An interpretable ensemble model framework for real-time anomaly detection and prediction of Ethereum blockchain transactions

The blockchain ecosystem is often referred to as a technology that ensures security. However, there have been concerns in the real world regarding the security of blockchain applications, like what happens with conventional database systems. The anonymous design of blockchain provides cyber-attackers with opportunities to commit crimes, resulting in an increase in scams, phishing, code manipulation of smart contracts, Ponzi schemes, and other fraudulent activities. Consequently, many individuals and national economies worldwide have suffered significant losses. Detecting fraudulent behavior in blockchain transactions manually is infeasible due to the enormous amount of data involved. Therefore, the optimal method for identifying abnormalities within the blockchain network is to combine a blockchain platform with a machine learning approach. This study employs filter method techniques such as mutual information (MI), analysis of variance (ANOVA), and recursive feature elimination (RFE) to identify the ideal set of features based on the maximum accuracy value, considering the feature dimension (k value). The study screens and ranks the top 10 feature sets using the feature importance random forest (RF) classifier, based on the dataset produced by the best filter approach (yielding higher accuracy). Subsequently, an ensemble methodology is used to create the final model, utilizing the final dataset consisting of 10 features. The purpose of this approach is to enhance the level of anomaly detection in the blockchain network. To determine the effectiveness of the proposed model, experiments are conducted, comparing it against individual classifiers such as extreme gradient boosting (XGB), decision tree (DT), logistic regression (LR), random forest (RF), and k-nearest neighbor (KNN). The study's findings reveal that the ensemble voting approach achieves a 96.78% accuracy rate, surpassing the accuracy of the individual classifier models that utilize optimal features. Additionally, the study's findings suggest that the selection of features and their quantity significantly impact the output of the model.