Recent research has focused on effectively combining blockchain with graph neural networks, utilizing the efficient computing methods of graph neural networks to promote the application and development of blockchain technology in more fields. However, traditional GNN methods only focus on the neighborhood information around the account, ignoring global topology and high-order neighborhood information. Additionally, the importance of different fraudulent accounts with the same neighborhood cannot be distinguished during the aggregation process. To address these above challenges, we propose a research model for anomaly detection in cryptocurrency transactions based on Chebyshev graph neural networks. Specifically, we use the Chebyshev kernel function to obtain high-order neighborhood information and differentiate the importance of accounts with the same neighborhood through the Laplace operator. We demonstrate through lemma that unreasonable coefficients lead to Chebyshev networks being inferior to other methods and penalize the learned outlier coefficients to mitigate the impact on higher-order neighborhood approximations. Subsequently, we aggregate multi-layer neighbor information through attention mechanisms to achieve comprehensive information integration. The effectiveness of the model has been verified through multiple experiments on the Elliptical dataset.