arXiv subjects

Wei-Bin Lee

Publications and source records attributed to Wei-Bin Lee.

14 recordsLinked to original sources

arXiv2026

Tracing Target Answers in Poisoned Retrieval Corpora via Token Influence Attribution

Retrieval-Augmented Generation (RAG) systems are vulnerable to corpus poisoning attacks that manipulate model outputs through malicious retrieved documents. Existing detection methods typically rely on auxiliary classifiers or additional LLM-based verification, introducing substantial computational overhead. We present TRACE, a lightweight detection framework that identifies poisoning attacks by tracing answer-related tokens through token influence attribution. TRACE first discovers recurrent high-influence keywords across retrieved documents and then performs a secondary verification to confirm their influence on model predictions. Experiments on three QA benchmarks and six LLMs demonstrate strong detection performance while simultaneously uncovering attacker-specified target answers.

cs.CR↗

arXiv2026

RAS: Measuring LLM Safety Through Refusal Alignment

Safety evaluation of large language models (LLMs) is commonly performed by querying models with unsafe or jailbreak prompts and judging whether their outputs violate a safety policy. Although useful, output-level evaluation is expensive, sensitive to judge choice, and easily tied to fixed question banks. We propose **SafeVec**, a white-box evaluation procedure that measures safety from internal representations rather than generated answers. **SafeVec** first extracts layer-wise refusal directions from a safety-aligned reference model, then selects stable layer windows where safe and unsafe behaviors are separable, and finally scores a target model by measuring whether its hidden states align with these refusal directions under unsafe and jailbreak prompts. The resulting metric, **RAS** (**R**efusal **A**lignment **S**core), maps representation-level refusal alignment to a calibrated 0-100 safety score. Across `Llama`, `Gemma`, and `Qwen` model families, RAS separates aligned models from uncensored and abliterated variants, tracks output-level attack success rate, and is substantially faster than judge-based evaluation. These results suggest that refusal alignment provides a compact and efficient signal for white-box LLM safety evaluation.

cs.CR↗

arXiv2026

CodeSentinel: A Three-Layer Defense Against Indirect Prompt Injection in Code Contexts

Code large language models increasingly retrieve external code context from repositories, documentation, issue threads, and coding-agent environments, creating an indirect prompt-injection surface where attackers hide instructions in comments, strings, identifiers, or decoy code. We propose CodeSentinel, a three-layer inference-time sanitizer. It uses Tree-sitter to extract high-risk model-facing CST nodes, then combines syntax-guided pre-filtering, CST-guided Dynamic Min-K\% scoring, and node perturbation analysis to detect adversarial and natural-looking semantic triggers. Detected nodes are removed or neutralized before reaching the downstream Code LLM. Across six recent attack families, \CodeSentinel achieves 0.80 average node-level F1, outperforming CodeGarrison, DePA, and KillBadCode.

cs.CR↗

arXiv2026

When VR Meets BCI: (Un)Observable Brainwave-aware Privacy Reconstruction in the Metaverse via Unrestricted Inbuilt Motion Sensors

Metaverse devices, such as virtual reality (VR), have seen substantial development and widespread applications in numerous areas. Although recent studies have revealed privacy leakages in VR, these vulnerabilities were limited in the scope of observable behaviors in virtual scenes (e.g., what a user is seeing). In this work, we uncover the feasibility of going beyond the scope of observable user behaviors to unobservable brain EEG-correlated representations (e.g., what a user is perceiving) by leveraging unrestricted motion sensors in VR headsets to reconstruct brain EEG signals, a seemingly neglected but promising vector. The insight is that the inbuilt motion sensors (e.g., accelerometers) in the VR headset can capture subtle vibrations induced by pupillary responses, which are highly correlated with users' visual stimuli and in-brain perceptions. Therefore, we design and implement BraVeSpy to systematically investigate and demonstrate the feasibility of this severe privacy leakage originating from brain EEG-correlated representations reconstructed from variations of inbuilt motion sensors. Our extensive evaluation results from different VR devices show that BraVeSpy, for the first time in the Metaverse, can reveal unobservable privacy, where we successfully unveiled perceptive images in the brain with 52.0%-67.2% accuracy. In particular, we also find that BraVeSpy outperforms the current approaches that are limited to coarse-grained inference of observable behaviors and achieves over 85.0% accuracy in inferring user activity-related sensitive information, such as fingerprinting websites, apps, and streaming videos, and over 96.0% accuracy in user de-anonymization, gaze movement tracking, and virtual keystroke inference.

cs.CR↗

arXiv2026

RedTopic: Toward Topic-Diverse Red Teaming of Large Language Models

As large language models (LLMs) are increasingly deployed as black-box components in real-world applications, red teaming has become essential for identifying potential risks. It tests LLMs with adversarial prompts to uncover vulnerabilities and improve safety alignment. Ideally, effective red teaming should be adaptive to evolving LLM capabilities and explore a broad range of harmful topics. However, existing approaches face two limitations: 1) topic-based approaches rely on pre-collected harmful topics, limited in flexibility and adaptivity. 2) topic-free methods use reinforcement learning (RL), but they lack an explicit reward signal for exploration and tend to over-optimize a narrow objective, reducing topic diversity. To address these limitations, we propose RedTopic, a novel red teaming framework that generates topic-diverse adversarial prompts through a contextualized generation pipeline, an aggregate reward design, and a multi-objective RL training loop. Experiments show that RedTopic produces more effective and diverse adversarial prompts than existing methods, with notable improvements in integrated evaluation metrics. We believe RedTopic represents a step toward more adaptive and topic-diverse red teaming for large language models.

cs.LG↗

arXiv2025

Asymmetry Vulnerability and Physical Attacks on Online Map Construction for Autonomous Driving

High-definition maps provide precise environmental information essential for prediction and planning in autonomous driving systems. Due to the high cost of labeling and maintenance, recent research has turned to online HD map construction using onboard sensor data, offering wider coverage and more timely updates for autonomous vehicles. However, the robustness of online map construction under adversarial conditions remains underexplored. In this paper, we present a systematic vulnerability analysis of online map construction models, which reveals that these models exhibit an inherent bias toward predicting symmetric road structures. In asymmetric scenes like forks or merges, this bias often causes the model to mistakenly predict a straight boundary that mirrors the opposite side. We demonstrate that this vulnerability persists in the real-world and can be reliably triggered by obstruction or targeted interference. Leveraging this vulnerability, we propose a novel two-stage attack framework capable of manipulating online constructed maps. First, our method identifies vulnerable asymmetric scenes along the victim AV's potential route. Then, we optimize the location and pattern of camera-blinding attacks and adversarial patch attacks. Evaluations on a public AD dataset demonstrate that our attacks can degrade mapping accuracy by up to 9.9%, render up to 44% of targeted routes unreachable, and increase unsafe planned trajectory rates, colliding with real-world road boundaries, by up to 27%. These attacks are also validated on a real-world testbed vehicle. We further analyze root causes of the symmetry bias, attributing them to training data imbalance, model architecture, and map element representation. To the best of our knowledge, this study presents the first vulnerability assessment of online map construction models and introduces the first digital and physical attack against them.

cs.CR↗

arXiv2025

RedRFT: A Light-Weight Benchmark for Reinforcement Fine-Tuning-Based Red Teaming

Red teaming has proven to be an effective method for identifying and mitigating vulnerabilities in Large Language Models (LLMs). Reinforcement Fine-Tuning (RFT) has emerged as a promising strategy among existing red teaming techniques. However, a lack of a unified benchmark hinders current RFT-based red teaming methods. Implementation details, especially in Proximal Policy Optimization (PPO)-based RFT, significantly affect outcome stability and reproducibility. To address this issue, we introduce RedRFT, a lightweight benchmark designed to simplify and standardize the implementation and evaluation of RFT-based red teaming. RedRFT combines the design strengths of both single-file CleanRL and highly modularized Tianshou, offering high-quality single-file red teaming implementations and modular PPO core components, such as the General Advantage Estimator. It supports a variety of token and sentence diversity metrics, featuring modularized intrinsic reward computation that facilitates plug-and-play experimentation. To clarify their influence on RFT performance, we conducted an extensive ablation study on key components, including Low-Rank Adaptation (LoRA), Kullback-Leibler (KL) divergence, and Lagrange Multiplier. We hope this work contributes to 1) gaining a comprehensive understanding of the implementation nuances of RFT-based red teaming algorithms, and 2) enabling rapid prototyping of innovative features for RFT-based red teaming. Code for the benchmark can be accessed at https://github.com/x-zheng16/RedRFT.git.

cs.LG↗

arXiv2025

CALM: Co-evolution of Algorithms and Language Model for Automatic Heuristic Design

Tackling complex optimization problems often relies on expert-designed heuristics, typically crafted through extensive trial and error. Recent advances demonstrate that large language models (LLMs), when integrated into well-designed evolutionary search frameworks, can autonomously discover high-performing heuristics at a fraction of the traditional cost. However, existing approaches predominantly rely on verbal guidance, i.e., manipulating the prompt generation process, to steer the evolution of heuristics, without adapting the underlying LLM. We propose a hybrid framework that combines verbal and numerical guidance, the latter achieved by fine-tuning the LLM via reinforcement learning based on the quality of generated heuristics. This joint optimization allows the LLM to co-evolve with the search process. Our method outperforms state-of-the-art (SOTA) baselines across various optimization tasks, running locally on a single 24GB GPU using a 7B model with INT4 quantization. It surpasses methods that rely solely on verbal guidance, even when those use significantly more powerful API-based models.

cs.NE↗

arXiv2025

Beyond Natural Language Perplexity: Detecting Dead Code Poisoning in Code Generation Datasets

The increasing adoption of large language models (LLMs) for code-related tasks has raised concerns about the security of their training datasets. One critical threat is dead code poisoning, where syntactically valid but functionally redundant code is injected into training data to manipulate model behavior. Such attacks can degrade the performance of neural code search systems, leading to biased or insecure code suggestions. Existing detection methods, such as token-level perplexity analysis, fail to effectively identify dead code due to the structural and contextual characteristics of programming languages. In this paper, we propose DePA (Dead Code Perplexity Analysis), a novel line-level detection and cleansing method tailored to the structural properties of code. DePA computes line-level perplexity by leveraging the contextual relationships between code lines and identifies anomalous lines by comparing their perplexity to the overall distribution within the file. Our experiments on benchmark datasets demonstrate that DePA significantly outperforms existing methods, achieving 0.14-0.19 improvement in detection F1-score and a 44-65% increase in poisoned segment localization precision. Furthermore, DePA enhances detection speed by 0.62-23x, making it practical for large-scale dataset cleansing. Overall, by addressing the unique challenges of dead code poisoning, DePA provides a robust and efficient solution for safeguarding the integrity of code generation model training datasets.

cs.CL↗

arXiv2025

Layer-Aware Task Arithmetic: Disentangling Task-Specific and Instruction-Following Knowledge

Large language models (LLMs) demonstrate strong task-specific capabilities through fine-tuning, but merging multiple fine-tuned models often leads to degraded performance due to overlapping instruction-following components. Task Arithmetic (TA), which combines task vectors derived from fine-tuning, enables multi-task learning and task forgetting but struggles to isolate task-specific knowledge from general instruction-following behavior. To address this, we propose Layer-Aware Task Arithmetic (LATA), a novel approach that assigns layer-specific weights to task vectors based on their alignment with instruction-following or task-specific components. By amplifying task-relevant layers and attenuating instruction-following layers, LATA improves task learning and forgetting performance while preserving overall model utility. Experiments on multiple benchmarks, including WikiText-2, GSM8K, and HumanEval, demonstrate that LATA outperforms existing methods in both multi-task learning and selective task forgetting, achieving higher task accuracy and alignment with minimal degradation in output quality. Our findings highlight the importance of layer-wise analysis in disentangling task-specific and general-purpose knowledge, offering a robust framework for efficient model merging and editing.

cs.CL↗

arXiv2025

A Survey on Backdoor Threats in Large Language Models (LLMs): Attacks, Defenses, and Evaluations

Large Language Models (LLMs) have achieved significantly advanced capabilities in understanding and generating human language text, which have gained increasing popularity over recent years. Apart from their state-of-the-art natural language processing (NLP) performance, considering their widespread usage in many industries, including medicine, finance, education, etc., security concerns over their usage grow simultaneously. In recent years, the evolution of backdoor attacks has progressed with the advancement of defense mechanisms against them and more well-developed features in the LLMs. In this paper, we adapt the general taxonomy for classifying machine learning attacks on one of the subdivisions - training-time white-box backdoor attacks. Besides systematically classifying attack methods, we also consider the corresponding defense methods against backdoor attacks. By providing an extensive summary of existing works, we hope this survey can serve as a guideline for inspiring future research that further extends the attack scenarios and creates a stronger defense against them for more robust LLMs.

cs.CR↗

arXiv2025

A Contemporary Survey of Large Language Model Assisted Program Analysis

The increasing complexity of software systems has driven significant advancements in program analysis, as traditional methods unable to meet the demands of modern software development. To address these limitations, deep learning techniques, particularly Large Language Models (LLMs), have gained attention due to their context-aware capabilities in code comprehension. Recognizing the potential of LLMs, researchers have extensively explored their application in program analysis since their introduction. Despite existing surveys on LLM applications in cybersecurity, comprehensive reviews specifically addressing their role in program analysis remain scarce. In this survey, we systematically review the application of LLMs in program analysis, categorizing the existing work into static analysis, dynamic analysis, and hybrid approaches. Moreover, by examining and synthesizing recent studies, we identify future directions and challenges in the field. This survey aims to demonstrate the potential of LLMs in advancing program analysis practices and offer actionable insights for security researchers seeking to enhance detection frameworks or develop domain-specific models.

cs.SE↗

arXiv2024

ICSFuzz: Collision Detector Bug Discovery in Autonomous Driving Simulators

With the increasing adoption of autonomous vehicles, ensuring the reliability of autonomous driving systems (ADSs) deployed on autonomous vehicles has become a significant concern. Driving simulators have emerged as crucial platforms for testing autonomous driving systems, offering realistic, dynamic, and configurable environments. However, existing simulation-based ADS testers have largely overlooked the reliability of the simulators, potentially leading to overlooked violation scenarios and subsequent safety security risks during real-world deployment. In our investigations, we identified that collision detectors in simulators could fail to detect and report collisions in certain collision scenarios, referred to as ignored collision scenarios. This paper aims to systematically discover ignored collision scenarios to improve the reliability of autonomous driving simulators. To this end, we present ICSFuzz, a black-box fuzzing approach to discover ignored collision scenarios efficiently. Drawing upon the fact that the ignored collision scenarios are a sub-type of collision scenarios, our approach starts with the determined collision scenarios. Following the guidance provided by empirically studied factors contributing to collisions, we selectively mutate arbitrary collision scenarios in a step-wise manner toward the ignored collision scenarios and effectively discover them. We compare ICSFuzz with DriveFuzz, a state-of-the-art simulation-based ADS testing method, by replacing its oracle with our ignored-collision-aware oracle. The evaluation demonstrates that ICSFuzz outperforms DriveFuzz by finding 10-20x more ignored collision scenarios with a 20-70x speedup. All the discovered ignored collisions have been confirmed by developers with one CVE ID assigned.

cs.CR↗

arXiv2024

A First Physical-World Trajectory Prediction Attack via LiDAR-induced Deceptions in Autonomous Driving

Trajectory prediction forecasts nearby agents' moves based on their historical trajectories. Accurate trajectory prediction is crucial for autonomous vehicles. Existing attacks compromise the prediction model of a victim AV by directly manipulating the historical trajectory of an attacker AV, which has limited real-world applicability. This paper, for the first time, explores an indirect attack approach that induces prediction errors via attacks against the perception module of a victim AV. Although it has been shown that physically realizable attacks against LiDAR-based perception are possible by placing a few objects at strategic locations, it is still an open challenge to find an object location from the vast search space in order to launch effective attacks against prediction under varying victim AV velocities. Through analysis, we observe that a prediction model is prone to an attack focusing on a single point in the scene. Consequently, we propose a novel two-stage attack framework to realize the single-point attack. The first stage of prediction-side attack efficiently identifies, guided by the distribution of detection results under object-based attacks against perception, the state perturbations for the prediction model that are effective and velocity-insensitive. In the second stage of location matching, we match the feasible object locations with the found state perturbations. Our evaluation using a public autonomous driving dataset shows that our attack causes a collision rate of up to 63% and various hazardous responses of the victim AV. The effectiveness of our attack is also demonstrated on a real testbed car. To the best of our knowledge, this study is the first security analysis spanning from LiDAR-based perception to prediction in autonomous driving, leading to a realistic attack on prediction. To counteract the proposed attack, potential defenses are discussed.

cs.CR↗