Searcharxiv⌕ Search

arXiv subjects

Tianhang Zheng

Publications and source records attributed to Tianhang Zheng.

At least 37 records · Page 2Linked to original sources
arXiv2026

AgentDoG 1.5: A Lightweight and Scalable Alignment Framework for AI Agent Safety and Security

Modern open-world agents such as OpenClaw exhibit powerful cross-environment execution capabilities yet introduce broad new safety risk sources. Meanwhile, advanced frontier AI models drastically lower attack barriers, rendering current agent alignment frameworks inadequate for real-world deployment. To tackle these emerging threats, we propose a lightweight and scalable agent safety alignment framework. Specifically, we update the agent safety taxonomy to accommodate emergent risks from Codex and OpenClaw execution scenarios. We further build a taxonomy-guided data engine with influence-function purification to train lightweight AgentDoG 1.5 variants (0.8B, 2B, 4B, and 8B parameters) using only around 1k samples, achieving comparable performance with leading closed-source models (e.g., GPT-5.4). Based on AgentDoG 1.5, we construct a highly efficient agentic safety SFT and RL training environment, which reduces deployment overhead in Docker-level environments by two orders of magnitude. Finally, we deploy AgentDoG 1.5 as a training-free online guardrail for real-time safety moderation. Extensive experimental results indicate that AgentDoG 1.5 achieves state-of-the-art performance in diverse and complex interactive agentic scenarios. All models and datasets are openly released.

cs.AI↗
arXiv2026

Reasoning as an Attack Surface: Adaptive Evolutionary CoT Jailbreaks for LLMs

Large Reasoning Models (LRMs) have demonstrated remarkable capabilities in reasoning and generation tasks and are increasingly deployed in real-world applications. However, their explicit chain-of-thought (CoT) mechanism introduces new security risks, making them particularly vulnerable to jailbreak attacks. Existing approaches often rely on static CoT templates to elicit harmful outputs, but such fixed designs suffer from limited diversity, adaptability, and effectiveness. To overcome these limitations, we propose an adaptive evolutionary CoT jailbreak framework, called AE-CoT. Specifically, the method first rewrites harmful goals into mild prompts with teacher role-play and decomposes them into semantically coherent reasoning fragments to construct a pool of CoT jailbreak candidates. Then, within a structured representation space, we perform multi-generation evolutionary search, where candidate diversity is expanded through fragment-level crossover and a mutation strategy with an adaptive mutation-rate control mechanism. An independent scoring model provides graded harmfulness evaluations, and high-scoring candidates are further enhanced with a harmful CoT template to induce more destructive generations. Extensive experiments across multiple models and datasets demonstrate the effectiveness of the proposed AE-CoT, consistently outperforming state-of-the-art jailbreak methods.

cs.AI↗
arXiv2026

Accelerating Suffix Jailbreak attacks with Prefix-Shared KV-cache

Suffix jailbreak attacks serve as a systematic method for red-teaming Large Language Models (LLMs) but suffer from prohibitive computational costs, as a large number of candidate suffixes need to be evaluated before identifying a jailbreak suffix. This paper presents Prefix-Shared KV Cache (PSKV), a plug-and-play inference optimization technique tailored for jailbreak suffix generation. Our method is motivated by a key observation that when performing suffix jailbreaking, while a large number of candidate prompts need to be evaluated, they share the same targeted harmful instruction as the prefix. Therefore, instead of performing redundant inference on the duplicated prefix, PSKV maintains a single KV cache for this prefix and shares it with every candidate prompt, enabling the parallel inference of diverse suffixes with minimal memory overhead. This design enables more aggressive batching strategies that would otherwise be limited by memory constraints. Extensive experiments on six widely used suffix attacks across five widely deployed LLMs demonstrate that PSKV reduces inference time by 40\% and peak memory usage by 50\%, while maintaining the original Attack Success Rate (ASR). The code has been submitted and will be released publicly.

cs.CR↗
arXiv2026

Mitigating Error Amplification in Fast Adversarial Training

Fast Adversarial Training (FAT) has proven effective in enhancing model robustness by encouraging networks to learn perturbation-invariant representations. However, FAT often suffers from catastrophic overfitting (CO), where the model overfits to the training attack and fails to generalize to unseen ones. Moreover, robustness oriented optimization typically leads to notable performance degradation on clean inputs, and such degradation becomes increasingly severe as the perturbation budget grows. In this work, we conduct a comprehensive analysis of how guidance strength affects model performance by modulating perturbation and supervision levels across distinct confidence groups. The findings reveal that low confidence samples are the primary contributors to CO and the robustness accuracy trade off. Building on this insight, we propose a Distribution-aware Dynamic Guidance (DDG) strategy that dynamically adjusts both the perturbation budget and supervision signal. Specifically, DDG scales the perturbation magnitude according to the sample confidence at the ground truth class, thereby guiding samples toward consistent decision boundaries while mitigating the influence of learning spurious correlations. Simultaneously, it dynamically adjusts the supervision signal based on the prediction state of each sample, preventing overemphasis on incorrect signals. To alleviate potential gradient instability arising from dynamic guidance, we further design a weighted regularization constraint. Extensive experiments on standard benchmarks demonstrate that DDG effectively alleviates both CO and the robustness accuracy trade off.

cs.LG↗
arXiv2026

Unveiling the Backdoor Mechanism Hidden Behind Catastrophic Overfitting in Fast Adversarial Training

Fast Adversarial Training (FAT) has attracted significant attention due to its efficiency in enhancing neural network robustness against adversarial attacks. However, FAT is prone to catastrophic overfitting (CO), wherein models overfit to the specific attack used during training and fail to generalize to others. While existing methods introduce diverse hypotheses and propose various strategies to mitigate CO, a systematic and intuitive explanation of CO remains absent. In this work, we innovatively interpret CO through the lens of backdoor. Through validations on pathway division, diverse feature predictions, and universal class distinguishable triggers in CO, we conceptualize CO as a weak trigger variant of unlearnable tasks, unifying CO, backdoor attacks, and unlearnable tasks under a common theoretical framework. Guided by this, we leverage several backdoor inspired strategies to mitigate CO: (i) Recalibrate CO affected model parameters using vanilla fine tuning, linear probing, or reinitialization-based techniques; (ii) Introduce a weight outlier suppression constraint to regulate abnormal deviations in model weights. Extensive experiments support our interpretation of CO and show the efficacy of the proposed mitigation strategies.

cs.LG↗
arXiv2026

Poisoning with A Pill: Circumventing Detection in Federated Learning

Without direct access to the client's data, federated learning (FL) is well-known for its unique strength in data privacy protection among existing distributed machine learning techniques. However, its distributive and iterative nature makes FL inherently vulnerable to various poisoning attacks. To counteract these threats, extensive defenses have been proposed to filter out malicious clients, using various detection metrics. Based on our analysis of existing attacks and defenses, we find that there is a lack of attention to model redundancy. In neural networks, various model parameters contribute differently to the model's performance. However, existing attacks in FL manipulate all the model update parameters with the same strategy, making them easily detectable by common defenses. Meanwhile, the defenses also tend to analyze the overall statistical features of the entire model updates, leaving room for sophisticated attacks. Based on these observations, this paper proposes a generic and attack-agnostic augmentation approach designed to enhance the effectiveness and stealthiness of existing FL poisoning attacks against detection in FL, pointing out the inherent flaws of existing defenses and exposing the necessity of fine-grained FL security. Specifically, we employ a three-stage methodology that strategically constructs, generates, and injects poison (generated by existing attacks) into a pill (a tiny subnet with a novel structure) during the FL training, named as pill construction, pill poisoning, and pill injection accordingly. Extensive experimental results show that FL poisoning attacks enhanced by our method can bypass all the popular defenses, and can gain an up to 7x error rate increase, as well as on average a more than 2x error rate increase on both IID and non-IID data, in both cross-silo and cross-device FL systems.

cs.LG↗
arXiv2026

Towards Identification and Intervention of Safety-Critical Parameters in Large Language Models

Ensuring Large Language Model (LLM) safety is crucial, yet the lack of a clear understanding about safety mechanisms hinders the development of precise and reliable methodologies for safety intervention across diverse tasks. To better understand and control LLM safety, we propose the Expected Safety Impact (ESI) framework for quantifying how different parameters affect LLM safety. Based on ESI, we reveal distinct safety-critical patterns across different LLM architectures: In dense LLMs, many safety-critical parameters are located in value matrices (V) and MLPs in middle layers, whereas in Mixture-of-Experts (MoE) models, they shift to the late-layer MLPs. Leveraging ESI, we further introduce two targeted intervention paradigms for safety enhancement and preservation, i.e., Safety Enhancement Tuning (SET) and Safety Preserving Adaptation (SPA). SET can align unsafe LLMs by updating only a few safety-critical parameters, effectively enhancing safety while preserving original performance. SPA safeguards well-aligned LLMs during capability-oriented intervention (e.g., instruction tuning) by preventing disruption of safety-critical weights, allowing the LLM to acquire new abilities and maintain safety capabilities. Extensive evaluations on different LLMs demonstrate that SET can reduce the attack success rates of unaligned LLMs by over 50% with only a 100-iteration update on 1% of model weights. SPA can limit the safety degradation of aligned LLMs within 1% after a 1,000-iteration instruction fine-tuning on different tasks. Our code is available at: https://github.com/ZJU-LLM-Safety/SafeWeights-ACL.

cs.CR↗
arXiv2026

HarmMetric Eval: Benchmarking Metrics and Judges for LLM Harmfulness Assessment

The potential of large language models (LLMs) to generate harmful content poses a significant safety risk for data management, as LLMs are increasingly being used as engines for data generation. To assess this risk, numerous harmfulness evaluation metrics and judges have been proposed. However, due to differences in their formats and scales, these metrics may yield inconsistent evaluation results on LLM-generated harmful data, undermining their credibility in practice. To address this gap, we present HarmMetric Eval, a systematic benchmark for assessing the quality of harmfulness metrics and judges with varying formats and scales. HarmMetric Eval includes a high-quality dataset comprising representative harmful prompts paired with harmful and non-harmful LLM outputs across multiple fine-grained categories, along with a unified scoring mechanism to reward the metrics for correctly ranking harmful outputs over non-harmful ones. Extensive experiments on HarmMetric Eval yield a surprising finding: conventional reference-based metrics such as ROUGE and METEOR can outperform LLM-based judges in fine-grained harmfulness evaluation, challenging prevailing assumptions about LLMs' superiority in this domain. To reveal the reasons behind this finding, we provide a fine-grained analysis to explain the limitations of LLM-based judges on rating irrelevant or useless LLM outputs. Motivated by these insights, we design an improved harmfulness judge that explicitly incorporates fine-grained harmfulness criteria in its prompt template and leverages reference-based metrics for lightweight fine-tuning of its base LLM. The resulting judge achieves state-of-the-art evaluation effectiveness on HarmMetric Eval.

cs.CL↗
arXiv2026

MAGIC: A Co-Evolving Attacker-Defender Adversarial Game for Robust LLM Safety

Ensuring robust safety alignment is crucial for Large Language Models (LLMs), yet existing defenses often lag behind evolving adversarial attacks due to their \textbf{reliance on static, pre-collected data distributions}. In this paper, we introduce \textbf{MAGIC}, a novel multi-turn multi-agent reinforcement learning framework that formulates LLM safety alignment as an adversarial asymmetric game. Specifically, an attacker agent learns to iteratively rewrite original queries into deceptive prompts, while a defender agent simultaneously optimizes its policy to recognize and refuse such inputs. This dynamic process triggers a \textbf{co-evolution}, where the attacker's ever-changing strategies continuously uncover long-tail vulnerabilities, driving the defender to generalize to unseen attack patterns. Remarkably, we observe that the attacker, endowed with initial reasoning ability, evolves \textbf{novel, previously unseen combinatorial strategies} through iterative RL training, underscoring our method's substantial potential. Theoretically, we provide insights into a more robust game equilibrium and derive safety guarantees. Extensive experiments validate our framework's effectiveness, demonstrating superior defense success rates without compromising the helpfulness of the model. Our code is available at https://github.com/BattleWen/MAGIC.

cs.AI↗
arXiv2026

Attack-Resistant Watermarking for AIGC Image Forensics via Diffusion-based Semantic Deflection

Protecting the copyright of user-generated AI images is an emerging challenge as AIGC becomes pervasive in creative workflows. Existing watermarking methods (1) remain vulnerable to real-world adversarial threats, often forced to trade off between defenses against spoofing and removal attacks; and (2) cannot support semantic-level tamper localization. We introduce PAI, a training-free inherent watermarking framework for AIGC copyright protection, plug-and-play with diffusion-based AIGC services. PAI simultaneously provides three key functionalities: robust ownership verification, attack detection, and semantic-level tampering localization. Unlike existing inherent watermark methods that only embed watermarks at noise initialization of diffusion models, we design a novel key-conditioned deflection mechanism that subtly steers the denoising trajectory according to the user key. Such trajectory-level coupling further strengthens the semantic entanglement of identity and content, thereby further enhancing robustness against real-world threats. Moreover, we also provide a theoretical analysis proving that only the valid key can pass verification. Experiments across 12 attack methods show that PAI achieves 98.43\% verification accuracy, improving over SOTA methods by 37.25\% on average, and retains strong tampering localization performance even against advanced AIGC edits. Our code is available at https://github.com/QingyuLiu/PAI.

cs.CR↗
arXiv2025

DualBreach: Efficient Dual-Jailbreaking via Target-Driven Initialization and Multi-Target Optimization

Recent research has focused on exploring the vulnerabilities of Large Language Models (LLMs), aiming to elicit harmful and/or sensitive content from LLMs. However, due to the insufficient research on dual-jailbreaking -- attacks targeting both LLMs and Guardrails, the effectiveness of existing attacks is limited when attempting to bypass safety-aligned LLMs shielded by guardrails. Therefore, in this paper, we propose DualBreach, a target-driven framework for dual-jailbreaking. DualBreach employs a Target-driven Initialization (TDI) strategy to dynamically construct initial prompts, combined with a Multi-Target Optimization (MTO) method that utilizes approximate gradients to jointly adapt the prompts across guardrails and LLMs, which can simultaneously save the number of queries and achieve a high dual-jailbreaking success rate. For black-box guardrails, DualBreach either employs a powerful open-sourced guardrail or imitates the target black-box guardrail by training a proxy model, to incorporate guardrails into the MTO process. We demonstrate the effectiveness of DualBreach in dual-jailbreaking scenarios through extensive evaluation on several widely-used datasets. Experimental results indicate that DualBreach outperforms state-of-the-art methods with fewer queries, achieving significantly higher success rates across all settings. More specifically, DualBreach achieves an average dual-jailbreaking success rate of 93.67% against GPT-4 with Llama-Guard-3 protection, whereas the best success rate achieved by other methods is 88.33%. Moreover, DualBreach only uses an average of 1.77 queries per successful dual-jailbreak, outperforming other state-of-the-art methods. For the purpose of defense, we propose an XGBoost-based ensemble defensive mechanism named EGuard, which integrates the strengths of multiple guardrails, demonstrating superior performance compared with Llama-Guard-3.

cs.CR↗
arXiv2025

MAJIC: Markovian Adaptive Jailbreaking via Iterative Composition of Diverse Innovative Strategies

Large Language Models (LLMs) have exhibited remarkable capabilities but remain vulnerable to jailbreaking attacks, which can elicit harmful content from the models by manipulating the input prompts. Existing black-box jailbreaking techniques primarily rely on static prompts crafted with a single, non-adaptive strategy, or employ rigid combinations of several underperforming attack methods, which limits their adaptability and generalization. To address these limitations, we propose MAJIC, a Markovian adaptive jailbreaking framework that attacks black-box LLMs by iteratively combining diverse innovative disguise strategies. MAJIC first establishes a ``Disguise Strategy Pool'' by refining existing strategies and introducing several innovative approaches. To further improve the attack performance and efficiency, MAJIC formulate the sequential selection and fusion of strategies in the pool as a Markov chain. Under this formulation, MAJIC initializes and employs a Markov matrix to guide the strategy composition, where transition probabilities between strategies are dynamically adapted based on attack outcomes, thereby enabling MAJIC to learn and discover effective attack pathways tailored to the target model. Our empirical results demonstrate that MAJIC significantly outperforms existing jailbreak methods on prominent models such as GPT-4o and Gemini-2.0-flash, achieving over 90\% attack success rate with fewer than 15 queries per attempt on average.

cs.CR↗
arXiv2025

Nearly Optimal Differentially Private ReLU Regression

In this paper, we investigate one of the most fundamental nonconvex learning problems, ReLU regression, in the Differential Privacy (DP) model. Previous studies on private ReLU regression heavily rely on stringent assumptions, such as constant bounded norms for feature vectors and labels. We relax these assumptions to a more standard setting, where data can be i.i.d. sampled from $O(1)$-sub-Gaussian distributions. We first show that when $\varepsilon = \tilde{O}(\sqrt{\frac{1}{N}})$ and there is some public data, it is possible to achieve an upper bound of $\tilde{O}(\frac{d^2}{N^2 \varepsilon^2})$ for the excess population risk in $(ε, δ)$-DP, where $d$ is the dimension and $N$ is the number of data samples. Moreover, we relax the requirement of $ε$ and public data by proposing and analyzing a one-pass mini-batch Generalized Linear Model Perceptron algorithm (DP-MBGLMtron). Additionally, using the tracing attack argument technique, we demonstrate that the minimax rate of the estimation error for $(\varepsilon, δ)$-DP algorithms is lower bounded by $Ω(\frac{d^2}{N^2 \varepsilon^2})$. This shows that DP-MBGLMtron achieves the optimal utility bound up to logarithmic factors. Experiments further support our theoretical results.

cs.LG↗
arXiv2025

Understanding and Mitigating Cross-lingual Privacy Leakage via Language-specific and Universal Privacy Neurons

Large Language Models (LLMs) trained on massive data capture rich information embedded in the training data. However, this also introduces the risk of privacy leakage, particularly involving personally identifiable information (PII). Although previous studies have shown that this risk can be mitigated through methods such as privacy neurons, they all assume that both the (sensitive) training data and user queries are in English. We show that they cannot defend against the privacy leakage in cross-lingual contexts: even if the training data is exclusively in one language, these (private) models may still reveal private information when queried in another language. In this work, we first investigate the information flow of cross-lingual privacy leakage to give a better understanding. We find that LLMs process private information in the middle layers, where representations are largely shared across languages. The risk of leakage peaks when converted to a language-specific space in later layers. Based on this, we identify privacy-universal neurons and language-specific privacy neurons. Privacy-universal neurons influence privacy leakage across all languages, while language-specific privacy neurons are only related to specific languages. By deactivating these neurons, the cross-lingual privacy leakage risk is reduced by 23.3%-31.6%.

cs.CL↗
arXiv2025

Pre-trained Encoder Inference: Revealing Upstream Encoders In Downstream Machine Learning Services

Pre-trained encoders available online have been widely adopted to build downstream machine learning (ML) services, but various attacks against these encoders also post security and privacy threats toward such a downstream ML service paradigm. We unveil a new vulnerability: the Pre-trained Encoder Inference (PEI) attack, which can extract sensitive encoder information from a targeted downstream ML service that can then be used to promote other ML attacks against the targeted service. By only providing API accesses to a targeted downstream service and a set of candidate encoders, the PEI attack can successfully infer which encoder is secretly used by the targeted service based on candidate ones. Compared with existing encoder attacks, which mainly target encoders on the upstream side, the PEI attack can compromise encoders even after they have been deployed and hidden in downstream ML services, which makes it a more realistic threat. We empirically verify the effectiveness of the PEI attack on vision encoders. we first conduct PEI attacks against two downstream services (i.e., image classification and multimodal generation), and then show how PEI attacks can facilitate other ML attacks (i.e., model stealing attacks vs. image classification models and adversarial attacks vs. multimodal generative models). Our results call for new security and privacy considerations when deploying encoders in downstream services. The code is available at https://github.com/fshp971/encoder-inference.

cs.LG↗
arXiv2025

Contextual Bandits for Unbounded Context Distributions

Nonparametric contextual bandit is an important model of sequential decision making problems. Under $α$-Tsybakov margin condition, existing research has established a regret bound of $\tilde{O}\left(T^{1-\frac{α+1}{d+2}}\right)$ for bounded supports. However, the optimal regret with unbounded contexts has not been analyzed. The challenge of solving contextual bandit problems with unbounded support is to achieve both exploration-exploitation tradeoff and bias-variance tradeoff simultaneously. In this paper, we solve the nonparametric contextual bandit problem with unbounded contexts. We propose two nearest neighbor methods combined with UCB exploration. The first method uses a fixed $k$. Our analysis shows that this method achieves minimax optimal regret under a weak margin condition and relatively light-tailed context distributions. The second method uses adaptive $k$. By a proper data-driven selection of $k$, this method achieves an expected regret of $\tilde{O}\left(T^{1-\frac{(α+1)β}{α+(d+2)β}}+T^{1-β}\right)$, in which $β$ is a parameter describing the tail strength. This bound matches the minimax lower bound up to logarithm factors, indicating that the second method is approximately optimal.

cs.LG↗
arXiv2024

AdvAnchor: Enhancing Diffusion Model Unlearning with Adversarial Anchors

Security concerns surrounding text-to-image diffusion models have driven researchers to unlearn inappropriate concepts through fine-tuning. Recent fine-tuning methods typically align the prediction distributions of unsafe prompts with those of predefined text anchors. However, these techniques exhibit a considerable performance trade-off between eliminating undesirable concepts and preserving other concepts. In this paper, we systematically analyze the impact of diverse text anchors on unlearning performance. Guided by this analysis, we propose AdvAnchor, a novel approach that generates adversarial anchors to alleviate the trade-off issue. These adversarial anchors are crafted to closely resemble the embeddings of undesirable concepts to maintain overall model performance, while selectively excluding defining attributes of these concepts for effective erasure. Extensive experiments demonstrate that AdvAnchor outperforms state-of-the-art methods. Our code is publicly available at https://anonymous.4open.science/r/AdvAnchor.

cs.LG↗
arXiv2024

Faithful Interpretation for Graph Neural Networks

Currently, attention mechanisms have garnered increasing attention in Graph Neural Networks (GNNs), such as Graph Attention Networks (GATs) and Graph Transformers (GTs). It is not only due to the commendable boost in performance they offer but also its capacity to provide a more lucid rationale for model behaviors, which are often viewed as inscrutable. However, Attention-based GNNs have demonstrated instability in interpretability when subjected to various sources of perturbations during both training and testing phases, including factors like additional edges or nodes. In this paper, we propose a solution to this problem by introducing a novel notion called Faithful Graph Attention-based Interpretation (FGAI). In particular, FGAI has four crucial properties regarding stability and sensitivity to interpretation and final output distribution. Built upon this notion, we propose an efficient methodology for obtaining FGAI, which can be viewed as an ad hoc modification to the canonical Attention-based GNNs. To validate our proposed solution, we introduce two novel metrics tailored for graph interpretation assessment. Experimental results demonstrate that FGAI exhibits superior stability and preserves the interpretability of attention under various forms of perturbations and randomness, which makes FGAI a more faithful and reliable explanation tool.

cs.LG↗