Reproduce “FaceMAE: Privacy-Preserving Face Recognition via Masked Autoencoders” on OpenResearch

About this paper

Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Jiankang Deng, Xinchao Wang, Hakan Bilen, Yang You

Researchers developed FaceMAE, a framework employing masked autoencoders, to create privacy-preserving face datasets that effectively balance data utility and individual privacy. It achieved a 50% reduction in error rate for face recognition tasks compared to prior methods, alongside a 20% lower risk of membership inference attacks.

Key insights

• Masked Autoencoders can be successfully adapted to generate privacy-preserving data by focusing reconstruction on maintaining feature-space utility for downstream tasks, rather than just visual exactness.
• The Instance Relation Matching (IRM) module, by minimizing feature distribution gaps, proved crucial in bridging the performance gap between models trained on original versus privacy-preserved data.
• A high mask ratio (75%) provides a good balance, significantly reducing privacy leakage risk while retaining enough semantic information for robust face recognition model training.

Summary by alphaXiv.