An Integrated Privacy-Preserving AI Framework for Real-Time Financial Fraud Detection

Author (s)

Alpha Koroma

Abstract

The research presents a practical solution for online financial fraud prevention grounded in operating under the operational, regulatory, and adversarial constraints of modern payment systems. The model characterizes the economic ecosystem as a dynamic, heterogeneous graph where entity types (e.g., accounts, devices, merchants) evolve and are connected by temporal message passing that captures both behavioral drift and short, bursty attack phases. It uses federated learning with secure aggregation and optional differential privacy to address cross-institutional blind spots while avoiding sharing raw data. Native explainability, such as subgraph rationales and temporal saliency, improves auditability and regulatory examination. The model is further strengthened via adversarial learning against fintech evasion tactics (e.g., amount splitting and time shifting) and hash-chain logging for decision tracking. We evaluate on a synthetic transaction dataset with less than 0.5% fraud and show the framework provides notable improvements in AUPRC, cost-sensitive F1, and expected loss over strong supervised and deep baselines while operating within p95 latency targets. Ablation and sensitivity analyses indicate that temporal reasoning, self-supervision, and federated cooperation are essential for stability, and privacy experiments show reduced membership-inference risk compared with centralized training. The accompanying graph–temporal, privacy-preserving, and explainable architecture advances early-warning capability and operational trust in online financial fraud defense.

Key words: Financial fraud detection, Graph neural networks, Temporal learning, Self-supervised learning, Federated learning.

Download PDF

Cite This Article:

Koroma, A. (2025). An Integrated Privacy-Preserving AI Framework for Real-Time Financial Fraud Detection, Journal of Scientific Reports, 11(1), 80-94. DOI: https://doi.org/10.58970/JSR.1137

About Author (s)

Alpha Koroma, Master’s in Information and Communication Engineering, Huazhong University of Science and Technology, China. ORCID: https://orcid.org/0009-0004-1176-4652

Download PDF

DOI: https://doi.org/10.58970/JSR.1137

Summarize this article with: