Abstract:

Aiming at the problem that the neural networks are easy to misclassify under the attack of adversarial examples in the task of image classification, which leads to the unreliability of deep learning models, this paper proposed a defense method against adversarial attacks based on neural architecture search. This method used reinforcement learning to model the search of defense network as the behavior of the agent. Through the definition of search space, the design of search strategy, and the evaluation of subnetwork performance, the search network can automatically obtain the best performance network to reconstruct adversarial images and restore them to natural images, achieving the purpose of defense against adversarial attacks. The experimental results show that the method can effectively reconstruct illegal examples, and make them lose aggressiveness, and consequently ensure the classification accuracy of the classifier.

Key words: neural architecture search, image classification, adversarial attack, deep learning