arXiv subjects

Zhengyu Zhao

Publications and source records attributed to Zhengyu Zhao.

At least 19 recordsLinked to original sources

arXiv2026

Extracting System Prompt of LLM Services in Complex (Agentic) Settings: Hierarchical Optimization and Reliable Evaluation Metrics

System prompt is a critical intellectual property of large language model (LLM) services. To analyze its leakage, prompt extraction attacks have been extensively studied. However, we validate that existing attacks substantially degrade on recent LLM services, whose complexity has evolved in both the prompt itself and system-level constraints. Specifically, (1) system prompts become longer and more structured to contain pre-defined tool schemas, execution logic, safety guardrails, etc., and (2) only the final post-processed outputs (e.g., agent's actions) but not the raw textual response can be observed. In this paper, we address the above increased complexity by proposing Extraction with Hierarchical Optimization (ExtraHerO), which (1) decomposes the prompt extraction into a hierarchical reconstruction process with a pre-defined structural prompt template, and (2) iteratively updates the adversarial query with guidance of feedback-template differences to observe diverse outputs. In particular, we monitor the information gain of attack optimization to early stop or reset it for a good trade-off between effectiveness and efficiency. We also find that existing evaluation metrics for prompt fidelity can be misleading because they do not account for structural information. Therefore, we rely on global textual similarity and prioritize structural keywords. Moreover, we introduce a system-level metric, adversarial utility, to measure the help of prompt extraction for real-world downstream attacks. Experimental results demonstrate the consistent superiority of ExtraHerO, with average improvements of 12.5\% (fidelity), 2.3\% (functionality), and 6.0\% (adversarial utility) on chatbots, and 12.2\%, 34.3\%, and 30.5\% on agents.

cs.CR↗

arXiv2026

Visual Input and Its Framing Affect Undermine the Stability of Attribute-based Descriptions Produced by Large Vision-Language Models

Large vision-language models (LVLMs) are commonly prompted with text alone or together with an image. In this paper, we find a new aspect of LVLM vulnerability in its response robustness: adding an image can unexpectedly affect responses even when text prompts ask about a concept rather than the instance of the concept depicted in the image. For example, adding an image of an animal can change the generated attribute descriptions of its breed, and framing of the image further affects those changes. We study this new problem on seven LVLMs with image sets covering 30 animal breeds and eight food categories, using two types of framings that focus on the subject (subject-focused) or place it in a broader context (subject-in-situation). We draw the following main findings: (1) adding images increases physical-term proportions for animals (16.7\% $\to$ 32.8\%) and decreases sensory-term proportions for food (83.6\% $\to$ 73.2\%) among the most frequent terms, averaged over seven models and two framings on correctly recognized images. (2) attribute-term distributions differ significantly between the two framing types for 41.4\% of animal and 71.4\% of food model--concept pairs, with generally higher physical-term proportions for animals and sensory-term proportions for food under subject-focused than subject-in-situation framing.

cs.CV↗

arXiv2026

Answer It or Not? Towards Prompt-Adapted Adversarial Images on Vision Language Models

With the advancement in language modeling, research on adversarial images has recently shifted from image classifiers to vision-language models (VLMs). Despite the new, text-prompt-driven nature of VLMs, we find that existing VLM adversaries have largely overlooked the dynamics of text prompts. In a typical use case, adversarial images can protect visual privacy by rejecting VLMs from answering questions about the image. However, existing protections are prompt-agnostic, i.e., indiscriminately rejecting any question, including those for potential legitimate use of the image. To address this limitation, we propose Prompt-Adapted (ProAda) adversarial images, which are optimized to balance the rejection of harmful prompts and the correct response of legitimate prompts. In particular, to achieve a more accurate adversarial direction, the text prompt is also iteratively updated with a simple classifier judging whether the harmfulness of the prompt remains. Extensive experiments on visual-privacy-related datasets validate that ProAda can reject over 90% of harmful prompts while allowing over 90% of legitimate ones, demonstrating superior overall performance compared to existing prompt-agnostic VLM adversaries. In addition, our ProAda achieves competitive overall cross-model transferability and improved robustness against countermeasures.

cs.CV↗

arXiv2026

Detecting Object Hallucinations in Large Vision-Language Models via Cross-Modal Attention Drifts and Mask-Based Verification

Despite recent advances in large vision-language models (LVLMs), object hallucination remains a major barrier to their reliable deployment. Existing detection methods often characterize visual grounding using attention from individual layers, leaving its evolution across layers underexplored. We propose CADMP, a lightweight object hallucination detection framework that combines adjacent-layer cross-modal attention drift with prediction sensitivity to targeted visual masking. During decoding, CADMP quantifies distributional changes between consecutive cross-modal attention maps to capture abrupt transitions in visual grounding. It then selects the transition with the largest drift, locates the corresponding visually relevant regions, and measures the change in prediction probability after masking these regions. These two signals provide complementary evidence: attention drift characterizes the stability of internal visual grounding, while probability variation verifies whether a prediction truly depends on the identified visual evidence. A lightweight detector integrates both signals to identify hallucinated predictions. Experiments on multiple benchmarks and representative open-source LVLMs demonstrate that CADMP achieves consistently competitive detection performance. Ablation studies further confirm the complementary contributions of adjacent-layer drift modeling and mask-based grounding verification.

cs.CV↗

arXiv2026

Detecting Backdoors in Object Detection via Pre-NMS Prediction Distribution Shift

Object detection models deployed in safety-critical applications remain vulnerable to backdoor attacks that cause targeted misbehaviors when a hidden trigger is present. Existing detection methods either rely on trigger inversion or exploit architecture-specific assumptions, and critically, representative existing methods fail to generalize reliably to scene-level attacks, where a single trigger induces anomalous behavior across all objects in the scene simultaneously. We present DistScan, a backdoor detection framework based on a simple but previously unexploited observation: backdoor injection systematically shifts a model's pre-NMS prediction class distribution away from its training class frequencies, even on clean inputs without any trigger present. DistScan aggregates intermediate class predictions over a clean validation set and flags a model as backdoored if the resulting distribution deviates significantly from the training class frequencies, requiring no model weight access, no trigger knowledge, and no additional training. Extensive experiments on MS-COCO and PASCAL VOC across two architectures and three scene-level attack scenarios demonstrate that DistScan substantially outperforms existing methods, improving average detection accuracy over the best-performing applicable baseline by 27.32 percentage points.

cs.CV↗

arXiv2026

Concept Unlearning by Modeling Key Steps of Diffusion Process

Text-to-image diffusion models remain susceptible to generating undesirable or harmful content. Although concept unlearning mitigates this risk, existing methods struggle with a critical optimization dilemma: thorough semantic erasure frequently induces the catastrophic forgetting of unrelated generative capabilities. To overcome this challenge, we propose Key Step Concept Unlearning (KSCU). Serving as an integrated methodological refinement deeply motivated by information theory, KSCU explores the profound impact of step scheduling order and reveals that traditional randomized timestep sampling severely disrupts trajectory dependency. We demonstrate that indiscriminately targeting the entire diffusion process is inefficient, as the optimal step range for unlearning inherently varies across different concepts. Rather than globally fine-tuning all timesteps, KSCU explicitly integrates a sequential-scheduling-based Key Step Table, CFG-aware leakage compensation, and prompt augmentation to dynamically isolate optimization to a concept-specific active region. This localized strategy successfully eradicates the target concept while preventing the structural collapse caused by early-step over-optimization. Consequently, KSCU significantly reduces computational overhead and establishes a state-of-the-art trade-off between concept erasure and utility retention. Comprehensive evaluations demonstrate that KSCU consistently delivers superior performance across diverse unlearning tasks, including nudity, style, object classes, and mass instance concepts. For example, in nudity removal, KSCU yields a 96.5% unlearning accuracy alongside a state-of-the-art FID of 14.1.

cs.CV↗

arXiv2026

On Success and Simplicity: A Second Look at Transferable Vision-Language Attack Pipeline

Vision-Language Pre-training Models (VLPMs) are known to be vulnerable to adversarial attacks. Recent transferable attacks on VLPMs have followed a common pipeline with complicated loss functions or multi-stage text/image attacks. However, in this paper, we demonstrate that such a sophisticated attack pipeline can be simpler yet more successful. Specifically, we identify three previously overlooked issues caused by inappropriate cross-modal interactions and excessive operations. To address them, we propose the Simple Vision-Language Attack (SimVLA) pipeline, which observably improves transferability and efficiency. Experiments on four datasets and three downstream tasks validate the superiority of our pipeline. For instance, on Flickr30k text-image retrieval dataset, our SimVLA outperforms the SOTA baseline in R@1 transferability by 8.01\%-14.71\%, while consuming only about 35.73\% of the time and 46.26\% of the max VRAM. Overall, the superiority of our SimVLA highlights the importance of leveraging domain knowledge (e.g., our proposed cross-modal word identification), while blindly pursuing intricate operations (e.g, complex loss functions and redundant multi-stage designs) may even be harmful. We hope our SimVLA can serve as a simple yet effective backbone for future extensions. Code is available at https://github.com/RYC-98/SimVLA.

cs.CV↗

arXiv2026

When Safe Concepts Become Unsafe: Multi-Concept Compositional Vulnerabilities in Text-to-Image Models

Text-to-image (T2I) models are increasingly optimized for following user instructions faithfully. However, we find that this capability introduces a safety vulnerability we call Multi-Concept Compositional Unsafety (MCCU). MCCU occurs when multiple individually safe concepts, if combined in a single generation request, lead to harmful or sensitive visual outputs. Unlike prior jailbreak settings, MCCU does not rely on adversarial prompts, model access, or explicitly disallowed content. Instead, the risk emerges from how the model composes multiple safe visual concepts into a single scene. To systematically measure this threat, we build TwoHamsters, a large-scale evaluation framework consisting of 20k prompts, 51 curated concept pairs, and six risk categories. We evaluate 13 T2I models under a black-box setting. Our results show a clear conflict between instruction-following and safety: models that follow prompts more faithfully tend to produce more MCCU failures. For example, FLUX.1 achieves a 99.35% Unsafe Alignment Rate while only reaching a 1.57% MCCU Defense Rate. We further evaluate three representative defenses, including safety filtering, MCCU-specific detector fine-tuning, and concept erasure, all of which fail against unseen concept combinations. Our findings suggest that compositional reasoning in T2I models creates an attack surface that is not captured by existing safety mechanisms. We anticipate the release of TwoHamsters will catalyze community development of advanced generative defense mechanisms.

cs.CR↗

arXiv2026

Revealing the Impact of Visual Text Style on Attribute-based Descriptions Produced by Large Visual Language Models

When the visual style of text is considered, a wide variety can be observed in font, color, and size. However, when a word is read, its meaning is independent of the style in which it has been written or rendered. In this paper, we investigate whether, and how, the style in which a word is visualized in an image impacts the description that a Large Visual Language Model (LVLM) provides for the concept to which that word refers. Specifically, we investigate how functional text styles (readability-oriented, e.g., black sans-serif) versus decorative styles (display-oriented, e.g., colored cursive/script) affect LVLMs' descriptions of a concept in terms of the attributes of that concept. Our experiments study the situation in which the LVLM is able to correctly identify the concept referred to by a visual text, i.e., by a word or words rendered as an image, and in which the visual text style should not influence the attribute-based description that the LVLM produces. Our experimental results reveal that even when the concept is correctly identified, text style influences the model's attribute-based descriptions of the concept. Our findings demonstrate non-trivial style leakage from text style into semantic inference and motivate style-aware evaluation and mitigation for LVLM-based multimedia systems.

cs.CV↗

arXiv2026

PoInit-of-View: Poisoning Initialization of Views Transfers Across Multiple 3D Reconstruction Systems

Poisoning input views of 3D reconstruction systems has been recently studied. However, we identify that existing studies simply backpropagate adversarial gradients through the 3D reconstruction pipeline as a whole, without uncovering the new vulnerability rooted in specific modules of the 3D reconstruction pipeline. In this paper, we argue that the structure-from-motion (SfM) initialization, as the geometric core of many widely used reconstruction systems, can be targeted to achieve transferable poisoning effects across diverse 3D reconstruction systems. To this end, we propose PoInit-of-View, which optimizes adversarial perturbations to intentionally introduce cross-view gradient inconsistencies at projections of corresponding 3D points. These inconsistencies disrupt keypoint detection and feature matching, thereby corrupting pose estimation and triangulation within SfM, eventually resulting in low-quality rendered views. We also provide a theoretical analysis that connects cross-view inconsistency to correspondence collapse. Experimental results demonstrate the effectiveness of our PoInit-of-View on diverse 3D reconstruction systems and datasets, surpassing the single-view baseline by 25.1% in PSNR and 16.5% in SSIM in black-box transfer settings, such as 3DGS to NeRF.

cs.CV↗

arXiv2026

Pay Less Attention to Function Words for Free Robustness of Vision-Language Models

To address the trade-off between robustness and performance for robust VLM, we observe that function words could incur vulnerability of VLMs against cross-modal adversarial attacks, and propose Function-word De-Attention (FDA) accordingly to mitigate the impact of function words. Similar to differential amplifiers, our FDA calculates the original and the function-word cross-attention within attention heads, and differentially subtracts the latter from the former for more aligned and robust VLMs. Comprehensive experiments include 2 SOTA baselines under 6 different attacks on 2 downstream tasks, 3 datasets, and 3 models. Overall, our FDA yields an average 18/13/53% ASR drop with only 0.2/0.3/0.6% performance drops on the 3 tested models on retrieval, and a 90% ASR drop with a 0.3% performance gain on visual grounding. We demonstrate the scalability, generalization, and zero-shot performance of FDA experimentally, as well as in-depth ablation studies and analysis. Code is available at https://github.com/michaeltian108/FDA.

cs.LG↗

arXiv2026

Finetune Like You Pretrain: Boosting Zero-shot Adversarial Robustness in Vision-language Models

Despite their impressive zero-shot abilities, vision-language models such as CLIP have been shown to be susceptible to adversarial attacks. To enhance its adversarial robustness, recent studies finetune the pretrained vision encoder of CLIP with adversarial examples on a proxy dataset such as ImageNet by aligning adversarial images with correct class labels. However, these methods overlook the important roles of training data distributions and learning objectives, resulting in reduced zero-shot capabilities and limited transferability of robustness across domains and datasets. In this work, we propose a simple yet effective paradigm AdvFLYP, which follows the training recipe of CLIP's pretraining process when performing adversarial finetuning to the model. Specifically, AdvFLYP finetunes CLIP with adversarial images created based on image-text pairs collected from the web, and match them with their corresponding texts via a contrastive loss. To alleviate distortion of adversarial image embeddings of noisy web images, we further propose to regularise AdvFLYP by penalising deviation of adversarial image features. We show that logit- and feature-level regularisation terms benefit robustness and clean accuracy, respectively. Extensive experiments on 14 downstream datasets spanning various domains show the superiority of our paradigm over mainstream practices. Our code and model weights are released at https://github.com/Sxing2/AdvFLYP.

cs.CV↗

arXiv2026

Adversarial Video Promotion Against Text-to-Video Retrieval

Thanks to the development of cross-modal models, text-to-video retrieval (T2VR) is advancing rapidly, but its robustness remains largely unexamined. Existing attacks against T2VR are designed to push videos away from queries, i.e., suppressing the ranks of videos, while the attacks that pull videos towards selected queries, i.e., promoting the ranks of videos, remain largely unexplored. These attacks can be more impactful as attackers may gain more views/clicks for financial benefits and widespread (mis)information. To this end, we pioneer the first attack against T2VR to promote videos adversarially, dubbed the Video Promotion attack (ViPro). We further propose Modal Refinement (MoRe) to capture the finer-grained, intricate interaction between visual and textual modalities to enhance black-box transferability. Comprehensive experiments cover 2 existing baselines, 3 leading T2VR models, 3 prevailing datasets with over 10k videos, evaluated under 3 scenarios. All experiments are conducted in a multi-target setting to reflect realistic scenarios where attackers seek to promote the video regarding multiple queries simultaneously. We also evaluated our attacks for defences and imperceptibility. Overall, ViPro surpasses other baselines by over $30/10/4\%$ for white/grey/black-box settings on average. Our work highlights an overlooked vulnerability, provides a qualitative analysis on the upper/lower bound of our attacks, and offers insights into potential counterplays. Code will be publicly available at https://github.com/michaeltian108/ViPro.

cs.CV↗

arXiv2026

Robustness Over Time: Understanding Adversarial Examples' Effectiveness on Longitudinal Versions of Large Language Models

Large Language Models (LLMs) undergo continuous updates to improve user experience. However, prior research on the security and safety implications of LLMs has primarily focused on their specific versions, overlooking the impact of successive LLM updates. This prompts the need for a holistic understanding of the risks in these different versions of LLMs. To fill this gap, in this paper, we conduct a longitudinal study to examine the adversarial robustness -- specifically misclassification, jailbreak, and hallucination -- of three prominent LLM families: GPT, Llama, and Qwen. Our study reveals that LLM updates do not consistently improve adversarial robustness as expected. For instance, a later version of GPT-3.5 degrades regarding misclassification and hallucination despite its improved resilience against jailbreaks. GPT-4 and GPT-4o demonstrate (incrementally) higher robustness overall. Larger Llama and Qwen models do not uniformly exhibit improved robustness across all three aspects studied. In addition, larger model sizes do not necessarily yield improved robustness. Minor updates lacking substantial robustness improvements can exacerbate existing issues rather than resolve them. We hope our study can offer valuable insights into navigating model updates and informed decisions in model development and usage.

cs.CR↗

arXiv2025

Privacy on the Fly: A Predictive Adversarial Transformation Network for Mobile Sensor Data

Mobile motion sensors such as accelerometers and gyroscopes are now ubiquitously accessible by third-party apps via standard APIs. While enabling rich functionalities like activity recognition and step counting, this openness has also enabled unregulated inference of sensitive user traits, such as gender, age, and even identity, without user consent. Existing privacy-preserving techniques, such as GAN-based obfuscation or differential privacy, typically require access to the full input sequence, introducing latency that is incompatible with real-time scenarios. Worse, they tend to distort temporal and semantic patterns, degrading the utility of the data for benign tasks like activity recognition. To address these limitations, we propose the Predictive Adversarial Transformation Network (PATN), a real-time privacy-preserving framework that leverages historical signals to generate adversarial perturbations proactively. The perturbations are applied immediately upon data acquisition, enabling continuous protection without disrupting application functionality. Experiments on two datasets demonstrate that PATN substantially degrades the performance of privacy inference models, achieving Attack Success Rate (ASR) of 40.11% and 44.65% (reducing inference accuracy to near-random) and increasing the Equal Error Rate (EER) from 8.30% and 7.56% to 41.65% and 46.22%. On ASR, PATN outperforms baseline methods by 16.16% and 31.96%, respectively.

cs.CR↗

arXiv2025

Use as Many Surrogates as You Want: Selective Ensemble Attack to Unleash Transferability without Sacrificing Resource Efficiency

In surrogate ensemble attacks, using more surrogate models yields higher transferability but lower resource efficiency. This practical trade-off between transferability and efficiency has largely limited existing attacks despite many pre-trained models are easily accessible online. In this paper, we argue that such a trade-off is caused by an unnecessary common assumption, i.e., all models should be \textit{identical} across iterations. By lifting this assumption, we can use as many surrogates as we want to unleash transferability without sacrificing efficiency. Concretely, we propose Selective Ensemble Attack (SEA), which dynamically selects diverse models (from easily accessible pre-trained models) across iterations based on our new interpretation of decoupling within-iteration and cross-iteration model diversity. In this way, the number of within-iteration models is fixed for maintaining efficiency, while only cross-iteration model diversity is increased for higher transferability. Experiments on ImageNet demonstrate the superiority of SEA in various scenarios. For example, when dynamically selecting 4 from 20 accessible models, SEA yields 8.5% higher transferability than existing attacks under the same efficiency. The superiority of SEA also generalizes to real-world systems, such as commercial vision APIs and large vision-language models. Overall, SEA opens up the possibility of adaptively balancing transferability and efficiency according to specific resource requirements.

cs.CV↗

arXiv2025

A Survey of Defenses Against AI-Generated Visual Media: Detection,Disruption, and Authentication

Deep generative models have demonstrated impressive performance in various computer vision applications, including image synthesis, video generation, and medical analysis. Despite their significant advancements, these models may be used for malicious purposes, such as misinformation, deception, and copyright violation. In this paper, we provide a systematic and timely review of research efforts on defenses against AI-generated visual media, covering detection, disruption, and authentication. We review existing methods and summarize the mainstream defense-related tasks within a unified passive and proactive framework. Moreover, we survey the derivative tasks concerning the trustworthiness of defenses, such as their robustness and fairness. For each defense strategy, we formulate its general pipeline and propose a multidimensional taxonomy applicable across defense tasks, based on methodological strategies. Additionally, we summarize the commonly used evaluation datasets, criteria, and metrics. Finally, by analyzing the reviewed studies, we provide insights into current research challenges and suggest possible directions for future research.

cs.CV↗

arXiv2025

Revisiting Transferable Adversarial Images: Systemization, Evaluation, and New Insights

Transferable adversarial images raise critical security concerns for computer vision systems in real-world, black-box attack scenarios. Although many transfer attacks have been proposed, existing research lacks a systematic and comprehensive evaluation. In this paper, we systemize transfer attacks into five categories around the general machine learning pipeline and provide the first comprehensive evaluation, with 23 representative attacks against 11 representative defenses, including the recent, transfer-oriented defense and the real-world Google Cloud Vision. In particular, we identify two main problems of existing evaluations: (1) for attack transferability, lack of intra-category analyses with fair hyperparameter settings, and (2) for attack stealthiness, lack of diverse measures. Our evaluation results validate that these problems have indeed caused misleading conclusions and missing points, and addressing them leads to new, \textit{consensus-challenging} insights, such as (1) an early attack, DI, even outperforms all similar follow-up ones, (2) the state-of-the-art (white-box) defense, DiffPure, is even vulnerable to (black-box) transfer attacks, and (3) even under the same $L_p$ constraint, different attacks yield dramatically different stealthiness results regarding diverse imperceptibility metrics, finer-grained measures, and a user study. We hope that our analyses will serve as guidance on properly evaluating transferable adversarial images and advance the design of attacks and defenses. Code is available at https://github.com/ZhengyuZhao/TransferAttackEval.

cs.CR↗