Distributed Knowledge Distillation with Spatial-Temporal Graph Attention for Financial Fraud detection

With current technological advancements, financial fraud cases that harm the interests of investors are becoming increasingly common. However, existing models face challenges due to the difficulty of adapting to evolving fraud patterns, inconsistent financial data and unbalanced data distribution. To overcome this problem, this paper proposed Distributed Knowledge Distillation with Spatial-Temporal Graph Attention (DKD-STGA) for financial fraud detection allows for highly effective and efficient identification of fraudulent activities-based Transformer. First, spatial-temporal graph attention used to intricate relationship across space and time by using graph and used to capture both spatial as well as temporal patterns. Next, neural networks are used to classify financial fraud and distributed knowledge distillation algorithm is used to overcome this complex structures, deep model depths and slow inference speeds. This algorithm gathers detection knowledge from multi-teacher network and transfers, it to a student network, which is used this shared knowledge to identify fraud in financial data from various industries. The experiments results show that the proposed DKD-STGA outperforms in terms of accuracy (98.97%), precision (99.89%), recall (97.99%) and F1-Score (98.90%) compared to existing method like transformer.