Risk Susceptibility of Brain Tumor Classification to Adversarial Attacks

Abstract

Discovery of adversarial attacks on deep neural networks, have exposed the vulnerabilities of these networks, wherein they often entirely fail to classify the attack generated images. While deep learning networks have generated promising results in performing brain tumor classification, there has been no analysis of their susceptibility to adversarial attacks. Vulnerability to adversarial attacks can render the deep neural networks useless for practical medical application. In this paper, a study has been performed to determine extent of white box adversarial attacks on convolutional neural networks used for brain tumor classification. Three different adversarial attacks are implemented on the network, namely Noise generated, Fast Gradient Sign, and Virtual Adversarial Training methods. The performance of the network under these attacks is analyzed and discussed. Furthermore, in the paper it is shown how these networks perform when trained on the adversarial attack generated images, which could be a possible solution to prevent the failure of the classification networks against adversarial attacks.