Federated learning (FL) systems are vulnerable to poisoning attacks, and numerous defenses have been developed for server-assisted FL. Byzantine robustness in decentralized FL (DFL), however, has received comparatively less attention. Without a central server that can inspect and filter client updates, each DFL client must defend its local aggregation against potentially malicious neighbors without incurring excessive communication overhead. We propose an iterative filtering approach for poisoning-resistant DFL. At each filtering step, the method computes the geometric median of the gradients received from neighboring clients and removes gradients whose distances from the median exceed a prescribed threshold. This procedure is repeated until the retained set stabilizes, after which the remaining gradients are averaged. The filter is attack-agnostic in the operational sense that it does not assume a particular attack distribution or construction. Its worst-case theoretical guarantee requires a certified upper bound on each honest client’s local Byzantine fraction that is strictly below one half. Experiments on MNIST and CIFAR-10 under five attacks show that the proposed method achieves the lowest or comparable-lowest maximum test error rate in the evaluated configurations.

Hierarchical federated learning (HFL) in Industrial Internet of Things (IIoT) systems often operates under strict synchronization deadlines, where heterogeneous client computation and wireless access conditions make repeated full local training difficult to execute reliably. Under such constraints, clients that fail to complete local training and transmission on time are excluded from aggregation, reducing client-side executability and exacerbating statistical bias under Non-IID data distributions. This paper studies deadline-constrained local training in IIoT-oriented HFL and proposes a workload-adaptive budgeted coreset framework. Specifically, the roundwise synchronization deadline is converted into a client-specific feasible local workload budget, such that clients with sufficient capability retain full local training while deadline-limited clients switch to budgeted training on a compact budgeted local coreset. To preserve learning effectiveness under reduced workloads, the proposed method employs a predictive-sensitivity scoring criterion that is informative for decision-boundary refinement, together with a class-aware allocation strategy that maintains class coverage under label skew. Experimental results on public IIoT datasets under multiple deadline settings demonstrate that the proposed method substantially improves client-side executability and achieves the best mean predictive performance among the baselines.

With the rapid advancement of technology and the growing prevalence of connected vehicle architectures, security concerns in vehicle networks have become increasingly critical, the vehicles are high-value assets. Thus, unauthorized intrusions can pose substantial financial and safety risks. Although many researchers have employed deep learning techniques for vehicle network intrusion detection, existing methods often struggle to efficiently extract network traffic features and handle unknown or evolving malicious attacks. In this paper, we propose a novel Internet of Vehicles intrusion detection system (IDS) that combines an autoencoder (AE) with a double random forest (DRF). Our two-layer architecture is designed to accurately distinguish between unknown attacks, known attacks, and normal traffic. We evaluated the performance of the proposed DRF algorithm against six traditional machine learning and deep learning methods and compared it with relevant state-of-the-art solutions. The experimental results demonstrate that our method outperforms existing approaches.

Tabular datasets containing sensitive personal attributes are widely used for classification but pose substantial privacy risks. Uniform differential privacy mechanisms such as DP-SGD apply a fixed privacy budget across features, often over-noising predictive attributes and under-protecting weak ones. This paper introduces an Attention-Based Adaptive Differential Privacy Noise Injection Framework that uses Transformer encoders and multi-head self-attention to estimate feature importance and allocate local privacy budgets. High-importance features receive smaller noise for utility preservation, whereas low-importance features receive stronger protection. The noisy embeddings are processed by a DP-SGD optimized MLP classifier with contrastive learning. Experiments on the Adult Census Income and Bank Marketing datasets show competitive privacy-utility trade-offs across multiple ϵ values while keeping membership-inference risk low.

Federated learning (FL) is a distributed learning framework that enables collaborative training of a global model across multiple Internet of Things (IoT) devices while preserving the privacy of local data. However, this collaborative paradigm also introduces security challenges to the global model. Compromised IoT devices can carry out data poisoning attacks on their local data, uploading malicious model gradients to the cloud server, and ultimately degrading the global model performance. Most existing data poisoning methods focus primarily on maximizing attack effectiveness, causing a significant drop in global model accuracy, while often neglecting the stealthiness of the attack. This leads to malicious updates being easily detected and filtered, reducing their threat in IoT deployments. To fill this gap, we propose a stealthy data poisoning method based on a dynamic generator. The approach dynamically generates perturbations with gradient-level stealthiness, achieving a tradeoff between poisoning effectiveness and stealthiness in realistic IoT settings. We conducted comprehensive experiments on three public IoT datasets: with 30% malicious clients, our attack reduces global accuracy by 10.78% on DeepHealth, 6.07% on ChestMNIST, and 7.81% on SVHN, achieving degradation comparable to baselines while preserving stealthiness at the gradient level. To counter this threat, we propose PCATrace, a trajectory-based defense that consistently detects and removes malicious clients across different attack ratios, effectively preventing poisoned updates from degrading the global model.

The continuous emergence of new attack classes challenges intrusion detection in edge Internet of Things (IoT) networks. Although federated learning enables distributed devices to collaboratively train a shared detector without exchanging raw traffic data, most federated intrusion detection systems assume a fixed label space. Retraining with all historical data incurs substantial storage and computation costs, whereas updating only with newly collected samples can cause catastrophic forgetting. The detector must mitigate catastrophic forgetting of previously observed attack classes while preserving sufficient new-class plasticity to learn emerging attacks under highly non-IID device data, intermittent client availability, constrained local memory, and repeated communication over bandwidth-limited and intermittently connected links. To address these challenges, this paper proposes EdgeFedCIL, a communication-efficient federated class-incremental intrusion detection framework. EdgeFedCIL preserves historical knowledge through client-local replay and knowledge distillation while reducing repeated model transmission through adaptive low-rank compression, quantization, and error feedback. A classifier-head protection strategy further limits compression-induced degradation of class discrimination. Experiments on public intrusion-detection datasets show that EdgeFedCIL achieves competitive or superior detection and historical-knowledge retention performance, particularly under highly heterogeneous client distributions, while reducing cumulative client-to-server model transmission by up to approximately 10.54 times relative to full-precision transmission. These results demonstrate the effectiveness of EdgeFedCIL for continual and communication-efficient intrusion detection in resource-constrained edge IoT networks.

Large Language Models (LLMs) are widely deployed in applications that accept user-submitted content, such as uploaded documents or pasted text, for tasks like summarization and question answering. In this paper, we identify a new class of attacks, prompt in content injection, where adversarial instructions are embedded in seemingly benign inputs. When processed by the LLM, these hidden prompts can manipulate outputs without user awareness or system compromise, leading to biased summaries, fabricated claims, or misleading suggestions. We demonstrate the feasibility of such attacks across popular platforms, analyze their root causes including prompt concatenation and insufficient input isolation, and discuss mitigation strategies. Our findings reveal a subtle yet practical threat in real-world LLM workflows.

The Internet of Things (IoT) has transformed consumer devices by connecting smart home appliances, wearable health monitors, and personal devices that gather and share data. Cloud-assisted consumer IoT enhances these networks with improved data processing and storage capabilities for real-time analytics and personalized services. However, this open communication exposes consumer devices to risks including data leakage, eavesdropping, and tracking by malicious actors who increasingly employ AI-enabled attack techniques such as intelligent phishing and automated vulnerability exploitation. To address these challenges, we propose a lightweight authentication protocol for resource-constrained consumer IoT devices using bitwise operations and hash functions. We incorporate Physical Unclonable Functions (PUFs) and blockchain-based smart contracts to enhance security and efficiency while defending against AI-powered hardware attacks. PUFs provide unique device fingerprints resistant to AI-based cloning attacks, while smart contracts secure transactions and detect AI-generated fraudulent attempts. Security evaluation through formal verification demonstrates that our scheme offers stronger protection for consumer devices against both traditional and AI-enabled threats while maintaining lower computational and communication costs compared to existing solutions.

With the rapid proliferation of mobile sensing in fields such as personal health monitoring in data processing are becoming more prominent. This paper introduces a decentralized DETR framework inspired by blockchain proof-of-work consensus. The framework trains models locally on each device and evaluates the device's reputation based on its historical performance. Only devices meeting predefined criteria are admitted to the update committee, which enhances security. This mechanism reduces reliance on centralized servers and minimizes infrastructure costs. While a supervisory operator ensures the smooth operation of the system. To further enhance trust, we propose a credibility assessment method that integrates risk metrics with data quality scores via a non-cooperative game-theoretic model. By achieving Nash equilibrium, this method not only guarantees local optimality but also prioritizes users who provide high-quality, low-risk data, thereby promoting timely committee updates to achieve global optimality. As a complement to DETR, we propose BAL-IDS, an advanced intrusion detection system (IDS) that extracts latent features using autoencoders and dynamically fine-tunes the hyperparameters of OCSVM using a Bayesian joint local agent optimization strategy. This dual approach enhances the system's resilience to complex threats, especially those that exploit requester feedback mechanisms. Experiments show that our research is superior to traditional schemes.

Authentication in the Internet of Vehicles (IoV) is challenged by significant communication delays and risks of attacker intrusion, attributed to the high mobility of vehicles. Current authentication schemes often fail to fully address the reliability of Road Side Units (RSUs) and typically overlook issues related to single points of failure. Additionally, these schemes do not adequately consider secure and efficient handover authentication during vehicle access, leading to increased computational and communication costs. To overcome these limitations, we introduce BAKP, a blockchain-assisted anonymous authentication and key agreement protocol designed specifically for IoV contexts. BAKP enables secure access authentication for vehicles connecting to network services, even in scenarios where RSUs are semi-trusted. When the vehicles re-request network services, the RSUs bypass the authentication center and directly query the blockchain-recorded authentication results, thereby streamlining the handover process. The protocol also supports the revocation of pseudo-identities and enhances the transparency in tracing malicious entities. We performed a provable security analysis of BAKP using the Standard Model, along with formal security evaluations employing Colored Petri Nets (CPN) and the DolevYao attacker model. Our results confirm that BAKP is robust against all relevant attacks and demonstrate its efficiency through comprehensive performance evaluations.

Federated learning (FL) enables collaborative model training across distributed devices without sharing raw data. However, frequent transmission of full-precision parameters incurs substantial communication overhead, posing challenges for deployment on resource-constrained edge and consumer devices. To address this limitation, we propose Layerwise Selective Quantization (LSQ), a communication-efficient FL scheme that adaptively applies quantization using a stochastic layer sampling strategy. Specifically, LSQ retains critical layers in full precision, while applying soft-threshold ternary quantization (STTQ) to less critical layers. We evaluate LSQ on MNIST and CIFAR-10 and compare it with the FedAvg with traditional quantization. Experimental results show that LSQ can substantially reduce communication cost while maintaining competitive model performance, demonstrating robustness and efficiency across typical FL scenarios.

Recent advances in decentralized systems and the widespread adoption of smart devices have enabled a new era of distributed machine learning, offering real-time data processing and personalized services. However, this paradigm brings significant challenges in preserving data privacy, particularly as regulations emphasize the “right to be forgotten.” In this paper, we present a novel and efficient federated unlearning method that successfully removes the influence of forgotten data while maintaining the overall performance of the global model. Additionally, we design a contribution-based client selection algorithm that leverages historical local update information stored on the server, thereby minimizing both computational and communication overhead. We validate our approach through comprehensive experiments on three datasets: MNIST, Fashion-MNIST, and CIFAR-10. These results demonstrate that our approach effectively removes the influence of forgotten data, maintains high global model accuracy, and reduces time overhead by over 60

With the continuous expansion of Industrial Internet of Things (IIoT) applications, secure cross-domain authentication has emerged as a critical technical challenge. Traditional authentication schemes based on trusted third parties (TTPs) suffer from single-point-of-failure risks, while blockchain-based decentralized approaches incur excessive computational overhead. Both methods share a common deficiency; authentication nodes must repeatedly execute authentication procedures at each target domain's gateway when performing cross-domain authentication. This repetitive mechanism significantly increases resource consumption and limits the system's ability to satisfy strict real-time requirements. To address these challenges, this article proposes a lightweight and reusable cross-domain authentication (LRCDA) scheme based on distributed architecture. First, the proposed scheme, based on Chinese remainder theorem (CRT), enables edge nodes to autonomously establish system keys and independently verify identities, completely eliminating dependence on centralized third parties and mitigating single-point-of-failure risks. Furthermore, the scheme introduces a cross-domain reusable signature mechanism, allowing signature to be securely reused across multiple security domains during its validity period, thereby reducing the frequency of signature generation and minimizing authentication latency. Moreover, through efficient task allocation, computationally intensive operations are offloaded to edge nodes, thereby reducing the computational burden on terminal devices. Theoretical analysis demonstrates that the LRCDA scheme meets the security and privacy requirements for cross-domain authentication in IIoT environments, encompassing anonymity, signature unforgeability, and resistance to man-in-the-middle (MITM) attacks. Experimental evaluations reveal that the LRCDA scheme reduces authentication latency by 77.4% compared to traditional schemes and by 88.2% compared to existing distributed approaches, demonstrating significant efficiency gains for resource-constrained IIoT deployments.

The Industrial Internet of Things requires authentication mechanisms that can adapt to dynamic contexts, preserve identity privacy, and support interoperable collaboration across security domains. This article presents BFL-DIAS, a blockchain- and federated-learning-assisted dynamic identity authentication scheme for cross-domain IIoT environments. BFL-DIAS integrates a permissioned consortium ledger, pseudonymous identity management, dynamic key and attribute-state updates, and off-path federated risk scoring into a unified workflow for registration, intradomain authentication, cross-domain authentication, and state maintenance. The federated risk score is used as a policy-assistance signal rather than as a prerequisite for cryptographic authentication, while synchronous on-chain writes are confined to registration, key update, and audit operations outside the user-facing critical path. We formalize the system and threat model and analyze the protocol under standard assumptions, showing that it provides mutual authentication, forward secrecy, impersonation resistance, replay protection, identity confidentiality, and conditional unlinkability under epoch-based pseudonym rotation. A prototype implementation on edge, mobile, and virtualized server platforms demonstrates low authentication latency, manageable ledger overhead, and stable off-path risk assessment. These results indicate that BFL-DIAS provides a secure, auditable, and deployable foundation for dynamic cross-domain authentication in industrial environments.

Deploying large language models (LLMs) directly at the edge offers significant benefits in reducing inference latency and enhancing privacy protection. However, this approach presents critical technical challenges due to the limited hardware resources available at the edge. In this paper, we explore the deployment of the Mixture-of-Experts (MoE) model, a popular LLM architecture, using weight quantization in a typical edge environment with multiple edge servers. Existing methods that uniformly quantize MoE models do not account for the varying importance of different experts or the dynamic nature of edge requests. To address these issues, we propose a hybrid expert quantization strategy, where each edge server quantizes experts to different bit-widths based on the preferences of received requests. Furthermore, we implement a comprehensive system optimization that jointly considers expert quantization and request assignment, aiming to achieve high-precision and low-latency inference services. To manage the complex, coupled optimization challenges, we design a reinforcement learning (RL)-based algorithm. To overcome the slow convergence typical of traditional RL, we pre-train the policy network using a system simulator that generates heuristic solutions, thus accelerating the learning process. Extensive experimental results on real datasets demonstrate that our approach significantly outperforms existing methods.

Privacy-preserving federated learning (FL) frameworks have been widely applied in the Internet of Vehicles (IoV) to enhance the privacy of local models and sensitive vehicle data. However, existing schemes cannot fully address challenges, such as large-scale data transmission efficiency, communication reliability, and single-point failures. To tackle these issues, we propose an efficient and secure privacy-preserving FL (ESPL) scheme for IoV. Specifically, we design a model compression method based on low-rank approximation and quantization, which effectively reduces transmission overhead while maintaining high accuracy. We further develop a secure and lightweight anonymous authentication and key agreement algorithm to ensure reliable interactions between clients and fog nodes. To rigorously evaluate security, we construct a colored Petri Net under the Dolev-Yao attacker model, verifying that the scheme can effectively prevent malicious attacks. In addition, we propose a Paillier-based hierarchical secure aggregation architecture that protects client privacy, reduces computational burden, and mitigates single-point failure risks. Experimental evaluation demonstrates that, compared with representative baselines CRS, FC-FL, and RUPT-FL, ESPL reduces client communication overhead by up to 44%, lowers computation overhead by up to 59%, and improves final model accuracy by nearly 1.8%. These results confirm the practicality, scalability, and robustness of ESPL in real-world IoV environments.

Large language models (LLMs) are promising for healthcare question answering, yet adapting them with sensitive clinical data can induce unintended memorization and privacy leakage. We present MedDP, an efficient privacy-budgeted adaptation framework that delivers an auditable utility-privacy trade-off under ($\epsilon,\delta$)-differential privacy. Our method, MedDP, applies DP-SGD only to low-rank adapter parameters while freezing the base model, using per-example gradient clipping and calibrated Gaussian noise to enable explicit privacy accounting with minimal compute and memory overhead. To improve the stability of noisy optimization, we introduce Adaptive Clipping (AC) that updates the clipping bound online via a quantile-based rule with momentum. To avoid spending additional privacy budget on late-stage accuracy degradation, we further propose Budget-Aware (BA) checkpoint selection, which chooses the best checkpoint on the training trajectory subject to a target budget $\epsilon _\rm{max}$, without re-training. For privacy evaluation, we adopt a canary-based suite and emphasize canary negative log-likelihood (NLL) as a more sensitive memorization signal than hit-rate. Experiments on public medical QA benchmarks derived from MedQA and MedMCQA, together with synthetic canary auditing, characterize a clear privacy-utility frontier in a controlled evaluation setting. Non-private LoRA improves average accuracy but increases memorization signals, whereas privacy-budgeted training mitigates memorization while maintaining stable performance under explicit privacy accounting.

Crowdsensing, as a powerful paradigm, enables the collection of large-scale sensor data from distributed participants. However, malicious contributors providing unreliable or tampered data pose significant challenges to data quality and task reliability. This paper proposes a Federated Trust Management (FTM) framework that integrates a trust assessment mechanism into Federated Learning to mitigate the impact of adversarial contributions. The proposed method dynamically adjusts the credibility of participants and introduces a weighted aggregation strategy in FL, thereby reducing the influence of low-quality or malicious data. Experimental results demonstrate that FTM significantly improves global model accuracy, adversarial impact, and enhances the robustness of crowdsensing applications.

The industrial Internet of Things (IIoT) requires robust intrusion detection systems (IDS) to ensure critical business continuity. However, existing solutions suffer from high false-positive rates, difficulty in interpreting, and cross-domain generalization. In particular, they fail to reliably infer context-dependent causal links from host logs under global graph-level context. To address these issues, we propose causal counterfactual graph-based IDS (CCG-IDS), an IIoT-focused interpretable graph neural network IDS based on conformal calibration and counterfactual reasoning. This system implements a unified provenance subgraph detection paradigm, and employs conformal anomaly detection and counterfactual reasoning to provide calibrated alerts and interpretable outputs. We also estimate predictive uncertainty via Fisher information to quantify decision confidence. We introduce a discrete counterfactual explainer with a counterfactual destructiveness score (CDS) to extract a minimal decision-critical evidence chain, and use this evidence to generate structured analyst-ready security reports. Experiments on real-world industrial log datasets, including the windows event log EVTX (Windows XML Event Log) and DARPA OpTC (Operationally Transparent Cyber) datasets, demonstrated that CCG-IDS achieved an $F1$ score of 92% and a near-zero false-positive rate, outperforming other state-of-the-art methods.

Federated Learning (FL) enables IoT devices to collaboratively train models without sharing raw data. However, privacy and accountability remain major challenges due to persistent identifiers and untrusted participants. We propose a lightweight FL framework integrating anonymous credentials and local differential privacy for unlinkable participation and secure model updates. A zk-SNARK-based revocation mechanism enables efficient verification and blacklisting of malicious clients. Experiments confirm low computational overhead and strong protection against inference and poisoning attacks.