Calibrating Uncertainty for Zero-Shot Adversarial CLIP
Researchers have developed a new method to improve the reliability of CLIP, a model used for zero-shot image classification. The proposed technique addresses the issue where adversarial attacks not only reduce accuracy but also cause the model to become over-confident by suppressing uncertainty. By treating CLIP's outputs as parameters of a Dirichlet distribution, the method aligns the model's confidence with input difficulty, thereby restoring calibrated uncertainty and enhancing adversarial robustness while maintaining clean accuracy. AI
IMPACT Enhances the robustness and trustworthiness of vision-language models against adversarial manipulations.