DOI: 10.67733/rlipe.2.1.2 ISSN: 3062-4487

A Unified Neural–Symbolic Compliance-Aware Framework for Real-Time Fraud Detection in Regulated Financial Systems

Akmalbek Abdusalomov, Alpamis Kutlimuratov, Cüneyt Şamil Oğurlu
Show PDF Cite
The rapid growth of digital banking and card-not-present transactions has significantly increased exposure to sophisticated cyber-enabled fraud, requiring real-time and regulation-compliant decision-making. Although recent advances in artificial intelligence have improved detection accuracy, most existing models remain incompatible with strict latency constraints and fail to satisfy regulatory requirements for explainability, auditability, and data governance. This study proposes NEO-STREAM, a compliance-aware transformer-based framework for real-time fraud detection. The model employs micro-episode tokenization to capture short-term behavioral patterns without relying on long-term user histories, ensuring both efficiency and data minimization. A latency-bound twin-path transformer integrates counterfactual reasoning directly into inference, providing deterministic and interpretable decision evidence. In addition, neural–symbolic learning aligns model behavior with regulatory rules, while conformal prediction enables uncertainty-aware decision-making through selective abstention. Experimental results on large-scale CNP transaction data demonstrate that the proposed framework improves fraud detection performance while maintaining sub-50 ms latency and reducing manual review workload. The findings indicate that compliance-aware AI architectures can effectively bridge the gap between predictive performance and regulatory accountability in modern digital financial systems.

More from our Archive