의료 이미지의 도메인 특성을 활용한 다중 작업 적대적 학습 기반 적대적 강건성 향상 방법

초록

While the performance of deep neural network-based medical diagnosis systems has significantly improved, they still fail to provide sufficient reliability from a security perspective. Particularly in the medical domain, adversarial vulnerabilities can directly lead to serious harm to human life and property, making the design of robust systems essential for preventing such consequences. This study aims to enhance the adversarial robustness of medical diagnosis systems by leveraging two characteristics of medical image: the ease of adversarial detection compared to natural image, and the frequent occurrence of class imbalance. The proposed method employs multi-task learning and adversarial training to simultaneously train the model on diagnosis, detection, and grouping tasks. Through experiments on three medical datasets, we demonstrate that the proposed method can effectively enhance the adversarial robustness of various medical diagnosis systems, and consequently significantly alleviate the accuracy gap and trade-off problem in adversarial training.

키워드