ID2image: Leakage of Non-ID Information into Face Descriptors and Inversion from Descriptors to Images

Embedding a face image to descriptor vector using deep CNN is widely used technique in recognition. Via several possible training strategies, such embeddings are supposed capture only identity information. Information about the environment (such as background and lighting) or changeable aspects of pose, expression, presence glasses, hat etc.) should be discarded since they not useful for In this paper, we present surprising result that case. We show non-ID attributes, well landmark positions histogram can recovered from ID embedding state-of-the-art networks (VGGFace2 ArcFace). fact, these attributes predicted with similar accuracy prediction original image. Going further, an optimisation strategy uses generative model (specifically StyleGAN2 faces) recover images embedding. photorealistic inversion which realistically reconstructed but lighting background/apparel some extent well.