CT-ARF: Detecting Ethereum Fraudulent Accounts with High Recall Rate and Low False Negative Rate Using ARFBoost and CTDA

Abstract

The development of blockchain technology has promoted the growth of cryptocurrencies, but it has also provided criminals with opportunities to engage in fraudulent activities. Currently, Ethereum fraud detection models based on machine learning face challenges such as feature redundancy and class imbalance, which lead to a high false negative rate. To address these issues, this paper proposes an improved CT-ARF detector to enhance the detection of fraudulent activities on Ethereum. This approach combines the ARFBoost feature enhancement mechanism and the CTDA (Conditioned Table Generative Adversarial Network-based) data augmentation strategy. First, we apply Recursive Feature Elimination (RFE) to select the most relevant and representative features from Ethereum transaction data, removing redundant or irrelevant features to improve data quality and obtain an optimal feature subset. Next, we design a multi-dimensional synchronized embedding technique that integrates performance evaluation metrics for each feature into the embedding model, thereby enhancing the contribution of each feature to target identification. Additionally, we use CTDA for data augmentation to address the class imbalance issue, further improving the model’s robustness and accuracy. Finally, we evaluate a CT-ARF detector with the Ethereum fraud detection dataset using LightGBM, XGBoost, CatBoost, AdaBoost, RF, DT, and KNN. The results demonstrate that a CT-ARF detector effectively improves the recall rate of Ethereum fraud detection while reducing the false negative rate.