An adversarial sample defense model based on computer attention mechanism

Authors:

Abstract
The traditional neural network can not realize the invariance of image rotation and distortion well, so the attacker can cheat the neural network by adding tiny disturbance on the image. If signposts are attacked, automatic driving will be misguided, leading to disastrous consequences. This paper proposes a new adversarial sample defense model based on attention mechanism, which used LSTM network to get the key features. Firstly, we describe our model which recognizes an object through enough main characteristics. It can discard large parts of the noise distributed throughout the image area. Secondly, we show recognition process of the model, receiving adversarial samples and then outputting the salient mask image which can be classified correctly by neural networks in deep learning. At last, we designed experiments to test model effect on the traffic sign and the general picture. The experimental results show that our defense model can strongly defend adversarial examples.

This publication has 13 references indexed in Scilit: