Sheng Zhong

Multimodal Large Language Models (MLLMs) have achieved remarkable success in tasks such as image captioning, visual question answering, and cross-modal reasoning by integrating visual and textual modalities. However, their multimodal nature also exposes them to adversarial threats, where attackers can perturb either modality or both jointly to induce harmful, misleading, or policy violating outputs. Existing defense strategies, such as adversarial training and input purification, face notable limitations: adversarial training typically improves robustness only against known attacks while incurring high computational costs, whereas conventional purification approaches often suffer from degraded image quality and insufficient generalization to complex multimodal tasks. In this work, we focus on defending the visual modality, which frequently serves as the primary entry point for adversarial manipulation. We propose a supervised diffusion based denoising framework that leverages paired adversarial clean image datasets to fine-tune diffusion models with directional, task specific guidance. Unlike prior unsupervised purification methods such as DiffPure, our approach achieves higher quality reconstructions while significantly improving defense robustness in multimodal tasks. Furthermore, we incorporate prompt optimization as a complementary defense mechanism, enhancing resistance against diverse and unseen attack strategies. Extensive experiments on image captioning and visual question answering demonstrate that our method not only substantially improves robustness but also exhibits strong transferability to unknown adversarial attacks. These results highlight the effectiveness of supervised diffusion based denoising for multimodal defense, paving the way for more reliable and secure deployment of MLLMs in real world applications.

In the realm of black-box jailbreak attacks on large language models (LLMs), the feasibility of constructing a narrow safety proxy, a lightweight model designed to predict the attack success rate (ASR) of adversarial prompts, remains underexplored. This work investigates the distillability of an LLM's core security logic. We propose a novel framework that incorporates an improved outline filling attack to achieve dense sampling of the model's security boundaries. Furthermore, we introduce a ranking regression paradigm that replaces standard regression and trains the proxy model to predict which prompt yields a higher ASR. Experimental results show that our proxy model achieves an accuracy of 91.1 percent in predicting the relative ranking of average long response (ALR), and 69.2 percent in predicting ASR. These findings confirm the predictability and distillability of jailbreak behaviors, and demonstrate the potential of leveraging such distillability to optimize black-box attacks.