1 paper · 1 filter
Chenlong Wei, Hanling Wang, Chong Zhang +3
Adversarial training is often plagued by "pseudo-attacks", perturbations that significantly increase the loss value but fail to truly alter the model's confidence structure. To add…