This article proposes CADiS, a causality-driven anomaly detection framework, to address the challenges of root cause identification in high-dimensional Industrial Internet of Things (IIoT) multivariate time series. The essential difference between CADiS and existing correlation-driven deep models lies in its core innovation: it fundamentally redefines anomalies as the structural decay of an underlying causal mechanism, rather than merely capturing symptomatic deviations or spurious correlations. Specifically, the framework first learns a directed and lag-aware causal prior from normal data, compiling it into a structured attention mask to constrain information flow. Then, a Causal-Phase Decomposition (CPD) technique treats each time window as a microexperiment, comparing an antephase with a postphase to explicitly capture the dynamics of causal attenuation. Inference relies on a unified Causal-Change Score (CCS), which quantifies the degradation of causal association strength, directly revealing the breakdown of the system's causal logic. Furthermore, the decomposed causal change matrix allows for fine-grained and auditable root cause diagnosis. Extensive experiments on real-world industrial datasets demonstrate that CADiS significantly outperforms strong baselines, achieving the $V_{ ext{ROC}}$ of 89.93% and $V_{ ext {PR}} $ of 76.93% on Secure Water Treatment (SWaT); The $A_{ ext {PR}} $ of 18.04% and $V_{ ext{ROC}}$ of 78.38% on Server Machine Dataset (SMD), thereby validating its robustness and diagnostic precision.

The issue of association privacy leakage has become increasingly critical during data release and usage. However, traditional privacy protection techniques often struggle to address privacy leakage resulting from implicit associations within the data. In this article, we propose a novel mechanism based on Cobweb Privacy to safeguard association privacy more comprehensively. First, we design the concept of epsilon-Cobweb Privacy (epsilon-CP) specifically to address association privacy leakage. This concept extends the traditional notion of differential privacy by incorporating associated prior knowledge, thereby offering more effective and comprehensive protection of association privacy. We further demonstrate its privacy guarantees through the theoretical analysis of the relationship between epsilon-CP, differential privacy, and pufferfish privacy. Second, we quantify the privacy leakage problem mathematically and examine the utility privacy trade-off under various priors. Additionally, we present a universal framework for association privacy protection in data aggregation scenarios using the epsilon-CP mechanism. Finally, this framework is integrated with three different noise addition methods and compared against mechanisms based on differential privacy and pufferfish privacy, and its utility is validated through experiments on both non-temporal and temporal real-world datasets. The results show that epsilon-CP provides distinct advantages in the utility privacy trade-off.

Multimodal data derived from 3D point clouds and RGB images provides the potential to provide richer contextual information, making it as an important tool for improving anomaly detection effectiveness in industrial fields. Due to the reliance on external memory and insufficient integration of complementary modal features, existing approaches often suffer from high computational costs. To address this challenge, we propose a pseudo knowledge driven and lightweight reverse distillation for multimodal anomaly detection (RDMAD) to achieve real-time, lightweight and high-precision anomaly detection in industrial settings. Specifically, RDMAD integrates a multimodal teacher encoder for distillation target generation, an OCBE module, and a learnable student decoder for reconstructing fused multimodal representations. The teacher encoder extracts complementary features from RGB images and auxiliary modalities, which are fused through a modal fusion module, while a pseudo-knowledge generator is introduced to simulate anomalies by injecting noise into RGB images during training. Furthermore, a multimodal projection layer is incorporated after each fusion module to ensure compact and anomaly-free representations. Comprehensive evaluations on the MVTec-3D AD and Eyecandies benchmarks demonstrate that RDMAD outperforms state-of-the-art methods in anomaly detection and localization. Notably, it achieves an AUPRO of 97.5% for anomaly localization, underscoring its robustness and practical effectiveness.

Cyber-physical power systems (CPPSs) are the deep integration of advanced information and other technologies applied to the grid to achieve fundamental changes in the power industry. Thus, it is very significant to enhance the robustness of the power communication system in order to ensure security. For instance, when an interdependent CPPSs is under attack, the failure can diffuse along interconnect topology to the whole system, even causing a system crash. Recall that most of the existing models are one-to-one coupling structures. Drawing close to reality, we develop a modified memetic algorithm to reconstruct CPPSs with multiple-to-multiple interlinks, called MA-Multiple, which is dedicated to improving the robustness of CPPSs. To improve the accuracy of the optimal solution of the MA-Multiple, we devise a crossover operator (CO) based on the null model to increase population diversity. Further improving the suitability of algorithms on CPPSs, we propose a new Kirchhoff centrality (KIC) based on the approximation of the inverse matrix to measure network connectivity and design a local search operator (LSO). In the experiment, we comprehensively compare the MA-Multiple with four closely relevant methods in different scenarios. The results imply that MA-Multiple performs better in enhancing network robustness and resisting attacks.

Real-time network anomaly detection is essential for maintaining security and service reliability, yet it remains challenging in modern deployments where traffic (or latency) data arrive as streams and the underlying tensor size varies over time due to node churn and adaptive measurement policies. Existing online tensor decomposition based detectors typically assume single-aspect growth (often only along time) and fixed multilinear ranks, which can lead to inconsistent factor updates, error accumulation, and degraded detection under multi-aspect evolution. This paper proposes MLOTD, a meta-learning and adaptiverank online Tucker decomposition framework for joint online low-rank modeling and anomaly separation on streaming network tensors. MLOTD introduces shared meta-factors that coordinate local subtensor updates through alternating local adaptation and lightweight global meta-updates, thereby aligning incremental learning with a global objective and mitigating forgetting when multiple modes expand. To cope with evolving normal-pattern complexity, we further develop a singular-value energy based mechanism that adaptively selects multilinear ranks online with negligible overhead. Anomalies are extracted via efficient residual screening within an alternating optimization loop. Extensive experiments on three real-world datasets (Abilene, GE´ANT, and the Seattle) under diverse random and temporally structured anomaly settings demonstrate that MLOTD consistently achieves a better TPR-FPR trade-off than representative baselines while maintaining practical per-batch runtime. These results verify that combining meta-learning based structural transfer with adaptive rank control yields accurate, stable, and scalable online anomaly detection in dynamic network environments.

Next point-of-interest (POI) recommendation aims to predict a user’s subsequent location based on their check-in sequence, thereby supporting real-time decision-making for mobility and local services. However, real-world scenarios are constrained by spatial, temporal, and contextual factors, which complicate unified modeling of user preferences and degrade predictive reliability. Additionally, these scenarios also face problems such as sparse user check-in data and nonstationary preferences. Therefore, we propose a framework based on Multi-Source Graphs and Dual KAN-Transformers for next POI recommendation (MSG-DKT). Concretely, MSG-DKT is composed of three modules. (1) Short-term preference learning. Based on multi-source graphs and contextual information, a short-term preference is constructed. The S-KAN-Transformer then jointly predicts the next POI, its category, and the visit time. (2) Long-term preference learning. The long-term sequence is enhanced by the L-KAN-Transformer and aggregated by long-term self-attention into a stable preference representation. (3) Hierarchical fusion. The outputs of the short-term and long-term branches are fused hierarchically to yield the final prediction. Extensive experiments on two real-world datasets demonstrate the superiority of MSG-DKT over state-of-the-art methods for next POI recommendation.

In energy-constrained intelligent transportation systems (ITS), ensuring efficient, secure, and adaptive network communications is critical for real-time decision-making and low-latency services. Conventional normalized graph cut models have been used in distributed software-defined networking (SDN) environments to address challenges of resource overhead, energy consumption, routing, and data privacy. However, they often suffer from high computational cost and limited adaptability to dynamic traffic. To this end, we propose EFL-ITS, an energy-efficient learning framework for load balancing in software-defined networking enabled intelligent transportation systems. The framework collects valid traffic by first partitioning the global network into subgraphs via normalized graph cuts and then performing anomaly detection within each subgraph using graph attention networks. Then, subgraph-level traffic prediction is employed for analysis, followed by Q-learning-based agents to perform multi-metric path selection. Finally, federated learning aggregates local models to reduce communication overhead and protect privacy. Simulation results on SDN topologies show that EFL-ITS significantly improves throughput, delay, packet loss, and load balancing, while maintaining low energy consumption—validating its efficiency and scalability for ITS.

Vision-language models (VLMs) have shown remarkable ability in aligning visual and textual representations, enabling a wide range of multimodal applications. However, their large-scale training data inevitably raises concerns about privacy, copyright, and undesirable content, creating a strong need for machine unlearning. While existing studies mainly focus on single-shot unlearning, practical VLM deployment often involves sequential removal requests over time, giving rise to continual machine unlearning. In this work, we make the first attempt to study continual unlearning for VLMs and identify three key challenges in this setting: effectiveness in removing target knowledge, fidelity in preserving retained model utility, and persistence in preventing knowledge re-emergence under sequential updates. To address these challenges, we propose CATA, a conflict-averse task arithmetic method that represents each forget request as an unlearning task vector. By maintaining historical task vectors and performing sign-aware conflict-averse aggregation, CATA suppresses conflicting update components that may weaken previous forgetting effects. Extensive experiments under both single-shot and continual settings show that CATA outperforms baselines in terms of forgetting effectiveness, model fidelity, and forgetting persistence.

With the development of High-Performance Computing (HPC) and Artificial Intelligence (AI) technologies, Mobile CrowdSensing (MCS) plays an important role in large-scale data processing and analysis. By combining the parallel computing capabilities of HPC, MCS can quickly process complex spatiotemporal data and utilize AI to optimize task recommendations and resource allocation. However, existing task allocation and recommendation models are inefficient due to limited consideration of users’ movement, location preferences, and collaboration needs. To address these issues, a High-performance computing Enhanced Task Recommendation Strategy based on Mobile Prediction (HEtrs-MP) is proposed in this paper. It combines HPC acceleration with deep learning models. Firstly, the User Trajectory Prediction algorithm based on Convolutional Neural Network - Long Short-Term Memory (UTPCL) uses Convolutional Neural Network - Long Short-Term Memory and HPC to predict the users’ location and achieve intelligent task allocation. Secondly, the Time Fuzzy Clustering algorithm based on User Time Preference (TFC-UTP) clusters users’ time preferences, optimizes task time recommendations, and reduces disruption to users. Thirdly, the Task Recommendation algorithm based on User Collaboration (TRUC) uses HPC to analyze user similarity and form collaborative groups, thereby improving task execution efficiency. Finally, extensive experiments are conducted on the GeoLife and T-Driver datasets to validate the effectiveness of the HEtrs-MP strategy. Compared with other task recommendation strategies, the prediction accuracy of the HEtrs-MP strategy can reach up to 96%. The accuracy and hit rate increase by more than 5%. Additionally, the sensing users’ mobility costs are reduced.

Federated learning (FL) is the standard paradigm for privacy-preserving model training across distributed Industrial Internet of Things (IIoT) devices; however, deployment remains hindered by non-IID data, high communication costs, and unstable asynchronous convergence. We present energy-aware asynchronous FL with adaptive layered personalization (EAFL-ALP), which achieves a 99.8% reduction in per-round traffic while improving accuracy and robustness. The framework comprises three coordinated modules: 1) adaptive fractal-wave personalization model (AFWPM), which for each client, grows an entropy-conditioned fractal branch and prunes it with wave-collapse, yielding a self-similar, capacity-adaptive head that captures data heterogeneity; 2) layerwise quantization-based reversible DP gradient compression model (LQGCM), a variance-driven block stratified that transmits 88-bit meta tuples only, enabling codebook resonance replay, invertible vector quantization and Laplace-private gradients without any numeric payload or sparsity mask; and 3) energy minimization aggregation model (EMAM), a closed-form update that mixes staleness weights, proxy-gradient correction and exponential moving average (EMA) momentum for stable convergence on lossy links. Experiments on five IIoT benchmarks show that EAFL-ALP increases accuracy by up to 32.1%, accelerates convergence 3.3x , lowers privacy leakage by 34.1%, and reduces communication volume by two orders of magnitude with no loss of model fidelity.

Federated learning is vulnerable to model poisoning attacks in which malicious participants compromise the global model by altering the model updates. Current defense strategies are divided into three types: aggregation-based methods, validation dataset-based methods, and update distance-based methods. However, these techniques often neglect the challenges posed by device heterogeneity and asynchronous communication. Even upon identifying malicious clients, the global model may already be significantly damaged, requiring effective recovery strategies to reduce the attacker's impact. Current recovery methods, which are based on historical update records, are limited in environments with device heterogeneity and asynchronous communication. To address these problems, we introduce FedHAN, a reliable federated learning algorithm designed for asynchronous communication and device heterogeneity. FedHAN customizes sparse models, uses historical client updates to impute missing parameters in sparse updates, dynamically assigns adaptive weights, and combines update deviation detection with update prediction-based model recovery. Theoretical analysis indicates that FedHAN achieves favorable convergence despite unbounded staleness and effectively discriminates between benign and malicious clients. Experiments reveal that FedHAN, compared to leading methods, increases the accuracy of the model by 7.86%, improves the detection accuracy of poisoning attacks by 12%, and enhances the recovery accuracy by 7.26%. As evidenced by these results, FedHAN exhibits enhanced reliability and robustness in intricate and dynamic federated learning scenarios.

Graph anomaly detection has achieved significant progress in recent years. However, traditional graph models are limited to binary relations and local structures, making it difficult to capture higher order interactions among nodes. To address this limitation, we propose a novel anomaly detection method based on dynamic hypergraph neural network (DHGAD). The method first integrates structural and attribute information to construct an initial graph that reflects dual similarity. To capture higher order and tightly knit node dependencies, a hypergraph is constructed based on representative motif structures extracted from the initial graph subsequently. A dynamic hypergraph update mechanism is introduced to refine the structure using reconstruction loss iteratively, enabling nodes to capture richer structural semantics. Finally, anomaly detection is enhanced by shifting from “node-neighbor” to “node-neighborhood” consistency, improving robustness by analyzing behavioral deviations within broader local contexts. Experiments on multiple datasets demonstrate the effectiveness and superiority of the proposed approach.

With the increasing development of mobile social networks, a large amount of data is generated by various social software. However, these datasets contain sensitive user information, and their direct release without proper privacy protection measures may lead to significant privacy breaches. Aiming at the problem of identity privacy leakage due to degree attacks on mobile social network, we proposes an privacy preserving scheme under the differential privacy model. First, we introduce a degree projection method to preserve the important nodes as much as possible. Second, we use an improved fuzzy clustering algorithm to divide the nodes, and realizes personalized differential privacy according to the variance of the node degrees in each cluster. The experiment shows that the scheme not only satisfies the differential privacy, but also has high data availability.

Network community detection is a crucial task in studying the structure of complex networks. However, traditional community detection methods primarily rely on the low-order structure of networks, overlooking high-order connection patterns, thus failing to fully capture the multi-scale structure of complex networks. The community detection problem can be regarded as a graph partitioning problem, aimed at partitioning the network into substructures with dense internal connections and sparse external connections by identifying the inter-cluster edge between communities. This paper investigates the general model of 3-uniform hypergraph subnetworks connected by random hyperedges. By leveraging hypergraph Laplacian spectral theory and inequality bounding, we demonstrate the existence of a critical phase transition when employing spectral clustering for community detection, contingent upon the probability of random inter-cluster hyperedge connections. Specifically, as the phase transition is approached, the community detection performance transitions from nearly perfect detectability to lower detectability. We derive upper and lower bounds for the phase transition and prove its exact location when the sizes of the two subnetworks are equal. Using simulated and real-world networks, we illustrate how to utilize empirical estimation of these bounds to validate the reliability of detected communities.

High-dimensional time-series anomaly detection remains challenging due to the limitation of dimensionality, complex spatiotemporal dependencies, and heterogeneous multimodal feature interactions. To address these limitations, we propose a novel framework called the Multimodal Hypergraph Generative Adversarial Network (MHG-GAN). This framework leverages a multimodal generator to extract temporal and spatial features from raw time-series data, while a dynamic hypergraph structure explicitly models higher-order correlations among data points, overcoming the pairwise relationship constraints of traditional graph models. A dual-network adversarial architecture, comprising time-series and frequency-series generators, is developed to enhance data learning. In the training phase, time-series data is processed through a multimodal builder and transformed into time-domain and frequency-domain data. These data are then encoded by time-domain and frequency-domain encoders and fed into the respective generators. The generated data are constructed into hypergraphs, which are fused and input into a discriminator for judgment, alongside the original time-series data. In the detection phase, the process is similar, with the final score derived from reconstruction and discriminator judgment scores. Extensive experiments on real-world datasets, including SWaT for industrial control systems and NSL-KDD for network traffic monitoring, demonstrate that MHG-GAN outperforms state-of-the-art baselines (LSTM-VAE, MAD-GAN) in terms of F1-score, precision, accuracy, and recall metrics. Specifically, MHG-GAN achieves an F1-score of 95.1

While being able to avoid single point failures, emerging decentralized security techniques are facing new challenges of reliability, robustness, and privacy preservation in blockchain-enabled Internet of Things (IoT) systems. To circumvent these issues, a zero-trust security scheme is proposed through distributed access control, enhanced authentication, dynamic authorization, and privacy preservation enabled by the consortium blockchain. The proposed scheme integrates three key components, i.e., a distributed recommendation mechanism, where multiple authorized nodes are utilized as referrers to efficiently confer their trust on a new public entity for enhanced authentication; an anonymous credential generation strategy, which is developed for the new entity to further protect its privacy from linking attacks; and an adaptive reputation update strategy, which is proposed for evaluating the nodes' behaviors in the system for accountability and dynamic multiple-level authorization. The proposed scheme is implemented in a Hyperledge Fabric and the results show that it significantly enhances security and protects private information.

The use of IoT-based semantic codecs to process complex contextual semantic information in holographic counterparts introduces significant privacy risks, as it may expose sensitive data, thereby increasing the likelihood of privacy disclosures. The diverse and dynamic nature of holographic counterparts in IoT environments exacerbates these challenges, making it more difficult for semantic codecs to effectively safeguard data privacy. This complexity further intensifies the need for privacy-preserving computation methods, as ensuring the confidentiality and security of the data processed by these codecs becomes a critical concern. However, current privacy protection strategy for multi-party training of semantic codecs relies heavily on the central server for gradient calculation, which may lead to gradient leakage issue. To address this issue, we propose PIMSeC (Privacy-Preserving Interactive Multi-Party Semantic Codec Training for IoT-Based Holographic Counterparts), a novel encryption-based technique that facilitates secure and efficient multi-party interactive training without the dependence on the central server, which enhances both data security and privacy resilience. PIMSeC not only proposes a full interactive secure multi-party deep learning model to protect data privacy during multi-party interactive training, but also, within the above deep learning model, establishes an encrypted additive gradient noise mechanism to ensure post-training semantic codec data privacy. Our theoretical analysis and experimental results demonstrate that PIMSeC promotes semantic codecs privacy protection effectively by interactive secure multi-party training. Compared to the state-of-art methods, PIMSeC achieves a 3% to 15% improvement in terms of accuracy, precision, F1-score, and recall at lower compression rates.

Secure data access control is widely concerned in Cyber Physical Systems (CPSs) to protect consumer rights. Ciphertext-Policy Attribute-Based Encryption (CP-ABE) is an ideal solution that supports fine-grained access control. It is a pity that they fail to consider the attribute control and the verifiable legitimacy when consumer's attributes are updated. Additionally, the existing revocable CP-ABE schemes are unsuitable for resource-constrained consumer electronic devices due to their high computational and storage overhead. This paper proposes a blockchain-based access control scheme with attribute update. A new revocation mechanism is designed to guarantee the forward/backward security of the attribute update. The lightweight ciphertext update protocol is proposed to optimize the computation cost. To realize secure and efficient attribute management, this paper designs the on-chain/off-chain separation storage mode, and proposes the traceability-based attribute set generation and validation algorithm. This paper performs the formal security proof. The experimental results demonstrate that the scheme is efficient in terms of encryption and decryption.

In this paper, we investigate the robust and secure transmission for a reconfigurable intelligent surface (RIS) assisted multi-user simultaneous wireless information and power transfer (MU-SWIPT) system in the presence of an eavesdropper (Eve) with hardware impairments (HWIs). An optimization problem is formulated with the goal of maximizing the average sum secrecy rate (ASSR) of all users subject to the constraints of unit-modulus, maximum transmit power budget, and minimum average harvested power (AHP) threshold, taking into account the distortion noise caused by the HWIs at the base station (BS), users, and RIS. To handle this problem, the alternate optimization (AO) method is utilized to decouple the formulated problem into several sub-problems which are solved by optimizing one set of variables with other sets being fixed. Specifically, we employ the successive convex approximation (SCA) method to obtain the BS's transmit beamforming vectors (TBVs), while obtaining the RIS phase shifts based on the combination of SCA and penalty convex-concave procedure (CCP) algorithms, and an optimal power splitting (PS) ratio of each user is derived in a closed-form solution. Numerical results validate the effectiveness of the proposed AO and penalty CCP algorithms, and reveal the ASSR performance gain of the proposed RIS-assisted networks with robust transmission deign (Robust-RIS) scheme over the conventional schemes without considering HWIs or RIS.

Federated learning, as a distributed learning framework, aims to develop a global model while preserving client privacy. However, heterogeneity of client data leads to fairness issues and reduced performance. Techniques like parameter decoupling and prototype learning appear promising, yet challenges such as forgetting historical data and limited generalization persist. These methods also lack local insights, with locally trained features prone to overfitting, which affects generalization in global parameter aggregation. To address these challenges, we propose FedCPD, a personalized federated learning framework. FedCPD maintains historical information, reduces information loss, and increases personalization through hierarchical feature distillation and cross-layer feature fusion. Moreover, we utilize representation techniques like prototype contrastive learning and prototype alignment to capture diverse client data features, thus improving model generalization and fairness. Experiments show FedCPD outperforms state-of-the-art models, enhancing generalization by up to 10.40% and personalization by up to 4.90%, highlighting its effectiveness and superiority.