Advancing 3D Face Morphing Detection Via Synthetic Data and Deep Neural Network

Face morphing attacks have posed great threats to the security of biometrics by mixing face features with the intention of passing identity verifications. In this regard, most of the current detection methods usually exhibit poor generalization to real variations due to limited data diversity. This work proposes a new deep learning framework for detecting attacks involving 3D face morphing, leveraging the advantages of geometric facial modeling for the generation of high-quality diverse synthetic 3D datasets with adversarial generative models. This is achieved by detecting morphing inconsistencies with the use of spatial and texture-level artifacts, allowing domain adaptation for robust performance of the solution across synthetic-real scenarios. Experimental results confirm its superior accuracy, adaptability to unseen techniques, and robustness against pose, lighting, and occlusion variations, thus turning out to be an efficient solution in biometric security.