Vulnerability Analysis of Chest X-Ray Image Classification Against\n Adversarial Attacks

Recently, there have been several successful deep learning approaches for\nautomatically classifying chest X-ray images into different disease categories.\nHowever, there is not yet a comprehensive vulnerability analysis of these\nmodels against the so-called adversarial perturbations/attacks, which makes\ndeep models more trustful in clinical practices. In this paper, we extensively\nanalyzed the performance of two state-of-the-art classification deep networks\non chest X-ray images. These two networks were attacked by three different\ncategories (ten methods in total) of adversarial methods (both white- and\nblack-box), namely gradient-based, score-based, and decision-based attacks.\nFurthermore, we modified the pooling operations in the two classification\nnetworks to measure their sensitivities against different attacks, on the\nspecific task of chest X-ray classification.\n