Improved Ethereum Fraud Detection Mechanism with Explainable Tabular Transformer Model

Blockchain technology has gained popularity due to its key features of decentralization, cryptographic verification, and immutability, which have proven extremely useful in various industries. However, despite their impressive security features, blockchain networks are not immune to cyber threats. In recent times, the blockchain system has been threatened by fraudulent attacks that require quick responses. Machine learning and deep learning models are increasingly leveraged to address these challenges. However, due to their black box nature, these models lack transparency, which is a major criticism. This study presents an approach to enhancing fraud detection mechanisms on Ethereum. This study presents an efficient and transparent fraud detection system on Ethereum known as IFS-TABPFN. An interpretable feature selection approach based on Shap values and optimized gradient boosting was introduced to develop five deep learning models built on neural networks. These models included Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), Convolutional Neural Networks and Long Short-Term Memory (CLSTM) and Tabular Prior-Data Fitted Network (TabPFN). A comparative analysis of our results indicates that IFS-TABPFN achieves 99.2% accuracy in just a few seconds, outperforming other neural networks and existing systems. This study highlights the importance of explainable AI in understanding how features influence decisions, performance and contribute to artificial intelligence models' transparency and trust.