What question did this study set out to answer?

Enhancing Zero-shot Adversarial Robustness of Vision-Language Models with Training-free Adaptive Feature Movement

View Full Paper

Ask AI

Bookmark

Authors

Randomized trial shows improved zero-shot adversarial robustness in vision-language models, indicating a novel training-free approach.

This research aims to improve the zero-shot adversarial robustness of vision-language models without traditional training techniques.
Proposed a training-free approach using Gaussian noise to enhance zero-shot robustness.
Used adversarial examples with Gaussian noise as anchors to explore a path in the embedding space.
Developed an adaptive parameter adjustment method based on distances in the embedding space.
Achieved an average improvement of 10.83% in top-1 robust accuracy across 16 datasets.
Preserved original zero-shot generalization abilities of vision-language models.
Demonstrated effectiveness without the risks of overfitting common in traditional adversarial training methods.

Tong et al. (2026) studied this question.

View Full Paper

Ask AI

Bookmark