Blockchain technology has transformed data management by offering a secure and decentralized platform for transactions. However, its inherent features, such as pseudonymity, can also enable illicit activities such as money laundering and fraud. Detecting anomalous transactions within these complex, high-dimensional networks remains a significant challenge. In this paper, we present Improved HybridNet, a sophisticated hybrid graph neural network tailored for identifying irregular transactions in Bitcoin networks. Improved HybridNet combines GraphSAGE convolutional layers, which aggregate node features via cosine similarity to capture local structures, with GATv2 convolutional layers, which utilize attention mechanisms to highlight important neighboring nodes. Furthermore, the model integrates temporal encoding and node degree features to recognize dynamic transaction patterns and structural connectivity. Extensive experiments on the Elliptic Bitcoin dataset show that Improved HybridNet achieves an F1 score of 0.9329, significantly surpassing traditional GNN models in terms of precision, recall, and accuracy. These results highlight the model’s effectiveness in strengthening the security and integrity of blockchain-based financial systems.