Design of an Improved Model for Blockchain Forensics Using Graph Convolutional Networks and Variational Autoencoders

Forensics is more needed to address the increasing sophistication of fraudulent activities within the ever-transacting and opaque nature of blockchains, where advanced and interpretable anomaly detection should ideally be present. Current methods struggle to scale with and handle the graph-structured nature of data in blockchains, usually failing in the interpretability that's necessary for either trust or accountability. The current solutions on anomaly detection are mostly heuristic in nature or at best a black-box machine learning model that offers minimal interpretability. Due to this fact, the stakeholders cannot understand or trust such results. In view of these limitations, this work proposes the development of explainable deep learning models with special appeal in blockchain forensic analysis. Some of the particular options considered are GCNs, VAEs, LSTMs, and a hybrid model of GCN- LSTM. GCNs are suitable to capture the graph structure present in blockchain data, while VAEs can detect anomalies through reconstruction error in high-dimensional spaces. LSTM networks are chosen for their capability to model temporal dependencies in blockchain transactions, while the hybrid GCN-LSTM model combines both graph-based and temporal information in order to improve the accuracy of detection. These models are integrated with state-of-the-art explainability techniques like SHAP, LIME, GNN Explainer, and Integrated Gradients that provide transparent insight into model decision-making processes. SHAP values give feature-level explanations, LIME provides local model interpretability, and GNN Explainer underlines the critical graph structure. Beyond that, these methods confer interpretability on the results of anomaly detection and further build trust in the system. Indeed, the performance of the proposed models was excellent: for GCN, the Fl-score ranged between 0.80 to 0.90, for VAEs it was between 0.85 to 0.90 as measured with AUROC, and for a hybrid GCN-LSTM model, precision and recall up to 0.90 was reached. It greatly enhances the capability of detection and explanation of anomalous activities in blockchains and opens the perspective towards more robust, transparent, and trustworthy forensic analysis tools.