Reducing residential energy consumption requires not only intelligent monitoring but also timely, personalized user engagement to drive meaningful behavioral change. However, many existing systems rely on centralized data processing, raising privacy concerns, increasing communication overhead, and limiting real-time responsiveness. They also often lack context-aware feedback mechanisms that help users make informed decisions. This study introduces a fully integrated smart plug system that combines micro-moment (MM) classification with a lightweight recommender system (RS) to deliver privacy-preserving, real-time energy-saving suggestions. Built on a semi-supervised federated learning (SSFL) framework, the system enables on-device inference without transmitting raw user data ensuring data privacy while supporting adaptive behavior modeling. The RS is triggered by transitions in MM classes and delivers recommendations enriched with real-time energy metrics, estimated cost savings, and environmental context such as room temperature and air quality. The system was implemented using Home Assistant (HA) for seamless sensor integration and appliance control and deployed in a 14-day pilot study across six residential households in Doha. During the deployment, the system achieved an average energy reduction of approximately 15.63% of total usage, with a 75% acceptance rate for anomaly-triggered recommendations. These results demonstrate the practicality and effectiveness of decentralized, edge-based intelligence for energy behavior modeling and user guidance. Overall, the proposed smart plug system shows strong promise for secure, responsive, and user-centric energy optimization in real-world smart home environments.
Human activity recognition (HAR) in Internet of Things (IoT) environments must operate under strict bandwidth, latency, energy, and privacy constraints, which makes continuous transmission of raw sensor streams impractical. This work aims to design an IoT-ready semantic HAR framework that reduces communication cost while preserving recognition reliability and enabling fast personalization across tenants and devices. To achieve this, we propose a task-aware semantic communication pipeline in which each device runs a lightweight encoder to transmit compact activity-oriented representations instead of raw signals. The framework further combines fidelity-aware semantic compression, few-shot prototype-based personalization in the latent space, calibrated uncertainty estimation, and service-level-objective-aware control of bitrate and batching. The proposed design allows HAR decisions to remain accurate and adaptable even under constrained network conditions and multi-tenant edge deployment. In addition, the method provides an auditable fidelity signal that helps determine when compressed communication is sufficient and when reliability may degrade. Experimental results on representative wearable HAR benchmarks show that the proposed framework substantially reduces communication overhead while maintaining strong recognition performance and supporting low-latency operation. These results demonstrate the practical value of semantic, personalized, and resource-aware HAR for real-world IoT systems.
In this paper, we propose the head-mounted individual monitoring and finding (HIMF) framework to achieve proactive individual safety and assistance based on the Internet of Things (IoT). HIMF is the novel framework that provides the following features: i) it collects and analyzes inertial data to monitor individual behavior, detects incidents in emergency situations, and recognizes and corrects body postures as necessary in daily use, ii) heterogeneous learning attention is explored to develop the head-mounted activity recognition model to accurately identify individual status, and iii) emergent alerts and location tracking for the target individual in hazardous situations can be actively performed in a crowdsourced sensing manner. In particular, the Android-based prototype with a head-mounted IoT device for individual monitoring and finding is implemented to verify the feasibility and performance of HIMF. Experimental results show that HIMF outperforms existing methods and can efficiently provide individual status monitoring and crowdsourced target localization. The developed heterogeneous learning model achieves peak recognition accuracies of approximately 98.2% on the HARSense dataset, 94.5% on the KU-HAR dataset, and 89.2% on the UMAFall dataset. Compared to the worst-performing baseline (CNN-LSTM) and the strongest competitor (DDC3N), the developed model yields the maximum accuracy improvement of up to 14.3% and 1.5%, respectively.
Traffic accidents frequently occur at intersections due to blind spots and unexpected situations, necessitating advanced solutions. To address these challenges, the proposed EagleSight system leverages existing intersection cameras and employs a cloud-edge-device collaborative approach to deliver pedestrian trajectory prediction, thereby enhancing drivers’ situational awareness. The EagleSight system introduces a series of modules for pedestrian trajectory prediction, including motion deblurring modules, pedestrian detection and tracking modules, as well as high-precision pedestrian trajectory prediction modules. Operating in real-time, the system is easy to deploy, ensures high privacy, and effectively manages and links various objects within vehicle-road cooperation frameworks. Extensive experiments on benchmark datasets and real-world demonstrate that EagleSight outperforms existing methods in prediction accuracy, computational efficiency, and long-horizon robustness, contributing significantly to the improvement of traffic safety.
The rapid evolution of cyber-physical systems (CPS) demands wireless architectures that simultaneously achieve ultra-low latency, high reliability, energy efficiency, and strong resilience against adversarial interference. However, conventional MIMO–OFDM frameworks rely on periodic channel state information (CSI) feedback and fixed-rate adaptation, which leads to computational overhead, slow reconfiguration under dynamic channels, and vulnerability to jamming and eavesdropping attacks. To address these limitations, this paper proposes a neuromorphic-assisted adaptive MIMO–OFDM architecture that leverages event-driven spiking neural networks (SNNs) for physical-layer intelligence and secure adaptation. Unlike dense deep-learning PHY controllers, the proposed model performs sparse spike-based CSI encoding, on-device beamforming control, adaptive modulation, and secure power allocation, achieving real-time adaptation with ultra-low operational cost. A security-aware extension integrates spike-driven anomaly detection and adaptive artificial noise (AN) injection, forming a self-defensive physical layer that can suppress high-power jammers. Results demonstrate that the proposed model achieves 32% higher achievable throughput, up to 30% BER reduction, 69% faster PHY reconfiguration latency, and 70% lower energy consumption compared to traditional MIMO–OFDM and deep learning–based baselines. Moreover, it maintains a positive secrecy capacity of 2.1 bps/Hz under 20 dB jammer power, highlighting strong resilience under adversarial RF conditions. The architecture is compatible with emerging neuromorphic hardware (Intel Loihi 2, SpiNNaker), enabling energy-efficient deployment at CPS edge nodes. This work introduces a new paradigm for intelligent and secure wireless physical layers, laying a foundation for AI-native 6G wireless systems.
In industrial manufacturing, production processes are usually in the form of workflows and support flexible manufacturing requirements. However, most tasks are subject to strict deadlines, and production orders often fluctuate over time. These tasks form complex workflows, and different tasks require specific computing capabilities. Owing to the arrival of time-critical workflows, scheduling decisions should be adapted continuously considering heterogeneous computing resources. Typical resources include CPUs and GPUs, which vary in both performance and cost. Industrial scheduling systems need to rapidly evaluate changing task urgency levels to reallocate resources in a reasonable amount of time. To address these challenges, an urgency-aware workflow scheduling method, UAWS, is proposed based on deep reinforcement learning, wherein resource selection is performed to satisfy deadline constraints and reduce cost. First, tasks are sorted by urgency to ensure that the highest– priority task is selected to avoid missing deadlines. Second, a cross-attention module extracts features from the selected task and heterogeneous resource set to construct a cost-aware state representation. Third, the selected task and the cost-aware representation are combined to form the state input for training the deep Q-network. The network outputs the optimal resource strategy that balances deadline satisfaction and execution cost. Finally, we carried out sufficient experiments and the results show that the proposed method outperforms well-known baselines, as deadline satisfaction is increased by 12%, execution cost is reduced by 2.03%, and resource utilization is improved by 2.32%, respectively.
The rapid expansion of the Industrial Internet of Things (IIoT) has enabled smart factory development through interconnected edge devices but also exposes industrial systems to cybersecurity vulnerabilities, including Distributed Denial-of-Service (DDoS) attacks, reconnaissance, man-in-the-middle (MitM) intrusions, injection attacks, and malware propagation. Deep learning (DL) techniques have shown promise in IIoT threat detection but rely on large-scale labeled data, which is often isolated across factories and protected by privacy constraints. Federated learning (FL) offers a privacy-preserving alternative by enabling collaborative model training across distributed data sources. Nevertheless, standard FL faces significant challenges in IIoT contexts, including non-IID data distributions, susceptibility to malicious updates, and lack of verifiable model integrity. This paper presents a secure and robust FL framework tailored for IIoT-enabled smart factories. The proposed approach integrates (i) a behavioral validation mechanism to identify and exclude unreliable clients, (ii) secure multi-party computation (SMPC) for privacy-preserving and tamper-resistant aggregation, and (iii) a hash-based model authentication step that allows clients to verify the integrity of the received global model before adoption. To the best of our knowledge, this is the first FL framework that jointly addresses privacy preservation, adversarial robustness, and model verifiability within a unified IIoT-specific pipeline. The primary contribution is this co-design—the coordinated integration of these three security dimensions into a single coherent pipeline—rather than new individual algorithms. Comprehensive experiments on the Edge-IIoTset dataset under both IID and non-IID conditions demonstrate the framework’s superior performance in accuracy, robustness, and security, making it a viable solution for real-world industrial deployment.
Platoon-based cellular vehicle-to-everything (C-V2X) networks require efficient resource allocation to maintain timely information exchange while optimizing energy consumption in autonomous vehicle operations. Age of Information (AoI) has emerged as a critical metric for quantifying information freshness in safety-critical vehicular applications, where outdated data can lead to catastrophic consequences. The resource allocation problem involves simultaneous optimization of multiple competing objectives: minimizing AoI for roadside unit communications, ensuring reliable delivery of Cooperative Awareness Messages (CAMs) within platoons, and managing power consumption across dynamic channel conditions. Existing reinforcement learning approaches, including the benchmark multi-agent multi-task framework, rely on traditional Deep Deterministic Policy Gradient (DDPG) algorithms that struggle with the inherent hybrid action space structure of V2X systems, where discrete decisions (communication mode and channel selection) must be jointly optimized with continuous parameters (transmission power). This limitation results in training instabilities and suboptimal performance in practical deployments. We adopt the Parameterized Deep Q-Network (PDQN) formulation of [1] to naturally handle hybrid action spaces through coordinated actor and critic networks, enabling joint optimization of discrete and continuous decision variables. The proposed framework is evaluated under a block-design protocol with five independent simulation seeds against eleven benchmarks covering nine non-learning heuristics (Random, Round-Robin, Pre-Scheduled, Load-Based, Max-SINR Greedy, Nearest-RSU-First, Proportional-Fair, AoI-Greedy, Power-Aware-Greedy) and two reinforcement-learning alternatives for hybrid action spaces: Hybrid-Action Deep Deterministic Policy Gradient (DDPG) and Independent Proximal Policy Optimization (IPPO). Reported metrics include the mean ± standard deviation of episode return and, as standalone physical quality-of-service measures, the mean and 95th-percentile system AoI in seconds, the lifetime CAM bit-delivery fraction, and the average transmit power. Differences between policies are assessed with a Friedman nonparametric rank test. Under the reported configuration, PDQN attains the largest mean episode return across all evaluated policies, whereas IPPO attains the lowest mean system-average AoI (1.41 s), the lowest 95th-percentile AoI, and the lowest mean transmit power among the learned policies; DDPG (1.79 s mean AoI) is intermediate and PDQN (2.35 s) is the highest of the three, although all three reduce the mean system-average AoI several-fold relative to the heuristic benchmarks (from 8.15 s under Round-Robin). The V2V-capable policies achieve a CAM bit-delivery fraction in the 13–15% range. PDQN is therefore best on the shaped episode-return objective rather than uniformly superior on the physical metrics, on which IPPO leads.
Internet of Things (IoT) networks are susceptible to intrusions, and intrusion detection is a challenging task due to the lack of labeled attack data, the non-stationary property of benign traffic, and the emergence of new attack variants. In order to address these challenges, this paper introduces ConFID (Conformal One-Class Framework for Unsupervised Zero-Day Intrusion Detection in IoT Networks), a one-class detection framework trained only on benign traffic and incorporating three synergistic mechanisms: a hierarchical autoencoder with a temporally aware regularization objective, a multi-layer Mahalanobis scoring mechanism estimated via the Ledoit-Wolf shrinkage procedure, and a conformal calibration stage that provides a finite-sample, distribution-free guarantee that the False Positive Rate (FPR) will not exceed any user-defined level α. On the large-scale benchmark CIC-IoT-2023, with more than 46 million flows in 33 attack categories, collected from 105 real IoT devices, the framework obtains an AUROC of 0.9921 and an F1 score of 0.9406, matching or outperforming Isolation Forest, One-Class SVM, Deep SVDD, Simple Autoencoder, and a flat Mahalanobis baseline in all evaluation metrics. ConFID has a comparable detection rate to standard baselines, but the difference is that it formally guarantees the FPR, which is not provided by any existing unsupervised Intrusion Detection System baselines. The conformal guarantee is validated empirically, as the realized FPRs match the target level within 0.008 over a wide range of operating points. The leave-one-attack-out zero-day protocol achieves near-perfect detection on Mirai-based attacks, while spoofing-type threats with flow-level signatures resembling legitimate traffic remain a challenge for all evaluated methods.
Internet of Things (IoT) devices face dual-natured security threats where compromised nodes exhibit both structural irregularities (unauthorized connections) and behavioral anomalies (abnormal traffic patterns). Existing graph neural network (GNN) detectors fail to address two critical issues. First, representation contamination: malicious nodes distort benign neighbors’ embeddings through message passing, leading to 34% false positives on adjacent clean devices in standard GCN approaches. Second, objective interference: unified loss functions cause gradient conflicts between topology and attribute anomaly objectives, reducing detection performance on mixed-attack scenarios. We propose DGAD (Disentangled Graph Anomaly Detection), a dual-decoupled framework addressing both issues. While the decoupling principle is general, DGAD is deliberately instantiated for IoT security, where compromised devices actively inject malicious traffic, device graphs are constructed from time-windowed communication flows, and mixed structural–behavioral attacks dominate. DGAD employs (1) an ego encoder for contamination-free representation learning (I(hie;N(i))=0) and a graph encoder for relational context, fused via correlation-guided weighting, and (2) independent topology-contrastive and attribute-reconstruction modules to eliminate gradient interference. Experiments on three IoT datasets (Bot-IoT, IoT-23, Smart Home) demonstrate that DGAD achieves 94.7% AUROC, surpassing the best baseline by 3.5% AUROC and 10.9% AUPRC, while reducing false positives to 8%. Edge deployment on a Raspberry Pi 4 confirms practicality with real-time inference (187 ms on GPU; 3.2 s on the Raspberry Pi 4) and 52 MB memory usage, making DGAD suitable for real-world IoT gateways.
