A Novel Time Series Approach to Anomaly Detection and Correction for Complex Blockchain Transaction Networks

The rapid rise in blockchain technology’s popularity has prompted numerous models to analyze patterns and detect anomalies in blockchain networks based on transaction history. However, most existing studies overlook transactions’ dynamic, nonlinear, and time-variant nature in a time series context. This paper introduces an innovative methodology for enhancing blockchain network performance through advanced time series analysis, anomaly detection, and correction. We propose a hybrid deep learning model integrating Long Short-Term Memory (LSTM) networks, Multi-Head Attention (MHA), and Fully Connected Network (FCN) layers to predict transaction volumes in blockchain networks. The LSTM network captures both short-term and long-term dependencies in blockchain time series data, while the MHA mechanism focuses on relevant input sequence segments. FCN layers perform final feature processing and map the output to predicted transaction volumes. To address overfitting, a Dropout layer is added between the FCN layers. Anomalies are identified using Gaussian Mixture Models (GMM) and corrected via Gaussian Process Regression (GPR). Applied to real-world blockchain transaction datasets, our methodology demonstrates superior efficacy in detecting and correcting anomalies, yielding a more accurate representation of the network’s true behavior. This leads to improved estimates of average and peak throughput and network volatility.