Abstract
The availability of publicly accessible datasets with identifiable facial images is essential for various research and development purposes. However, this wide access also underscores an increasing need for robust privacy protection. Regulations like the General Data Protection Regulation (GDPR) impose strict requirements for safeguarding personal data, yet effectively anonymizing facial images while maintaining their research utility remains a significant challenge. Most current methods use either latent space or feature space for face anonymization. Each has its strengths, but relying on just one can result in incomplete anonymization or less realistic images. This paper introduces a new two-stage anonymization method that combines Deep Face Latent space Anonymization (FLA) with Face Feature space Anonymization (FFA) guided by semantic masks. In the first stage, FLA hides identity by modifying the latent space of facial images. In the second stage, FFA uses semantic masks to preserve important facial features like expressions and head poses while still hiding identity. Extensive experimentation on the CelebA-HQ and LFW datasets demonstrates that our approach achieves strong identity obfuscation while preserving facial attributes, as evidenced by quantitative metrics. Our pipeline generates high-quality facial images that protect identities while preserving non-identifying features of the original images, ensuring the utility of the anonymized images.