Authors: K. Sai Teja, M.Surya Teja, S.Bharath Simha Rao, Y.Hemanth Kumar
Abstract: Face morphing attacks represent a critical vulnerability in biometric authentication sys- tems, where two or more facial images are digitally blended to create a forged identity. Such morphed images can successfully deceive automated face verification systems, leading to severe risks in applications like passport issuance, border control, and iden- tity management. Traditional detection techniques, relying on handcrafted features or differential meth- ods, often fail to generalize across diverse morphing techniques and image qualities. To overcome these limitations, we propose MorphDetect, a deep learning-based Single- Image Morphing Attack Detection (S-MAD) system powered by the EfficientNet-B7 model. The system first preprocesses face images for normalization and then extracts high- dimensional features using EfficientNet-B7’s advanced convolutional blocks. These features are passed through a classification layer that determines whether an input is genuine or morphed, producing a reliable confidence score for decision-making. MorphDetect eliminates the need for a trusted reference image and provides a scal- able, real-time solution for morph detection. By leveraging a strong pretrained back- bone, it ensures robustness against unseen morphing techniques and diverse imaging conditions. This makes the system well-suited for deployment in high-security appli- cations such as e-passport verification, financial KYC procedures, and secure access systems.
