Let Relations Speak: An End-to-End LLM-GNN Soft Prompt Framework for Fraud Detection

Researchers propose LGSPF, an LLM-GNN framework using soft prompts to improve fraud detection without relying on textual data. The method combines language models with graph neural networks to capture multi-relational complexity in fraud patterns, achieving state-of-the-art results across benchmarks.

This research addresses a fundamental challenge in applying large language models to fraud detection: the scarcity of rich textual attributes in transaction and relationship data. Traditional approaches either force-fit graph structures into text through hard prompts—risking information distortion—or struggle with multi-relational complexity inherent in fraud schemes. LGSPF introduces soft prompts as a bridge between graph topology and semantic understanding, allowing the LLM to process structural relationships directly without lossy textualization.

The framework's dual-encoder architecture represents a meaningful evolution in fraud detection methodology. By running GNN and LLM components in parallel, with end-to-end optimization, the system achieves semantic alignment between structural pattern recognition (GNN's strength) and contextual reasoning (LLM's strength). This approach is particularly relevant for cryptocurrency and financial crime prevention, where transaction graphs exhibit complex relational patterns across multiple entity types—exchanges, wallets, users, and temporal sequences.

For the broader security landscape, this work suggests that hybrid AI approaches combining symbolic (graph) and semantic (language) processing outperform single-modality systems. Crypto platforms and fintech companies face increasing pressure to deploy sophisticated anti-fraud measures that scale with transaction volume while maintaining interpretability for compliance teams. LGSPF's demonstrated improvement in semantic interpretability addresses a critical regulatory requirement: explaining why a transaction was flagged.

The framework's validated performance across diverse benchmarks suggests practical deployment potential. Future development likely focuses on real-time inference optimization and integration with existing blockchain monitoring infrastructure, which could materially improve detection rates for sophisticated fraud rings that evolve tactics to evade simpler rule-based systems.

• →Soft prompt methodology eliminates reliance on textual data, addressing a key limitation in applying LLMs to graph-based fraud detection
• →Parallel GNN-LLM architecture captures both structural patterns and semantic context for superior fraud comprehension
• →State-of-the-art benchmark results suggest practical viability for deployment in financial and cryptocurrency platforms
• →Enhanced semantic interpretability of fraud behaviors supports regulatory compliance and explainability requirements
• →Framework advances hybrid AI approaches combining symbolic and semantic processing for security applications