Influence of Control Parameters and the Size of Biomedical Image\n Datasets on the Success of Adversarial Attacks

Abstract

In this paper, we study dependence of the success rate of adversarial attacks\nto the Deep Neural Networks on the biomedical image type, control parameters,\nand image dataset size. With this work, we are going to contribute towards\naccumulation of experimental results on adversarial attacks for the community\ndealing with biomedical images. The white-box Projected Gradient Descent\nattacks were examined based on 8 classification tasks and 13 image datasets\ncontaining a total of 605,080 chest X-ray and 317,000 histology images of\nmalignant tumors. We concluded that: (1) An increase of the amplitude of\nperturbation in generating malicious adversarial images leads to a growth of\nthe fraction of successful attacks for the majority of image types examined in\nthis study. (2) Histology images tend to be less sensitive to the growth of\namplitude of adversarial perturbations. (3) Percentage of successful attacks is\ngrowing with an increase of the number of iterations of the algorithm of\ngenerating adversarial perturbations with an asymptotic stabilization. (4) It\nwas found that the success of attacks dropping dramatically when the original\nconfidence of predicting image class exceeds 0.95. (5) The expected dependence\nof the percentage of successful attacks on the size of image training set was\nnot confirmed.\n

Ähnliche Arbeiten