Search arXiv⌕ Search

arXiv subjects

Hangtao Zhang

Publications and source records attributed to Hangtao Zhang.

At least 19 recordsLinked to original sources
arXiv2026

Sleeping Secrets: How Fine-Tuning Reawakens Privacy Risks in Language Models

Beyond adapting Large Language Models (LLMs) to specialized applications, fine-tuning has recently been shown to recover private information that is no longer accessible through direct queries. Previous fine-tuning recovery attacks, however, require genuine private supervision drawn from the same distribution, i.e., the previous training dataset. We argue that such recovery remains possible without such impractical knowledge. We show that LLM-generated candidates can provide sufficient supervision to recover previously learned private associations. Based on this, we propose ReGap, a data-free attack that recovers private associations using task structure, filters them by answer-token likelihood, and updates the target model via low-rank adaptation. Specifically, ReGap requires neither target answers nor auxiliary genuine private supervision. Across six GPT-2, OPT, and Qwen3 models, ReGap improves target-association recovery by 6-21 percentage points over the post-training target model. Recovery remains substantial even when the adaptation identities are disjoint from all memorized and evaluation identities, with no exact target answers appearing in the generated or selected supervision. Moreover, the same trained adapters increase recovery from 42\% to 63\% on a previously exposed checkpoint, but produce no gain on a matched checkpoint that never encountered the targets. This contrast shows that adaptation alone is insufficient to explain the observed recovery and that prior target exposure strongly affects post-adaptation recoverability. Our findings highlight that routine model customization can reawaken latent privacy risks, warranting urgent attention from the academic and industrial communities.

cs.CR↗
arXiv2026

JevAdvBench: A Benchmark and Black-Box Attacks for Reinforcement Learning for Calibrated Decisions Models

Models trained with reinforcement learning for calibrated decisions (RLCD), such as Jev, answer a typed question about an input, the state, with a probability, a choice, or a score, and software acts on the answer without a person reading it. Their robustness has not been measured: adversarial benchmarks score what a model generates or executes, whereas a typed model generates nothing and returns a well-formed answer even when manipulated. Measurement is also hard, because identical requests can return different answers, most available labels come from the model itself, and the API preprocesses each request out of view. Our key idea is to score each attacked decision against the model's own clean decision rather than against labels, and to read it against the change caused by an identical re-run. Building on this, we introduce JevAdvBench, to our knowledge the first adversarial benchmark for RLCD models, with 812 typed questions over 66 scenarios, and a black-box attack suite of 9,744 single-edit variants that each edit one part of a request, with billed input tokens confirming that the edit reached the model. On jev-1.13.0, rewording stays within 1.2 percentage points of the re-run baseline, and fields outside the schema never reach the model. In contrast, one unverified opinion appended to the state flips 12.1% of decisions, statistically tied with the strongest injected command (10.1%), and pushes 38% of confident answers below the 0.8 confidence threshold that routes them to human review. Applications built on RLCD models should therefore treat the state as untrusted, argued input. Project website: https://JevAdvBench.github.io/JevAdvBench/

cs.CR↗
arXiv2026

ODPure: Backdoor Purification for Object Detection via Ensemble Corruption Consensus

With the development of applications like autonomous driving, object detection has gained significant attention, while also highlighting critical vulnerabilities like backdoor attacks that severely compromise model integrity. Specifically, such attacks involve altering the categories of objects (i.e., object misclassification), removing bounding boxes (i.e., object disappearance), or generating bounding box proposals for non-existent objects (i.e., object generation) when a predefined trigger is present in the input. Although backdoor defenses for image classification are well-established, the research for object detection remains comparatively underexplored. Existing defenses address these threats by scanning outputs or models for potential backdoors but require discarding either malicious data or models. This remedy fails to enable a continuous and accurate perceptual stream for the object detection pipeline. To address such limitations, we propose ODPure, a novel input-stage black-box defense for object detection, which is based on input purification that ensures stable perception flows. Tailored to the dense prediction nature of object detectors, our Corruption-Reconstruction-Selection (CRS) paradigm operates by neutralizing triggers through a diverse portfolio of corruptions to generate a massive pool of redundant proposals, then recovering fine-grained structural cues via generative priors, and finally employing voting to reach a consensus on the resulting detections. Comprehensive experiments demonstrate that our method provides robust defense against diverse backdoor attacks and trigger types while preserving baseline accuracy. Our code is available at https://github.com/Alex66366/ODPure.

cs.CV↗
arXiv2026

Safety in Self-Evolving Agents: A Survey

Large language models (LLMs) exhibit strong general capabilities, yet their parameters typically remain fixed after deployment, limiting learning from new interactions. In open-ended environments, this motivates self-evolving agents that continually update reusable state-including model parameters, memories, tool definitions, skills, and workflows-from data, feedback, and accumulated experience. This shift changes the safety problem: once experience becomes reusable state, past events become future causes, and information harmless in one context may later influence decisions with greater persistence, authority, or scope. Self-evolving agent safety therefore asks not only whether a response is aligned or an action authorized, but whether safety properties survive the accumulation, generalization, and cross-context reuse of locally useful experience. We introduce SAVER, a transition-centered framework in which Substrate locates reusable influence, Adaptation captures how it changes, Violation identifies compromised safety attributes, Exposure marks where failures become observable, and Response assesses containment, repair, or revocation. Our survey reveals that failures need not originate from harmful information: legitimate state can become unsafe when adaptation expands its persistence, authority, or scope beyond the conditions under which it was valid. Existing work provides comparatively strong evidence for admission, retrieval, activation, exposure, and local containment, but much less for descendant repair and evaluation after adaptation resumes. We therefore argue for longitudinal evaluation that traces unsafe influence to its originating transition, verifies repair across descendants, and tests whether it can re-emerge under continued evolution.

cs.CR↗
arXiv2026

Test-Time Backdoor Detection for Object Detection Models

Object detection models are vulnerable to backdoor attacks, where attackers poison a small subset of training samples by embedding a predefined trigger to manipulate prediction. Detecting poisoned samples (i.e., those containing triggers) at test time can prevent backdoor activation. However, unlike image classification tasks, the unique characteristics of object detection -- particularly its output of numerous objects -- pose fresh challenges for backdoor detection. The complex attack effects (e.g., "ghost" object emergence or "vanishing" object) further render current defenses fundamentally inadequate. To this end, we design TRAnsformation Consistency Evaluation (TRACE), a brand-new method for detecting poisoned samples at test time in object detection. Our journey begins with two intriguing observations: (1) poisoned samples exhibit significantly more consistent detection results than clean ones across varied backgrounds. (2) clean samples show higher detection consistency when introduced to different focal information. Based on these phenomena, TRACE applies foreground and background transformations to each test sample, then assesses transformation consistency by calculating the variance in objects confidences. TRACE achieves black-box, universal backdoor detection, with extensive experiments showing a 30% improvement in AUROC over state-of-the-art defenses and resistance to adaptive attacks.

cs.CV↗
arXiv2026

TYPO: Instruction-Dense Visual Jailbreaks against Commercial Closed-Source Image-Generation Models

Recent commercial image-generation models can generate high-quality images with readable text (e.g., posters, infographics, and manuals), attracting considerable attention. Yet we first show that this same capability also introduces a previously unreported safety vulnerability: these systems may refuse to generate harmful text directly, yet permit the same content when rendered as text within generated images, i.e., safety alignment does not reliably transfer from textual outputs to text embedded in images. In this paper, unlike existing visual jailbreaks against image-generation models, which primarily induce models to generate harmful visual objects or scenes, we introduce the concept of instruction-dense visual jailbreaks, in which image-generation models produce detailed, readable, and actionable harmful instructions within images. Such outputs can amplify harm because the rendered instructions can be readily read and widely spread. To instantiate this threat, we propose TYPO, a black-box framework that exploits this safety gap by automatically generating adversarial TYPOgraphy prompts, which covertly steer image-generation models to express harmful intent as highly legible, typographically structured text. Specifically, TYPO decomposes prompt generation into two channels: a textual channel for reframing the target intent, and a visual channel for specifying its presentation form. We formulate these two channels as a dual-channel textual-visual strategy space and optimize candidate strategy combinations through an adaptive combinatorial search. Extensive experiments across four commercial models (i.e., GPT-Image-2, Nano Banana Pro, Qwen-Image-2, and Seedream 5.0 Lite) show that TYPO substantially outperforms nine representative jailbreak attacks by 50.2% in ASR on average, while incurring an average query cost of only $0.04.

cs.CR↗
arXiv2026

GhostPrompt: Cross-Image Adversarial Prompt for Vision-Language Models

Vision-Language Models (VLMs) are known to be vulnerable to adversarial attacks, where subtle perturbations to images or texts induce erroneous outputs. However, most text-based attacks are adapted from language-model-centric methods, in which the visual input is fixed during optimization, resulting in adversarial prompts that are tied to specific images and thus limiting their attack effectiveness. To this end, we first introduce a new research perspective: cross-image transferability for adversarial prompts. We then propose GhostPrompt, an adversarial prompt that is optimized once and reused to steer VLM outputs toward attacker-specified responses across diverse images. GhostPrompt employs a joint optimization that distills image-invariant adversarial features into the prompt by "worst-case" generation. Specifically, it alternates between constructing hard visual conditions for the current prompt and updating the prompt to remain effective under these conditions. Extensive experiments on prevalent VLMs verify that \ourmethod achieves an improvement of over 30% in attack success rates compared to state-of-the-art (SoTA) baselines, while reducing computation time by ~70%. Our code is avalable at https://github.com/Ye-ze-yu/GhostPrompt.

cs.CR↗
arXiv2026

PVDetector: Detecting Prompt Injection Attacks on Purpose-Specific LLM Agents through Policy-Violation Concept Analysis

Large language models (LLMs) are increasingly deployed as purpose-specific agents to handle domain-specific tasks such as customer service and code generation. These agents are expected to comply with not only generic safety guardrails but also purpose-specific restrictions tailored to their designated roles. Such additional restrictions enlarge the attack surface, particularly to prompt injection (PI) attacks. To defend against such attacks, existing detection methods primarily rely on analyzing input-output patterns, yet yield limited effectiveness. To address this limitation, we turn to analyzing the hidden activation space and discover that LLMs inherently retain latent policy-violation (PV) concepts when prompted with requests beyond their designated purpose. Particularly, PV concepts capture the semantics of conflicts between user queries and predefined restrictions, implicitly reflecting LLMs' intrinsic awareness of recognizing policy violations. Building on this insight, we propose PVDetector, a training-free framework that detects PI attacks during LLM inference by measuring hidden-state alignment with PV concepts, which are derived offline from the contrastive pairs of policy-violating and policy-compliant prompts. Experiments across multiple LLMs and datasets show that PVDetector achieves <1\% false negative rate with minimal auxiliary overhead, consistently outperforming state-of-the-art methods. Our code is available at https://github.com/Claresigle/PVDetector .

cs.CR↗
arXiv2026

Dual-branch Robust Unlearnable Examples

Unlearnable examples (UEs) aim to compromise model training by injecting imperceptible perturbations to clean samples. However, existing UE schemes exhibit limited robustness against advanced defenses due to their heuristic design or narrowly scoped domain perturbations. To address this, we propose \texttt{DUNE}, a \underline{\textbf{D}}ual-branch \underline{\textbf{UN}}learnable \underline{\textbf{E}}nsemble perturbation optimization approach. Specifically, \texttt{DUNE} separately optimizes perturbations in the spatial and color domains to establish the mapping between perturbations and shift-induced labels. This design extends the perturbation domain to increase noise intensity for improving robustness and drives the models to learn perturbation-oriented features with degraded generalization, thereby achieving unlearnability. To strengthen \texttt{DUNE}'s performance, we further propose an unlearnability-enhancing ensemble strategy that aggregates diverse pre-trained models during the dual-branch optimization. Extensive experiments on benchmark datasets CIFAR-10 and ImageNet verify that \texttt{DUNE}'s robustness outperforms 12 SOTA UE schemes under 7 mainstream defenses, yielding a lower average test accuracy of 14.95% to 50.82%.

cs.CV↗
arXiv2026

BadRobot: Jailbreaking Embodied LLM Agents in the Physical World

Embodied AI represents systems where AI is integrated into physical entities. Large Language Model (LLM), which exhibits powerful language understanding abilities, has been extensively employed in embodied AI by facilitating sophisticated task planning. However, a critical safety issue remains overlooked: could these embodied LLMs perpetrate harmful behaviors? In response, we introduce BadRobot, a novel attack paradigm aiming to make embodied LLMs violate safety and ethical constraints through typical voice-based user-system interactions. Specifically, three vulnerabilities are exploited to achieve this type of attack: (i) manipulation of LLMs within robotic systems, (ii) misalignment between linguistic outputs and physical actions, and (iii) unintentional hazardous behaviors caused by world knowledge's flaws. Furthermore, we construct a benchmark of various malicious physical action queries to evaluate BadRobot's attack performance. Based on this benchmark, extensive experiments against existing prominent embodied LLM frameworks (e.g., Voxposer, Code as Policies, and ProgPrompt) demonstrate the effectiveness of our BadRobot. Our code is available at https://github.com/Rookie143/BadRobot.

cs.CY↗
arXiv2026

Defending Jailbreak Attacks on Large Language Models via Manifold Trajectory Kinetics

Jailbreak prompts can bypass alignment guardrails in large language models (LLMs) and elicit unsafe outputs, making reliable deployment-time detection critical. Prior detection approaches largely rely on a fixed metric space, e.g., raw inputs, gradients, or hidden features, in which benign and jailbreak prompts are linearly separable. We show this assumption breaks under (i) pseudo-malicious prompts that are benign by intent but contain safety-related keywords, and (ii) adaptive attacks that explicitly optimize against the deployed detector. To overcome this limitation, we shift our focus from identifying a universal metric space to analyzing the more robust neighborhood structure of the underlying data manifold. We present Manifold Trajectory Kinetics (MTK), which treats an LLM as a kinetic system transforming inputs into outputs and detects jailbreaks by tracking how a prompt's neighborhood structure evolves across layers. Benign prompts remain close to benign neighborhoods throughout inference, whereas jailbreak prompts exhibit a characteristic trajectory that begins near malicious seeds and later strategically shifts toward benign neighborhoods to evade refusal.Across four LLMs and ten jailbreak attacks, MTK achieves strong robustness to both failure modes: on pseudo-malicious prompts, it attains a jailbreak true positive rate of 95% at a false positive rate of 5% on benign prompts and 2% on pseudo-malicious prompts, and under adaptive attacks, it maintains a true positive rate of 85%. We further demonstrate the superior performance of MTK for jailbreak detection in vision-language models. Our code is available at https://github.com/Rookie143/mtk.

cs.CR↗
arXiv2026

Image-to-Video Diffusion: From Foundations to Open Frontiers

Diffusion-based \textit{image-to-video} (I2V) generation has become a central direction in generative models by turning a reference image, with optional conditions, into a temporally coherent video. Compared with broader video generation settings, this task places stricter demands on content consistency, identity preservation, and motion coherence. Although the literature grows rapidly, existing works mostly discuss I2V generation within broader topics and still lack a dedicated taxonomy together with a systematic analysis centered on this field. This work addresses that gap by treating diffusion I2V generation as a standalone subject. It first reviews the task formulation, model architectures, datasets, and evaluation metrics, and then organizes existing methods through a taxonomy based on architecture and training paradigm. It further distills four core designs, namely condition encoding, temporal modeling, noise prior design, and spatial-temporal upsampling, and discusses representative application scenarios together with major open challenges.

cs.CV↗
arXiv2026

Robot Collapse: Supply Chain Backdoor Attacks Against VLM-based Robotic Manipulation

Robotic manipulation policies are increasingly empowered by \textit{large language models} (LLMs) and \textit{vision-language models} (VLMs), leveraging their understanding and perception capabilities. Recently, inference-time attacks against robotic manipulation have been extensively studied, yet backdoor attacks targeting model supply chain security in robotic policies remain largely unexplored. To fill this gap, we propose \texttt{TrojanRobot}, a backdoor injection framework for model supply chain attack scenarios, which embeds a malicious module into modular robotic policies via backdoor relationships to manipulate the LLM-to-VLM pathway and compromise the system. Our vanilla design instantiates this module as a backdoor-finetuned VLM. To further enhance attack performance, we propose a prime scheme by introducing the concept of \textit{LVLM-as-a-backdoor}, which leverages \textit{in-context instruction learning} (ICIL) to steer \textit{large vision-language model} (LVLM) behavior through backdoored system prompts. Moreover, we develop three types of prime attacks, \textit{permutation}, \textit{stagnation}, and \textit{intentional}, achieving flexible backdoor attack effects. Extensive physical-world and simulator experiments on 18 real-world manipulation tasks and 4 VLMs verify the superiority of proposed \texttt{TrojanRobot}

cs.RO↗
arXiv2026

TSCAN: Context-Aware Uplift Modeling via Two-Stage Training for Online Merchant Business Diagnosis

A primary challenge in ITE estimation is sample selection bias. Traditional approaches utilize treatment regularization techniques such as the Integral Probability Metrics (IPM), re-weighting, and propensity score modeling to mitigate this bias. However, these regularizations may introduce undesirable information loss and limit the performance of the model. Furthermore, treatment effects vary across different external contexts, and the existing methods are insufficient in fully interacting with and utilizing these contextual features. To address these issues, we propose a Context-Aware uplift model based on the Two-Stage training approach (TSCAN), comprising CAN-U and CAN-D sub-models. In the first stage, we train an uplift model, called CAN-U, which includes the treatment regularizations of IPM and propensity score prediction, to generate a complete dataset with counterfactual uplift labels. In the second stage, we train a model named CAN-D, which utilizes an isotonic output layer to directly model uplift effects, thereby eliminating the reliance on the regularization components. CAN-D adaptively corrects the errors estimated by CAN-U through reinforcing the factual samples, while avoiding the negative impacts associated with the aforementioned regularizations. Additionally, we introduce a Context-Aware Attention Layer throughout the two-stage process to manage the interactions between treatment, merchant, and contextual features, thereby modeling the varying treatment effect in different contexts. We conduct extensive experiments on two real-world datasets to validate the effectiveness of TSCAN. Ultimately, the deployment of our model for real-world merchant diagnosis on one of China's largest online food ordering platforms validates its practical utility and impact.

cs.LG↗
arXiv2025

SegTrans: Transferable Adversarial Examples for Segmentation Models

Segmentation models exhibit significant vulnerability to adversarial examples in white-box settings, but existing adversarial attack methods often show poor transferability across different segmentation models. While some researchers have explored transfer-based adversarial attack (i.e., transfer attack) methods for segmentation models, the complex contextual dependencies within these models and the feature distribution gaps between surrogate and target models result in unsatisfactory transfer success rates. To address these issues, we propose SegTrans, a novel transfer attack framework that divides the input sample into multiple local regions and remaps their semantic information to generate diverse enhanced samples. These enhanced samples replace the original ones for perturbation optimization, thereby improving the transferability of adversarial examples across different segmentation models. Unlike existing methods, SegTrans only retains local semantic information from the original input, rather than using global semantic information to optimize perturbations. Extensive experiments on two benchmark datasets, PASCAL VOC and Cityscapes, four different segmentation models, and three backbone networks show that SegTrans significantly improves adversarial transfer success rates without introducing additional computational overhead. Compared to the current state-of-the-art methods, SegTrans achieves an average increase of 8.55% in transfer attack success rate and improves computational efficiency by more than 100%.

cs.CV↗
arXiv2025

DarkHash: A Data-Free Backdoor Attack Against Deep Hashing

Benefiting from its superior feature learning capabilities and efficiency, deep hashing has achieved remarkable success in large-scale image retrieval. Recent studies have demonstrated the vulnerability of deep hashing models to backdoor attacks. Although these studies have shown promising attack results, they rely on access to the training dataset to implant the backdoor. In the real world, obtaining such data (e.g., identity information) is often prohibited due to privacy protection and intellectual property concerns. Embedding backdoors into deep hashing models without access to the training data, while maintaining retrieval accuracy for the original task, presents a novel and challenging problem. In this paper, we propose DarkHash, the first data-free backdoor attack against deep hashing. Specifically, we design a novel shadow backdoor attack framework with dual-semantic guidance. It embeds backdoor functionality and maintains original retrieval accuracy by fine-tuning only specific layers of the victim model using a surrogate dataset. We consider leveraging the relationship between individual samples and their neighbors to enhance backdoor attacks during training. By designing a topological alignment loss, we optimize both individual and neighboring poisoned samples toward the target sample, further enhancing the attack capability. Experimental results on four image datasets, five model architectures, and two hashing methods demonstrate the high effectiveness of DarkHash, outperforming existing state-of-the-art backdoor attack methods. Defense experiments show that DarkHash can withstand existing mainstream backdoor defense methods.

cs.CV↗
arXiv2025

ADVEDM:Fine-grained Adversarial Attack against VLM-based Embodied Agents

Vision-Language Models (VLMs), with their strong reasoning and planning capabilities, are widely used in embodied decision-making (EDM) tasks in embodied agents, such as autonomous driving and robotic manipulation. Recent research has increasingly explored adversarial attacks on VLMs to reveal their vulnerabilities. However, these attacks either rely on overly strong assumptions, requiring full knowledge of the victim VLM, which is impractical for attacking VLM-based agents, or exhibit limited effectiveness. The latter stems from disrupting most semantic information in the image, which leads to a misalignment between the perception and the task context defined by system prompts. This inconsistency interrupts the VLM's reasoning process, resulting in invalid outputs that fail to affect interactions in the physical world. To this end, we propose a fine-grained adversarial attack framework, ADVEDM, which modifies the VLM's perception of only a few key objects while preserving the semantics of the remaining regions. This attack effectively reduces conflicts with the task context, making VLMs output valid but incorrect decisions and affecting the actions of agents, thus posing a more substantial safety threat in the physical world. We design two variants of based on this framework, ADVEDM-R and ADVEDM-A, which respectively remove the semantics of a specific object from the image and add the semantics of a new object into the image. The experimental results in both general scenarios and EDM tasks demonstrate fine-grained control and excellent attack performance.

cs.CV↗
arXiv2025

PB-UAP: Hybrid Universal Adversarial Attack For Image Segmentation

With the rapid advancement of deep learning, the model robustness has become a significant research hotspot, \ie, adversarial attacks on deep neural networks. Existing works primarily focus on image classification tasks, aiming to alter the model's predicted labels. Due to the output complexity and deeper network architectures, research on adversarial examples for segmentation models is still limited, particularly for universal adversarial perturbations. In this paper, we propose a novel universal adversarial attack method designed for segmentation models, which includes dual feature separation and low-frequency scattering modules. The two modules guide the training of adversarial examples in the pixel and frequency space, respectively. Experiments demonstrate that our method achieves high attack success rates surpassing the state-of-the-art methods, and exhibits strong transferability across different models.

cs.CV↗