The rapid expansion of the cryptocurrency market has led to soaring transaction volumes and heightened transaction-related risks, necessitating robust data-driven security analysis to measure and mitigate emerging threats. Detecting abnormal activities such as fraud, money laundering, and market manipulation remains challenging due to the anonymous and decentralized nature of cryptocurrency transactions, which complicates the measurement of suspicious behavior. Traditional anomaly detection methods often struggle with insufficient labeled data and limited accuracy, hindering effective security analysis in this domain. To address these challenges, this paper proposes a novel data-driven framework for cryptocurrency transaction anomaly detection that combines semi-supervised learning with Graph Attention Networks (GAT). By leveraging labeled data to train a random forest classifier and iteratively predicting labels for unlabeled data, the framework expands the training dataset, mitigates data imbalance issues, and enhances GAT’s ability to interpret graph-structured transaction data. This approach not only improves detection accuracy but also provides actionable insights for measuring subtle patterns in transaction behavior, which are critical for identifying security risks. The framework, applied to fraud detection and behavior analysis, offers a scalable solution for analyzing and securing cryptocurrency ecosystems.