Adversarial Neural Pruning with Modified Latent Vulnerability utilizing Feature Robustness

Abstract

Today, vision recognition with Convolutional Neural Network (CNN) shows performance good enough to be employed in safety-critical autonomous driving. However, CNN models are vulnerable to adversarial attacks. To overcome this vulnerability, many methods have been studied Pruning is regarded as one of the effective methods to make the network more robust against adversarial attacks. In this paper, we introduce a new vulnerability loss to suppress the vulnerability better when the pruning is used to counter adversarial attacks. With this vulnerability suppression, we achieve up to 1.12% better accuracy against the adversarial examples compared to a previous study called ANP-VS.