Fraud detection based on GNNs with local augmentation and adaptive relation aggregation

Abstract

Fraud detection based on Graph Neural Networks (GNNs) relies on aggregating information from the local neighborhoods, but this mechanism is vulnerable to two adversarial tactics: feature camouflage where fraudsters manipulate node attributes to mimic benign users, and relation camouflage where they establish connections with benign entities to dilute suspicious signals. These camouflage strategies compromise GNNs’ discriminative capability by exploiting the neighborhood aggregation mechanism itself. To address this vulnerability, we propose a fraud detection method based on GNNs with Local Augmentation and Adaptive Relation Aggregation (GNN-LAARA). GNN-LAARA integrates three synergistic components: a conditional variational autoencoder (CVAE) that generates discriminative node representations to expose camouflaged patterns, a reinforcement learning-based neighbor selector that dynamically filters noisy connections, and a multi-relational attention aggregator that adaptively fuses heterogeneous relationships. The effectiveness of GNN-LAARA is validated by two real-world fraud detection datasets. Experimental evaluation on two real-world fraud detection datasets demonstrates that GNN-LAARA achieves significant performance improvements, with up to 2.24% enhancement in AUC over state-of-the-art methods. Ablation studies further confirm the individual contributions of each module to the overall detection capability.