Multilevel Fusion of Deep Features for Face Morphing Attack Detection

Face biometric systems are widely deployed in a magnitude of security-related applications, including border control. However, the vulnerability of the Face Recognition Systems(FRS) to various types of attacks is well demonstrated. This work presents a novel approach for face morphing attack detection for a single image scenario by performing a multi-level fusion of deep features. The features are extracted using the pre-trained deep CNNs such as AlexNet, ResNet50. These extracted features are combined at both features and score levels to conclude if the given face image is a morph. The proposed single image Morph Attack Detection (S-MAD) approach is extensively evaluated on the face morphing dataset constructed using five different face morphing generation techniques and three different data mediums. The data mediums including digital, print-scan (re-digitised), print-scan compression (re-digitised and compressed.) Extensive experiments are carried out with intra (same datatype used for training and testing) and inter-evaluation scenarios (cross datatype used for training and testing). Further, the proposed method is compared with the State-Of-The-Art (SOTA) approaches for No reference-based/Single image Morph Attack Detection (S-MAD). The statistical analysis indicates the best performance of the proposed approach in all three different mediums.