An Adversarial Example Defense Algorithm for Intelligent Driving

Abstract

In terms of intelligent driving, the adversarial example of an attack against traffic signs will cause the vehicle to make wrong judgments and decisions. However, the existing adversarial examples of defense algorithms generally have problems such as high training costs and poor defense effects and struggle to adapt to the environment of intelligent driving. In order to reduce the training cost while improving the accuracy of example classification, we propose a novel defense algorithm for adversarial examples combining micro-network structure and a generative adversarial network (GAN). The algorithm compresses the classification model and the Discriminator. And the Generator is designed to make the reconstructed sample generated closer to the real example distribution so as to solve the common problems of the existing adversarial example defense algorithm. Experiments on the collected traffic sign data set show that the proposed algorithm can achieve a better defense effect on the premise of lower training costs. The example classification accuracy can reach more than 97.9%, and the similarity between the reconstructed samples and the real examples reaches 96.26%. Moreover, the number of computations and parameters for training a single example is far lower than that of other commonly used defense methods, and the response speed is approximately doubled, which can greatly improve the safety of intelligent driving.