A Deepfake Detection Framework for Securing Facial Biometrics Using Attention-Based Deep Learning and Blockchain

Facial biometric authentication is widely deployed for identity verification, yet it is increasingly exposed to deepfakebased presentation and injection attacks and to large-scale breaches of centrally stored biometric templates. This paper proposes a deepfake-resilient facial biometric security framework that combines attention-based deep learning with blockchain technology. At registration, face images are pre-processed, augmented with a generative adversarial network (GAN), encoded by a convolutional neural network (CNN) with self-attention, and classified as real or fake by an artificial neural network (ANN). Only images classified as real are encrypted, bound to a unique user identifier and committed to a private blockchain through a cryptographic hash, whereas fake attempts are written to an immutable event log. At authentication, a fresh probe passes through the same pipeline and is verified by smart contracts that check the integrity of the registered commitment and the template similarity using zero-knowledge proofs (ZKPs), so that raw biometric features are never disclosed. We formalise the threat model, specify the architecture, the smart-contract interface and the security properties, and define an evaluation protocol based on the CIPLAB Real and Fake Face Detection and UTKFace datasets that covers detection accuracy, authentication error rates, adversarial robustness, demographic fairness and system overhead. The framework provides a basis for privacy-preserving, auditable and scalable facial authentication in digital identity management, financial services, healthcare and e-governance.