Tianwei Zhang

Large vision-language models (LVLMs) have achieved impressive success across multimodal tasks, but their reliance on visual inputs exposes them to significant adversarial threats. Existing encoder-based attacks perturb the input image by optimizing solely on the vision encoder, rather than the entire LVLM, offering a computationally efficient alternative to end-to-end optimization. However, their transferability across different LVLM architectures in realistic black-box scenarios remains poorly understood. To address this gap, we present the first systematic study towards encoder-based adversarial transferability in LVLMs. Our contributions are threefold. First, through large-scale benchmarking over eight diverse LVLMs, we reveal that existing attacks exhibit severely limited transferability. Second, we perform in-depth analysis, disclosing two root causes that hinder the transferability: (1) inconsistent visual grounding across models, where different models focus their attention on distinct regions; (2) redundant semantic alignment within models, where a single object is dispersed across multiple overlapping token representations. Third, we propose Semantic-Guided Multimodal Attack (SGMA), a novel framework to enhance the transferability. Inspired by the discovered causes in our analysis, SGMA directs perturbations toward semantically critical regions and disrupts cross-modal grounding at both global and local levels. Extensive experiments across different victim models and tasks show that SGMA achieves higher transferability than existing attacks. These results expose critical security risks in LVLM deployment and underscore the urgent need for robust multimodal defenses.

Large Vision-Language Models (LVLMs) are increasingly deployed in real-world intelligent systems for perception and reasoning in open physical environments. While LVLMs are known to be vulnerable to prompt injection attacks, existing methods either require access to input channels or depend on knowledge of user queries, assumptions that rarely hold in practical deployments. We propose the first Physical Prompt Injection Attack (PPIA), a black-box, query-agnostic attack that embeds malicious typographic instructions into physical objects perceivable by the LVLM. PPIA requires no access to the model, its inputs, or internal pipeline, and operates solely through visual observation. It combines offline selection of highly recognizable and semantically effective visual prompts with strategic environment-aware placement guided by spatiotemporal attention, ensuring that the injected prompts are both perceivable and influential on model behavior. We evaluate PPIA across 10 state-of-the-art LVLMs in both simulated and real-world settings on tasks including visual question answering, planning, and navigation, PPIA achieves attack success rates up to 98%, with strong robustness under varying physical conditions such as distance, viewpoint, and illumination. Our code is publicly available at https://github.com/2023cghacker/Physical-Prompt-Injection-Attack.

Visual token compression is widely used to accelerate large vision-language models (LVLMs) by pruning or merging visual tokens, yet its adversarial robustness remains unexplored. We show that existing encoder-based attacks can substantially overestimate the robustness of compressed LVLMs, due to an optimization-inference mismatch: perturbations are optimized on the full-token representation, while inference is performed through a token-compression bottleneck. To address this gap, we propose the Compression-AliGnEd attack (CAGE), which aligns perturbation optimization with compression inference without assuming access to the deployed compression mechanism or its token budget. CAGE combines (i) expected feature disruption, which concentrates distortion on tokens likely to survive across plausible budgets, and (ii) rank distortion alignment, which actively aligns token distortions with rank scores to promote the retention of highly distorted evidence. Across diverse representative plug-and-play compression mechanisms and datasets, our results show that CAGE consistently achieves lower robust accuracy than the baseline. This work highlights that robustness assessments ignoring compression can be overly optimistic, calling for compression-aware security evaluation and defenses for efficient LVLMs.