Double Visual Defense: Adversarial Pre-training and Instruction Tuning for Improving Vision-Language Model Robustness
It addresses the vulnerability of vision-language models to adversarial attacks, which is a critical issue for applications in AI safety and reliability, representing a strong specific gain rather than a foundational breakthrough.
This paper tackles the problem of improving the robustness of vision-language models against adversarial visual perturbations by introducing a 'double visual defense' method, resulting in models that achieve state-of-the-art adversarial defense with improvements of ~20% on ImageNet-1k and ~20-30% on tasks like image captioning and visual question answering.
This paper investigates the robustness of vision-language models against adversarial visual perturbations and introduces a novel ``double visual defense" to enhance this robustness. Unlike previous approaches that resort to lightweight adversarial fine-tuning of a pre-trained CLIP model, we perform large-scale adversarial vision-language pre-training from scratch using web-scale data. We then strengthen the defense by incorporating adversarial visual instruction tuning. The resulting models from each stage, $Δ$CLIP and $Δ^2$LLaVA, show substantially enhanced zero-shot robustness and set a new state-of-the-art in adversarial defense for vision-language models. For example, the adversarial robustness of $Δ$CLIP surpasses that of the previous best models on ImageNet-1k by ~20%. %For example, $Δ$CLIP surpasses the previous best models on ImageNet-1k by ~20% in terms of adversarial robustness. Similarly, compared to prior art, $Δ^2$LLaVA brings a ~30% robustness improvement to image captioning task and a ~20% robustness improvement to visual question answering task. Furthermore, our models exhibit stronger zero-shot recognition capability, fewer hallucinations, and superior reasoning performance compared to baselines. Our project page is https://doublevisualdefense.github.io/.