A Multi-Modal Framework for Fraud Detection and Weakly Supervised Risk Stratification
TL;DR
This study proposes a reliability-aware dynamic multimodal framework, termed MM-DRTR-Fraud, which integrates temporal and relational dependencies for fraud detection and weakly supervised risk stratification.
Abstract
Fraud detection and fine-grained risk stratification under weak supervision are critical for financial security, telecommunication services, and online platform risk management, as they play a key role in mitigating economic losses caused by fraudulent activities. However, existing approaches are limited by suboptimal detection performance, ineffective utilization of multimodal information due to static fusion strategies, and high misclassification rates in fine-grained risk grading, which hinder their applicability in complex real-world scenarios. To address these challenges, this study proposes a reliability-aware dynamic multimodal framework, termed MM-DRTR-Fraud, which integrates temporal and relational dependencies for fraud detection and weakly supervised risk stratification. The framework consists of a Modality-Specific Feature Enhancement Module (MSFEM) and three core components: dynamic reliable cross-modal fusion, temporal relational joint modeling, and hierarchical risk stratification.