Face Morphing Attack Detection Using Deep Learning

Early detection of face morphing attacks plays a pivotal role in providing security and integrity in biometric systems, particularly in high-risk applications such as border crossing and identity verification. Traditional detection methods often fail to provide accurate detection for subtle morphing artifacts even in dynamic conditions and demographic differences. The research overcomes such limitations using advanced deep learning approaches and residual noise analysis for the detection of morphing artifacts from face images. We have tested and compared five of the latest deep learning architectures: UNetXception, UNetResNet18, UNetResNet50, MobileNetV2, and DenseNet121. These models were tested on the Synthetic Morphing Attack Detection Development (SMDD) dataset, a large, ethically gathered dataset of bona fide and morphed images. Our highest performing model, UNetXception, achieved a test BPCER of 0.0044 and an EER of 0.0031, demonstrating its robustness in face morphing artifact detection. The other models also demonstrated robustness, which is an indicator of the effectiveness of residual noise analysis in morphing artifact detection. The proposed models not only enhance the power of detection of existing biometric systems but also provide a robust and scalable solution for real-world use. This contribution helps in the ongoing endeavor to safeguard biometric authentication processes from increasingly sophisticated face morphing attacks, offering a robust, effective, and versatile approach that may substantially improve the resilience of identity authentication systems.