Break the Visual Perception: Adversarial Attacks Targeting Encoded Visual Tokens of Large Vision-Language Models

Yubo Wang, Chaohu Liu, yanqiuqu, Haoyu Cao, Deqiang Jiang, Linli Xu

Abstract: Large vision-language models (LVLMs) integrate visual information into large language models, showcasing remarkable multi-modal conversational capabilities. However, the visual modules introduces new challenges in terms of robustness for LVLMs, as attackers can craft adversarial images that are visually clean but may mislead the model to generate incorrect answers. In general, LVLMs rely on vision encoders to transform images into visual tokens, which are crucial for the language models to perceive image contents effectively. Therefore, we are curious about one question: Can LVLMs still generate correct responses when the encoded visual tokens are attacked and disrupting the visual information? To this end, we propose a non-targeted attack method referred to as VT-Attack (Visual Tokens Attack), which constructs adversarial examples from multiple perspectives, with the goal of comprehensively disrupting feature representations and inherent relationships as well as the semantic properties of visual tokens output by image encoders. Using only access to the image encoder in the proposed attack, the generated adversarial examples exhibit transferability across diverse LVLMs utilizing the same image encoder and generality across different tasks. Extensive experiments validate the superior attack performance of the VT-Attack over baseline methods, demonstrating its effectiveness in attacking LVLMs with image encoders, which in turn can provide guidance on the robustness of LVLMs, particularly in terms of the stability of the visual feature space.

Primary Subject Area: [Content] Vision and Language

Secondary Subject Area: [Generation] Social Aspects of Generative AI, [Generation] Multimedia Foundation Models

Relevance To Conference: The large multimodal models such as large vision-language models have garnered considerable attention owing to their remarkable multimodal interaction capabilities. Our work introduces a novel adversarial attack method against LVLMs, which achieves strong attack performance by extensively compromising the encoded visual information from multiple perspectives. And this reveals the vulnerability of large multimodal models. We hope that our work can provide guidance and assistance in enhancing their robustness.

Submission Number: 1886