TIMA: Text-Image Mutual Awareness for Balancing Zero-Shot Adversarial Robustness and Generalization Ability
The episode discusses the paper TIMA, which improves CLIP models by balancing zero-shot adversarial robustness and generalization ability. The authors propose Text-Image Mutual Awareness, using tuning mechanisms to increase class separation in both text and image spaces. Results show TIMA achieves better clean accuracy and significantly higher robustness against strong attacks compared to previous methods.