Ünsal Öztürk

Quantifying Training Membership Information in the Hyperspherical Embedding Geometry of Face Recognition Models

Ünsal Öztürk, Sébastien Marcel

Face recognition models represent each face as an embedding vector on the unit hypersphere by clustering embeddings of the same identity while pushing different identities apart through angular-margin losses. Because these losses act only on training identities, non-member identi…Read more