Face-DeID-Net: Generative Face De-Identification with Identity Removal and Attribute Preservation for Latent Diffusion Model Training
Yilin Zeng, Miao Zhang, Xin Hu · 2025
To address the challenge of privacy leakage in training data for personalized portrait generation using Latent Diffusion Model, we propose Face-DeID-Net, a novel end-to-end approach for generative face de-identification. This method is designed not only to preserve privacy by removing facial identity features but also to retain essential non-identity attributes of the face, ensuring the portraits remain suitable for Latent Diffusion Model training tasks. Specifically, Face-DeID-Net leverages a multimodal fine-grained approach that extracts detailed identity information from facial regions and applies an accurate deidentification loss to progressively remove this information during training. Importantly, the method ensures that nonidentity facial attributes, which are crucial for model training, are preserved. Experimental results on the MyStyle dataset demonstrate that Face-DeID-Net outperforms existing face deidentification methods, achieving superior functional quality while effectively safeguarding privacy.