GenOpt-FedGNN: GAN-Augmented Federated Graph Learning with GWO and ASIO-Inspired Adaptive Aggregation for Imbalanced Fraud Detection

Detecting fraud in distributed financial transaction systems is challenging. It is extremely rare to find a fraudulent record. The interactions among transactional records have a relational structure and the data can’t be stored centrally by every participating institution. The paper introduces the federated graph learning framework i.e. GenOpt-FedGNN. The framework combines Generative Adversarial Network (GAN) with client local learning, graph neural classification, and the aggregation of the trained models through adaptive process of Grey Wolf Optimization (GWO) and Asteroid Satellite Inspired Optimization (ASIO). Each client preprocesses their own transactions independently, produces more minority fraud samples to train a class weighted Graph Neural Network (GNN), and builds a transaction graph. Each client sends updates regarding model and validator to the server. The framework is tested with a set of 200,000 PaySim transactions. Transactions were randomly sampled and distributed among the five distributed clients with 30 communication rounds. Six configurations (FedAvg, GWO and ASIO with and without GAN augmentation), have been compared by Accuracy, Balanced Accuracy, Precision, Recall, F1-score, ROC-AUC, PR-AUC, Specificity, False Positive Rate, False Negative Rate and Confusion-matrix measures. The ablation analysis demonstrated that GAN augmentation doubled local fraud training samples and reduced class-imbalance ratio. But, the local fraud data has been extremely imbalanced. Overall, at the final round, the accuracy was highest with ASIO-GAN (96.40%), the specificity (96.41%), precision (1.58%), F1-score (3.11%), and the lowest False Positive Rate (3.59%). The best Balanced Accuracy (94.16%) was found for FedAvg-GAN, and the best PR-AUC (60.02%) was found for ASIO-NoGAN. The GWO GAN-based configurations got best PR-AUC of 59.20%. The ablation results on the other hand further showed that using GAN for augmentation resulted in a consistent increase in the Accuracy, Precision, F1-Score, Specificity and false alarm control compared to using it without GAN. So augmentation with GAN and aggregation with Adaptive Aggregation can affect the sensitivity–specificity trade-off but not necessarily all the other aggregation metrics. The proposed framework is reproducible and data local to the raw data for the graph-based fraud detection application which features extreme class imbalance.