SC-GAN: A GAN-Based Data Augmentation Approach for Stablecoin Fraud Detection on Imbalanced Transaction Data

Abstract

Stablecoins are becoming more common in the Fin-Tech (Financial Technology) ecosystem as they keep their value stable and combines easily with decentralized finance inherent in the financial technology ecosystem due to their price stability and simplicity of integration in decentralized finance (DeFi), cross-border payments, and automated trading systems. However, the same characteristics that propel utility transaction speed, pseudonymity, and automation through smart contracts have also made them vulnerable to financial manipulation. Tactics such as wash trading, spoofing, and pump-and-dump schemes have become more prevalent, compromising market integrity significantly. However, major technical challenge in detecting these fraudulent activities and behaviors, especially under conditions of extreme class imbalance even the legitimate transactions vastly outnumber fraudulent ones.This paper introduces SC-GAN, a conditional Generative Adversarial Network that addresses the scarcity of fraudulent samples by synthesizing realistic blockchain-based fraud instances. The model conditions on key financial and transactional features native to blockchain systems, enabling the generation of high-fidelity synthetic data. We then compare SC-GAN with traditional oversampling techniques like SMOTE and Borderline-SMOTE on a variety of supervised classification models. Our experiments on a real-world stablecoin transaction dataset show with the help of SC-GAN improves both the F1 Score and overall accuracy that is resulting in more efficient detection of rare but crucial fraudulent transactions. This approach also provides the possibility for stronger fraud prevention methods and risk management policies within FinTech platforms.