Transformer-Based Financial Fraud Detection with Cloud-Optimized Real-Time Streaming

TL;DR

This paper introduces a cloud-enabled Transformer model for real-time financial fraud detection, leveraging graph self-attention to improve accuracy and adapt to evolving threats in digital transactions.

Contribution

It presents a novel Transformer-based approach that directly excavates fraud features from transaction networks without complex feature engineering, optimized for cloud deployment.

Findings

Outperforms 7 baseline models on all evaluation metrics.

Achieves 20% higher average accuracy (AP).

Increases AUC by 2.7% compared to GAT.

Abstract

As the financial industry becomes more interconnected and reliant on digital systems, fraud detection systems must evolve to meet growing threats. Cloud-enabled Transformer models present a transformative opportunity to address these challenges. By leveraging the scalability, flexibility, and advanced AI capabilities of cloud platforms, companies can deploy fraud detection solutions that adapt to real-time data patterns and proactively respond to evolving threats. Using the Graph self-attention Transformer neural network module, we can directly excavate gang fraud features from the transaction network without constructing complicated feature engineering. Finally, the fraud prediction network is combined to optimize the topological pattern and the temporal transaction pattern to realize the high-precision detection of fraudulent transactions. The results of antifraud experiments on credit card transaction data show that the proposed model outperforms the 7 baseline models on all evaluation indicators: In the transaction fraud detection task, the average accuracy (AP) increased by 20% and the area under the ROC curve (AUC) increased by 2.7% on average compared with the benchmark graph attention neural network (GAT), which verified the effectiveness of the proposed model in the detection of credit card fraud transactions.