A Survey of Attacks on Large Vision-Language Models: Resources, Advances, and Future Trends

With the significant development of large models in recent years, LargeVision-Language Models (LVLMs) have demonstrated remarkable capabilities acrossa wide range of multimodal understanding and reasoning tasks. Compared totraditional Large Language Models (LLMs), LVLMs present great potential andchallenges due to its closer proximity to the multi-resource real-worldapplications and the complexity of multi-modal processing. However, thevulnerability of LVLMs is relatively underexplored, posing potential securityrisks in daily usage. In this paper, we provide a comprehensive review of thevarious forms of existing LVLM attacks. Specifically, we first introduce thebackground of attacks targeting LVLMs, including the attack preliminary, attackchallenges, and attack resources. Then, we systematically review thedevelopment of LVLM attack methods, such as adversarial attacks that manipulatemodel outputs, jailbreak attacks that exploit model vulnerabilities forunauthorized actions, prompt injection attacks that engineer the prompt typeand pattern, and data poisoning that affects model training. Finally, wediscuss promising research directions in the future. We believe that our surveyprovides insights into the current landscape of LVLM vulnerabilities, inspiringmore researchers to explore and mitigate potential safety issues in LVLMdevelopments. The latest papers on LVLM attacks are continuously collected inhttps://github.com/liudaizong/Awesome-LVLM-Attack.

Further reading#

• Access Paper in arXiv.org