Who are the evil backstage manipulators: Boosting graph attention networks against deep fraudsters

Abstract

Telecommunications fraud causes vast economic losses all over the world every year. Although Graph Neural Networks (GNNs) brings new possibilities to the solution of fraud detection problem, their performance in telecom fraud detection is not satisfying. An important reason is that little prior work has noticed the deep fraudster phenomenon, which makes it difficult to be eradicated. Another reason may be the shortage of available datasets for researchers. After all, few operators are willing to publish telecom fraud detection benchmark datasets due to the sensitivity and privacy of subscriber data, resulting in less related research work. In this paper, we rearrange and publish two real-world telecom fraud detection datasets based on publicly available sources on the Internet. And we propose a GNN-based semi-supervised telecom fraud detection method by Boosting graph Attention Networks Against Deep frAudsters (BANADA). Specifically, we first learn different neighbor weights with the help of inner-layer attention, and then design simplified inter-layer attention to achieve inter-layer feature aggregation. Finally, we use GNNs with different depths as different base classifiers and perform ensemble learning through AdaBoost to discriminate deep fraudsters. Extensive experiments on two real-world telecom fraud datasets demonstrate the effectiveness of our proposed BANADA, which outperforms all state-of-the-art GNNs and GNN-based fraud detectors. The BANADA code and datasets are available at https://github.com/xxhu94/BANADA.