AI-Enhanced Cybersecurity for Financial Networks: A Federated Learning Implementation

Abstract

During a 14-month deployment across four financial institutions, including a tier-1 bank in the Northeast US, we developed a hybrid threat detection system that integrates Transformer models with Graph Neural Networks. The system was implemented using Python 3.8.10 and PyTorch 1.12.1 on NVIDIA RTX 3090 GPUs (24GB VRAM). Our team, despite frequent methodological disagreements, achieved a detection accuracy of 86.7%, which fell short of the anticipated 95% or higher. The federated learning component, initially planned for six months, was extended due to regulatory compliance requirements. This component enables collaborative threat intelligence while preserving data privacy. Under normal operating conditions, the system processes approximately 1.1 million events per second, with throughput decreasing to around 400,000 events per second during periods of market volatility, such as Q4 2023. The architecture reduces false positives to 2.1%. Implementation costs exceeded the original $127,000 NSF grant by roughly 40%, necessitating additional university cost-sharing. Three preliminary approaches were abandoned before the current architecture was finalized. Real-world deployment highlighted hardware bottlenecks that were not evident in simulations, requiring compromises in system design. The system is now operational in production, although stability issues persist during high-frequency trading periods.