UniDetect: LLM-Driven Universal Fraud Detection across Heterogeneous Blockchains
New AI system detects fraudulent crypto accounts across multiple blockchains with over 94% accuracy.
A research team led by Shuyi Miao has introduced UniDetect, a novel AI system that leverages large language models (LLMs) to detect fraudulent cryptocurrency accounts across multiple, different blockchains. The core innovation lies in using domain-specific knowledge to guide an LLM to generate standardized, textual summaries of transaction histories from otherwise incompatible blockchain data. These summaries serve as a universal 'evidence' layer, allowing the system to analyze and compare activity patterns across Ethereum, Bitcoin, and other networks. This approach directly addresses a critical weakness in current regulatory frameworks, where illicit funds can be laundered by moving them between chains to evade single-chain monitoring tools.
UniDetect employs a sophisticated two-stage alternating training strategy that continuously refines its ability to reason jointly over both the generated textual evidence and the underlying transaction graph patterns. This multimodal approach allows the model to dynamically adapt and improve its fraud detection capabilities. In rigorous testing, UniDetect demonstrated a significant performance boost, outperforming existing state-of-the-art methods by 5.57% to 7.58% on the Kolmogorov-Smirnov (KS) metric. Most impressively, in cross-chain zero-shot detection scenarios—where the model identifies fraud on a blockchain it wasn't specifically trained on—it successfully flagged over 94.58% of fraudulent accounts. The researchers have made the dataset and source code publicly available, encouraging further development in this crucial area of blockchain security.
- Uses LLMs to create universal transaction summaries from heterogeneous blockchain data, enabling cross-chain analysis.
- Employs a two-stage training strategy for multimodal reasoning, improving fraud detection by 5.57-7.58% over existing methods.
- Achieves 94.58% accuracy in zero-shot detection of fraudulent accounts across unseen blockchains.
Why It Matters
This provides a powerful tool for regulators and exchanges to track illicit crypto flows that jump between blockchains, closing a major security gap.