DeepBiometricFusion: Multi-Modal Defense Against DeepFake Payment Attacks

Abstract

DeepFake-driven payment fraud has rapidly evolved into one of the most serious threats in biometric authentication ecosystems, particularly across digital banking, fintech wallets, eCommerce checkout systems, and remote KYC workflows. Modern generative models can synthesize highly convincing facial videos, overlay spoofed identities, manipulate voiceprints, and exploit biometric vulnerabilities to trigger fraudulent financial transactions. This research introduces DeepBiometricFusion, a unified multi-modal defense architecture that integrates facial dynamics, voiceprint signatures, micro-texture liveness cues, and cross-modal temporal coherence to detect DeepFake payment attacks with high reliability. The proposed system leverages multi-stream deep neural networks, cross-attention fusion, and adversarial anomaly detection to jointly validate biometric consistency across modalities. Experiments conducted on three large-scale datasets—DFSecurePaySet, MultiModal-LiveID, and VidVoice-FraudBench—demonstrate the superior robustness of our framework compared to existing uni-modal and dual-modal solutions. The results show that DeepBiometricFusion achieves a 98.4% attack detection accuracy with a substantially reduced false acceptance rate (FAR), establishing it as a practical and secure solution for financial-grade authentication.