Controllable facial protection against malicious translation-based attribute editing

Yiyi Xie, Yuqian Zhou, Tao Wang, Zhongyun Hua, Wenying Wen, Shuang Yi, Yushu Zhang

Abstract: Benefiting from the rapid development of AI-generated content, face attribute editing has achieved realism that is indistinguishable from reality while meeting users’ demands for social sharing and personalization. However, it also triggers users’ concerns about arbitrary modifications to their facial images. Existing schemes effectively prevent facial images from being tampered with, but they fail to simultaneously accommodate nonmalicious attribute editing outputs. Herein, we propose a controllable facial protection scheme to counter malicious translation-based facial attribute editing models. Our scheme supports the editing of target attributes but prevents protected attributes from being tampered with. It also employs iteratively optimized adversarial perturbations to divert attribute editing. Target attribute edits can be ensured to be correctly output by the model, while outputs of other protected attribute edits cannot achieve the desired results. Furthermore, our scheme utilizes low-frequency information to control image content characteristics, thereby constraining the output of denied access attribute editing while also maintaining consistency in attribute classification with the original image. Extensive experiments validate the effectiveness of our scheme in controlling access, maintaining image quality, and controlling attribute classification.