Cryptocurrency networks face escalating illicit activities that exploit blockchain’s pseudo-anonymous nature, costing billions annually. We present RADIAL (Robust Adversarial Discrepancy-aware Illicit Account Locator), a framework addressing four critical challenges in illicit account detection through key innovations: Edge2Seq+ for transaction sequence modeling, graph transformer layers with discrepancy-aware message passing, adversarial training, and self-supervised pre-training. Experiments on four large-scale datasets demonstrate RADIAL’s superiority over 15 state-of-the-art methods, achieving F1 scores of 94.19–97.98% across Bitcoin and Ethereum networks. The framework maintains above 90% F1-scores under targeted adversarial attacks and requires only 50% of transaction history for accurate early detection. With sub-millisecond inference time per node, RADIAL enables financial institutions to deploy effective anti-money laundering systems that balance security with computational efficiency. To ensure reproducibility, we release our code, preprocessed datasets, and model checkpoints (RADIAL is available at https://github.com/anonymous/RADIAL ); with detailed documentation and scripts to replicate all experiments.