Enhancing Ethereum Fraud Detection Accuracy with Sparse-Attention-Based Model

Abstract

As decentralized finance (DeFi) expands, Ethereum’s role as the backbone for digital asset exchange, smart contracts, and financial protocols has grown—but so has its exposure to fraud. Phishing, money laundering, and malicious contracts exploit its openness. Existing ML and deep learning models often lack the balance between speed, accuracy, and explainability needed for real-time blockchain analysis. High-performing models like transformers are accurate but too resource-heavy and opaque. This study leverages TabNet—a sparse-attention deep learning model optimized for tabular Ethereum transaction data. It dynamically selects relevant features during training, enhancing both interpretability and efficiency. With an accuracy of 0.86, precision of 0.80, and F1-score of 0.79, TabNet outperforms traditional models in fraud detection while remaining lightweight and transparent. Its feature-level insights make it ideal for environments where trust, latency, and transparency are crucial. The results position TabNet as a scalable, practical alternative for fraud detection in blockchain ecosystems.