Investigating Feature Leakage in Synthetic Faces

Attention to synthetic datasets as an alternative to authentic datasets has increased. The state-of-the-art generative models for those datasets, including IDiff-Face, UIFace, DCFace, and Arc2Face, can generate large numbers of fake faces, but it remains unclear whether these identities are related to the authentic training datasets. In this thesis, we examine whether these diffusion models leak identity-related information from their real training datasets. From the samples we manually created, we compare the real-real, real-generated, and similarity distributions between generated datasets and evaluate them. In addition, we examine another possible form of leakage: bias transfer. The experiments demonstrate that no identity leakage occurs, but the bias transfer from the training dataset is clearly visible in the generated samples.