Task offloading is critical for optimizing resource allocation in edge computing systems. In practical scenarios, user applications often comprise multiple interdependent tasks, where both task dependencies and parallelism strongly affect offloading decisions. This paper presents a novel dependent task offloading framework for multi-edge server environments. The task offloading problem is formulated as a Markov Decision Process (MDP) to minimize computational delay. Task dependencies are modeled using a Directed Acyclic Graph (DAG), and a Graph Convolutional Network (GCN) encoder is employed to extract DAG features as inputs for a Deep Reinforcement Learning (DRL) model. The proposed DRL-based method applies the Proximal Policy Optimization (PPO) algorithm to simultaneously select subtasks and determine their offloading decisions. Experimental evaluations across varying numbers of subtasks confirm the effectiveness of the approach, demonstrating superior performance compared to state-of-the-art solutions.

With the rapid development of Artificial Intelligence of Things (AIoT) technologies, the security of Industrial Internet of Things (IIoT) data faces increasing challenges, particularly in time series anomaly detection. IIoT data are typically scarce in abnormal samples and noisy, making unsupervised learning a common solution. The security challenges of IIoT data in AIoT environments require robust unsupervised anomaly detection methods. While Variational Autoencoders (VAEs) excel in noise resilience, they face two critical challenges in IIoT data: difficulties in single-variable time-series modeling and conflicts between static prior assumptions and dynamic temporal features. To address these challenges, we propose the Greater Cane Rat Algorithm-enhanced Fourier-Wavelet Conditional Variational Autoencoder (GCRA-FWVAE). Our method introduces a time-frequency dual-branch architecture that synergistically combines wavelet transforms for localized transient feature extraction and Fourier transforms for global spectral characterization. These complementary representations jointly regulate the Conditional Variational Autoencoder (CVAE) reconstruction process, effectively preserving critical anomaly signatures while suppressing noise interference. The architecture is further optimized through bioinspired Greater Cane Rat Algorithm (GCRA) to improve adaptive learning capabilities. Extensive validation on the Yahoo benchmark indicates state-of-the-art performance, achieving an F1-score of 93.6% (an improvement of 4.5% over baseline VAEs) and a precision of 95.1%. These improvements significantly increase anomaly detection accuracy and robustness, particularly in the AIoT environment, where it effectively handles more complex and dynamic industrial data.

In the Internet-of-Things (IoT) scenarios, Continual Graph Learning (CGL) has become a key technique for capturing the complex relational structures underlying diverse applications including sensor networks, road systems and biomedical networks. However, the structural changes in these evolving graphs introduce instability, making catastrophic forgetting a primary challenge for CGL. Experience replay is currently a promising method, as it strikes a balance between new and old knowledge. It also provides CGL models with a human-like memory capability. However, prior work rarely leverages the graph’s intrinsic properties to proactively identify the critical patterns hiding in the evolving graphs. To this end, we propose a unified framework that integrates cohesion-subgraph awareness into existing CGL mechanisms. We propose a novel cohesive structure-aware experience replay framework that leverages intrinsic graph properties, such as 𝑘-core and 𝑘-truss metrics, to guide the selection of representative historical nodes for replay. Unlike conventional replay strategies that rely on random sampling or task-driven node selection, our approach systematically identifies structurally significant nodes that encapsulate the evolving patterns of streaming graphs. By integrating these cohesive subgraph properties into the experience replay process, our method effectively preserves critical historical knowledge while adapting to new graph structures with low computational overhead. The experimental results demonstrate that our method consistently outperforms existing replay strategies in mitigating catastrophic forgetting and maintaining classification performance. On the PubMed dataset, our 𝑘-core-based replay strategy improves the 𝐹1 score by 3.7% compared to random sampling, while reducing training time by up to 85% compared to full retraining. Similarly, on the Cora dataset, our approach achieves a 98.3% 𝐹1 score, surpassing baseline methods by 4.5%.

Integrated Internet of Things (IoT) brings novel opportunities for pervasive smart services, as these systems allow for seamless information and resource sharing among IoT devices. Meanwhile, Federated Learning (FL) has emerged as a new framework for distributed deployment of machine learning models and a promising approach for implementing intelligent IoT systems. However, integrated IoT systems are usually composed of diverse IoT devices from different systems, and thus their ownership, roles, data distribution, and capabilities are heterogeneous. Current FL algorithms mainly focus on handling non-Independent and Identically Distributed (Non-IID) issues, but often result in reduced and unsustainable performance in integrated IoT systems. Therefore, we investigate in this paper the problem of personalized and sustainable FL in integrated IoT systems. First, we argue that different parties in integrated IoT are heterogeneous and limited in available resources for FL, and these parties are also selfish and expect rational outcomes during cooperation, which is essential for guaranteeing the sustainability of integrated IoT. Then, this paper provides a novel framework for device selection in FL. It first sets one instance of the model for each device, and iteratively selects devices to participate in model training based on the joint consideration of local model accuracy, similarity of parameters, and remaining resources per device. The proposed method guarantees the rational allocation of resources to maintain balanced model performance across all devices. In this way, the sustainability of the whole IoT system is improved such that no devices will suffer extreme resource exhaustion or poor performance. Finally, extensive evaluation is conducted to validate the advanced performance of the proposed method in integrated IoT systems.

Smart Grid infrastructures have enhanced energy distribution efficiency, reliability, and sustainability, but their proper operation requires robust anomaly detection to mitigate risks from equipment failures, cyberattacks, and natural disasters. Federated Learning (FL) offers a privacy-preserving solution by allowing power plants and grid sectors to collaboratively train models without sharing raw data, addressing privacy concerns, regulatory compliance, and single points of failure that often emerge in centralized approaches. FL also improves real-time anomaly detection and scalability by adapting dynamically to different grid topologies while incurring minimal communication overhead. Within our FL framework, Transformer models excel in anomaly detection due to their self-attention mechanisms that capture intricate temporal dependencies in sensor data. Unlike traditional models, Transformers effectively learn long-range patterns, enhancing detection accuracy and responsiveness. This work conducts a comparative study of two state-of-the-art Tranformer models in an FL environment, evaluating their anomaly detection performance across four diverse smart grid datasets. To assess robustness, we introduce a GAN-based Anomaly Injection Attack (GAIA) that generates and injects realistic syntheitc anomalies. Our results indicate that both federated Transformer models achieve high detection performance across seven metrics, even under adversarial conditions, offering valuable insights into their capabilities in decentralized smart grid applications.

An increasing number of medical institutions and patients are adopting the practice of storing their data on medical cloud servers, which greatly facilitates the sharing of medical data. However, when many nodes are connected to the medical big data architecture, the overall system’s security can be compromised. To address this, an explainable dynamic security protection scheme for medical healthcare, named EHKP-Res, has been proposed. This scheme aggregates data relevance features and integrates them with a deep residual network to predict the behavior of medical staff, addressing the issue of sudden changes in their behavioral credibility. To tackle the black-box problem, a model post-interpretation scheme using Bayesian networks to generate perturbed datasets has been proposed. It calculates sample weights and performs nonlinear fitting, deriving model interpretation through the eigenvalues of the instances. Experimental results show that the proposed Hidden KP-ABE method reduces time overhead by 14.95% compared to other access control methods and achieves a prediction accuracy of 98.34% for doctor trust metrics, effectively preventing malicious behavior by doctors.

Efficient and accurate anomaly detection in a network is of great significance for maintaining network and device security. Most anomaly detection methods assume that different anomalous network data distributions are the same or similar and ignore data privacy preservation. In this paper, a novel Federated Learning (FL) is proposed that it can quickly detect different types of anomalies in Non-Independent and Identically Distributed (Non-IID) data. First, we design a multi-domain machine learning model for multi-domain data, named Aegean, which consists of two modules: an ensemble AutoEncoder (AE) and a Generative Adversarial Network (GAN). Second, because data from different domains are non-IID, we model the anomaly detection problem as a dual problem, which can be recast as a robust optimization problem. The robust optimization problem is non-convex and therefore difficult to solve. As a remedy, we formulate and solve a dual problem by taking the Lagrangian dual function of the original problem. Experiments demonstrate that Aegean significantly outperforms the current state-of-the-art methods, with a 16% F1 score improvement over that of a One-Class Support Vector Machine (OCSVM). The designed FL significantly reduces the communication overhead of FedAvg without sacrificing anomaly detection performance.

The Industrial Internet of Things (IIoT) has revolutionized conventional manufacturing industries by improving control, monitoring, and management, resulting in increased agility and long-term sustainability. The demand for processing and storage resources increases as the number of IoT devices and the data they generate increases. Cloud computing often serves these needs, but in certain scenarios, real-time data processing near the source is necessary to guarantee low latency, minimize bandwidth usage, and enhance data security. This work proposes a novel edge computing-based IIoT architecture for delay-sensitive monitoring and control in manufacturing industries. In contrast to previous work, this work is designed to enhance system performance by ensuring the availability of edge nodes and distributing task execution loads evenly across multiple edge computing nodes. The proposed architecture is further optimized for efficient allocation of resources and scheduling of tasks by the addition of a task scheduling algorithm by using Round Trip Time (RTT) and Resource Logger (RL). In an effort to optimize the response time of both the complete system and individual Device Nodes (DN), the work also considers the deployment of heterogeneous Edge Nodes (EN). The evaluation results show that the proposed architecture reduces the communication delay and overhead by 88.28% compared to the traditional cloud-based approach and 12.32% compared to the previously proposed methods ATS-FOA and RBEC/PENB, respectively, for executing IIoT applications. The proposed Resource Logger, Round Trip Time-Based Edge Node Selection (RLRTT-ESA), and task scheduling algorithms further improved the total application response time by 47.25% in comparison to the previously proposed methods ATS-FOA and RBEC/PENB, respectively.

Link failures are a critical issue that disrupts consistent communications in the fast-evolving Internet-of-Vehicles ( IoV) environment. In response to the aforementioned limitation, we introduce the rapid discovery algorithm for routes ( RADAR), a novel scheme designed for rapid route discovery in the event of abrupt link failures or topological changes. Built on a heuristic 𝑇ℎ𝑒𝑡𝑎^∗ algorithm, RADAR operates within a three-layered Software-Defined Networking ( SDN) architecture to ensure seamless communications. The first layer collects vital vehicle information, including location, speed, direction, and vehicle ID. The second layer uses an SDN edge controller to carry out local and intra-zone path discovery with the 𝑇ℎ𝑒𝑡𝑎^∗ heuristic algorithm. The third layer introduces a global controller that oversees inter-zone route discovery, thereby expanding the network’s reach and capabilities. To validate RADAR’s effectiveness, we compare its performance to those of Dijkstra-based and 𝐴^∗ path planning algorithms across key metrics—Route Discovery Messages ( RDM), Path Length ( PL), and Route Discovery Time ( RDT). Our analysis shows significant improvements with RADAR over Dijkstra, achieving up to an 80% increase in RDM, 20% in PL, and 70% in RDT. It also surpasses 𝐴^∗ by up to 70% in RDM and 60% in RDT. These results indicate that RADAR enhances routing reliability in IoV systems, thereby improving overall network performance.

In industrial Internet of Everything (IoE) environments, the precise detection of tiny foreign fibers on the surface of bobbin yarns is crucial for ensuring the quality of textile products. However, detecting these fibers often exceeds the capabilities of both human vision and existing automation equipment. To address this challenge, this research proposes a novel foreign fiber segmentation method that integrates Generative Adversarial Networks (GANs) with an enhanced encoder-decoder architecture, significantly improving detection accuracy in industrial IoE scenarios. Specifically, we develop a dual-path attention encoding network that synergistically combines MobileNetV2’s computational efficiency with ContextNet’s multi-scale contextual awareness, thereby enhancing the extraction of contextual features for tiny foreign fibers. A hybrid channel-spatial attention mechanism is designed by parallel integration of channel-wise excitation and spatial attention mapping, which substantially strengthens the capture of discriminative features for tiny foreign fibers in high-resolution images. The decoding stage employs dense skip-connections to construct multi-scale feature propagation paths, optimizing detail preservation during upsampling processes. To tackle the severe class imbalance in fiber-background pixel distribution, this research introduces a Weighted Binary Cross-Entropy (WBCE) loss function with adaptive focal weighting. Experimental results demonstrate that the proposed DeepLab-DPA framework achieves 98.77% Accuracy, 85.93% MIoU, and balanced performance metrics (87.01% Precision, 86.84% Recall, 86.92% F1-Score), confirming its effectiveness for industrial fiber detection tasks.

Autonomous Underwater Vehicle (AUV) has become an important tool to accomplish various path planning tasks due to its high intelligence and good maneuverability. Aiming at the problem of data collection at underwater Internet of Everything (IoE) nodes, this paper constructs a complex 3D marine environment based on real marine current data, and proposes a path planning algorithm based on reinforcement learning to ensure that the AUV completes the data collection with a short path length. In particular, in order to address the problem of complex path planning tasks, the Parallel Dense neural Network (PDNet) is proposed to improve the performance of the agent by extracting the core features of the input state. In addition, to simplify the reward shaping, we constructed a marine environment with sparse rewards. Sparse rewards can greatly interfere with the agent’s exploration and learning. To solve the sparse reward problem, the Hindsight Experience Replay (HER) is introduced, which not only solves the sparse reward problem, but also improves the sampling efficiency and convergence of the algorithm.

Physical layer security has emerged as a promising approach to counter eavesdropping threats in the Internet of Things (IoT). With the advancement of technology, eavesdroppers (Eves) are increasingly sophisticated, exacerbating the difficulties in ensuring the security of uplink transmissions. In situations where transmitters are unable to deter Eves from approaching at close range, these intelligent Eves can attain substantial channel gains, thereby posing significant security threats. To address these challenges, we first propose a time-domain artificial noise (AN) scheme to eliminate interference at the legitimate receiver in Multiple-Input Multiple-Output Orthogonal Frequency Division Multiplexing (MIMO-OFDM) systems. Then, we model the adversarial interaction between legitimate transceivers and intelligent Eves as a game-theoretic problem. To solve the dynamic game problem, we introduce a deep deterministic policy gradient (DDPG)-based framework, where an optimal beamforming fractional programming algorithm (OBFP) and a lower-complexity zero-forcing beamforming difference of convex algorithm (ZFDC) are designed to ascertain the most effective strategy for legitimate transceivers. Simulation results validate the effectiveness of our proposed algorithms and highlight the critical role of time-domain AN in combating intelligent eavesdropping.

In recent years, the Internet of Everything (IoE) has been developing rapidly; however, there are currently issues with efficiency and sustainability within IoE. To address this problem, model lightweighting can be employed by constraining the size of deep learning models, thereby reducing the demand for computational resources and ensuring the efficiency and sustainability of IoE devices. In this regard, we propose a model information evaluation method based on DeepDream. This method does not require real samples to participate; instead, it evaluates the importance of each neuron based on the model’s own information. Additionally, we introduce a method for automatically selecting high-information neurons, which can identify neurons that have a significant impact on the model. We also present a visualization method for neuron class information, which can visualize information related to various classes within neurons. Through experiments, we demonstrate the effectiveness of the methods we have proposed.

Video frame interpolation focuses on directly synthesizing intermediate frames by utilizing inter-frame changes, relying heavily on large, high-frame-rate video datasets for supervised training, which imposes significant demands on bandwidth and computational resources in the Internet of Everything (IoE) environments. Since video frame rate down and up conversion are inverse processes, an Invertible Neural Network (INN) provides an efficient solution by ensuring lossless and symmetrical information transfer in forward and backward processes. This paper introduces a self-supervised video frame rate conversion method based on an INN to reconstruct missing intermediate frames. By leveraging an invertible coupling structure, the model encodes the spatio-temporal features of sparse input frames into a Gaussian distribution, which effectively simulates the frame rate downsampling process. Through the network’s invertibility and lossless processing capabilities, the intermediate frames are then reconstructed through reverse inference. This approach captures missing information from a Gaussian prior, ensuring stability and realism in the generated frames. Extensive experiments on public video datasets show that the proposed method surpasses existing state-of-the-art algorithms in accuracy and efficiency, offering superior visual quality, faster processing speeds, and reduced model parameters, especially for high-frame-rate recovery.

Recently, the technique of Integrated Sensing and Communication (ISAC) has gained great attentions and is expected to enable more advanced applications in cellular Vehicle-to-Everything (V2X). We intend to use the periodical reference signals, such as the synchronization signals in 5G, to actively detect targets at different azimuths via beam scanning. This paper mainly aims to optimize the beamwidth of these reference signals in cellular-V2X to facilitate both sensing and communication functions. Firstly, to address the issue of insufficient beam duration for accurate estimation of Doppler and the corresponding velocity as well, we combine multiple beam scanning cycles to support improved sensing performance. In this context, an innovative beamwidth optimization algorithm is proposed. Specifically, we deduce the Fisher Information Matrix (FIM) of the sufficient statistic associated with target azimuths. Considering the randomness of the target position, we build the objective function based on the expected trace of the derived FIM. Additionally, we formulate the constraint of channel coherent time for efficient communications. It is worth noting that this is the first time to optimize the beamwidth of periodic reference signals for joint communication and sensing in cellular V2X. Experimental results show that the optimal beamwidth varies with the steering azimuth of antennas and reaches its maximum when the beam points at the direction perpendicular to the roadside.