인공지능 개인정보 보호 기술과 개인정보 보호 법제의 과제

The current personal information protection legislation faces the challenge of accommodating social demands that advocate for safeguarding the right to self-determination of personal information, while permitting the use of personal information for the purpose of artificial intelligence (AI) training. Although measures such as anonymization and pseudonymization have been implemented to tackle this issue, critics argue that these approaches are inadequate in safeguarding personal information and excessively restrictive towards data usage, diminishing their utility. As an alternative, new technical approaches that transcend these de-identification techniques have been gaining attention. This study classifies these novel technical approaches into four types that strengthen personal information protection while enabling AI training and inference: (1) data modification approaches (differential privacy), (2) data minimization approaches (federated learning), (3) data encryption approaches (homomorphic encryption), and (4) virtual data generation approaches (synthetic data). It then evaluates the significance and limitations of each approach. Based on this analysis, this study proposes that personal information protection legislation should incorporate these new technologies and evolve in a manner that advances both personal information protection levels and data utility