Recent works have demonstrated the feasibility of inverting face recognition
systems, enabling to recover convincing face images using only their
embeddings. We leverage such template inversion models to develop a novel type
ofdeep morphing attack based on inverting a theoretical optim