변형과 적대적 학습에 강건한 얼굴 사진 보호 기법 연구

A Study on Face Photo Protection Robust to Transforms and Adversarial Training

초록

Publicly shared face photos can be collected and used to train facial recognition models without the owner's consent. Unlearnable examples counter this by adding imperceptible noise to photos before sharing. In practice, however, collected photos undergo transforms before training; face recognition in particular crops the face and resizes it. Moreover, the attacker may adopt adversarial training to erase the effect of the noise. We show that existing unlearnable examples designed to withstand adversarial training lose their protection after this transform alone. We then propose SafeFace, which confines the noise to the face region, generates it through a differentiable model of the attacker's transform, and gradually increases the adversarial perturbation budget assumed during generation. On 50 identities, SafeFace is the only method among the five compared that withstands both conditions, suppressing the attacker to 1.8% under the transform and 12.0% under adversarial training , where prior methods allow 33.2-42.2%

키워드

얼굴 프라이버시; AI 안전; 적대적 학습; 데이터 보호; Unlearnable Examples; Face Privacy; Adversarial Training; Data Protection; Robustness
제목
변형과 적대적 학습에 강건한 얼굴 사진 보호 기법 연구
제목 (타언어)
A Study on Face Photo Protection Robust to Transforms and Adversarial Training
저자
오병훈; 이재우
DOI
10.36623/KSMA.2026.88.3
발행일
2026-09
유형
Y
저널명
Korean Security Journal
호
88
페이지
51 ~ 70