A Frequency Domain Adversarial Attack in Medical Image Analysis System

Abstract Deep neural networks (CNNs) have gained popularity in medical image analysis tasks, such as cancer diagnosis and lesion detection. However, recent research has revealed that medical deep learning systems are vulnerable to adversarial examples. In this paper, We propose a novel attack method from a spatial-domain perspective, named TextureDrop. The method crafts adversarial examples by dropping the high-frequency components (HFC) of medical images, thus making the medical image classification model misclassified. Additionally, we have developed a training method for medical neural networks. Extensive experiments have demonstrated the effectiveness of our method, which is not only more destructive but also stealthy. Furthermore, our approach introduces a new assessment framework for evaluating the robustness of medical networks from a frequency-domain perspective.