인공지능 모델 대상 백도어 공격을 응용한 저작권 보호 목적의 워터마킹 연구

With the advancement of artificial intelligence(AI) technology, backdoor attacks have emerged as one of various cybersecurity threats, causing damage in various AI application domains, including autonomous driving. In this paper, we propose a backdoor attack detection method specialized to ensemble model training environments, achieving quick detection by separating the processes of model construction and attack detection. Additionally, we propose a method to prevent data abuse and protect data copyrights by utilizing backdoors as an image watermarking technique.