Semi-supervised GANs for Fraud Detection, 2020 International Joint Conference on Neural Networks

The research proposes a generative adversarial framework for detecting fraud in online gambling.

Categories

The research introduces a novel fraud detection framework using semi-supervised generative adversarial networks (SSGANs) combined with sparse auto-encoders (SAEs) to address the challenge of imbalanced data in online gambling and other domains. Key findings:

• New Architecture for Imbalanced Classification: The proposed system avoids traditional oversampling techniques by leveraging latent feature representations from SAEs, improving classification performance across multiple datasets.
• Superior Performance Across Benchmarks: Compared to logistic regression, random forest, and multilayer perceptron—both standalone and combined with SMOTE/ADASYN—the SSGAN+SAE model consistently achieved higher F1 scores on Breast Cancer, Diabetes, and Credit Card Fraud datasets.
• Real-World Application to Gambling Fraud: Applied to a gambling dataset with 4,700 samples, the framework improved the F1 score by 3.64% over the existing knowledge-based anti-money laundering system, demonstrating practical value in detecting suspicious behavior.
• Complementary Generator Enhances Stability: The use of a complementary generator in the GAN architecture led to better convergence and classification accuracy compared to regular GANs, especially in structured data environments.
• Sparse Auto-Encoders Improve Feature Separation: Encoding data into higher-dimensional sparse representations increased the separation between fraudulent and non-fraudulent cases, boosting recall and overall model robustness.

These insights support the adoption of semi-supervised GANs with sparse auto-encoders as a powerful tool for fraud detection in domains with imbalanced and structured data. For full details, please refer to the complete document.