A2062
Title: Multi-modal enhanced graph transfer learning for digital finance fraud detection
Authors: Chenguang Yang - University of California, Riverside (United States) [presenting]
Abstract: Fraudulent activities on blockchain networks pose serious threats to the integrity and reliability of decentralized finance ecosystems. Detecting malicious nodes, such as phishing or ransomware addresses, in large-scale blockchain transaction graphs remains challenging because these graphs are dynamic, sparse, and continuously evolving. Transfer Learning is promising for fraud detection, as different fraud types are often conducted by overlapping actor groups and may share behavioral and structural patterns across networks. By transferring knowledge from known fraud types to emerging ones, models can improve detection under limited supervision. However, digital financial systems are inherently Multi-mOdal, involving Graph structures, transaction sequences, temporal price dynamics, and textual metadata. In addition, distribution shifts across time and platforms make it difficult for existing Graph Transfer Learning methods to align source and target domains effectively. To address these challenges, the Multi-mOdal Enhanced Graph Transfer Learning framework, or MOE-GTL, integrates Graph, temporal, and textual modalities for fraudulent node detection. The framework further introduces Temporal-aware Maximum Mean Discrepancy, or TMMD, a regularization mechanism that aligns Multi-mOdal feature distributions between source and target graphs over time. Experiments on Ethereum and Solana transaction graphs demonstrate that MOE-GTL improves fraudulent node classification accuracy.
