Retention Score: Quantifying Jailbreak Risks for Vision Language Models
This addresses security risks for users of vision-language models by providing a certified robustness metric, though it is incremental as it builds on existing adversarial attack methods.
This paper tackles the problem of assessing vision-language models' vulnerability to jailbreak attacks by proposing a novel metric called Retention Score, which quantifies robustness through toxicity score margins and shows that most VLMs with visual components are less robust than plain VLMs, with specific findings like GPT4V's robustness being similar to Gemini's medium settings.
The emergence of Vision-Language Models (VLMs) is a significant advancement in integrating computer vision with Large Language Models (LLMs) to enhance multi-modal machine learning capabilities. However, this progress has also made VLMs vulnerable to sophisticated adversarial attacks, raising concerns about their reliability. The objective of this paper is to assess the resilience of VLMs against jailbreak attacks that can compromise model safety compliance and result in harmful outputs. To evaluate a VLM's ability to maintain its robustness against adversarial input perturbations, we propose a novel metric called the \textbf{Retention Score}. Retention Score is a multi-modal evaluation metric that includes Retention-I and Retention-T scores for quantifying jailbreak risks in visual and textual components of VLMs. Our process involves generating synthetic image-text pairs using a conditional diffusion model. These pairs are then predicted for toxicity score by a VLM alongside a toxicity judgment classifier. By calculating the margin in toxicity scores, we can quantify the robustness of the VLM in an attack-agnostic manner. Our work has four main contributions. First, we prove that Retention Score can serve as a certified robustness metric. Second, we demonstrate that most VLMs with visual components are less robust against jailbreak attacks than the corresponding plain VLMs. Additionally, we evaluate black-box VLM APIs and find that the security settings in Google Gemini significantly affect the score and robustness. Moreover, the robustness of GPT4V is similar to the medium settings of Gemini. Finally, our approach offers a time-efficient alternative to existing adversarial attack methods and provides consistent model robustness rankings when evaluated on VLMs including MiniGPT-4, InstructBLIP, and LLaVA.