How Resilient Are Deep Learning Models in Medical Image Analysis? The Case of the Moment-Based Adversarial Attack...

In the past years, deep neural networks (DNNs) have become popular in many disciplines such as computer vision (CV). One of the most important challenges in the CV area is Medical Image Analysis (MIA). However, adversarial attacks (AdAs) have proven to be an important threat to vision systems by significantly reducing the performance of the models. This paper proposes a new black-box adversarial attack, which is based οn orthogonal image moments named Mb-AdA. Additionally, a corresponding defensive method of adversarial training using Mb-AdA adversarial examples is also investigated, with encouraging results. The proposed attack was applied in classification and segmentation tasks with six state-of-the-art Deep Learning (DL) models in X-ray, histopathology and nuclei cell images. The main advantage of Mb-AdA is that it does not destroy the structure of images like other attacks, as instead of adding noise it removes specific image information, which is critical for medical models' decisions. The proposed attack is more effective than compared ones and achieved degradation up to 65% and 18% in terms of accuracy and IoU for classification and segmentation tasks, respectively, by also presenting relatively high SSIM. At the same time, it was proved that Mb-AdA adversarial examples can enhance the robustness of the model.