Conditional Identity Disentanglement for Differential Face Morph Detection
Face morphing can circumvent biometric recognition by matching a single morphed image to two distinct identities. Such attacks can be used in photo-based identification documents, and pose a serious concern in real-world scenarios. Differential face morph attack detection (MAD) uses a trusted non-morphed reference image to deduce whether the document image is morphed or bonafide (non-morphed). We propose a novel framework that leverages the underlying conditional probabilistic nature of the differential MAD problem.
We use a conditional GAN to disentangle the second identity from the morphed document image using the trusted reference image belonging to the first identity. The output of the conditional GAN and the reference image are fed to a biometric comparator; the scores are then used to decide whether the document image is morphed or not. Therefore, we can use the same framework for both morph detection as well as recovering the participant subject in the morph with promising results.
Check out the highlight video of our morphing work published in IJCB 2021 here.
S. Banerjee and A. Ross, “Conditional Identity Disentanglement for Differential Face Morph Detection,” Proc. of International Joint Conference on Biometrics (IJCB), (Shenzhen, China), August 2021.