Assessing the vulnerability of Ai-based solutions in histopathology of cancer

Abstract

In this paper, we experimentally study the robustness of the Convolutional Neural Networks (CNNs) to adversarial attacks in different scenarios of computerized disease diagnosis. In order to disclose practically-relevant solutions, we attempt to compare the final CNN vulnerability scores under the condition of the use of different kinds of adversarial attacks as well as defense methods. On all occasions, we attempt to compare the basic and the most advanced solutions being available in every direction of the inquiry. In order to achieve this, we investigate EfficientNet CNN as one of the most popular convolutional networks. Also, we study the following three types of adversarial attacks: the FGSM Attacks, the Carlini-Wagner attacks as well as the AutoAttacks. After that, we examined three types of adversarial defenses including Adversarial Training, High-Level Representation Guided Denoiser, and the MagNet. The experiments have been performed on medical images typically used for computerized disease diagnosis in oncology (the whole-slide histology)