Robustification of Segmentation Models Against Adversarial Perturbations In Medical Imaging

This paper presents a novel yet efficient defense framework for segmentation\nmodels against adversarial attacks in medical imaging. In contrary to the\ndefense methods against adversarial attacks for classification models which\nwidely are investigated, such defense methods for segmentation models has been\nless explored. Our proposed method can be used for any deep learning models\nwithout revising the target deep learning models, as well as can be independent\nof adversarial attacks. Our framework consists of a frequency domain converter,\na detector, and a reformer. The frequency domain converter helps the detector\ndetects adversarial examples by using a frame domain of an image. The reformer\nhelps target models to predict more precisely. We have experiments to\nempirically show that our proposed method has a better performance compared to\nthe existing defense method.\n