Cross-modality Information Check for Detecting Jailbreaking in Multimodal Large Language Models
Yue Xu, Xiuyuan Qi, Zhan Qin, Wenjie Wang
TL;DR
CIDER is a plug‑and‑play detector that protects Multimodal LLMs from optimization-based jailbreaking by exploiting cross‑modal semantic shifts between text queries and perturbed images. It denoises the image modality and uses the relative change in text–image cosine similarity before and after denoising to decide whether an input is adversarial, enabling model-agnostic deployment with minimal inference overhead. The approach achieves strong detection and robustness gains across multiple MLLMs and attacks, with transferability to white-box and black-box settings, while incurring a manageable utility cost on normal tasks. Overall, CIDER advances practical safety for MLLMs by leveraging cross‑modal information and diffusion-based denoising to identify crafted perturbations at the input stage.
Abstract
Multimodal Large Language Models (MLLMs) extend the capacity of LLMs to understand multimodal information comprehensively, achieving remarkable performance in many vision-centric tasks. Despite that, recent studies have shown that these models are susceptible to jailbreak attacks, which refer to an exploitative technique where malicious users can break the safety alignment of the target model and generate misleading and harmful answers. This potential threat is caused by both the inherent vulnerabilities of LLM and the larger attack scope introduced by vision input. To enhance the security of MLLMs against jailbreak attacks, researchers have developed various defense techniques. However, these methods either require modifications to the model's internal structure or demand significant computational resources during the inference phase. Multimodal information is a double-edged sword. While it increases the risk of attacks, it also provides additional data that can enhance safeguards. Inspired by this, we propose Cross-modality Information DEtectoR (CIDER), a plug-and-play jailbreaking detector designed to identify maliciously perturbed image inputs, utilizing the cross-modal similarity between harmful queries and adversarial images. CIDER is independent of the target MLLMs and requires less computation cost. Extensive experimental results demonstrate the effectiveness and efficiency of CIDER, as well as its transferability to both white-box and black-box MLLMs.