One-Shot Synthetic Disguised Face Generation to improve Robustness against Disguise Attacks

Disguised face recognition can be challenging in real-world circumstances because the disguise add-ons are usually used to hide the key facial features. It is very important to solve this problem especially at sensitive places. In order to design disguised facial recognition system good quality disguised face database is required. We propose to solve this problem by presenting a one-shot disguised face generation method that can augment normal non-disguised faces and help improve the model accuracy when tested on disguised faces. The proposed method utilizes patchwise contrastive loss and cycle-consistency loss to take benefit of both approaches. The patchwise contrastive loss allows training with as few as one sample per image however, to further condition the networks for better results cycle-consistency among domains is applied. The results are satisfactory given that the models were trained with only one image per subject per add-on.