L2IR: LLM Framework Reveals Latent Intent in Graph Fraud Detection
Researchers propose L2IR, an LLM-driven framework for graph fraud detection that uncovers latent intent behind suspicious connections. Graph Neural Networks (GNNs) traditionally propagate information across relational data, but fraudsters often forge connections with benign users, diluting fraud signals. L2IR uses Large Language Models to extract semantic cues from user behaviors and connections, addressing the scarcity of annotated fraud samples. The framework aims to improve detection reliability under heavy camouflage.
Key facts
- L2IR is an LLM-driven Latent Intent Revealing framework for graph fraud detection.
- Graph fraud detection relies on Graph Neural Networks (GNNs) for information propagation.
- Fraudsters disguise themselves by forging connections with benign users.
- Fraud signals are diluted during neighborhood aggregation.
- Large Language Models provide rich semantic cues for fraud detection.
- Scarcity of annotated fraud samples hinders robust detector training.
- L2IR uncovers latent intent from user behaviors and suspicious connections.
- The framework is proposed to address gaps in existing fraud detection methods.
Entities
—