Abstract With the rapid development of cloud computing, enterprises have increasingly migrated their servers and services from internal networks to cloud environments. Traditional boundary-based protection mechanisms are no longer sufficient for addressing the security requirements of modern cloud-based systems. As cyberattacks continue to evolve, ensuring the long-term, secure, and reliable operation of increasingly complex IT systems has become a significant challenge. Consequently, researchers have proposed various fine-grained access control approaches. Fine-grained access control remains a significant challenge in complex cloud-edge-endpoint collaboration scenarios. Existing approaches often rely on intricate policy definitions to achieve granular control, which can lead to increased computational overhead and degraded system performance. In this paper, we propose a lightweight, Zero-Trust-based Cloud-Edge-Endpoint Access Control System—TLBAC. By embedding trusted labels into TCP packets at the security endpoint (Endpoint), TLBAC enables traffic blocking and forwarding through access control policies issued by the Security Cloud Brain (Cloud) via the Edge Decision Gateway (Edge). This framework achieves fine-grained access control through trusted labels in a cloud-edge-endpoint architecture. To evaluate the effectiveness of TLBAC, four experiments are designed. The first experiment examines the impact of trusted labels insertion on TCP traffic packet transmission behavior. The experimental results show that even under high-latency, significant jitter, and high packet loss conditions in an LTE network, TCP packets with embedded trusted labels maintain stable and reliable data transmission. The second experiment assesses the resource consumption of TLBAC, focusing on CPU and memory overhead. The experimental results indicate that TLBAC’s resource consumption rate is only ± 0.2%. The third experiment simulates a realistic attack environment by launching attacks against the protected system from an adversarial perspective to validate TLBAC’s defensive capabilities. The experimental results demonstrate that TLBAC successfully detected and blocked all attacks in over thirty vulnerability cases involving different components, versions, and exploitation methods. The fourth experiment evaluates the practical operational capability of TLBAC in a multi-tenant environment. The results show that TLBAC can achieve connection-oriented fine-grained access control with low latency, while maintaining policy accuracy and isolation in multi-tenant scenarios.

Code obfuscation is a process that complicates reverse engineering, protects intellectual property, and conceals malware. Existing deobfuscation approaches often lack generality or struggle with complex, mixed, or unknown transformations. To address this issue, this paper proposes LUCID, a Large Language Model (LLM) based Universal Control-flow Integrated Deobfuscation framework. We first formalizes the control-flow deobfuscation task and introduce the Topologically Feasible Path Set (TFPS) as a new evaluation metrics. Building upon this foundation, LUCID leverages an LLM to infer Predicate Mapping Rules between basic blocks in linear time, which then guide the precise expansion of the Runtime Feasible Path Set to identify and eliminate spurious control flows. Finally, semantically equivalent paths are merged to reconstruct a clean, compilable, and behaviorally faithful control-flow graph, from which security-analyst-friendly C-like pseudocode is generated. Comprehensive evaluation on 780 binaries employing 13 distinct obfuscation techniques demonstrates that our method reduces average cyclomatic complexity by 52.4%, achieves full deobfuscation in 53.8% of cases, and suppresses TFPS inflation caused by bogus control flow by over 99%. The framework demonstrates superiority over existing state-of-the-art tools in terms of both generality and semantic consistency, thus evidencing the transformative potential of LLMs in facilitating scalable malware reverse engineering.

The burgeoning prevalence of Ethereum phishing behavior has iCSUR-2025-0155mposed substantial constraints on the advancement of blockchain finance, resulting in losses of more than $7.7 billion to date, so it is urgent to detect it in time. Currently, available detection methods usually focus on the spatial features within transaction graphs. These methods often employ shallow mining techniques on small samples. As a result, they may overlook certain aspects of interaction patterns, such as temporal behavior. Additionally, their data mining capability is limited due to the small sample sizes. In this study, we propose a graph contrastive learning framework to enrich features of accounts behavior patterns with restricted samples to overcome these limitations. Firstly, we construct an Ethereum interaction graph with the multi-graph involving more temporal information centered with labeled nodes and lighten it with our strategy. Secondly, to comprehensively characterize the accounts pattern, we design the encoder part with the GAT-LSTM model based on attention mechanism fusing statistical features , fine-grained temporal behavioral features and graph structural semantic features . Thirdly, to moderate the sparsity of phishing nodes, we employ data augmentation and contrastive learning to fully mine sparse node information. Moreover, we carried out an in-depth experimental evaluation. The CMD-EPD approach, boasting an F 1 -score of 0.87, outperformed all comparison methods. We also executed a thorough case study to analyze phishing accounts phenomenological indicators which back up the superiority of our framework.

Backdoor attacks pose a serious threat to large language models (LLMs), but existing defenses remain fragmented, failing to pro?vide unified defense against both dirty-label and clean-label attacks. To investigate why such fragmentation arises, we present the first systematic feature-level mechanistic analysis of LLM backdoors using sparse autoencoders (SAEs). Starting from a 2 x 2 comparison of clean and poisoned models on clean and triggered inputs, we trace backdoor-induced logit shifts to high-contributing SAE features and categorize them into four roles: interac?tion, suppressed, mixed, and weight-modified features. This taxonomy reveals system?atic encoding differences: dirty-label back?doors are dominated by isolated interaction features, whereas clean-label backdoors rely more on heterogeneous mixtures of mixed and weight-modified features. These differ?ences explain why existing defenses remain fragmented across attack paradigms. We val?idate this hypothesis through inference-time feature clamping, which reduces ASR to at most 10.8

Path planning constitutes a critical component of penetration testing for 5G industrial cyber-physical system (5G-ICPS) networks. The diversity of interfaces and protocols necessitates deep analysis of vulnerability exploitation methods and cross-protocol combination strategies, significantly increasing attack-path-planning complexity. Furthermore, dynamic network slice configurations and physical information coupling effects drive continuous topological evolution, causing state-space explosion and challenging path planning under uncertainty with incomplete information. To address these issues, we construct a temporal attack graph modeling 5G-ICPS attack processes and design a GraphSAGE-based environment representation encoder. This encoder undergoes multitiered self-supervised pretraining, employing node-level and graph-level training to encode diverse reinforcement learning environments across attack scenarios into fixed-dimensional vector representations. This achieves decoupling from underlying topology, vulnerability specifics, and security configurations, effectively mitigating state-space explosion in large-scale networks. Subsequently, we cluster highly similar vulnerabilities and filter invalid attack actions using three typical 5G-ICPS attack constraints, compressing the agent's exploration space. Especially, we design a customized reward function that dynamically incentivizes/penalizes actions based on compromised assets. Experimental results demonstrate significant improvements: penetration testing invalid action rates decrease from 22.3% to 7.5%, while average steps to achieve attack targets reduce by >54%. These advancements effectively reduce penetration testing costs and increase attack success rates.

Abstract Extracting Chinese Cyber Threat Intelligence (CTI) under increasingly complex advanced persistent threat scenarios is crucial, yet challenging due to domain-specific term ambiguity and frequent long, nested entities. To address polysemy, nested-label conflicts, and cross-sentence semantic discontinuity, we propose an enhanced Transformer-based entity recognition method formulated as a pointer network. On the encoder side, we build a RoBERTa model with Rotary Positional Embeddings. To handle complex positions and boundaries of heterogeneous entity types, we introduce tokenization compensation and positional-parameter compression to sharpen boundary sensitivity. In the decoder, we refine GlobalPointer and model recognition as 2D head–tail span matching, enabling direct detection of overlapping and nested entities. To mitigate long-tail bias, we introduce an entity-frequency-aware dynamic threshold and a reweighted zero-boundary log-loss to improve recall for rare entities. Experiments demonstrate an overall F1 improvement of 6.32% over baselines on Chinese CTI datasets, with absolute gains reaching 19.7% specifically on nested and long entities. These results validate the model’s effectiveness in Chinese-specific named entity recognition and its utility for high-accuracy automated CTI analysis.

The rapid proliferation of IoT devices enlarges the network attack surface, making encrypted traffic classification essential for device identification and security management. Existing supervised models rely heavily on labeled data and often generalize poorly under device-state or domain shifts, while many self-supervised schemes do not explicitly capture the bidirectional request-response structure reflected by packet-direction alternations. We propose Directional-Context Masked Autoencoder (DCMA), a self-supervised framework that learns transferable representations from unlabeled sessions and enables accurate device-type classification with limited annotations. DCMA constructs a compact fixed-length block sequence by selecting the session-initiating packet and packets at uplink/downlink transition points, preserving interaction structure while suppressing redundancy. We pre-train a Transformer-based masked autoencoder to reconstruct randomly masked non-padding blocks and fine-tune the encoder for downstream classification. On CICIoT2022-Active/Idle, DCMA achieves macro-F1 scores of 98.23% and 97.28%, respectively, and improves robustness under class imbalance and device-state shifts.

Existing host intrusion detection systems typically utilize deep learning techniques to detect network attack behaviors from large volumes of log data and generate alerts. However, these methods often fail to balance usability and accuracy, as the generated alerts typically only include the alert type and its mapping to raw logs. This information is insufficient for subsequent processing and response tasks, and accurate diagnosis still requires substantial manual effort to determine the specific causes. To enhance the utility of detection alerts, this study proposes a novel alert analysis method, 𝐀^3𝐋𝐋𝐌 , based on Large Language Model (LLM) technology. The method fine-tunes a smaller model through supervision and applies GRPO-based reinforcement learning to optimize the model’s reasoning outputs. The proposed method builds reasoning from raw logs to alert logs, determining the attack alert and providing explanations of the causes of alerts. Experimental validation on public datasets shows significant improvements in the quality of generated content after reinforcement learning, and the proposed method outperforms state-of-the-art general LLMs in the accuracy of alert authenticity judgment.

While virtualization and resource pooling empower cloud networks with structural flexibility and elastic scalability, they inevitably expand the attack surface and challenge cyber resilience. Reinforcement Learning (RL)-based defense strategies have been developed to optimize resource deployment and isolation policies under adversarial conditions, aiming to enhance system resilience by maintaining and restoring network availability. However, existing approaches lack robustness as they require retraining to adapt to dynamic changes in network structure, node scale, attack strategies, and attack intensity. Furthermore, the lack of Human-in-the-Loop (HITL) support limits interpretability and flexibility. To address these limitations, we propose CyberOps-Bots, a hierarchical multi-agent reinforcement learning framework empowered by Large Language Models (LLMs). Inspired by MITRE ATT&CK's Tactics-Techniques model, CyberOps-Bots features a two-layer architecture: (1) An upper-level LLM agent with four modules--ReAct planning, IPDRR-based perception, long-short term memory, and action/tool integration--performs global awareness, human intent recognition, and tactical planning; (2) Lower-level RL agents, developed via heterogeneous separated pre-training, execute atomic defense actions within localized network regions. This synergy preserves LLM adaptability and interpretability while ensuring reliable RL execution. Experiments on real cloud datasets show that, compared to state-of-the-art algorithms, CyberOps-Bots maintains network availability 68.5% higher and achieves a 34.7% jumpstart performance gain when shifting the scenarios without retraining. To our knowledge, this is the first study to establish a robust LLM-RL framework with HITL support for cloud defense. We will release our framework to the community, facilitating the advancement of robust and autonomous defense in cloud networks.

In recent years, insider threat incidents have occurred with increasing frequency, leading to severe data breaches and substantial economic losses. Most existing insider threat detection methods rely primarily on single-modal features, such as system logs and registry data, while failing to fully exploit the rich semantic information embedded in instant messaging and email content of insider users. To address the above issues, we propose FusionITD, a cross-modal insider threat perception enhancement framework based on the fusion of behavioral and semantic features. This framework combines users' temporal behavioral characteristics such as file operations and login device patterns with the semantic information derived from web browsing and email content. By modeling user behavior baselines from multiple dimensions, FusionITD enables more accurate anomaly detection when deviations from the baseline occur. Firstly, based on the temporal distribution of user behaviors, the behavior data is segmented and aggregated according to the time window to form a user behavior graph. We propose WR-GNN based on graph representation learning to capture temporal behavioral features, and introduce the Focal MSE loss function to address the data imbalance problem caused by sparse abnormal behavior data. Secondly, we propose a retrieval-augmented generation-based semantic analysis algorithm. We use cosine similarity to perform semantic matching and ranking between behavioral contents and historical behaviors. We extract features such as emotion, intention, and focus to achieve fine-grained anomaly detection for user behavior. Finally, we designed an adaptive weighting mechanism based on logistic regression to dynamically integrate the outputs of the previous two parts, enhancing the generalization ability for different threat scenarios. Experimental results conducted on the CERT datasets show that FusionITD outperforms other methods by achieving a 5% increase in AUC, a higher TPR, and a lower false positive rate.

Detecting stealthy attack traffic in large-scale, high-speed networks is challenging because weak malicious signals are sparsely embedded in long flows with irregular packet timings. Existing deep models either truncate flows or assume uniformly sampled sequences, which mismatches real traffic. We propose FlowMamba, a time-aware Memory-Decaying State Space Model (MD-SSM) for flow-level stealthy attack detection. FlowMamba adopts a hierarchical architecture that extracts packet semantics using Mamba blocks and attention pooling, and then models flow-level dependencies with a memory-decaying SSM whose state transitions are modulated by inter-packet time intervals. By treating inter-arrival times as continuous control signals driving exponential memory decay, FlowMamba emphasizes recent suspicious behaviors while preserving long-range context. Experiments on CICIoT2023 and USTC-TFC2016 show that FlowMamba consistently outperforms traditional and state-of-the-art detectors, achieving accuracies of 93.67% and 99.46%, respectively.

With enterprise networks expanding, lateral movement attacks on critical assets pose growing threats. Prior researches employ pattern-matching techniques for detection but suffer from high false positives and poor scalability. This paper proposes a novel unsupervised method LMBE to address these challenges. LMBE first extracts successful login behaviors from multiple sources as triples to constitute a login graph. Based on this graph, LMBE trains to acquire node embeddings by capturing the computational semantics and structural semantics among nodes. Then node embeddings are calculated and used to differentiate benign logins and anomalous logins by an unsupervised model to produce initial anomalies. Finally, a filtering mechanism is proposed to mitigate false positives and raise final alerts. Extensive experimental evaluation on two datasets demonstrates that the proposed method achieves 95

Aligning Common Vulnerability Exposures(CVE) with ATT\&CK is a crucial step in achieving active defense. Existing machine learning and deep learning-based methods have received attention due to their high detection accuracy and strong generalization capabilities, but they struggle to effectively capture the deep semantic associations between CVE and ATT\&CK. Large language models(LLMs), with their ability to understand complex contexts, are likely to become the mainstream approach for CVE-ATT\&CK alignment. However, low-resource environments offer limited computation and scarce labeled samples. Prompt-based methods let large-scale LLMs (such as GPT) perform well, but their high computational cost and low inference efficiency limit their practical use in CVE-ATT\&CK alignment. In contrast, under the same constrained conditions, small-scale LLMs (such as Qwen-2.5-7B), although they can adapt to specific tasks through fine-tuning and offer better practicality, are limited by insufficient understanding of professional knowledge, making it difficult to comprehensively grasp complex attack behavior patterns. To address this, this paper proposes a multi-stage knowledge enhancement framework called CATBridge, whose core lies in constructing a ''teacher-student'' large-small model collaboration mechanism, through three progressive stages of targeted data enhancement, domain knowledge enhancement, and reasoning knowledge enhancement, to gradually guide lightweight LLMs to complete CVE-ATT\&CK alignment tasks. Experimental results show that the framework performs well in low-resource environments, with the deployed lightweight LLM (7B) achieving an accuracy of 83.50%, an absolute improvement of 33.29 percentage points over SOTA, and an average improvement of 46.24 percentage points over zero-shot large-scale general-purpose LLMs such as GPT-5, DeepSeek-R1 (671B), and Qwen3 (235B).

Retrieval-Augmented Generation (RAG) significantly mitigates the hallucinations and domain knowledge deficiency in large language models by incorporating external knowledge bases. However, the multi-module architecture of RAG introduces complex system-level security vulnerabilities. Guided by the RAG workflow, this paper analyzes the underlying vulnerability mechanisms and systematically categorizes core threat vectors such as data poisoning, adversarial attacks, and membership inference attacks. Based on this threat assessment, we construct a taxonomy of RAG defense technologies from a dual perspective encompassing both input and output stages. The input-side analysis reviews data protection mechanisms including dynamic access control, homomorphic encryption retrieval, and adversarial pre-filtering. The output-side examination summarizes advanced leakage prevention techniques such as federated learning isolation, differential privacy perturbation, and lightweight data sanitization. To establish a unified benchmark for future experimental design, we consolidate authoritative test datasets, security standards, and evaluation frameworks. To the best of our knowledge, this paper presents the first end-to-end survey dedicated to the security of RAG systems. Distinct from existing literature that isolates specific vulnerabilities, we systematically map the entire pipeline-providing a unified analysis of threat models, defense mechanisms, and evaluation benchmarks. By enabling deep insights into potential risks, this work seeks to foster the development of highly robust and trustworthy next-generation RAG systems.

Federated learning (FL), a promising paradigm for distributed machine learning, facilitates collaborative model training while preserving data locally for enhanced privacy. Nevertheless, its widespread application in edge computing scenarios is increasingly challenged by poisoning attacks, wherein malicious adversaries compromise model convergence through contamination of local datasets or model parameters. Currently, existing Byzantine-robust FL methods predominantly rely on idealized assumption of independent and identically distributed (IID) data or necessitate access to raw gradients-introducing significant privacy risks. To overcome these limitations, we propose privacy-preserving FL method against poisoning attacks on non-IID data (PFLAP). Inspired by community detection principles, we adaptively partition edge nodes with non-IID data into clusters exhibiting similar data distributions. Building upon this foundation, we develop a novel poisoning defense scheme that integrates scaled dot-product attention with linear homomorphic encryption, enabling detection of encrypted poisoning gradients. Security analysis proves that PFLAP provides rigorous privacy preservation, and extensive experiments demonstrate that compared with existing defense schemes, PFLAP improves accuracy by 8.57%-34.42% under highly non-IID settings, exhibits stronger resistance against both targeted and untargeted attacks, and reduces computational overhead by 6.1%-10.2% and communication overhead by 11.4%-17.4%.

The proliferation of Android multimedia applications highlights the critical role of mobile sensors. Inherent manufacturing defects enable implicit device identification without consent, facilitating covert tracking. This bolsters security through reliable malicious actor tracking, unlike spoof-vulnerable explicit methods. However, prior sensor-based identification suffers from signal noise and device degradation, compromising robustness.In this paper, we propose AMSensorFP, a novel Android implicit device tracking framework. We develop an application to collect device information and multi-sensor data to construct device fingerprints. Subsequently, we build a dataset by performing pairwise difference calculations on the collected fingerprints. And then enhance the dataset with Gaussian noise to improve data diversity and robustness. An autoencoder reduces feature dimensionality, and the processed features are fed into a BiLSTM model with a multi-head attention mechanism, enabling effective fingerprint recognition. Experimental results show that AMSensorFP achieves 99.88% accuracy and 97.34% true positive rate(TPR), significantly outperforming existing methods. Ablation analysis further highlights the contributions of each module and feature in the framework. AMSensorFP delivers a reliable solution for device tracking and security enhancement.

Binary code search is critical for applications such as plagiarism detection and security analysis, but it is challenging due to compiler-induced transformations at different optimization levels. Existing function similarity methods often fail in large-scale search scenarios, particularly pairwise approaches that struggle with scalability. To address this, we propose BASSET, a novel framework that leverages multilevel hybrid semantic features for efficient large-scale binary function clone search. BASSET decomposes functions into five semantic units and applies various embedding strategies to generate indexing vectors for similarity measurement. Notably, it integrates an expression tree-based representation to capture robust features across compiler optimization levels. By utilizing a learning-to-rank approach with convolutional neural networks, BASSET combines similarity scores from different semantic units to generate a final ranking. Experimental results show that BASSET outperforms existing methods, achieving an AUC of 0.992, an nDCG@10 of 0.853, and a stable MRR of 59%, even as the search space grows.

In recent years, DNS over Encrypted (DoE) methods have been regarded as a novel trend within the realm of the DNS ecosystem. In these DoE methods, DNS over HTTPS (DoH) provides encryption to protect data confidentiality while providing better obfuscation to avoid censorship by multiplexing port 443 with web services. This development introduced certain inconveniences in discovering publicly available DoH services. In this paper, we propose the E-DoH method for elegant and efficient DoH service detection. First, we optimized the probing mechanism to enable a single DoH connection to accomplish multiple tasks including service discovery, correctness validation and dependency construction. Second, we propose an efficient DoH detection tool. This tool can enhance probing efficiency while significantly reduce the required traffic volume. Third, based on the above optimization methods, we conducted an exploration of the IPv4 space and performed an in-depth analysis of DoH based on the collected information. Through experiments, our approach demonstrates a remarkable 80% improvement in time efficiency, and only requires 4%-20% traffic volume to complete the detection task. In wild detection, our approach discovered 46k DoH services, which nearly doubles the number discovered by the state-of-the-art. Based on the collected data, we present several intriguing conclusions about the current DoH service ecosystem.