English

Double Visual Defense: Adversarial Pre-training and Instruction Tuning for Improving Vision-Language Model Robustness

Computer Vision and Pattern Recognition 2025-04-09 v2

Abstract

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\DeltaCLIP and Δ2\Delta^2LLaVA, 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\DeltaCLIP surpasses that of the previous best models on ImageNet-1k by ~20%. %For example, Δ\DeltaCLIP surpasses the previous best models on ImageNet-1k by ~20% in terms of adversarial robustness. Similarly, compared to prior art, Δ2\Delta^2LLaVA 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/.

Keywords

Cite

@article{arxiv.2501.09446,
  title  = {Double Visual Defense: Adversarial Pre-training and Instruction Tuning for Improving Vision-Language Model Robustness},
  author = {Zeyu Wang and Cihang Xie and Brian Bartoldson and Bhavya Kailkhura},
  journal= {arXiv preprint arXiv:2501.09446},
  year   = {2025}
}
R2 v1 2026-06-28T21:08:11.581Z