Yuling Liu | Searcharxiv

arXiv2026

Why Are LLM Backdoor Defenses Fragmented? A Feature-Level Explanation with Sparse Autoencoders

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% in most dirty-label settings and at most 15.4% in the majority of clean-label settings, while preserving benign-task perfor?mance. These results show that SAE-based analysis can explain defense fragmentation and guide interpretable backdoor mitigation.

cs.CR↗

arXiv2026

Evidence-Grounded Trustworthy Multimodal Reasoning and Evaluation Benchmark in Complex Urban Scenes

While Multimodal Large Language Models (MLLMs) demonstrate impressive performance in benign scenarios, their cognitive reliability deteriorates significantly in complex scenes under adverse conditions. In these settings, models often rely on implicit inference without sufficient visual evidence, leading to a disconnect between perception and reasoning. Meanwhile, existing outcome-oriented benchmarks evaluate only final predictions and fail to diagnose failures in the underlying reasoning process. To address this gap, the authors propose AD2-Bench, which introduces a Hierarchical Visual Diagnosis framework that decomposes reasoning into a structured Chain of Evidence (CoE). This fine-grained diagnosis reveals that robust multimodal reasoning fundamentally depends on accurate evidence acquisition. Building on this perspective, the authors formulate reasoning from a probabilistic viewpoint and identify two primary causes of reasoning failure: Spatial Ambiguity, where models fail to distinguish target objects from background clutter, resulting in localization errors; and Semantic Uncertainty, where degraded visual features lead to incorrect semantic interpretation, resulting in understanding errors. To overcome these evidence deficiencies, they further propose Evidence-grounded Visual Reasoning (EGVOR), which replaces implicit reasoning with the explicit generation of Evidence Atoms - structured spatial-semantic triplets that enforce tight alignment between localization and semantic understanding. The model is trained through a hierarchical curriculum that progresses from reflective supervision construction to reinforcement learning, where reducing reasoning variance is explicitly rewarded. Extensive experiments demonstrate that EGVOR substantially improves reasoning stability under adverse conditions, providing a more robust framework for trustworthy multimodal cognition.

cs.CV↗

arXiv2026

MirageBackdoor: A Stealthy Attack that Induces Think-Well-Answer-Wrong Reasoning

While Chain-of-Thought (CoT) prompting has become a standard paradigm for eliciting complex reasoning capabilities in Large Language Models, it inadvertently exposes a new attack surface for backdoor attacks. Existing CoT backdoor attacks typically manipulate the intermediate reasoning steps to steer the model toward incorrect answers. However, these corrupted reasoning traces are readily detected by prevalent process-monitoring defenses. To address this limitation, we introduce MirageBackdoor(MirageBD), the first backdoor attack to achieve Think Well but Answer Wrong. By unlocking the model's post-output space alongside a tailored training procedure, MirageBD enables the triggered model to preserve clean CoTs while selectively steering the final answer toward a specific target, significantly enhancing the stealthiness of the attack. Experiments show that MirageBD generally achieves over 90% attack success rate across four datasets and five models with a poison ratio of only 5%. Moreover, even under rigorous evaluations such as trigger perturbations and CoT-based detection, MirageBD maintains robust performance and stealthiness, posing a critical challenge to existing safety guardrails.

cs.CR↗

arXiv2026

Towards Secure Retrieval-Augmented Generation: A Comprehensive Review of Threats, Defenses and Benchmarks

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.

cs.CR↗

arXiv2026

Enhancing Cloud Network Resilience via a Robust LLM-Empowered Multi-Agent Reinforcement Learning Framework

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.

cs.CR↗

arXiv2025

A Content-Preserving Secure Linguistic Steganography

Existing linguistic steganography methods primarily rely on content transformations to conceal secret messages. However, they often cause subtle yet looking-innocent deviations between normal and stego texts, posing potential security risks in real-world applications. To address this challenge, we propose a content-preserving linguistic steganography paradigm for perfectly secure covert communication without modifying the cover text. Based on this paradigm, we introduce CLstega (\textit{C}ontent-preserving \textit{L}inguistic \textit{stega}nography), a novel method that embeds secret messages through controllable distribution transformation. CLstega first applies an augmented masking strategy to locate and mask embedding positions, where MLM(masked language model)-predicted probability distributions are easily adjustable for transformation. Subsequently, a dynamic distribution steganographic coding strategy is designed to encode secret messages by deriving target distributions from the original probability distributions. To achieve this transformation, CLstega elaborately selects target words for embedding positions as labels to construct a masked sentence dataset, which is used to fine-tune the original MLM, producing a target MLM capable of directly extracting secret messages from the cover text. This approach ensures perfect security of secret messages while fully preserving the integrity of the original cover text. Experimental results show that CLstega can achieve a 100\% extraction success rate, and outperforms existing methods in security, effectively balancing embedding capacity and security.

cs.CR↗

arXiv2025

QLPro: Automated Code Vulnerability Discovery via LLM and Static Code Analysis Integration

We introduce QLPro, a vulnerability detection framework that systematically integrates LLMs and static analysis tools to enable comprehensive vulnerability detection across entire open-source projects.We constructed a new dataset, JavaTest, comprising 10 open-source projects from GitHub with 62 confirmed vulnerabilities. CodeQL, a state-of-the-art static analysis tool, detected only 24 of these vulnerabilities while QLPro detected 41. Furthermore, QLPro discovered 6 previously unknown vulnerabilities, 2 of which have been confirmed as 0-days.

cs.SE↗

arXiv2025

Evidence-Grounded Trustworthy Multimodal Reasoning and Evaluation Benchmark in Complex Urban Scenes

While Multimodal Large Language Models (MLLMs) demonstrate impressive performance in benign scenarios, their cognitive reliability deteriorates significantly in complex scenes under adverse conditions. In these settings, models often rely on implicit inference without sufficient visual evidence, leading to a disconnect between perception and reasoning. Meanwhile, existing outcome-oriented benchmarks evaluate only final predictions and fail to diagnose failures in the underlying reasoning process. To address this gap, the authors propose AD2-Bench, which introduces a Hierarchical Visual Diagnosis framework that decomposes reasoning into a structured Chain of Evidence (CoE). This fine-grained diagnosis reveals that robust multimodal reasoning fundamentally depends on accurate evidence acquisition. Building on this perspective, the authors formulate reasoning from a probabilistic viewpoint and identify two primary causes of reasoning failure: Spatial Ambiguity, where models fail to distinguish target objects from background clutter, resulting in localization errors; and Semantic Uncertainty, where degraded visual features lead to incorrect semantic interpretation, resulting in understanding errors. To overcome these evidence deficiencies, they further propose Evidence-grounded Visual Reasoning (EGVOR), which replaces implicit reasoning with the explicit generation of Evidence Atoms - structured spatial-semantic triplets that enforce tight alignment between localization and semantic understanding. The model is trained through a hierarchical curriculum that progresses from reflective supervision construction to reinforcement learning, where reducing reasoning variance is explicitly rewarded. Extensive experiments demonstrate that EGVOR substantially improves reasoning stability under adverse conditions, providing a more robust framework for trustworthy multimodal cognition.

cs.CV↗

arXiv2024

Winemaking: Extracting Essential Insights for Efficient Threat Detection in Audit Logs

Advanced Persistent Threats (APTs) are continuously evolving, leveraging their stealthiness and persistence to put increasing pressure on current provenance-based Intrusion Detection Systems (IDS). This evolution exposes several critical issues: (1) The dense interaction between malicious and benign nodes within provenance graphs introduces neighbor noise, hindering effective detection; (2) The complex prediction mechanisms of existing APTs detection models lead to the insufficient utilization of prior knowledge embedded in the data; (3) The high computational cost makes detection impractical. To address these challenges, we propose Winemaking, a lightweight threat detection system built on a knowledge distillation framework, capable of node-level detection within audit log provenance graphs. Specifically, Winemaking applies graph Laplacian regularization to reduce neighbor noise, obtaining smoothed and denoised graph signals. Subsequently, Winemaking employs a teacher model based on GNNs to extract knowledge, which is then distilled into a lightweight student model. The student model is designed as a trainable combination of a feature transformation module and a personalized PageRank random walk label propagation module, with the former capturing feature knowledge and the latter learning label and structural knowledge. After distillation, the student model benefits from the knowledge of the teacher model to perform precise threat detection. We evaluate Winemaking through extensive experiments on three public datasets and compare its performance against several state-of-the-art IDS solutions. The results demonstrate that Winemaking achieves outstanding detection accuracy across all scenarios and the detection time is 1.4 to 5.2 times faster than the current state-of-the-art methods.

cs.CR↗

arXiv2024

Are Watermarks Bugs for Deepfake Detectors? Rethinking Proactive Forensics

AI-generated content has accelerated the topic of media synthesis, particularly Deepfake, which can manipulate our portraits for positive or malicious purposes. Before releasing these threatening face images, one promising forensics solution is the injection of robust watermarks to track their own provenance. However, we argue that current watermarking models, originally devised for genuine images, may harm the deployed Deepfake detectors when directly applied to forged images, since the watermarks are prone to overlap with the forgery signals used for detection. To bridge this gap, we thus propose AdvMark, on behalf of proactive forensics, to exploit the adversarial vulnerability of passive detectors for good. Specifically, AdvMark serves as a plug-and-play procedure for fine-tuning any robust watermarking into adversarial watermarking, to enhance the forensic detectability of watermarked images; meanwhile, the watermarks can still be extracted for provenance tracking. Extensive experiments demonstrate the effectiveness of the proposed AdvMark, leveraging robust watermarking to fool Deepfake detectors, which can help improve the accuracy of downstream Deepfake detection without tuning the in-the-wild detectors. We believe this work will shed some light on the harmless proactive forensics against Deepfake.

cs.CV↗

arXiv2024

E-DoH: Elegantly Detecting the Depths of Open DoH Service on the Internet

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.

cs.CR↗

arXiv2023

CBSeq: A Channel-level Behavior Sequence For Encrypted Malware Traffic Detection

Machine learning and neural networks have become increasingly popular solutions for encrypted malware traffic detection. They mine and learn complex traffic patterns, enabling detection by fitting boundaries between malware traffic and benign traffic. Compared with signature-based methods, they have higher scalability and flexibility. However, affected by the frequent variants and updates of malware, current methods suffer from a high false positive rate and do not work well for unknown malware traffic detection. It remains a critical task to achieve effective malware traffic detection. In this paper, we introduce CBSeq to address the above problems. CBSeq is a method that constructs a stable traffic representation, behavior sequence, to characterize attacking intent and achieve malware traffic detection. We novelly propose the channels with similar behavior as the detection object and extract side-channel content to construct behavior sequence. Unlike benign activities, the behavior sequences of malware and its variant's traffic exhibit solid internal correlations. Moreover, we design the MSFormer, a powerful Transformer-based multi-sequence fusion classifier. It captures the internal similarity of behavior sequence, thereby distinguishing malware traffic from benign traffic. Our evaluations demonstrate that CBSeq performs effectively in various known malware traffic detection and exhibits superior performance in unknown malware traffic detection, outperforming state-of-the-art methods.

cs.CR↗