Trans-FD: Transformer-Based Representation Interaction for Face De-Morphing

Abstract

Face morphing attacks aim to deceive face recognition systems by using a facial image that contains multiple biometric information. It has been demonstrated to pose a significant threat to commercial face recognition systems and human experts. Although a large number of face morphing detection methods have been proposed in recent years to enhance the security of face recognition systems, little attention has been paid to restoring the identity of the accomplice from a morphed image. In this paper, Trans-FD, a novel model that uses Transformer representation interaction to restore the identity of the accomplice, is proposed. To effectively separate the identity of an accomplice, Trans-FD applies Transformer to perform representation interaction in the separation network. Additionally, it utilizes CNN encoders to extract multi-scale features, and it establishes skip connections between the encoder and generator through the Transformer-based separation network to provide detailed information for the generator. Experiments demonstrate that Trans-FD can effectively restore the accomplice’s face and outperforms previous works in terms of restoration accuracy and image quality.