Why not using clean samples during training?

[coordxyz]

Hi, Does anyone can help me to understand why not using clean samples during training? Will it reduce the performance? Thanks~

[leslierice1]

You could try TRADES, which is a commonly used loss function aimed at reducing the trade-off between robustness and accuracy (using clean and adversarial images), and compare the performance. Training on clean examples during training can sometimes reduce the final model's PGD test accuracy.