, and 240 tasks demonstrate that the proposed trust-aware optimization strategy consistently outperforms the evaluated baseline approaches. The Simulated Annealing solution achieved average improvements of 12.53% over Random Feasible Assignment and 11.36% over the RBAC-like strategy, while remaining slightly superior to the stronger Risk-Aware Access-Control baseline, with an average improvement of 0.03%. Furthermore, the proposed optimization procedure maintained an average optimality gap of only 1.21% relative to the exact optimization solution. The proposed framework provides an explainable, data-driven, and optimization-based approach for integrating human trust assessment into Industrial IoT access management, thereby strengthening resilience against insider threats while improving the security, transparency, and adaptability of critical industrial infrastructures. Full article

(This article belongs to the Special Issue IoT: Security, Privacy and Best Practices (3rd Edition))

► Show Figures

Figure 1

17 pages, 808 KB

Open AccessFeature PaperArticle

Prior-Informed Graph Skeleton Learning for ncRNA–Drug Resistance Association Prediction

by Liye Zhu and Ping Zhang

Computers 2026, 15(9), 615; https://doi.org/10.3390/computers15090615 - 14 Sep 2026

Viewed by 227

Abstract

Identifying the associations between non-coding RNAs and drug resistance (RDRAs) is crucial for uncovering resistance mechanisms and screening effective biomarkers. However, existing graph-based methods typically rely on purely data-driven learning and commonly assume a simplified noise independence hypothesis, overlooking the complex dependencies among [...] Read more.

Identifying the associations between non-coding RNAs and drug resistance (RDRAs) is crucial for uncovering resistance mechanisms and screening effective biomarkers. However, existing graph-based methods typically rely on purely data-driven learning and commonly assume a simplified noise independence hypothesis, overlooking the complex dependencies among feature, structural, and label noise in biomedical networks. This leads to issues such as spurious associations, poor generalization, and lack of interpretability for noisy association prediction tasks. To address these challenges, we propose Prior-RDRGSE, a prior-knowledge-guided dependency-aware graph learning framework. This framework integrates both dependency-aware graph noise modeling and domain knowledge into graph representation learning. Specifically, we first construct a heterogeneous bipartite graph and employ a deep generative inference encoder to jointly infer the underlying clean graph structure and the association signals, thereby explicitly modeling and purifying the intertwined complex noise within the network. Next, we design a resistance-semantics-conditioned interaction module that injects disease-specific and mechanism-related semantic priors into attention queries, explicitly guiding subnetwork interactions in a biologically plausible manner. Furthermore, we introduce a resistance consistency constraint based on KL divergence, which regularizes model training by aligning the learned association distribution with prior distributions derived from clinical and literature data. Comprehensive experiments demonstrate that Prior-RDRGSE achieves state-of-the-art performance in RDRA prediction and significantly outperforms existing methods. Full article

(This article belongs to the Section AI-Driven Innovations)

► Show Figures

Figure 1

15 pages, 4357 KB

Open AccessArticle

A Bearing Fault Diagnosis Method for Variable Operating Conditions Based on MMDSC-CBAM-BiLSTM

by Yufang Wang, Cairong Li and Jianan Wang

Computers 2026, 15(9), 614; https://doi.org/10.3390/computers15090614 - 14 Sep 2026

Viewed by 232

Abstract

Bearing fault diagnosis under variable operating conditions is challenging because changes in rotational speed and load alter the vibration response and cause substantial distribution shifts between operating domains. To improve cross-condition fault recognition, this study proposes an MMDSC-CBAM-BiLSTM model that combines multi-scale depthwise [...] Read more.

Bearing fault diagnosis under variable operating conditions is challenging because changes in rotational speed and load alter the vibration response and cause substantial distribution shifts between operating domains. To improve cross-condition fault recognition, this study proposes an MMDSC-CBAM-BiLSTM model that combines multi-scale depthwise separable convolution (MMDSC), convolutional block attention (CBAM), and bidirectional long short-term memory (BiLSTM). The vibration signals are first transformed into time–frequency representations using a wavelet transform. MMDSC then extracts complementary fault features at multiple spatial scales with reduced convolutional cost, CBAM adaptively reweights informative channel and spatial responses, and BiLSTM models bidirectional temporal dependencies before feature fusion and Softmax classification. The physical interpretation of the diagnosis is linked to the characteristic vibration responses generated by localized defects on the inner race, outer race, and rolling element, while the network itself learns discriminative representations rather than explicitly reconstructing defect morphology. Cross-condition experiments on the CWRU and Jiangnan University bearing datasets yield average accuracies of 98.22% and 93.26%, respectively, demonstrating improved robustness and generalization under varying operating conditions. The results indicate that the proposed architecture provides an effective data-driven solution for variable-condition bearing fault diagnosis. Full article

(This article belongs to the Special Issue Advances in Failure Detection and Diagnostic Strategies: Enhancing Reliability and Safety)

► Show Figures

Figure 1

28 pages, 1568 KB

Open AccessArticle

Collaboratively Designing Curriculum-Aligned Bee-Bot Learning Resources: Insights from a Qualitative Case Study of a Professional Development Workshop for Primary Teachers

by Kalliopi Kanaki, Michail Kalogiannakis and Nicholas Zaranis

Computers 2026, 15(9), 613; https://doi.org/10.3390/computers15090613 - 14 Sep 2026

Viewed by 245

Abstract

Although educational robotics is widely recognised as an effective approach to supporting curriculum-aligned learning in primary education, comparatively little attention has been paid to how teachers develop the learning resources needed for meaningful classroom activities. This descriptive qualitative case study, situated within the [...] Read more.

Although educational robotics is widely recognised as an effective approach to supporting curriculum-aligned learning in primary education, comparatively little attention has been paid to how teachers develop the learning resources needed for meaningful classroom activities. This descriptive qualitative case study, situated within the interpretive research paradigm, examined primary school teachers’ experiences in collaboratively designing curriculum-aligned learning resources during a three-hour practical workshop focused on integrating the Bee-Bot^® robot into formal instructional settings. Data collection methods included participant observation, field notes, semi-structured interviews, photographic documentation, and participant-produced learning resources, all analysed using Reflexive Thematic Analysis. The findings indicate that participants generally found Bee-Bot^® easy to programme and recognised its educational potential for fostering pupil engagement and collaboration in interdisciplinary learning environments. However, significant challenges emerged during the collaborative design of curriculum-based learning resources, particularly in translating curriculum objectives into pedagogically meaningful and technically feasible Bee-Bot mats. The study concludes that professional development in educational robotics should extend beyond technical familiarisation and prioritise supporting teachers as designers of curriculum-aligned educational activities that actively engage pupils as co-creators of robotics-enhanced learning experiences. Full article

(This article belongs to the Special Issue STEAM Literacy and Computational Thinking in the Digital Era)

► Show Figures

Graphical abstract

13 pages, 621 KB

Open AccessArticle

Unsupervised Machine Learning Reveals Heterogeneous Acoustic Phenotypes in Autistic Adult Speech

by Georgios P. Georgiou

Computers 2026, 15(9), 612; https://doi.org/10.3390/computers15090612 - 11 Sep 2026

Viewed by 318

Abstract

Autistic speech is highly heterogeneous, yet group-level comparisons may obscure meaningful individual acoustic patterns. This study used unsupervised machine learning to identify data-driven acoustic profiles in native speakers of Cypriot Greek, including autistic and neurotypical adults. Participants produced disyllabic pseudowords across controlled phonetic [...] Read more.

Autistic speech is highly heterogeneous, yet group-level comparisons may obscure meaningful individual acoustic patterns. This study used unsupervised machine learning to identify data-driven acoustic profiles in native speakers of Cypriot Greek, including autistic and neurotypical adults. Participants produced disyllabic pseudowords across controlled phonetic and stress conditions. Sixteen acoustic measures, including fundamental frequency, formants, duration, cepstral peak prominence, Mel-frequency cepstral coefficients, jitter, shimmer, harmonics-to-noise ratio, and intensity, were summarized at the participant level and normalized appropriately. Principal component analysis retained eight components explaining 81.4% of total variance, followed by k-means clustering. A three-cluster solution provided the best silhouette coefficient among tested solutions and showed good bootstrap stability. Cluster membership was significantly associated with diagnostic group: one profile was exclusively autistic, one was relatively balanced, and one was predominantly neurotypical. The dominant acoustic dimension was driven primarily by voice-quality and spectral measures, particularly cepstral peak prominence, intensity, shimmer, harmonics-to-noise ratio, and jitter, whereas pitch and formant measures contributed comparatively little. These findings demonstrate that unsupervised acoustic profiling can reveal stable, diagnostically relevant speech phenotypes that are not captured by conventional binary group comparisons, highlighting substantial within-group heterogeneity in autistic speech and supporting more individualized approaches to characterizing vocal variation. Full article

(This article belongs to the Special Issue Machine Learning and Statistical Learning with Applications (2nd Edition))

► Show Figures

Graphical abstract

Journal Menu

Journal Browser

Computers, Volume 15, Issue 9 (September 2026) – 91 articles

Further Information

Guidelines

MDPI Initiatives

Follow MDPI