Interpretable Federated Learning for Fraud Detection in Non-IID Blockchain Transactions

Detecting fraud in decentralized blockchain transactions presents significant challenges due to privacy constraints, scalability issues, and the need for interpretability in decision-making. Traditional machine learning and deep learning approaches rely on centralized data aggregation, which compromises privacy and introduces single points of failure. Federated Learning (FL) offers a decentralized, privacy-preserving alternative but struggles with highly non-IID data distributions and lacks built-in interpretability. To address these limitations, this study introduces a novel framework that integrates FL with Explainable AI (XAI) to enable secure, scalable, and transparent fraud detection in blockchain transactions. Our contributions include the development of federated optimization strategies specifically designed for non-IID environments, ensuring fairness and efficiency across diverse client nodes. Experimental results demonstrate that the Tabular Transformer model with the Adam optimizer achieves over 99% accuracy, precision, recall, and F1 score under IID settings. In contrast, a 2-layer Neural Network model trained with SGD shows superior robustness under non-IID conditions, achieving 96.24% to 98.98% accuracy across both partial and full client participation with minimal computational cost. XAI techniques such as SHAP and LIME further enhance the system’s transparency by identifying token-specific statistical features (e.g., maximum, minimum, average, and uniqueness of ‘ERC20’ tokens) as primary contributors to model decisions. This work highlights the promise of combining FL and XAI for developing privacy-preserving and interpretable fraud detection systems and sets the foundation for future research on lightweight, ethical AI solutions in decentralized, resource-constrained environments.