Deep Learning-Based Face Morphing Detection for Identity Verification

Face morphing attacks present a real threat to biometric-based identity verification systems, allowing attackers to manipulate facial images and bypass authentication mechanisms in mission-critical applications like border control, passport issuance, and online banking. Conventional facial recognition systems do not detect such subtle manipulations because they are based on shallow or static feature descriptions. To address this, we introduce a secure deep learning-based detection framework with the combination of Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), and a Multi-Scale Feature Fusion mechanism to increase sensitivity to morphing artifacts. Explainability AI modules like Focused Layer-Wise Relevance Propagation (FLRP) provide interpretability by highlighting key facial features that guide classification. Apart from that, advanced verification architectures such as Siamese and Double Siamese Networks are used to compare face identities and highlight inconsistencies. Experimental benchmarking of public and private datasets shows the enhanced performance of the proposed hybrid model over stand-alone architectures with 94.6% accuracy, 93.6% F1-score, and 96.8% AUC-ROC, demonstrating improved accuracy and robustness to varied image quality and morphing complexity. Additional testing shows a 45% hallucination reduction and minimal mean token utilization (~1200), demonstrating effectiveness and real-time applicability of the system. The model architecture is also modular and scalable with flexibility to enable the inclusion of new morphing methods and security demands. Verification in real scenarios under diversified illumination, resolution, and expression conditions demonstrates its utility for secure authentication of identity in mission-critical applications. The work elevates trustworthy and interpretable biometric authentication to another level against increased digital identity threat scenarios.