Data privacy mandates necessitate efficient machine unlearning protocols capable of excising specific training data without the prohibitive cost of full retraining. We introduce ScrubRef-Rep, a fast, data-free class unlearning framework. Moving beyond superficial logit manipulation, our approach jointly updates a localized deep-feature tail and the classifier head. It synthesizes proxy distributions for the forgotten class via generative gradient ascent, combining this with masked knowledge distillation and a novel representation-aware objective. Evaluated on CIFAR-10, CIFAR-100, and ImageWoof-320 against state-of-the-art baselines (UNSIR, Bad Teacher, SCRUB, SSD), ScrubRef-Rep consistently achieves an absolute 0.00% forgotten-class accuracy while maintaining robust retained-data utility (89.18%, 70.70%, and 92.30%). Furthermore, it ensures strong empirical privacy preservation, driving membership inference attack (MIA) ROC-AUCs to near-random chance (0.505, 0.508, and 0.527) with up to a 22.7× speed-up over full retraining. Crucially, representation audits confirm that our framework achieves near-complete erasure of the target concept in the latent space, with only negligible structural traces remaining. This explicit delineation between robust empirical privacy and theoretical absolute erasure establishes ScrubRef-Rep as a highly practical, transparent, and representation-aware unlearning solution.
Network intrusion detection for the Internet of Things increasingly relies on deep-learning classifiers, yet their reported adversarial robustness rests on an untested networking assumption: that a feature-space perturbation corresponds to a packet or flow an attacker can actually transmit on the wire. We propose PAVE (Protocol-Validity-Aware Evaluation), a framework that judges robustness at the realizable network artifact (CAN frames, MQTT packets, IP flows) rather than the normalized feature tensor, coupling a protocol-space validity test, a gradient-masking audit, and seventeen imbalance-aware metrics with corrected paired tests over three IoT datasets, five architectures, and six defenses. PAVE revises standard conclusions: adversarial examples produced by standard unconstrained continuous feature-space attacks (FGSM, PGD, Square) retain only 0%–3% validity in protocol space, so most of the vulnerability these attacks report cannot occur on the wire (enforcing validity shows that unconstrained evaluation overstates the MCC degradation of a Transformer under FGSM by 0.18–0.20: 0.39 vs. 0.56 and 0.30 vs. 0.50 on two benchmarks, i.e. a measurement of evaluation bias, not a defense); its imbalance-aware gradient-masking audit screens for majority-class collapse, a risk we observe for an untuned defense configuration and which principled penalty tuning resolves; and two of three common benchmarks are saturated or leakage-prone (CICIoV2024 reduces to ≈3.6k unique flows from 1.4M). Among low-cost defenses, a feature-space Mahalanobis detector gives strong non-adaptive separation (PR-AUC ≈1.0, FPR@95%TPR ≈0), though it degrades under an adaptive attacker, which PAVE recommends as a complement to, not a replacement for, adversarial training. PAVE provides a realizability-grounded protocol and a more accurate estimate of adversarial risk for IoT intrusion detection.
On-device personalization has emerged as a promising paradigm to fully exploit all available computing resources to adapt AI models to user-specific needs. This computing paradigm aims to provide user-centric, efficient, fault-tolerant, and personalized services while maintaining robust security and privacy. Meanwhile, Large Language Models (LLMs) are among the most widely used AI applications and have shown exceptional capabilities in tasks such as natural language understanding, language generation, and complex reasoning. As LLMs evolve into personal AI agents, there is a growing demand to personalize these models directly on users’ devices to enable fast responses, protect sensitive data, and tailor behaviors to individual preferences. However, performing such on-device LLM personalization on resource-constrained platforms introduces substantial challenges in computation, memory, energy, and system design. These challenges pose an urgent and fundamental question: How can we efficiently personalize LLM on constrained hardware for user-centric, privacy-preserving, and reliable AI agent realizations? Current research on this topic is fragmented and lacks a comprehensive overview. This paper seeks to fill this gap by summarizing and organizing previous studies in this area. We offer a thorough review of on-device LLM personalization research, categorizing the existing approaches into four main categories: data-centric, model-centric, optimizer-centric, and system-centric perspectives. Our goal is to serve as a valuable resource – helping researchers and developers gain a comprehensive understanding of the state-of-the-art on-device LLM personalization and guide the development of next-generation user-centric, privacy-preserving, and reliable AI agents.
The widespread application of encrypted traffic, while protecting user privacy, also provides a covert communication channel for malware. Existing machine learning-based detection methods suffer from insufficient feature representation capability and weak generalization performance when facing frequent malware variants and unknown attacks. To address this issue, this paper proposes TCBI (Three-Channel Behavioral Image). This method first proposes a novel encrypted traffic representation approach, which reconstructs network sessions into tri-channel behavioral images. According to the distribution characteristics of different features, nonlinear or piecewise mapping strategies are designed to preserve information completeness while enhancing the discriminability of key behavioral features, thereby effectively capturing the structural similarity and behavioral homology of attack traffic. Secondly, a dual-branch parallel detection model combining Swin Transformer and Bi-LSTM is designed. The image branch extracts local details and global structures of behavioral images through a window-based self-attention mechanism, while the sequence branch utilizes a Bi-LSTM to mine the temporal dependencies of packet length and inter-arrival time, enabling collaborative modeling of spatial and temporal features and improving the accurate identification capability of malicious traffic. Finally, learnable adaptive weighted fusion mechanisms are introduced both within the sequence and between the image and sequence, dynamically adjusting the contribution weights of different feature branches. This allows the model to flexibly integrate multi-modal information according to sample characteristics, significantly enhancing its generalization ability and robustness against variant and unknown malicious traffic. Experimental evaluation results show that TCBI achieves excellent performance in known malicious traffic detection and demonstrates superior generalization ability and robustness in unknown malicious traffic detection, outperforming existing state-of-the-art methods.
Mobile Edge Computing (MEC) provides an effective solution to meet the high bandwidth and low latency requirements of 360-degree panoramic video streaming. However, existing research still exhibits deficiencies in the synergistic optimization of caching and bitrate allocation. On one hand, most schemes optimize caching and transmission as independent processes, lacking effective integration. On the other hand, some joint optimization approaches fail to fully exploit the hierarchical characteristics of Scalable Video Coding (SVC), or merely treat the cache status as a static constraint for bitrate allocation decisions, leading to limited resource efficiency during network congestion. To address these issues, this paper proposes a Joint Optimization of Caching and Allocation Scheme (JOCAS), aiming to maximize the Quality of Experience (QoE) through deep synergy between caching and resource allocation. The framework adopts a multi-dimensional popularity-based caching strategy that combines content popularity with SVC structures to construct a fine-grained value evaluation system, thereby improving the cache hit rate. Furthermore, a policy-aware game-theoretic bitrate allocation algorithm is designed, which utilizes the cache status as a utility weight parameter for the Nash Bargaining Solution (NBS), enabling the bitrate allocation decisions to proactively perceive and fully utilize the cached content. Simulation results demonstrate that JOCAS achieves significant improvements in both cache hit rate and average QoE compared to representative existing schemes.
With the widespread adoption of 360-degree video in virtual reality environments, users exhibit complex and dynamic gaze behaviors. Existing viewport prediction methods suffer from insufficient spatio-temporal feature extraction, weak modeling of global dependencies, and poor prediction stability during long-term forecasting. To address these challenges, this paper proposes a long-term viewport prediction model based on a convolutional multi-head attention mechanism. Trained offline using historical viewing data from other users, this model incorporates convolutional multi-head attention to simultaneously model contextual dependencies, global features, and temporal dynamics, thereby enhancing long-term prediction accuracy and stability. Additionally, a dilated SE convolutional module is designed to expand the receptive field and adaptively recalibrate channels, strengthening the model’s multi-scale feature representation and local detail capture capabilities. Experimental results demonstrate that the proposed model maintains an average prediction accuracy of 96% across different prediction windows, exhibiting excellent stability and reliability.
Explainable Artificial Intelligence (XAI) has emerged as a critical enabler for improving the transparency, trustworthiness, and operational usability of Machine Learning (ML) and Deep Learning (DL)-based Intrusion Detection Systems (IDS). While ML/DL models have significantly advanced detection capabilities, their black-box nature poses challenges for security analysts, regulatory compliance, model debugging, and adversarial robustness. This survey presents a comprehensive review of XAI techniques applied to IDS, with particular emphasis on research published over the past five years. We develop a unified taxonomy of XAI-enabled ML-based IDS methods, categorizing methods based on model transparency, explanation scope, and operational objectives. The survey systematically analyzes IDS frameworks across diverse deployment environments–including cloud, IoT/IIoT/CPS, SDN, and IoMT–and reviews major explanation mechanisms such as SHAP, LIME, attention-based models, saliency methods, and rule-extraction techniques. We consolidate evaluation practices by examining both detection-performance metrics and explanation-quality measures, highlighting their interpretations and limitations in IDS contexts. Furthermore, we investigate XAI-enabled ML-based IDS from an adversarial perspective, identifying vulnerabilities to evasion, mimicry, poisoning, and attacks targeting explanation mechanisms. We also review the dataset landscape, discussing representativeness gaps and generalization challenges. Finally, we identify critical open issues and outline future research directions, including adversarially robust explainability, scalable explanation mechanisms, realistic datasets, deployment-aware design, and human-in-the-loop IDS. This survey aims to provide researchers and practitioners with a structured foundation for designing transparent, robust, and operationally effective XAI-enabled ML-based IDS.
AI-driven visual cheats in competitive gaming—such as aimbots built on real-time object detectors—pose a growing threat to fair play that current anti-cheat systems fail to reliably address. This article presents a gradient-based adversarial attack methodology that neutralizes such cheats by generating Universal Adversarial Perturbations (UAPs) to suppress character detection without visibly degrading the game image. To achieve this, we first introduce a novel multi-game benchmark dataset of 10,000 annotated images (from CS2, Rust, Marvel Rivals, and Apex) to enable standardized evaluation. Building on this, we propose a UAP generation pipeline that utilizes Momentum Integrated Gradient (MIG) initialization to improve convergence and transferability across heterogeneous detection architectures. We further enhance this methodology with directional noise refinement and gradient filtering, which simultaneously improve attack effectiveness and preserve perceptual image quality. Finally, we introduce a zero-knowledge black-box perturbation search that operates without access to proxy models, relying solely on visual feedback. Extensive validation against real-world commercial cheats demonstrates that our approaches consistently outperform the existing Invisibility Cloak framework in both transferability and image fidelity.
Background: With approximately one in six people experiencing a significant disability, the accessibility of software is of critical importance. However, developers lack the knowledge and skills needed to fill the gap in creating universally accessible software. Objective: Inspired by the recent success of large language models (LLMs) in code generation, we investigate if compact language models (CLMs) can detect and correct motor accessibility related issues in code to support developers in meeting accessibility guidelines within on-premises development environments. Methods: We compared eight off-the-shelf CLMs’ abilities to resolve motor-related accessibility errors in sample code. We used four prompt variations (two persona-based and two non-persona-based) for each code sample to determine how the prompt impacts the CLMs’ outputs. Three LLMs were used to evaluate and score these outputs. Results: We see that CLMs have the capability to detect and correct a number of the code errors we tested, with several models providing correct code with at least one prompt per metric 40% of the time. However such success appears inconsistent as models rarely pass with every prompt on a given metric. We also observe that the prompt variations had minimal impact on output. Conclusions: The CLMs may not yet be ready to reliably support developers with improving motor accessibility in their code, despite showing promise. We urge the community to work towards resolving the gaps so that these CLMs may become a more helpful, reliable part of the developers’ workflow.
With the increased use of mobile networks these days, it is important to protect our personal data passing through these unprotected radio channels against AI-facilitated attacks. To offer a high level of data protection, we proposed in this work a dynamic version of the ZUC algorithm designed by the Data Assurance and Communication Security Research Center of the Chinese Academy of Sciences and standardized by the 3GPP (3rd Generation Partnership Project) organization to ensure the LTE (Long Term Evolution of radio networks) confidentiality and integrity. The principle relies on combining two chaotic maps to generate dynamic S-boxes that replace the static S-boxes of the original algorithm (S0 and S1). To keep the same performance of the standardized algorithm and eliminate any additional latency, the dynamic S-boxes are generated in parallel during its initialization mode, where the Cipher Key (CK) and the Initialization Vector (IV) are loaded to initialize the internal registers. To improve the randomness of the generated keystream, we bitwise XORed the standard output with a chaotic sequence using a third chaotic map. To test the feasibility and suitability of our proposal for real applications, we performed an FPGA-based implementation using the Xilinx XC7Z020 PYNQ-Z2 platform, then evaluated its resistance to cryptanalysis attacks using the most well-known security tests (keystream distribution, randomness and correlation, key sensitivity, plaintext sensitivity, keyspace complexity, and NIST statistical tests). As expected, the experimental results demonstrated the enhanced security level of the ZUC algorithm and its suitability for new generations of mobile communications.
Large Language Models (LLMs) have demonstrated immense potential in the medical field, yet they face inherent limitations such as hallucination and knowledge obsolescence. Although retrieval-augmented generation (RAG) effectively mitigates these issues by incorporating external knowledge, traditional approaches are primarily limited to textual data, failing to leverage the rich multimodal nature of clinical information, which potentially compromises diagnostic accuracy. Multimodal RAG extends information processing capabilities to encompass diverse data types, thus overcoming the limitations of its text-only counterpart. This paper aims to provide a comprehensive survey of medical Multimodal RAG to facilitate a thorough understanding of its current landscape and future potential. Specifically, we introduce its related technical foundations, review the various types of multimodal data in the medical field, analyze its core challenges and corresponding solutions, explore the diverse applications and representative works of medical Multimodal RAG, and conclude by outlining evaluation methods and future research directions.
Objective:: Body Mass Index (BMI) is a critical indicator for assessing obesity, playing a key role in public health surveillance, chronic disease prevention, and personalized medicine. Traditional BMI estimation methods rely on manual measurements of weight and height; although computationally straightforward, they have significant limitations in large-scale monitoring and unobtrusive assessment scenarios. This study aims to systematically review recent advances in BMI estimation methods based on Artificial Intelligence (AI) and sensing technologies, summarizing their technical characteristics and development trends. Methods:: Through a systematic retrieval and analysis of relevant literature from multiple sources, this study categorizes existing BMI estimation methods according to their technical principles and application scenarios. Furthermore, it provides a comparative discussion of the strengths and limitations of the related research. Results:: This study indicates that current research primarily focuses on three technical approaches: wireless sensing, contact sensors, and vision-based technologies. All three categories can achieve varying degrees of BMI estimation accuracy across different application scenarios. Wireless sensing methods offer advantages in unobtrusive monitoring; sensor-based methods excel in stability and long-term monitoring; and vision-based methods demonstrate greater potential in information richness and accuracy, albeit with challenges related to privacy and data diversity. In addition, these approaches differ significantly in terms of performance, deployment conditions, and target populations. Conclusion:: AI-based BMI estimation methods provide novel solutions for smart health monitoring. However, further research is needed to address challenges such as model generalization, privacy protection, and large-scale real-time deployment. Future development trends point toward multi-modal fusion, enhanced robustness, and the implementation of deployable, unobtrusive health monitoring systems.
Unified Large Language Models (LLMs) have transformed diverse recommendation tasks into a shared “text-to-text” paradigm. However, existing methods primarily face two challenges: 1) joint fine-tuning across highly heterogeneous tasks (e.g., discriminative sequential recommendation versus generative explanation generation) often suffers from the “seesaw effect” due to severe gradient conflicts; and 2) parameter-efficient methods struggle to isolate these task-specific conflicts while preserving the collaborative sharing of underlying general recommendation knowledge. To address these limitations, we propose Rec-MoELoRA, a hybrid fine-tuning framework that seamlessly integrates the mixture-of-experts architecture with low-rank adaptation. Specifically, we devise a soft-decoupling strategy that maintains the updates of most backbone network layers to continuously absorb globally shared knowledge, while exclusively introducing task-motivated low-rank experts into the self-attention layers for physical task isolation. Furthermore, to effectively overcome the prevalent “expert collapse” dilemma, we incorporate a dual regularization strategy of diversity and entropy. Combined with an asymmetric structural prior (e.g., allocating three experts for two tasks), this forces the model to spontaneously evolve a “Shared-Specific” representation pattern. Extensive experiments on real-world e-commerce datasets (Amazon Sports and Beauty) demonstrate that Rec-MoELoRA significantly outperforms the full fine-tuning baseline across core ranking and text generation metrics while introducing only a marginal number of additional parameters (e.g., HR@10 and BLEU-4 improve by 4.3% and 3.7% on the Sports dataset, respectively).
Traditional Centralized Key Management Systems (KMS) fundamentally conflict with the core philosophies advocated by decentralized storage systems, such as decentralization, censorship resistance, and user autonomy. Their inherent single points of failure and trust bottlenecks severely constrain overall system security. To systematically explore secure key management pathways without trusted third parties, this paper conducts a systematic review of research progress in the field of Decentralized Key Management (DKM). We propose an analytical framework covering four key stages: key generation, distributed distribution, dynamic rotation, and revocation verification. Furthermore, we distill four core evaluation dimensions supporting this framework: security, efficiency, dynamic adaptability, and hierarchical access control. Through the classification, comparison, and in-depth analysis of representative technical schemes, this paper aims to provide clear theoretical references and design guidelines for key security management in decentralized storage systems. Finally, this paper summarizes the major technical challenges currently facing the field and provides an outlook on future research directions.
Dynamic asynchronous Byzantine Fault-Tolerant (BFT) protocols provide strong consistency guarantees in open networks with unbounded communication delays and evolving membership, making them a fundamental building block for long-lived distributed systems. However, efficiently and securely bootstrapping new nodes remains a key scalability obstacle, as existing solutions rely on checkpoint chains with linearly growing collections of threshold signatures and bandwidth-intensive full-state replication.This paper presents Aurelia, a bandwidth-efficient dynamic asynchronous BFT protocol that overcomes these limitations. Aurelia introduces a compressible checkpoint mechanism that aggregates accumulated threshold signatures into a constant-sized proof, effectively compressing the total proof chain size to approximately one-third of the baseline. Furthermore, it employs a distributed state synchronization protocol based on Verifiable Information Dispersal, enabling new nodes to reconstruct historical consensus results without full replication. We show that Aurelia preserves the safety and liveness properties of Atomic Broadcast under a hybrid fault model, while significantly improving the communication efficiency of node joining. These results demonstrate that dynamic asynchronous BFT can be deployed efficiently in long-term, large-scale systems.
Entity linking in visually rich documents aims to identify semantic relationships between entities (e.g., key–value pairs) by jointly leveraging textual, visual, and spatial information. Despite the success of pre-trained document models such as LayoutLMv3, two challenges remain for relation extraction: (1) spatial position signals injected only at the input embedding layer tend to decay in deeper transformer layers, weakening the model’s ability to capture layout-dependent entity associations; and (2) in long documents, softmax attention distributes weights across many irrelevant tokens, diluting the focus on informative regions. To address these issues, we propose Gated Spatial Attention (GSA), a lightweight, plug-in framework on top of LayoutLMv3 that comprises two complementary modules: Spatial Position Enhancement (SPE), which injects ALiBi-style linear biases into every attention layer with head groups specialized for reading-order, horizontal, vertical, and semantic proximity, and Gated Attention (GA), which applies a per-token scalar gate after the scaled dot-product attention to suppress outputs from irrelevant tokens.Experiments on FUNSD and CORD demonstrate that GSA consistently improves both semantic entity recognition and relation extraction, achieving state-of-the-art results with negligible computational overhead.
Mobile systems are revolutionizing healthcare delivery through their integration with artificial intelligence (AI) and Internet-of-Things (IoT) technologies, ushering in the era of Healthcare 5.0. This emerging healthcare paradigm leverages mobile technologies to enable personalized, intelligent, and ubiquitous medical services. Mobile devices, serving as the primary interface between patients and healthcare systems, facilitate continuous health monitoring, real-time medical interventions, and personalized treatment optimization. Through AI-powered mobile applications, Healthcare 5.0 enhances diagnostic accuracy, treatment efficacy, and patient engagement beyond traditional digital health solutions. Key innovations include mobile-based precision diagnostics, AI-driven health monitoring, and personalized medical interventions. Nevertheless, Healthcare 5.0 still faces significant challenges, including the delivery of accurate secure, and robust personalized healthcare solutions.In this paper, we present a systematic review of mobile healthcare applications across major disease domains, where mobile healthcare systems facilitate continuous and remote care by enabling real-time monitoring, early symptom detection, and personalized interventions, leveraging multimodal data from smartphones, wearable sensors, and IoT-enabled platforms. We provide in-depth insights into the principles, challenges, and research opportunities in mobile Healthcare 5.0. We hope this paper inspires further research into mobile systems for Healthcare 5.0, an area of immense potential for safeguarding public health.
The rapid growth of Edge Intelligence (EI) and heterogeneous user demands has led to the widespread generation of multimodal data at the network edge. Multimodal Federated Learning (MFL) provides a promising solution for collaborative, privacy-preserving model training across distributed clients. However, existing MFL frameworks often assume homogeneous environments and fail to account for disparities in client data distributions, modality characteristics, and computational resources, limiting their effectiveness in real-world edge deployments. To address these challenges, we propose Multimodal Federated Edge Learning (MFEL), a flexible framework that supports resource-adaptive deployment through variable-capacity submodels. Building upon MFEL, we introduce MFEL-H2B, a heterogeneous-aware approach that integrates three core mechanisms: (1) Prototype Networks for cross-client modality alignment, mitigating representation divergences caused by non-IID data and heterogeneous sensing conditions; (2) Rebalanced Modality Gradient Modulation (R-MGM), which adaptively amplifies gradients of underrepresented modalities while suppressing dominant ones to alleviate intra-client modality imbalance; and (3) Ensemble Momentum-based Knowledge Distillation (E-MKD), which constructs a dynamic ensemble teacher from client predictions and leverages a momentum mechanism to facilitate efficient and robust knowledge transfer among clients with heterogeneous model capacities. Extensive experiments on heterogeneous multimodal datasets demonstrate that MFEL-H2B consistently outperforms state-of-the-art baselines in accuracy, convergence speed, and training stability, while maintaining strong generalization across diverse client architectures and resource profiles.
Decentralized finance (DeFi) has emerged as a transformative paradigm, leveraging programmable blockchains to innovate upon traditional financial services without centralized intermediaries. However, DeFi introduces a unique and highly adversarial security landscape characterized by immutable transactions, complex protocol composability, and transparent execution environments. This survey provides a comprehensive systematization of DeFi security, categorizing vulnerabilities across three distinct layers: technical and code layer, economic and protocol layer, and infrastructure and crosschain layer. Furthermore, we structure the defense mechanisms according to the protocol lifecycle, including pre-deployment prevention strategies, runtime mitigation techniques, and post-incident response and recovery mechanisms. We also delve into specific phenomena such as maximal extractable value, analyzing its dual role as both a market efficiency tool and a security vector. By synthesizing existing literature and incident reports, this survey establishes a holistic framework for understanding the interplay between code and finance. Finally, we identify critical open challenges and propose future research directions aimed at maturing the discipline of DeFi security and mitigating systemic risks. (c) 2026 The Author(s). Published by Elsevier B.V. on behalf of Shandong University. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
Embodied AI systems (e.g., autonomous vehicles, service robots, and LLM-driven interactive agents) are rapidly transitioning from controlled environments to safety-critical real-world deployments. Unlike disembodied AI, failures in embodied intelligence lead to irreversible physical consequences, raising fundamental questions about security, safety, and reliability. While existing research predominantly analyzes embodied AI through the lenses of Large Language Model (LLM) vulnerabilities or classical Cyber–Physical System (CPS) failures, this survey argues that these perspectives are individually insufficient to explain many observed breakdowns in modern embodied systems. We posit that a significant class of failures arises from embodiment-induced system-level mismatches, rather than from isolated model flaws or traditional CPS attacks. Specifically, we identify four core insights that explain why embodied AI is fundamentally harder to secure: (i) semantic correctness does not imply physical safety, as language-level reasoning abstracts away geometry, dynamics, and contact constraints; (ii) identical actions can lead to drastically different outcomes across physical states due to nonlinear dynamics and state uncertainty; (iii) small errors propagate and amplify across tightly coupled perception–decision–action loops; and (iv) safety is not compositional across time or system layers, enabling locally safe decisions to accumulate into globally unsafe behavior. These insights suggest that securing embodied AI requires moving beyond component-level defenses toward system-level reasoning about physical risk, uncertainty, and failure propagation.
