针对多种交通标识分类模型的攻击及鲁棒性分析

Attack and Robustness Analysis of Multiple Traffic Sign Classification Models

为了保证自动驾驶系统的安全性符合预期功能安全的要求,有必要对交通标识牌分类模型实施对抗攻击。基于此,采用两种基于图形学噪声的攻击方法对交通标识牌分类模型实施对抗攻击。同时,在4种交通标识牌分类模型上开展对抗攻击实验,通过分析普遍逃逸率(Universal Evasion Rate,UER)的值来分析此方法的可行性和神经网络的鲁棒性。

In order to ensure that the safety of the automated driving system meets the requirements of the expected functional safety,It is necessary to implement an adversarial attack on the traffic sign classification model.Based on this,two attack methods based on graphics noise are used to counter the traffic sign classification model.At the same time,carry out anti-attack experiments on four traffic sign classification models,and analyze the feasibility of this method and the robustness of neural network by analyzing the value of Universal Evasion Rate(UER).