Website fingerprinting (WF) attacks aim to infer user-visited websites by analyzing traffic patterns in encrypted Tor traffic. Traditional deep-learning-based WF attacks typically require large, comprehensive datasets, which is often impractical for resource-constrained adversaries in real-world environments. To address this limitation, herein, we propose a novel few-shot WF attack framework, Trackfficer. This framework transforms prelabeled Tor traffic packet sequences into multichannel spatiotemporal feature tensors that capture both global contextual patterns and fine-grained traffic characteristics. Leveraging these representations, the framework adopts a multilayer convolutional neural network architecture with dual-attention mechanisms within a quintuplet network for contrastive pretraining, enabling the extraction of highly discriminative traffic-feature embeddings. It further incorporates a dual-loop c learning strategy that enables rapid adaptation to previously unseen websites using only limited support samples. Extensive experiments conducted under various network conditions in both closed-world and open-world settings demonstrate that Trackfficer achieves effective fingerprinting of Tor traffic with minimal labeled data. The framework also demonstrates strong resilience against various WF defense techniques and consistently outperforms existing WF attacks, even in the presence of temporal drift. These findings validate the practicality and effectiveness of Trackfficer in real-world WF attack scenarios.

Large language models (LLMs) have recently achieved notable progress in long-range reasoning and long-text generation. However, these capabilities introduce new vulnerabilities in adversarial settings. In multi-turn dialogues, model outputs are influenced by prior interactions and self-generated responses. Adversaries can exploit this through carefully crafted prompts and gradual context injection, leading to model drift and instability. Based on this observation, we propose Evolva, a jailbreak framework that formalizes multi-turn prompt attacks as a multi-stage structured process. Evolva is composed of three interdependent modules and utilizes auxiliary LLMs to automatically construct hierarchical and progressively complex prompt sequences. Through the gradual introduction of adversarial context across dialog rounds and by exploiting interaction dynamics, Evolva can reliably guide target models toward unsafe generations, thereby bypassing existing prompt-level safety controls. To more reliably evaluate multi-turn prompt attacks, we introduce the Attack_Robust, which complements success-rate-based measures by characterizing the robustness of a model’s behavior along the interaction trajectory. We evaluate several state-of-the-art LLMs on REDTask, a red-teaming dataset tailored for multi-turn adversarial prompting. Experimental results show that Evolva consistently outperforms existing approaches, revealing that long-range reasoning and multi-turn interaction significantly expand the model’s attack surface and lead to a degradation in robustness.

Malware detection faces growing challenges due to sophisticated obfuscation techniques that undermine the robustness of single-modal approaches relying solely on static code analysis or dynamic behavioral profiling. To address this issue, we propose MalCAFF, a cross-attention-based framework for fine-grained fusion of static assembly semantics and dynamic API behaviors. Static features are refined through program slicing to preserve critical semantics, while dynamic behaviors are represented by API Semantic Block Sequence (ABS), which aggregate API calls into parameter-aware, semantically enriched units aligned with static functions. A Cross Attention-based Feature Enhancement (CAFE) module then achieves bidirectional semantic complementation across modalities. Furthermore, contrastive pre-training mitigates inter-modal distributional discrepancies and enhances generalization. Extensive experiments on the VirusShare dataset demonstrate that MalCAFF outperforms state-of-the-art methods.

Resource-constrained edge intelligence increasingly requires lightweight and trustworthy access to heterogeneous knowledge, where decisions often involve reasoning over structured tables and associated text. In this paper, trustworthiness is operationalized as robustness under weak supervision, imperfect evidence selection, and noisy hybrid contexts. Table–text Hybrid Question Answering (HQA) is representative of this need, yet existing methods often face a trade-off between specialization-driven systems with high engineering cost and generalization-driven LLM pipelines that are flexible but may lag behind the strongest specialized systems. We propose RRS, a parameter-efficient retriever–reader framework that formulates HQA as two dialog-style tasks, row retrieval and answer generation, handled by a shared pretrained language model with task-specific low-rank adapters. RRS combines modular decomposition, noise-robust training against weak-supervision errors, and a natural-language interface without specialized feature engineering. Sharing over 99.8% of backbone parameters, it introduces no more than 0.2% task-specific trainable parameters. With modest-scale public models of 7B and 4B parameters, RRS achieves 71.9%/79.6% and 71.6%/79.2% EM/F1 on the HybridQA official blind test, surpassing the previous best result of 67.9%/75.5%. These results demonstrate a lightweight and noise-robust table–text reasoning approach that is well suited to resource-constrained deployment and provides a useful building block for agentic edge intelligence.

Encryption technology has become ubiquitous in network communication and encrypted malicious traffic detection becomes an important part of malware detection and cyber attack detection. Existing machine learning models and deep learning models are mainly trained based on packet length sequence information and time series information. Recent studies have shown that these models perform poorly in real network environments. In response to this challenge, this paper proposes a novel malicious traffic detection method based on certificate information extracted during the TLS (Transport Layer Security) encrypted handshake protocol. Our approach demonstrates that certificate information exhibits a strong correlation with the maliciousness of traffic, while remaining unaffected by the complexities of the real network environment. The experimental results illustrate that our method has high accuracy and low time overheading.

Collaborative treatment has emerged as a crucial approach for improving the quality and efficiency of medical services in modern healthcare systems. To support this paradigm, secure access control over electronic health records (EHR) and flexible delegation of patient permissions are essential for enabling efficient, privacy-preserving data sharing. This paper proposes a fine-grained access control scheme with decentralized permission delegation tailored for medical collaboration scenarios. To ensure fine-grained access control, we adopt a hybrid encryption scheme that combines a dual-key regression tree with Identity-Based Encryption with Wildcard Key Derivation (WKD-IBE) for efficient and scalable key management. This integration enables access control based on data attributes rather than user identities and adheres to the principle of minimal data disclosure. To support decentralized permission delegation, we extend the WKD-IBE scheme to enable patients to authorize multiple doctors to grant access permissions collaboratively. This extension ensures controlled delegation by enforcing a predefined threshold of doctors and requiring consensus on the requested access scope. Additionally, we provide both theoretical and practical security analyses, along with an implementation to demonstrate the scheme's real-world applicability. Experimental results demonstrate that our scheme achieves lower authorization latency and better scalability in collaborative healthcare scenarios while maintaining comparable encryption efficiency.

Malware detection has become increasingly challenging due to the evolving complexity and diversity of malicious software. Federated learning (FL) provides a promising framework for collaborative model training without sharing sensitive data, making it particularly suitable for malware detection across distributed and heterogeneous data sources. This paper proposes an FL-based malware detection approach that addresses the challenges posed by data heterogeneity. By converting binary malware samples into grayscale images and leveraging convolutional neural networks (CNN), our method achieves effective feature extraction while ensuring cross-platform consistency. To mitigate the impact of non-independent and identically distributed (non-IID) data, we introduce an improved loss function that incorporates a regularization term to align local and global models. Experimental results on the Big2015 dataset demonstrate that the proposed approach outperforms traditional FL methods, such as FedAvg, in scenarios with heterogeneous data distributions. Our findings highlight the effectiveness of combining FL with image-based malware detection techniques in addressing realworld challenges in distributed environments.

The increasing volume and variants of malware are posing substantial risks to the security of personal privacy and property, which gradually makes malware classification a hot research topic in machine learning. Compared with traditional malware classification methods requiring much prior knowledge to explicitly construct feature representation, CNN-based methods have more strength for implicitly exploring deep representation and achieved better performance. However, there is an obvious long-tailed distribution problem in existing malware datasets, i.e., a small fraction of families occupies most of the samples, while the sample distribution of most families is relatively sparse. The long-tailed distribution will lead to poor generalization performance on families with fewer sample numbers in conventional CNN-based malware classification. In this paper, we propose a malware classification method based on representation learning, called RL4Mal, which utilizes metric learning for better generalization on the tail families while keeping performance on the head families. Besides designing a weighted metric loss, we also introduce data augmentation and a memory bank to expand negative samples for the loss optimization, further promoting a better construction of the representation space for malware classification. We conduct sufficient experiments on the Malimg dataset, proving that RL4Mal effectively solves the long-tailed problem with 99.47% accuracy in tail families while keeping 99.78% accuracy on average.

The evolution of malware has led to the development of increasingly sophisticated evasion techniques, significantly escalating the challenges for researchers in obtaining and labeling new instances for analysis. Conventional deep learning detection approaches struggle to identify new malware variants with limited sample availability. Recently, researchers have proposed few-shot detection models to address the above issues. However, existing studies predominantly focus on model-level improvements, overlooking the potential of domain adaptation to leverage the unique characteristics of malware. Motivated by these challenges, we propose a few-shot learning-based malware family detection framework (MalFSLDF). We introduce a novel method for malware representation using structural features and a feature fusion strategy. Specifically, our framework employs contrastive learning to capture the unique textural features of malware families, enhancing the identification capability for novel malware variants. In addition, we integrate entropy graphs (EGs) and gray-level co-occurrence matrices (GLCMs) into the feature fusion strategy to enrich sample representations and mitigate information loss. Furthermore, a domain alignment strategy is proposed to adjust the feature distribution of samples from new classes, enhancing the model's generalization performance. Finally, comprehensive evaluations of the MaleVis and BIG-2015 datasets show significant performance improvements in both 5-way 1-shot and 5-way 5-shot scenarios, demonstrating the effectiveness of the proposed framework.

Remote Attestation (RA) is an effective security service that allows a trusted party (verifier) to initiate the attestation routine on a potentially untrusted remote device (prover) to verify its correct state. Despite their usefulness, traditional challenge-response remote attestation protocols suffer from certain limitations, such as challenges in scaling attestation collection and the forced suspension of normal operation during attestation. Self-attestation tackles these issues by enabling the prover to measure its own state asynchronously with the verifier’s attestation request. Existing self-attestation methods rely on hybrid architectures to provide the required security properties, which may not be compatible with low-end Internet of Things (IoT) devices due to hardware limitations. In addition, these protocols currently lack formal verification of design correctness. In this paper, we present FlashAttest, a formally verified self-attestation protocol for low-end IoT devices. FlashAttest leverages the flash device to fulfill the security properties required by self-attestation, eliminating the requirement for hardware modifications. In particular, FlashAttest allows the prover to initiate the attestation routine and guarantee the trustworthiness of the results based on the verified software-based security architecture. By collaborating with the flash device during attestation to generate timestamped reports, FlashAttest enables the verifier to collect and verify the legitimacy of the attestation results. More importantly, FlashAttest achieves strong security guarantees supported by a formally verified design using the Tamarin prover. We implement and evaluate FlashAttest on MSP430 architecture, showing a reasonable overhead in terms of memory footprint, communication overhead, runtime and power consumption. Compared with state-of-the-art self-attestation schemes, our approach achieves similar runtime overhead, low energy consumption, and reasonable memory overhead while eliminating the need for hardware modifications. The results confirm the suitability of FlashAttest for low-end devices.

Machine learning (ML) and deep learning (DL) have significantly improved the detection accuracy of contemporary Network Intrusion Detection Systems (NIDS), yet they remain susceptible to adversarial attacks. Current attacks against ML/DL-based NIDS primarily focus on altering feature vectors, thereby overlooking the discrete and irreversible nature of network traffic packets, which significantly limits its practical applicability. To address these challenges, we propose TransfficFormer to generate adversarial attack traffic that combines heuristic algorithm and transformer. We train a Transformer-based generator by transforming source-space features into discrete sequence autoregressive models. The three-layer particle swarm optimization algorithm with random and perception factor is utilized to optimize the generation of adversarial mutation malicious traffic with reversible metadata feature vectors. Furthermore, the discriminator feedback probability is fine-tuned using reinforcement learning strategies, ensuring the preservation of both malicious intent and normal communication functionality within the generated traffic. Comprehensive experiments demonstrate that Transfficformer can autonomously generate mutant malicious traffic, effectively evading various ML/DL-based NIDS with minimal overhead. The practicality of the generated mutant traffic is validated in the NSFOCUS cyber range.

In High-Performance Computing (HPC) environments, a comprehensive understanding of cybersecurity threats and their underlying attack strategies is essential. However, current research predominantly focuses on maliciousness determination, typically emphasizing the code’s operational behaviors rather than the attack strategies employed. The advancements in multimedia computing, particularly Large Language Models (LLMs), have paved the way for innovative solutions to the aforementioned bottleneck. This work proposes MalGTA (Guided Malware Tactical Analysis), an LLM-based system that automates ATT CK (Adversarial Tactics, Techniques, and Common Knowledge)-aligned malware tactical analysis through Cuckoo Sandbox-driven dynamic profiling. Specifically, we construct a multi-source knowledge base integrated with Retrieval-Augmented Generation (RAG), which mitigates hallucinations in LLMs through context-sensitive threat intelligence retrieval. In addition, we propose a query optimization strategy to address challenges related to input information overload and attention dispersion in LLMs, enabling context-aware data refinement from Cuckoo reports. Finally, this study conducts dynamic analysis on classical VirusShare and Advanced Persistent Threat (APT) samples and constructs an evaluation dataset based on the authoritative malware analysis platform HybridAnalysis. Experimental results show the effectiveness of the method.

Recently, redactable blockchain has been proposed and leveraged in a wide range of real systems for its unique properties of decentralization, traceability, and transparency while ensuring controllable on-chain data redaction. However, the development of redactable blockchain is now obstructed by three limitations, which are data privacy breaches, high communication overhead, and low searching efficiency, respectively. In this paper, we propose PriChain, the first efficient privacy-preserving fine-grained redactable blockchain in decentralized settings. PriChain provides data owners with rights to control who can read and redact on-chain data while maintaining downward compatibility, ensuring the one who can redact will be able to read. Specifically, inspired by the concept of multi-authority attribute-based encryption, we utilize the isomorphism of the access control tree, realizing fine-grained redaction mechanism, downward compatibility, and collusion resistance. With the newly designed structure, PriChain can realize O（n） communication and storage overhead compared to prior O（n 2 ） schemes. Furthermore, we integrate multiple access trees into a tree-based dictionary, optimizing searching efficiency. Theoretical analysis proves that PriChain is secure against the chosen-plaintext attack and has competitive complexity. The experimental evaluations show that PriChain realizes 10× efficiency improvement of searching and 100× lower communication and storage overhead on average compared with existing schemes.

While deep learning techniques have been extensively employed in malware detection, there is a notable challenge in effectively embedding malware features. Current neural network methods primarily capture superficial characteristics, lacking in-depth semantic exploration of functions and failing to preserve structural information at the file level. Motivated by the aforementioned challenges, this paper introduces MalHAPGNN, a novel framework for malware detection that leverages a hierarchical attention pooling graph neural network based on enhanced call graphs. Firstly, to ensure semantic richness, a Bidirectional Encoder Representations from Transformers-based (BERT) attribute-enhanced function embedding method is proposed for the extraction of node attributes in the function call graph. Subsequently, this work designs a hierarchical graph neural network that integrates attention mechanisms and pooling operations, complemented by function node sampling and structural learning strategies. This framework delivers a comprehensive profile of malicious code across semantic, syntactic, and structural dimensions. Extensive experiments conducted on the Kaggle and VirusShare datasets have demonstrated that the proposed framework outperforms other graph neural network (GNN)-based malware detection methods.

Intrusion Detection Systems (IDS) must cope with concept drift and zero-day attacks under evolving, partially labeled traffic. While AOC-IDS enables self-adaptive detection via autoencoders and online pseudo-labeling, its pairwise objective yields insufficient abnormal separation, noisy label propagation, and weak open-set ability. We present QHCL-AOC-IDS, an uncertainty-aware online IDS that couples Quadruplet-Hard Contrastive Learning (QHCL) with confidence-based adaptation and open-set rejection. QHCL enforces (i) anchor–hard-negative margins and (ii) negative–negative diversity, producing discriminative embeddings. An uncertainty-aware pseudolabeler filters low- confidence samples to curb error accumulation. A lightweight open-set module models normal embeddings and rejects out-of-distribution traffic. On NSL-KDD and UNSW-NB15, QHCL-AOC-IDS improves overall F1 and zero-day detection rate over strong baselines (AOC-IDS, AE, supervised CNN/MLP). It also exhibits faster and more stable online adaptation. Code, configs, and scripts for reproducibility are provided.

Large language models (LLMs) exhibit remarkable text-generation capabilities, yet struggle with factual consistency, motivating growing interest in factuality verification. Existing factuality verification methods typically follow a Decompose-Then-Verify paradigm, which improves granularity but suffers from poor scalability and efficiency. We propose a novel Decompose-Embed-Interact paradigm that shifts factuality verification from costly text-level reasoning to efficient alignment in embedding space, effectively mitigating the scalability bottlenecks and computational inefficiencies inherent to prior approaches. While the proposed paradigm promises scalable verification, its implementation faces three practical challenges: efficient decomposition, factually faithful embedding, and accurate verification in embedding space. To address these challenges, we introduce E-Verify , a lightweight framework that resolves them through three specially designed modules, each aligned with a specific stage of the paradigm and designed to preserve scalability and efficiency. Experiments demonstrate that E-Verify significantly improves both decomposition and verification efficiency while maintaining competitive accuracy. These results confirm that the proposed paradigm enables scalable and fine-grained factuality verification with minimal performance trade-offs.

Large Language Models (LLMs) exhibit extraordinary competence in language comprehension and generation. However, as these models find increasing adoption in communication systems, they also become vulnerable to adversarial threats. Such attacks exploit model response mechanisms to generate malicious content or execute jailbreak attempts. Given the paramount importance of reliability and security in communication networks, this issue has garnered considerable attention. Of particular concern are black-box attacks, which circumvent conventional defense strategies by exploiting input-output interactions without requiring internal model knowledge or parameter access. These attacks are highly clandestine and demonstrate substantial practical feasibility, with possible repercussions such as data compromise, the production of harmful content, and the interruption of standard operations. Although relevant research efforts have achieved notable breakthroughs, a comprehensive examination of the topic, especially a systematic review within the realm of communication networks, remains insufficient. This article seeks to offer a comprehensive survey of contemporary black-box attack strategies aimed at LLMs. We begin by retracing the development and applications of such attacks across diverse fields, then propose a taxonomy pertinent to black-box attacks within communication networks, classifying them into three principal categories: scenario and context manipulation attacks, transformation and evasion attacks, and automated and optimized generation attacks. In addition, we examine the associated impacts and potential risks, synthesize limitations in existing research, and present prospective directions and challenges for future research. Our overarching goal is to provide substantial insights that advance the security and reliability of LLMs while promoting the stable evolution of these models within communication network environments.

With the rapid development of network technologies, the threat posed by malicious software has become increasingly complex and diverse, evolving from traditional viruses and worms to sophisticated ransomware, advanced persistent threats (APTs), and supply chain attacks. Traditional malware detection methods, such as signature-based detection and traditional machine learning techniques, face significant challenges in addressing the polymorphism and obfuscation of modern malware. Recent advances in generative artificial intelligence (GAI), particularly large language models (LLMs) like GPT-4 and CodeBERT, offer new opportunities for malware detection. GAI leverages self-supervised pretraining to understand code syntax and semantics, enabling automated feature extraction and high-level semantic pattern recognition from raw code. By analyzing the structure and behavior of malicious code, GAI can detect hidden threats, predict attack trends, and process complex multimodal data. This paper provides a comprehensive overview of the current challenges in malware detection, including issues with data imbalance, adversarial attacks, and the high cost of expert-labeled samples. It also explores how GAI-powered LLMs enhance static and dynamic analysis, enable multimodal detection, and improve explainability in malware analysis. By summarizing the contributions of GAI in this field, this paper highlights its transformative potential for malware detection methodologies and addresses future research directions.