Abstract:

[Objective] Current mainstream methods for detecting malicious behavior on blockchains rely on graph neural networks and primarily focus on feature aggregation. This paper improves traditional approaches by exploring the threat environment perception in blockchain transactions and leveraging large language models for feature enhancement. [Methods] The proposed Graph Community-Aware Augmentation with Large Language Models model first divides transactions into subgraphs using a community detection algorithm, treating the resulting communities as the environmental context of node addresses. The node features and their associated communities are then converted into textual descriptions, which are processed by a large language model to generate enhanced features. Finally, the original node features and the LLM-enhanced features are fused to form the final node representation. [Results] The model demonstrates an improved ability to capture and learn behavioral characteristics of malicious node addresses, even in scenarios with limited labeled samples. [Conclusions] Experimental results show that the proposed model outperforms traditional algorithms across multiple metrics on two public benchmark datasets.

Key words: Community Detection, Large Language Models, Blockchain