THE DISTRIBUTED KNOWLEDGE DISTILLATION FRAMEWORK FOR THE FINANCIAL FRAUD DETECTION BASED ON TRANSFORMER.

Financial fraud is a persistent and evolving challenge in the digital economy, causing substantial monetary losses and security threats. Traditional fraud detection systems often rely on centralized models that face scalability, privacy, and computational constraints. To overcome these limitations, we propose a Distributed Knowledge Distillation (DKD) framework for financial fraud detection, leveraging Transformerbased models. The framework enables decentralized learning by distributing knowledge across multiple nodes, ensuring enhanced model performance, privacy preservation, and adaptability to diverse fraud patterns. Our approach integrates self-supervised learning techniques to extract high-level features from transactional data while maintaining minimal computational overhead. The DKD framework distills essential knowledge from a pre-trained Transformer model into lightweight models deployed across various nodes, facilitating efficient fraud detection with reduced latency. Experimental evaluations demonstrate that the proposed framework achieves superior accuracy and robustness compared to conventional machine learning and deep learning-based fraud detection systems. The DKD framework's key contributions include adaptive learning mechanisms, reduced dependency on centralized data storage, and enhanced resilience to adversarial attacks. Comparative studies with existing centralized architectures highlight the advantages of distributed learning in financial fraud detection. Future enhancements may explore federated learning extensions and hybrid Transformer models for improved generalization across financial domains.