CVPR 2023 – Attribute-preserving Face Dataset Anonymization via Latent Code Optimization

In this episode we discuss Attribute-preserving Face Dataset Anonymization via Latent Code Optimization
by Simone Barattin, Christos Tzelepis, Ioannis Patras, Nicu Sebe. The paper presents a task-agnostic approach for anonymizing the identities of faces in a dataset of images while retaining the facial attributes necessary for downstream tasks. The proposed method optimizes the latent representation of images in the latent space of a pre-trained GAN, ensuring the desired distance between the original image and its anonymized version, with an identity obfuscation loss. A novel feature-matching loss is used to preserve facial attributes, and experiments show that the method better preserves these attributes compared to existing approaches. Code and pre-trained models are publicly available.