This paper introduces EAGLE (Ensemble Adaptive Graph Learning for Enhanced Ethereum Fraud Detection), a framework for detecting fraudulent transactions across Ethereum stablecoin networks. EAGLE addresses limitations in existing approaches through an ensemble architecture that integrates specialized token-specific models with a cross-token graph neural network component, augmented by temporal-spatial attention mechanisms and confidence calibration. Experimental evaluation on 36.7 million transactions demonstrates strong performance, achieving 91.3% precision and 90.0% F1-score an improvement of 8.2% in precision over current methods. Our GPU-accelerated implementation processes 9,763 transactions per second, though requiring substantial computational resources. The system maintains 85.9% F1-score under adversarial perturbations ( \(\epsilon =0.2\) ), showing robust performance against evasion attempts. Comprehensive ablation studies quantify each component’s contribution, with the integrated approach outperforming individual models by 5.5% in F1-score. While focused on multi-token interactions within Ethereum, EAGLE advances detection of complex fraud spanning stablecoin networks. Deployment and ethical considerations, along with ground truth labeling methodology, are briefly discussed.