PaperManta Sign in
Active · from 2025 onwards (ongoing) · Shared by Marek

Adversarial robustness of vision-language models

Adversarial robustness of vision-language models

Want to keep this search?

Sign in to save it to your profile and mark papers as good, seen or bookmarked.

Sign in to save
322 results
  • TOPIC
    This paper investigates visual jailbreak attacks designed to bypass safety alignment in Multimodal Large Language Models (MLLMs).
    NOVELTY
    It introduces the Jailbreak Connectivity (JC) framework, which optimizes adversarial images as connected low-loss regions using quadratic Bézier curves rather than isolated points.
    INSIGHT
    The key takeaway is that targeting these connected manifolds significantly improves attack transferability and the linguistic fluency of the harmful outputs.
    RELEVANCE
    Since you are focused on the adversarial robustness of vision-language models, this work is directly relevant as it exposes structural vulnerabilities in MLLM safety guards across both open-source and proprietary models.
    Found Oct 30
  • EMNLP 2026

    Same Answer, Different Representations: Hidden instability in VLMs

    Farooq Ahmad Wani; Alessandro Suglia; Rohit Saxena; Aryo Pradipta Gema; Wai-Chung Kwan; Fazl Barez; Maria Sofia Bucarelli; Fabrizio Silvestri; Pasquale Minervini

    95 Match
    TOPIC
    This paper examines the adversarial robustness of Vision-Language Models (VLMs) by analyzing the gap between output stability and internal representation consistency.
    NOVELTY
    It introduces a white-box diagnostic framework using Dirichlet energy and Cohen's d to quantify latent state drift under semantically-invariant visual perturbations.
    INSIGHT
    The key takeaway is that stable predictions often mask significant internal representation drift, meaning models can arrive at the correct answer despite their latent states being pushed into unstable regions.
    RELEVANCE
    This is directly applicable to your work on VLM robustness as it provides a method to detect hidden vulnerabilities that standard accuracy-based benchmarks miss.
    Found Oct 28
  • EMNLP 2026

    DIVA: Exploiting Cross-Step Conditional Propagation for Visual Jailbreaks in Discrete Diffusion Vision-Language Models

    Guorui Song; Runqing Tang; Jingye Zhang; Luyuan Zhang; Feice Huang; Cong Ray; Guocun Wang; Dake Zhong; Choo Sin Wai; Bingquan Dai; Chuming Wang; Tongxu Lin; Wanyu Guo; Haoqian Wang

    95 Match
    TOPIC
    This paper investigates visual jailbreaking vulnerabilities specifically within discrete diffusion vision-language models (dVLMs).
    NOVELTY
    It introduces DIVA, a white-box framework that optimizes adversarial perturbations across multiple denoising timesteps to bypass safety filters.
    INSIGHT
    The key insight is that dVLMs are more susceptible to visual attacks than autoregressive models because the visual embedding is repeatedly propagated throughout the iterative generation process.
    RELEVANCE
    Since you focus on the adversarial robustness of VLMs, this work is critical as it identifies a unique structural vulnerability in diffusion-based architectures that differs from standard ARMs.
    Found Oct 28
  • EMNLP 2026

    MAViQ: Memory-Augmented Visual Querying for Multi-Turn MLLM Jailbreaks

    JinYun Lu; Xiaodan Li; Guanghao Zhou; Cen Chen; YueFeng Chen

    95 Match
    TOPIC
    This paper investigates multi-turn jailbreaking attacks specifically designed to bypass safety guardrails in Multimodal Large Language Models.
    NOVELTY
    It introduces MAViQ, a framework that treats attacks as a trace-to-query realization process, using a memory-augmented allocator to dynamically split harmful intent between text and images.
    INSIGHT
    The key takeaway is that adaptive, multi-turn visual grounding is significantly more effective at breaking high-capability closed-source models than static prompt-based attacks.
    RELEVANCE
    Since you are focused on the adversarial robustness of vision-language models, this work provides a direct benchmark for how current MLLMs fail under sophisticated, iterative multimodal pressure.
    Found Oct 26
  • ICML 2026

    You Don't Protect if You Don't Expect: Breaking the Key Assumption behind CLIP's Test-Time Defenses

    Ruize Zhang; Yu Li; Zhang Wan; Juan Cao; Jie Zhang; Sheng Tang

    95 Match
    TOPIC
    This paper investigates the adversarial robustness of zero-shot classification in vision-language models, specifically focusing on CLIP.
    NOVELTY
    The authors introduce CLIP-MAD, an adaptive attack that bypasses test-time defenses by manipulating the indicative measurements used to detect adversarial perturbations.
    INSIGHT
    It reveals that many current defenses for VLMs are fragile because they rely on easily spoofable distributional traits rather than true robustness.
    RELEVANCE
    Since you are working on the adversarial robustness of VLMs, this is a critical read as it exposes systemic vulnerabilities in state-of-the-art test-time defense mechanisms.
    Found Oct 23
  • TOPIC
    This paper examines the physical adversarial robustness of Vision-Language Models (VLMs) compared to traditional object detectors in person detection tasks.
    NOVELTY
    They develop a white-box optimization pipeline using Expectation Over Transformation (EOT) to create printed adversarial patches that evade or misdirect these models.
    INSIGHT
    Crucially, the authors find that VLMs are just as vulnerable as detectors, but their failure modes differ based on output format, with free-text VLMs tending toward dangerous 'invisible' failures like displacement and substitution.
    RELEVANCE
    Since you are focusing on VLM adversarial robustness, this work is essential as it provides a unified evaluation framework and highlights how distribution shifts exacerbate these security risks in real-world deployments.
    Found Oct 23
  • ICML 2026

    On the Adversarial Robustness of Large Vision-Language Models under Visual Token Compression

    Xinwei Zhang; Hangcheng Liu; Li Bai; Hao Wang; Qingqing Ye; Tianwei Zhang; Haibo Hu

    95 Match
    TOPIC
    This paper examines the adversarial robustness of Large Vision-Language Models (LVLMs) that utilize visual token compression to reduce computational costs.
    NOVELTY
    They introduce CAGE, a gray-box attack that uses Expected Feature Disruption and Rank Distortion Alignment to ensure adversarial perturbations survive the token pruning or merging process.
    INSIGHT
    The key takeaway is that standard attacks are too optimistic because they ignore the compression pipeline, which can actually concentrate distortions on the most informative tokens.
    RELEVANCE
    Since you're focused on the robustness of VLMs, this is a critical read as it highlights a major vulnerability in efficient multimodal architectures that current benchmarks likely miss.
    Found Oct 21
  • ICML 2026

    Localize and Neutralize: Gradient-Guided Token Suppression Against Visual Prompt Injection Attack

    Dongpeng Zhang; Ke Ma; Yangbangyan Jiang; Gaozheng Pei; Longtao Huang; Qianqian Xu; Qingming Huang

    95 Match
    TOPIC
    This paper investigates the vulnerability of Large Vision-Language Models (LVLMs) to perturbation-based visual prompt injection and multimodal jailbreak attacks.
    NOVELTY
    The authors introduce Gradient Token Masking (GTM), an inference-time defense that identifies and masks a sparse subset of high-saliency image tokens using a representation-aware gradient norm metric.
    INSIGHT
    They demonstrate that adversarial influence is concentrated in a few critical tokens and that using hidden-state gradients is more effective for localization than traditional output-probability methods.
    RELEVANCE
    Since you are focusing on the adversarial robustness of vision-language models, this work provides a direct method for neutralizing visual triggers without sacrificing benign model utility.
    Found Oct 21
  • ArXiv 2026

    Pushing the Frontier of Black-Box LVLM Attacks via Fine-Grained Detail Targeting

    Xiaohan Zhao; Zhaoyi Li; Yaxin Luo; Jiacheng Cui; Zhiqiang Shen

    95 Match
    TOPIC
    This paper investigates black-box transfer-based adversarial attacks on Large Vision-Language Models (LVLMs) to induce targeted semantic alignment.
    NOVELTY
    They introduce M-Attack-V2, a framework that uses multi-crop alignment and auxiliary target anchors to stabilize gradients and improve transferability across diverse ViT encoders.
    INSIGHT
    The key takeaway is that ViT-based LVLMs exhibit significant gradient instability due to translation sensitivity, which can be mitigated by decoupling source-side view aggregation from target-side semantic expansion.
    RELEVANCE
    Since you are focused on the adversarial robustness of VLMs, this work is critical as it demonstrates that even frontier models like GPT-o3 remain vulnerable to targeted hallucinations despite their advanced reasoning capabilities.
    Found Oct 21
  • TOPIC
    This paper investigates targeted black-box adversarial attacks on closed-source Multimodal Large Language Models (MLLMs) to achieve semantic hijacking.
    NOVELTY
    They introduce DART, which uses a GNN-based router for surrogate model selection and a distribution-aware alignment method using Energy Distance to prevent overfitting to specific surrogate features.
    INSIGHT
    The key takeaway is that aligning the latent semantic distribution via data augmentations significantly boosts attack transferability to commercial models like GPT-4o and Claude-3.7.
    RELEVANCE
    Since you are focused on the adversarial robustness of vision-language models, this work is critical as it demonstrates how vulnerabilities in visual encoders persist even in the most advanced reasoning-capable MLLMs.
    Found Oct 20
  • ICML 2026

    Self-Calibrated Consistency can Fight Back for Adversarial Robustness in Vision-Language Models

    Jiaxiang Liu; Jiawei Du; Xiao Liu; Shangyang Li; Songchen Ma; Changshuo Wang; Prayag Tiwari; Mingkun Xu

    95 Match
    TOPIC
    This paper investigates test-time adversarial robustness for zero-shot image classification in Vision-Language Models like CLIP.
    NOVELTY
    It introduces Self-Calibrated Consistency (SCC), a plug-and-play defense that uses a cross-modal margin objective and multi-view aggregation to correct perturbed image embeddings.
    INSIGHT
    The key takeaway is that combining a soft semantic anchor with spatial consistency prevents the 'semantic drift' typically seen in test-time repair methods, significantly boosting robust accuracy without retraining.
    RELEVANCE
    Since it directly addresses adversarial perturbations in VLMs using a white-box threat model, this is a critical baseline for any work on VLM robustness.
    Found Oct 20
  • ICML 2026

    Calibrating Uncertainty for Zero-Shot Adversarial CLIP

    Wenjing Lu; Zerui Tao; Yuning Qiu; Dongping Zhang; Yang Yang; Qibin Zhao

    95 Match
    TOPIC
    This paper investigates zero-shot adversarial robustness (ZSAR) in vision-language models, specifically focusing on the overconfidence of CLIP under attack.
    NOVELTY
    They introduce the Uncertainty-Calibrated Adversarial fine-Tuning (UCAT) framework, which reparameterizes logits as Dirichlet distribution parameters to explicitly decompose aleatoric and epistemic uncertainty.
    INSIGHT
    The key takeaway is that aligning the entire Dirichlet distribution—rather than just softmax probabilities—prevents the collapse of predictive uncertainty during adversarial perturbations.
    RELEVANCE
    Since it directly addresses the reliability gap in CLIP's adversarial robustness through a novel calibration method, it is a critical reference for your work.
    Found Oct 19
  • ICML 2026

    MADA-Attack: Transferable Multi-modal Attention Distraction Adversarial Attack against Vision Language Models

    Zhihan Qin; Jiahao Chen; Chunyi Zhou; Yuwen Pu; Chunqiang Hu; Xiaolei Liu; Shouling Ji

    95 Match
    TOPIC
    This paper investigates the generation of transferable universal adversarial perturbations (UAPs) to compromise the security of CLIP-based vision-language models.
    NOVELTY
    It introduces MADA-Attack, which uses semantic token manipulation and fused embedding training to disrupt cross-modal alignment at the embedding level rather than relying on simple image-only noise.
    INSIGHT
    The key takeaway is that exploiting the layer-wise asymmetry between visual and textual tokens significantly boosts attack transferability across diverse tasks like VQA and image captioning.
    RELEVANCE
    Since you are focused on the adversarial robustness of VLMs, this work is critical as it demonstrates a high-success black-box attack that generalizes to closed-source models like GPT-4o and Gemini.
    Found Oct 19
  • ICML 2026

    Attacking Gray-Box Large Vision-Language Models with Adaptive SVD-Structured Adversarial Alignment

    Daizong Liu; Xiaowen Cai; Junhao Dong; Zhongliang Guo; Xiaoye Qu; Runwei Guan; Xiang Fang; Dengpan Ye

    95 Match
    TOPIC
    This paper investigates targeted adversarial attacks on Large Vision-Language Models (LVLMs) using a gray-box setting where only the vision encoder is accessible.
    NOVELTY
    The authors introduce a multi-level alignment strategy that uses SVD-structured projection for global semantics and Optimal Transport for fine-grained visual-textual alignment.
    INSIGHT
    Their approach demonstrates that projecting adversarial image embeddings onto a text-induced subspace significantly increases attack success rates and transferability to closed-source models like GPT-4o.
    RELEVANCE
    Since you are focused on the adversarial robustness of VLMs, this work is critical as it provides a highly effective baseline for testing model vulnerabilities against targeted textual prompts.
    Found Oct 19
  • TOPIC
    This paper investigates test-time adaptation (TTA) to improve the adversarial robustness of CLIP-based vision-language models.
    NOVELTY
    It introduces SS-TPT, which optimizes textual prompts at inference time using stability and suitability scores to weight the influence of different augmented image views.
    INSIGHT
    The key takeaway is that quantifying the quality of augmented views allows for a consistency loss that suppresses adversarial outliers without needing a large, computationally expensive ensemble of views.
    RELEVANCE
    Since it specifically tackles adversarial perturbations in VLMs and addresses the trade-off between robustness and throughput, it is directly applicable to your work.
    Found Oct 19
  • TOPIC
    This paper investigates adversarial robustness in Vision-Language Models (VLMs) through the lens of test-time adaptation for fine-grained classification.
    NOVELTY
    It introduces A-TPT, a framework that uses a robustified gradient-attention mechanism to identify semantic anchors and guide spatially-varying augmentations during test-time prompt tuning.
    INSIGHT
    The key takeaway is that protecting high-attention semantic regions while aggressively augmenting low-attention areas prevents the destruction of discriminative features during defense.
    RELEVANCE
    Since it specifically targets the adversarial robustness of CLIP-based models without requiring retraining, it directly aligns with your focus on VLM security.
    Found Oct 18
  • TOPIC
    This paper investigates adversarial transferability in Vision-Language Models (VLMs), focusing on creating multimodal perturbations that mislead black-box target models.
    NOVELTY
    It introduces Dirichlet Distributional Gradient Aggregation (DDGA), which replaces heuristic sampling with a learnable Dirichlet distribution to optimize the mixing weights of adversarial perturbations.
    INSIGHT
    The key takeaway is that explicitly modeling the perturbation distribution and utilizing its covariance for orthogonal exploration significantly increases the transferability of attacks across different VLM architectures.
    RELEVANCE
    Since you are focused on the adversarial robustness of VLMs, this work is critical as it demonstrates how perturbations optimized for retrieval tasks can effectively compromise generative captioning models and MLLMs like Qwen-VL.
    Found Oct 18
  • ICML 2026

    Robustifying Vision-Language Models via Test-Time Prompt Adaptation

    Xingyu Zhu; Huanshen Wu; Shuo Wang; Beier Zhu; Jiannan Ge; Jiaheng Zhang; Long Chen

    95 Match
    TOPIC
    This paper investigates improving the adversarial robustness of pre-trained Vision-Language Models (VLMs) like CLIP through Test-Time Adaptation (TTA).
    NOVELTY
    It introduces RITA, a framework that replaces standard point-to-point feature matching with distribution-to-distribution alignment using entropy-regularized Optimal Transport.
    INSIGHT
    The key takeaway is that aligning distributions of augmented views via the Wasserstein distance is significantly more resilient to adversarial perturbations than relying on sample-level confidence heuristics.
    RELEVANCE
    Since you are focused on VLM adversarial robustness, this is a critical read as it provides a plug-and-play method to mitigate white-box attacks without needing labeled training data.
    Found Oct 18
  • 95 Match
    TOPIC
    This paper investigates test-time adversarial robustness for large-scale Vision-Language Models, specifically targeting CLIP.
    NOVELTY
    It introduces Contrastive Spectral Rectification (CSR), a training-free defense that uses a Gaussian low-pass filter as a semantic anchor to actively push adversarial embeddings back toward the natural image manifold via a contrastive objective.
    INSIGHT
    The key takeaway is that CLIP's vulnerability stems from a spectral bias toward high-frequency components, which can be mitigated without retraining by optimizing a rectification perturbation at inference time.
    RELEVANCE
    Since it generalizes to LLaVA-based VQA and captioning, this provides a direct framework for improving the robustness of the multimodal encoders you are studying.
    Found Oct 14
  • ICML 2026

    Dissecting the Safety Circuit: Neuronal Intervention for Transferable Adversarial Attacks on VLMs

    Chunlong Xie; Kangjie Chen; Shangwei Guo; Shudong Zhang; Jiamou Liu; Tianwei Zhang; Tao Xiang

    95 Match
    TOPIC
    This paper investigates the generation of transferable black-box jailbreak attacks against Vision-Language Models (VLMs).
    NOVELTY
    It introduces the Safety Circuit Intervention Attack (SCIA), which uses linear probing to disentangle model-specific defensive circuits from model-agnostic transferable circuits.
    INSIGHT
    The key takeaway is that safety representations are localized in specific neuronal circuits, and surgically suppressing these while amplifying transferable ones significantly boosts attack success rates across diverse architectures.
    RELEVANCE
    Since you are focused on the adversarial robustness of VLMs, this work is critical as it demonstrates a mechanistic approach to bypassing safety alignments that generalizes even to proprietary models.
    Found Oct 12
  • Loading more results…