Enhancing Online Transaction Fraud Detection via Heterogeneous Source Models

Obtaining dedicated fraud detection models is important for financial risk management. Online transaction platforms traditionally rely on local data to accumulate domain knowledge and establish fraud detection models to detect fraud. This naturally makes those online trading platforms with limited data and hardware resources more susceptible to fraudulent transactions. In this article, we propose GMDK, which elegantly integrates credible knowledge generation and collaborative model training on weak transaction platforms. Specifically, we decouple the local model parameters into base and specialized layers to learn different levels of knowledge. Local platforms milk different source models with delicately designed two-factor cues to acquire both cross-domain and credible domain-specific knowledge. To this end, source models first aggregate local base layer parameters of each local model to obtain cross-domain base layers for all local models. Then, distinct semantic distribution prototypes generated by heterogeneous source models on desensitized and sampled local business data are aligned to update local models. We conduct extensive experiments on various real-world online transaction platforms, and the results demonstrate that local models trained with GMDK can achieve state-of-the-art fraud detection accuracy in each target transaction area.