Robustness Feature Adapter for Efficient Adversarial Training

Quanwei Wu, Jun Guo, Wei Wang +1

Adversarial training (AT) with projected gradient descent is the most popular method to improve model robustness under adversarial attacks. However, computational overheads become…