A Hierarchical Dual-Role Interaction Network for Telephone Conversation Fraud Detection

The rapid increase in telecommunication fraud has led to substantial financial losses. A critical challenge in detecting such fraud based on call content lies in accurately modeling and interpreting the dual-role interactions between fraudsters and victims, often manifested in complex dialog structures. To address this issue, we propose the Hierarchical Dual-Role Interaction Network (HDRIN), a novel approach for telecom fraud detection. Our method categorizes raw text into two types: two-person dialog text and single text. For two-person conversational texts, we introduce DialogBert, which performs word embedding transformations followed by local feature extraction using 1DCNN. For single texts, we develop the Multi-Level Dynamic TextRank (MDTR) algorithm to extract concise summaries from lengthy single dialogs, while Hierarchical Attention Networks (HAN) are employed to enhance global features. Additionally, a synergistic attention mechanism is proposed to integrate features from both text types. We constructed the ZhCCVi dataset, using real Chinese call speech from platforms like Jitterbug and YouTube, to evaluate our approach. Extensive experiments on ZhCCVi and the publicly available KoCCVi2 dataset demonstrate that our approach achieves an accuracy of 99.88% and an F1 score of 99.88%, surpassing the performance of existing state-of-the-art models. The code is available at https://github.com/Gaopayvin/HDRIN.