Electronic commerce or E-commerce is an emerging key component in the global economy due to its adaptability and it is convenient for both service providers and customers. In this competitive realm, business strategies are crucial for addressing market trends, enhancing customer satisfaction, and ensuring sustainable growth. To support this, the research presents a Deep Learning (DL)-based Explainable Recommender System (ERS) designed to effective business strategies for the e-commerce sector. The proposed ERS leverages a Deep Semantic Attention Graph Convolutional Network (DSA-GCNet), which includes a Deep Graph Convolutional Network (DGCN), integrating with a Collaborative Knowledge Graph (CKG) to generate effective business strategy. Specifically, the Global Attention-Graph Convolution (GA-GC) is integrated into the DSA-GCNet for capturing the semantic relations between users, products, and strategies through attention-weighted feature aggregation for learning the context aware relationships. The strategies offered by this research focus on improving customer satisfaction, brand awareness, and revenue. The complex patterns and dependencies in data are captured by the proposed DGCN, which is used to provide precise and appropriate suggestions. Subsequently, the CKG combines with diverse data sources to enhance the relevant information for providing recommendations. Furthermore, the Local Interpretable Model Agnostic Explanation (LIME) method is used to evaluate interpretability and user trust. The performance of the proposed ERS is evaluated using precision, recall, and Normalized Discounted Cumulative Gain (NDCG) metrics. The performance of the ERS is examined against baseline models such as NeuMF and ConvMF using various datasets such as Last-FM, Yelp2018, TripAdvisor, and a collected dataset. The precision of ERS using DSA-GCNet for the 80 recommendations on the Last-FM dataset is 0.9477, which is higher than the NeuMF, ConvMF, and A2-GCN models.
Modern healthcare analytics increasingly relies on complex AI-driven workflows operating over sensitive datasets across institutional boundaries. Ensuring secure, auditable, and policy-compliant execution of these workflows remains a significant challenge. Existing systems typically treat workflow generation, policy enforcement, and execution auditing as separate concerns, resulting in fragmented governance mechanisms. This article proposes a governance-aware workflow lifecycle architecture for AI-driven analytics workflows that integrates natural language workflow synthesis, blockchain-based policy enforcement, and auditable execution, embedding governance directly within the workflow lifecycle. In our framework, large language models translate user requests into executable analytics workflows represented using Business Process Model and Notation (BPMN). Access control policies defined in XACML are compiled into smart contracts that enforce policy compliance and record immutable audit logs on a blockchain. Sensitive datasets remain off-chain while metadata and compliance events are securely recorded. We implement a prototype platform for healthcare analytics demonstrating the generation and execution of survival analysis workflows. We further evaluate the robustness of LLM-generated workflows through an ablation study examining how prompt structure affects BPMN diagram correctness and completeness. Results highlight both the promise and limitations of automated workflow generation in regulated analytics environments. Our work demonstrates the feasibility of integrating governance mechanisms directly into AI workflow infrastructures, providing a foundation for secure and auditable analytics in sensitive domains.
Clinical workflows require segmentation of the ribs in chest radiographs (X-rays), but expert-annotated data is scarce, and semi-supervised methods are prone to confirmation bias, limiting the efficacy of automated diagnostics. This paper proposes the Dual Uncertainty-Gated Switch Mean Teacher (DUGS-MT), a dual-stream semi-supervised framework that addresses semi-supervised rib segmentation issues and promotes computer-aided diagnosis in Sustainable Development Goal 3 (SDG 3). An Uncertainty-Gated Switch mechanism with intra-consistency and cross-consistency filters out unreliable pseudo-labels and prevents errors from spreading across streams by limiting each student to learning from the other teacher's predictions when pixels have high confidence and low entropy. The model was evaluated on the VinDr-RibCXR public dataset using labelled data subsets ranging from 5% to 100% to assess its anatomical segmentation performance. With only 5% labelled data, the framework achieved a Dice Similarity Coefficient (DSC) of 0.8442 and a 95th Percentile Hausdorff Distance (HD95) of 6.7034, outperforming the fully supervised baseline (DSC 0.7971). Dual consistency losses act as a strong regularizer in the fully supervised setting, reaching a dataset record DSC of 0.8861 and HD95 of 3.5780. DUGS-MT outperformed fully and semi-supervised baselines and can segment data effectively even with few labels, indicating its potential for scalable, accessible clinical use.
Adversarial deep learning has emerged as a critical area of research within computer vision, revealing both the vulnerabilities and resilience of modern neural networks. This study provides a comprehensive survey of adversarial attacks and corresponding defense mechanisms, bridging foundational concepts with recent advancements. The taxonomy of adversarial attacks is thoroughly examined, including white-box, black-box, real-world, and modern attacks, each exploiting specific vulnerabilities in deep learning models. On the defense front, this paper explores an extensive range of strategies, from adversarial training and architectural modifications to certifiable robustness frameworks and semantic-level defenses, emphasizing their strengths and limitations. Despite significant progress, the dynamic and evolving nature of adversarial threats underscores the need for adaptive, human-centered solutions. Integrating robust defenses such as input reconstruction, ensemble-based models, and dynamic mechanisms highlights the field's innovation, but challenges persist in achieving universal robustness and scalability. This survey emphasizes the importance of interdisciplinary collaboration, blending human perception, computational security, and explainable AI to design defenses that align with real-world complexities and ethical considerations. By critically analyzing current methodologies and proposing future directions, this paper contributes to advancing the security and reliability of computer vision applications. It advocates for a forward-thinking approach that balances technical sophistication with human-centric principles, ensuring that adversarial deep learning evolves into a discipline not just of technical defense, but also of trust, transparency, and accountability.
Accurate scene understanding is a key requirement for autonomous driving systems, especially when operating under varying and partially observable scenes. However, it is also known that most of the perception systems are heavily dependent on large-scale annotated datasets, making them vulnerable to sensor degradation, illumination, and weather. In this regard, we propose a unified multimodal framework combining self-supervised representation learning and supervised proposal initialization for robust scene understanding using RGB, thermal, LiDAR, and radar sensors. The proposed framework is based on modality-specific self-supervised encoders, a transformer-based multimodal fusion component, structured scene graphs, and graph neural networks for relational understanding. The experiments are performed on nuScenes, nuImages, and FLIR datasets, demonstrating the effectiveness of the proposed framework in improving robustness under varying illumination, occlusions, and sensor degradation, and also outperforming other baselines for object detection, semantic segmentation, and trajectory prediction tasks. The experiments demonstrate that multimodal self-supervised learning, when coupled with relational understanding, is a reliable and scalable framework for autonomous driving perception systems.
Industrial Internet of Things (IIoT) environments generate highly distributed and statistically heterogeneous data, making centralized intrusion detection impractical and conventional federated learning unstable under Non-IID conditions. To address this challenge, we propose GE2F-IDS, a Generalization- Guided Multimodal Federated Intrusion Detection System that integrates adaptive multimodal representation learning with a reliability-aware aggregation strategy. Unlike standard size-based aggregation, the proposed method incorporates a local generalization gap into the federated weighting mechanism, down-weighting overfitted client updates to enhance convergence stability and cross-client robustness. Additionally, a learnable gating mechanism dynamically balances network-level and system-level telemetry features to improve attack-invariant representation learning. Extensive experiments on the TON_IoT multimodal dataset under varying Dirichlet-based heterogeneity regimes demonstrate consistent superiority over Centralized training, FedAvg, and FedProx. Under severe Non-IID conditions (20 clients, $\alpha = 0.1$), GE${}^{2}$F-IDS achieves 97.63% accuracy and 0.972 macro-F1, outperforming FedAvg by 2.75 percentage points. In the best-performing configuration, the framework attains 98.91% accuracy and 0.987 macro-F1, exceeding even centralized training. Zero-day evaluation further confirms strong cross-attack generalization, achieving Cross-Attack Transfer Scores up to 0.964 and robustness ratios as high as 0.977, indicating near-preservation of indistribution performance under unseen attacks. Ablation analysis reveals that the combination of adaptive fusion and generalization-aware aggregation yields the highest robustness (CATS = 0.983; $R = 0.991$). These results demonstrate that embedding generalization-awareness directly into federated optimization substantially enhances stability, transferability, and minority-class robustness in realistic IIoT deployments
The proliferation of IoT devices has increased the demand for robust security to protect sensitive data and mitigate threats. Traditional attestation methods face scalability and efficiency challenges, requiring more adaptive solutions. This article presents ENVOT, an attestation framework for IoT swarms using Variational Autoencoders (VAEs) and ensemble classifiers based on RandomForest and XGBoost. ENVOT processes critical 10% region of RAM to reduce time and energy overhead, while statistical and frequency-domain features provide a compact representation of runtime memory behavior. The extracted features are standardized and augmented using SMOTE, encoded through a VAE, and classified using weighted soft voting. Experiments involving firmware logic errors, parameter tampering, and sensor-data manipulation achieve 90% detection accuracy. In ns-3 evaluations with 250 devices, ENVOT achieved an end-to-end attestation latency of 431.0 ms and aggregate communication energy of 0.377 mJ in the circular swarm topology. Security analysis discusses how ENVOT complements integrity-protected communication by detecting memory-observable deviations associated with firmware tampering, parameter manipulation, and sensor-data attacks.
Alaryngeal speech enhancement seeks to restore intelligibility and naturalness for individuals who have undergone laryngectomy. Existing methods based on direct spectral mapping or frame-level voice conversion fail to exploit speech's linguistic structure and are constrained by scarce pathological data. We propose a multi-stage framework integrating discrete representation learning and attention-driven feature alignment. WavLM embeddings are quantized via $k$-means and refined by a Speech Language Model that maps alaryngeal tokens toward normal-speech tokens, followed by Conformer-based reconstruction. An Attention-Driven Feature Matcher (ADFM) then refines features through three coupled modules: a residual Feature Enhancer with learnable scaling, a Multi-Head Attention Matcher performing soft alignment over $k$-nearest normal-speech neighbors, and a Quality Predictor that estimates per-frame confidence as an inference-time reliability indicator. A HiFi-GAN vocoder synthesizes the waveform. To overcome data scarcity, we introduce a Multi-Mask GAN augmentation strategy extending MaskCycleGAN-VC with a timbre similarity loss for speaker-identity preservation. Evaluations on Thai and English corpora show best speaker similarity in both languages (0.858 Thai, 0.850 English) with competitive or state-of-the-art FAD (5.252 Thai, 5.385 English) and MCD (5.644 Thai, 5.322 English), yielding up to 27.3% FAD reduction over kNN-VC. Subjective tests confirm significant listener preference (77% English, 67% Thai).
Large Language Models (LLMs) encode social biases in their internal representations, but most mitigation methods require costly retraining or rely on pruning designed for compression, not fairness. We present a fairness-first layer pruning framework that treats transformer blocks as the intervention unit and ranks them with a hybrid importance score that combines utility, redundancy, and bias. Unlike prior structured pruning methods focused on perplexity or efficiency, our approach explicitly optimizes fairness metrics during layer selection and imposes strict gates that allow only configurations with bounded utility loss and improved scores on multiple bias measures. We evaluate open-weight causal LLMs across small, medium, and larger settings TinyLlama-1.1B, Qwen2.5-0.5B, Llama-3.2-1B, Gemma-2-2B, Mistral-7B-Instruct, Gemma-2-9B, and Llama-2-7B on counterfactual benchmarks spanning eight demographic axes, using complementary fairness metrics (pseudo-log-likelihood stereotype preference, counterfactual total variation distance (TVD), stereotype KL divergence, demographic parity gap, and toxicity propensity) with neutral-text perplexity as the main utility constraint. In our experiments, the method achieves up to a 0.467 absolute TVD reduction on Qwen2.5-0.5B with no perplexity penalty and a 0.250 TVD reduction on TinyLlama-1.1B with only a 12% perplexity increase, while also lowering stereotype KL. These results show that targeted, fairness-aware block pruning can mitigate bias without catastrophic utility loss across diverse architectures, enabling more equitable LLM deployment without full retraining, guided by a hybrid, multi-criteria importance scoring framework.
Negative emotions can negatively affect how people drive, increasing risky behaviors and the likelihood of accidents. This paper presents an Affective Driving Assistant (ADA), a system designed to detect drivers' emotions and help manage them to improve road safety. ADA collects physiological signals such as heart rate, skin conductance, and temperature using a specially designed glove. These signals are processed by a lightweight arousal-valence estimation module trained on the CASE dataset and integrated into a closed-loop regulation pipeline. To help regulate emotions, ADA uses a combination of adaptive music, colored lighting, and expressive feedback. The system was tested in a virtual reality driving simulator with real users. The results show that ADA can lower emotional arousal and improve mood, which leads to better driving performance. The greatest benefit was observed when music, lighting, and expressive feedback were used together. This work highlights the value of affective computing in cars and shows that combining multiple methods can effectively manage drivers' emotions and enhance safety.
Federated learning (FL) is a decentralized machine learning paradigm in which multiple clients train models locally without sharing private data. Although FL is designed to preserve privacy, recent studies have shown that shared model updates remain vulnerable to powerful inference attacks, where adversaries intercept gradients or parameters to recover sensitive information. A prominent example is the reconstruction attack. In this paper, we propose FedGB, a generator-based FL framework that leverages Conditional Generative Adversarial Networks (CGANs) to reduce the exposure of sensitive features. Unlike classical FL and prior GAN-based approaches, FedGB shares only the generator with the server, thereby avoiding leakage through classifiers or feature extractors. FedGB improves privacy by minimizing information exchanged with the server while maintaining competitive accuracy. Extensive experiments on FMNIST, CIFAR-10, and Digit5 demonstrate that FedGB achieves competitive accuracy-privacy trade-offs. In particular, FedGB reduces the image reconstruction quality, measured by PSNR, from 19.79 to 9.31 on CIFAR-10 and from 20.17 to 6.67 on Digit5.
Driver digital twins (DTs) provide a structured mechanism for capturing individualized behavioral characteristics in connected vehicle systems. However, most existing approaches rely on centralized data access and discriminative learning pipelines, limiting privacy preservation, generative capability, and deployment scalability. In addition, the role of generative latent spaces in preserving driver identity (DI) for downstream inference remains insufficiently explored. In this work, the latent space is explicitly interpreted as a persistent DT state that captures driver-specific behavioral signatures rather than serving solely as a reconstruction intermediate. This paper presents a federated driver digital twin (FDDT) framework composed of three components: SIGNet, which learns driver-aware latent DT representations; IDInferNet, which evaluates DI separability from real latent embeddings; and DT-GDIN, which assesses DI consistency using digital-twin-generated samples. A structured evaluation protocol containing 21 multi evaluation metrics examines latent consistency, inter-driver separability, and inference reliability across both observed and generated contexts. Deployment-oriented profiling further investigates latency, model capacity, and accuracy–efficiency tradeoffs to assess feasibility on automotive edge devices. Experimental results indicate that the proposed pipeline supports DI preservation under federated constraints while remaining compatible with deployment-oriented vehicular inference requirements
Hyperscale cloud data centers are rapidly expanding to support the growing computational demands of AI/ML workloads. This growth increases power density, where even small inefficiencies result in significant energy consumption and cooling overhead. A key challenge in such environments is idle CPU-memory fragmentation, where scattered and unusable resource portions prevent efficient workload placement. As a result, underutilized machines remain active, leading to unnecessary energy usage.To address this problem, this study proposes an Idle Fragmentation-Aware Scheduling (FAS) framework. The framework introduces fragmentation-aware resource modeling and an intelligent scheduling mechanism to reduce residual CPU-memory imbalance at both node and cluster levels. Workload characteristics and fragmentation patterns are analyzed to design effective multi-resource scheduling policies that improve packing efficiency.The proposed FAS framework is implemented and evaluated in a Kubernetes environment, where container workloads are scheduled based on real-time fragmentation conditions. An energy-aware consolidation strategy is applied to safely shut down underutilized nodes. Experimental results show that FAS reduces the Resource Fragmentation Index (RFI) by 33.14%, improves average utilization by 24.8%, and decreases energy consumption by 36.8%, while maintaining 100% job acceptance and SLA reliability. These results demonstrate that fragmentation-aware scheduling significantly improves resource utilization and energy efficiency. Overall, the proposed approach provides a practical solution for sustainable cloud management and supports energy-aware orchestration aligned with UN SDG 7 (Affordable and Clean Energy) and SDG 12 (Responsible Consumption and Production).
Sarcasm detection in Bengali is challenging due to implicit sentiment, contextual dependence, and limited linguistic resources. This study proposes a novel multi-task learning framework that jointly predicts sarcasm polarity and estimates sarcasm intensity using a Cross-Task Attention mechanism to enable effective feature sharing and capture fine-grained sarcasm patterns. We preprocess a balanced Bengali dataset of 25,636 comments with a language-specific normalization pipeline and generate intensity scores using a transformer-based scoring model. The framework is evaluated on several pre-trained transformers, including BanglaBERT, Bangla-ELECTRA, mBERT, XLM-RoBERTa, and DistilBERT. BanglaBERT achieves the best performance with 77.38% accuracy and an F1-score of 0.7761, while ELECTRA and BERT-Base Uncased perform competitively. For interpretability, we apply LIME to Bengali predictions and SHAP to English-translated outputs. Overall, the proposed multi-task learning framework uses a shared transformer encoder with a Cross-Task Attention mechanism to capture sarcasm patterns more effectively.
Natural disasters pose recurring threats to human life and infrastructure, demanding intelligent systems that can process heterogeneous data streams and provide actionable insights in real time. Existing approaches often treat textual signals from social media and emergency communications separately from spatial hazard attributes, limiting their effectiveness in capturing the full complexity of evolving crises. This paper proposes an AI-driven geo–textual intelligence framework that integrates disaster-related text with GIS-based hazard features for real-time risk prediction and evacuation planning. The framework employs contextual text encoders and a neural GIS encoder, fused through an attention mechanism that dynamically weights cross-modal signals. Experiments on three datasets; social media disaster tweets, multi-label disaster response messages, and flood hazard GIS attributes. It shows that the proposed model achieves an accuracy of 0.94, macro-F1 of 0.89, IoU of 0.81, and kappa of 0.78, surpassing all classical and deep baselines. Evacuation simulations further demonstrate that the model produces safer and faster routes, reducing travel time from 42.1 to 41.5 minutes and increasing the safety index from 0.61 to 0.81. These results highlight the novelty of aligning linguistic cues with spatial context and underscore the potential of deep geo–textual fusion for enhancing situational awareness and supporting critical decision-making in disaster response.
Background: Current automated external defibrillators require pauses in cardiopulmonary resuscitation (CPR) for reliable ECG rhythm analysis, reducing chest compression fraction and delaying defibrillation. Methods: We propose a skip-connection BiLSTM autoencoder (SBAE) that removes CPR artifacts directly from one-dimensional ECG signals without auxiliary reference signals or time-frequency transformation. A cascade architecture employs a balanced model for initial denoising followed by biased models whose reconstruction is optimized toward one rhythm class. A conservative routing strategy labels cases where the two stages disagree as indeterminate. Performance was evaluated through 10-fold cross-validation with artifact-level data separation using an FDA-approved shock advisory algorithm. Results: The cascade SBAE achieved 97.51 $\pm$ 1.91% sensitivity for ventricular fibrillation, 94.58 $\pm$ 3.39% for ventricular tachycardia, 99.30 $\pm$ 0.28% specificity for normal sinus rhythm, 98.02 $\pm$ 3.02% for asystole, and 97.21 $\pm$ 1.05% for other non-shockable rhythms, exceeding American Heart Association thresholds for all categories. The indeterminate rate was 3.63 $\pm$ 1.10%. Conclusion: The proposed framework provides a compact, reference-free solution for shock advisory decision support in automated external defibrillators. Deployment on embedded AED hardware would require additional model optimization and platform-specific engineering.
Medical cities serve large populations and experience high parking demand due to heterogeneous visitor profiles and time constraints. This paper addresses the parking allocation problem in such environments by proposing a grouping-based optimization framework that allocates spaces according to the total number of passengers per vehicle rather than vehicle count alone. The proposed system integrates passenger-count information and real-time occupancy data through a coordinated architecture involving booking interfaces, IoT-based sensing, and a centralized allocation engine. Six novel algorithms are developed to achieve equitable load distribution across parking zones. Their performance is evaluated using nine instance classes and 2,790 test cases. Experimental results demonstrate that the $GVF$ algorithm achieves the highest best-solution percentage (68.2%) and the lowest average gap among the proposed algorithms, confirming its scalability and robustness in large-scale medical city environments
Federated Learning (FL) enables distributed clients, from organizational silos to mobile devices, to collaboratively train a global model. While clients only share model updates, malicious aggregators can infer sensitive information from these updates. Secure Aggregation (SA) protects individual updates, but recent work demonstrates a critical vulnerability where adversarial aggregators manipulate client selection to bypass SA protections, constituting a Biased Selection Attack (BSA). Applying pseudorandom functions like verifiable random functions (VRFs) can mitigate BSA, but degrades model performance by precluding informed client selection. Hence, ensuring both informed and secure client selection is an open challenge. To address this, we propose Adversarial Robust Federated Learning (AdRo-FL), which enables: (1) informed client selection based on client utility, and (2) a robust defense against BSA. AdRo-FL implements two client selection frameworks for distinct FL settings. The first framework assumes clients are grouped into clusters managed by a trusted cluster head (e.g., different branches of an organization). The second framework handles distributed clients where no trust relationships exist. AdRo-FL further incorporates quantization to reduce communication overhead and enforces transmission deadlines to improve efficiency. Extensive experiments on MNIST, FMNIST, SVHN, and CIFAR-10 show that AdRo-FL converges up to 1.85× faster while delivering comparable or higher final accuracy than insecure baselines.
Vision Transformer (ViT) models are rapidly emerging as a preferred choice for solving various computer vision challenges. ViTs have a high parameter count, leading to substantial computational and memory demands, often exceeding those of traditional convolutional neural networks (CNNs). While compression techniques have proven effective in reducing these burdens, fundamental differences between CNNs and ViTs, such as architectural structure, feature extraction mechanisms, and input dependency, mean that CNN-based pruning schemes cannot be directly applied to ViTs. In this paper, a comprehensive examination of ViT pruning techniques is presented, making this the first survey focused exclusively on pruning in ViTs. Existing methods are analyzed from multiple perspectives, including pruning targets, granularity, pruning schedules, acceleration objectives, and selection criteria, to provide a structured understanding of their effectiveness. The differences between CNN and ViT pruning methods are elucidated, highlighting why traditional CNN strategies are insufficient for ViTs. A quantitative comparison of representative approaches is conducted to evaluate trade-offs among computational cost, model size reduction, and accuracy preservation. Finally, practical pruning guidelines are offered for deployment on resource-constrained devices, and future research directions are discussed to encourage further advancements in efficient, architecture-aware pruning strategies.
