Search

As deep learning advances, Large Language Models (LLMs) and their multimodal counterparts, Multimodal Large Language Models (MLLMs), have shown exceptional performance in many real-world tasks. However, MLLMs face significant security challenges, such as jailbreak attacks, where attackers attempt to bypass the model's safety alignment to elicit harmful responses. The threat of jailbreak attacks on MLLMs arises from both the inherent vulnerabilities of LLMs and the multiple information channels that MLLMs process. While various attacks and defenses have been proposed, there is a notable gap in unified and comprehensive evaluations, as each method is evaluated on different dataset and metrics, making it impossible to compare the effectiveness of each method. To address this gap, we introduce \textit{MMJ-Bench}, a unified pipeline for evaluating jailbreak attacks and defense techniques for MLLMs. Through extensive experiments, we assess the effectiveness of various attack methods against SoTA MLLMs and evaluate the impact of defense mechanisms on both defense effectiveness and model utility for normal tasks. Our comprehensive evaluation contribute to the field by offering a unified and systematic evaluation framework and the first public-available benchmark for MLLM jailbreak research. We also demonstrate several insightful findings that highlights directions for future studies.

Randomized smoothing is the primary certified robustness method for accessing the robustness of deep learning models to adversarial perturbations in the l2-norm, by adding isotropic Gaussian noise to the input image and returning the majority votes over the base classifier. Theoretically, it provides a certified norm bound, ensuring predictions of adversarial examples are stable within this bound. A notable constraint limiting widespread adoption is the necessity to retrain base models entirely from scratch to attain a robust version. This is because the base model fails to learn the noise-augmented data distribution to give an accurate vote. One intuitive way to overcome this challenge is to involve a custom-trained denoiser to eliminate the noise. However, this approach is inefficient and sub-optimal. Inspired by recent large model training procedures, we explore an alternative way named PEFTSmoothing to adapt the base model to learn the Gaussian noise-augmented data with Parameter-Efficient Fine-Tuning (PEFT) methods in both white-box and black-box settings. Extensive results demonstrate the effectiveness and efficiency of PEFTSmoothing, which allow us to certify over 98% accuracy for ViT on CIFAR-10, 20% higher than SoTA denoised smoothing, and over 61% accuracy on ImageNet which is 30% higher than CNN-based denoiser and comparable to the Diffusion-based denoiser.