A Multi-detector Solution Towards an Accurate and Generalized Detection of Face Morphing Attacks

Face morphing attack images are built to be verifiable to multiple identities. Associating such images to identity documents leads to building faulty identity links, causing vulnerabilities in security critical processes. Recent works have studied the face morphing attack detection performance over variations in morphing approaches, pointing out low generalization. This work introduces a multi-detector fusion solution that aims at gaining both, accuracy and generalization over different morphing types. This is performed by fusing classification scores produced by detectors trained on databases with variations in morphing type and image pairing protocols. This work develop and evaluate the proposed solution along with baseline solutions by building a database with three different pairing protocols and two different morphing approaches. This proposed solution successfully lead to decreasing the Bona Fide Presentation Classification Error Rate at 1.0% Attack Presentation Classification Error Rate from 15.7% and 3.0% of the best performing single detector to 2.7% and 0.0%, respectively on two face morphing techniques, pointing out a highly generalized performance.