Privacy-Preserving Deep Multimodal Behavioral Intelligence Framework for Continuous Authentication in Real-Time FinTech Payment Systems

Abstract

The rapid digitalization of global financial ecosystems has increased both the speed and volume of mobile and online financial transactions, consequently heightening the risk of payment fraud and unauthorized access. Traditional rule-based and transaction-only fraud detection systems are increasingly ineffective against sophisticated attacks such as social-engineering–assisted transfers, account takeovers and cross-device fraud. To address these gaps, this paper proposes a Privacy-Preserving Deep Multimodal Behavioral Intelligence Framework for continuous authentication in real-time FinTech payment systems. The framework integrates multimodal behavioral biometrics-keystroke dynamics, mouse/touch interaction patterns, device intelligence and transaction metadata-into an attention-based deep learning architecture. A federated learning pipeline enhanced with calibrated differential privacy ensures the protection of biometric signatures and compliance with financial regulations. Experiments conducted on a hybrid dataset comprising real behavioral biometric samples (HMOG, keystroke datasets) and a synthetically constructed payment-session dataset demonstrate significant improvements in fraud detection and user verification accuracy compared to conventional transaction-only baselines. The proposed solution achieves robust performance under device heterogeneity, sparse-session conditions and adversarial mimicry simulations. The results show the feasibility of deploying the system in real-time FinTech environments with inference latency under 180 ms. The study presents a novel fusion of continuous authentication, multimodal deep learning and privacy-preserving training tailored specifically for payment flows-an area largely underexplored in contemporary research.