Adversarial self-training for robustness and generalization

Zhuorong Li, Minghui Wu, Canghong Jin, Daiwei Yu, Hongchuan Yu

Abstract: Highlights•An adversarial training technique using self-training is proposed.•Consistency regularization is applied to suppress the distortion of representations in latent space.•The proposed technique can be easily generalized to other application scenarios or alternative training objectives.•Various adversaries have been employed for a broad security guarantee.•Results show the success of adversarial self-training to improve robust accuracy and generalization.