Adversarial Robustness via Label-Smoothing

Morgane Goibert, Elvis Dohmatob

We study Label-Smoothing as a means for improving adversarial robustness of supervised deep-learning models. After establishing a thorough and unified framework, we propose several…