딥러닝 기반 의료 영상 인공지능 모델의 취약성: 적대적 공격

Abstract

Due to rapid developments in the deep learning model, artificial intelligence (AI) models are expected to enhance clinical diagnostic ability and work efficiency by assisting physicians. Therefore, many hospitals and private companies are competing to develop AI-based automatic diagnostic systems using medical images. In the near future, many deep learning-based automatic diagnostic systems would be used clinically. However, the possibility of adversarial attacks exploiting certain vulnerabilities of the deep learning algorithm is a major obstacle to deploying deep learning-based systems in clinical practice. In this paper, we will examine in detail the kinds of principles and methods of adversarial attacks that can be made to deep learning models dealing with medical images, the problems that can arise, and the preventive measures that can be taken against them.