Parameter-Interpolated Adversarial Training for Robust Image Classification

Parameter Interpolation Adversarial Training (PIAT) interpolates the model parameters of the previous round and the current round between each epoch, making decision boundary changes smoother and alleviating robustness oscillation and overfitting in adversarial training. The method also uses normalized mean squared error (NMSE) to align the relative magnitude of clean-sample and adversarial-sample logits, rather than the absolute magnitude. Experiments on multiple benchmark datasets show that PIAT can significantly improve the image classification robustness of CNNs and ViTs.