Analysis on Adversarial Robustness of Deep Learning Model LeNet-5 Based on Data Perturbation

Yudi Liu, Minyan Lu, Di Peng, Jie Wang, Jun Ai

Abstract: At present, deep learning technology is widely used in daily life. From recommendation algorithms to autonomous driving, deep learning models play an important role. However, once these models face perturbation, especially in case of adversarial attack perturbation, depending on the situation, the wrong output of the model may cause adverse consequences, such as property damage or personal safety accidents. Therefore, the ability of the model to resist the perturbation of adversarial attacks, that is, adversarial robustness, remains as a problem worthy of attention. In the present study, a deep learning model based on the convolutional neural network LeNet-5 was used as the experimental object, and adversarial examples are formed by adversarial attacks on the input data of the model, in order to observe the changing law of the adversarial robustness of the deep learning model.