CSR: Small: Multi-FPGA System for Real-time Fraud Detection with Large-scale Dynamic Graphs

The rise of international financial and cybercrime, such as fraud and money laundering, has led to billions of dollars in losses annually in the United States. To address this urgent issue, there is a need for real-time fraud detection algorithms and systems with extremely low latency. Graph-based machine learning algorithms, specifically Graph Neural Networks (GNNs), have emerged as a promising solution.

Financial activity graphs possess two crucial characteristics: they are extremely large, comprising vast amounts of transactions, financial institutions, and customers, and they evolve dynamically over time as new transactions occur. These characteristics present significant challenges for real-time, low-latency fraud detection, as it requires a scalable and parallelizable distributed system.

Additionally, the dynamic nature of the graphs poses challenges for system control and optimization, as the partitioning of the problem on a distributed system may quickly become suboptimal due to graph updates. To tackle these challenges, this project proposes a distributed FPGA (Field-Programmable Gate Array) system for real-time fraud detection on large-scale dynamic graphs.

The project aims to achieve microsecond-level latency, which has not been explored in real-time distributed dynamic GNNs. The research involves three main tasks: system construction, dynamic optimization, and uncertainty analysis and optimization. The impacts of this project are significant.

Successful implementation will enhance the effectiveness of fraudulent transaction alerts for millions of people globally and help businesses reduce fraud losses and increase revenue. The project's outcomes can have applications in various domains, such as cybercrime detection, insurance fraud, national security infrastructure protection, and identifying anomalies and terrorist attacks.

The constructed system will be made publicly accessible, and all codes will be open-sourced to benefit the community. Furthermore, the project presents an opportunity to involve students, including those from underrepresented groups, in research and education through collaborations and engaging competitions.

The project aims to address the rising challenges of international financial and cybercrime by developing a distributed FPGA system for real-time fraud detection on large-scale dynamic graphs. Current fraud detection systems lack low-latency capabilities, making real-time detection difficult. To overcome this, the project proposes the use of Graph Neural Networks (GNNs) and introduces three key tasks: system construction, dynamic optimization, and uncertainty analysis and optimization.

The financial graphs involved in fraud detection are characterized by their immense size and dynamic nature. These attributes pose significant obstacles to real-time detection and system optimization. To tackle these challenges, we plan to utilize a distributed FPGA system that can handle large-scale dynamic graphs efficiently.

By leveraging Smart Network Interface Cards (SmartNICs) and multi-agent reinforcement learning (MARL), the system will dynamically repartition evolving graphs across FPGAs for optimal performance. Additionally, we propose to use Bayesian Neural Networks (BNNs) to model and analyze system predictability and uncertainty. This information is crucial for real-time systems.

The BNN will guide active learning strategies, allowing the system to make informed decisions when faced with high uncertainty.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.