Presentation Information
[N-1-40]Identity-Keeping anonymization using Deepfake and VAE
◎△Ryo Yamatomi1, Hiroshi Ninomiya1 (1. Shonan Institute of Technology)
Keywords:
Identity-Keeping Anonymization,Deepfake,Variational Autoencoder
The rapid proliferation of social networking services (SNS) has escalated the risk of unintended facial image dissemination. Since traditional methods like pixelation damage facial expressions and image context, we propose a novel identity-keeping anonymization method combining GHOST, a Deepfake technology, with a Variational Autoencoder (VAE). The method transforms a target face into a non-existent person's face while preserving attributes like age and gender. GHOST's Attribute Encoder extracts the target's orientation and expression, while the VAE uses a scale factor to convert an ArcFace-derived vector into a non-existent face's feature vector. The Generator merges these outputs to produce the anonymized image. Setting the scale factor greater than 1 enhances anonymity while maintaining attributes. Evaluated on the UTKFace dataset, increasing the scale factor reduced cosine similarity. At a scale factor of 2, similarity dropped to 0.5734, ensuring high anonymity, while maintaining a 90.41% gender consistency and suppressing the age error (RMSE) to 7.8757 years. This demonstrates that our method alters local identification features, such as eye shapes, while highly preserving attributes, expressions, and context.