Privacy-Preserving Behavioral Anomaly Detection in Dynamic Graphs for Card Transactions
Résumé
Anomaly detection in financial transactions presents significant challenges while ensuring the privacy of individuals. This paper introduces a federated learning (FL) framework for Privacy-Preserving Behavioral Anomaly Detection, utilizing Graph Neural Networks (GNNs) on dynamic graphs to represent cardholder transactions in vector space. We propose a novel negative sampling technique to train anomaly detection models without relying on labeled data, making it applicable to real-world applications. We benchmarked various graph pooling models on two datasets, synthetic data from a United States (US) bank, and real-world data from a Brazilian bank. Our results demonstrate that deep learning-based methods outperform clustering-based approaches for anomaly detection, where ASAPooling and DiffPool achieved F1-scores of 0.91±0.02 on the US dataset and 0.87±0.04 on the Brazilian Bank dataset, respectively, outperforming clustering-based graph pooling methods. Finally, we show how the anomaly score can be used as a feature for centralized and federated fraud detection. The experimental results demonstrate a 1.76\%±0.54\% improvement in F1-score compared to a supervised federated model, showcasing the proposed framework to enhance fraud detection performance while preserving privacy.