Authors:
Tanaka Kenta, Sagarika Mishra, Y. Anantha Vishwa Priya, Aditya Nawle, Noe Hernández Hernández, Syamsu Rijal

Addresses:
Department of Sales, Fuji Trading Co., Tokyo, Japan. Department of Artificial Intelligence and Machine Learning, SRM Institute of Science and Technology, Ramapuram, Chennai, Tamil Nadu, India. Department of Engineering and Technology, Autonomous University of Tlaxcala, Tlaxcala, Mexico. Faculty of Economics and Business, Universitas Negeri Makassar (UNM), Kota Makassar, Sulawesi Selatan, Indonesia.

Abstract:

Temporal distribution shifts, class imbalance, and adaptive adversarial approaches make it hard to detect blockchain corruption. Researchers demonstrate that full-batch GraphSAGE training fails to achieve F1 = 0.000 under class imbalance, inspiring every component of our proposed solution. The five-stage FAITH-AML (Failure-Aware Intelligent Temporal Hybrid AML) requires architectural decisions to address failure modes explicitly. The framework uses a novel ResHybridSAGE-GRU architecture to resolve gradient domination and sparse-node failures, Population Stability Index monitoring to detect temporal drift without ground-truth labels, REINFORCE-based adversarial simulation to quantify the evasion surface, and a hybrid XGBoost meta-model to decide that GNN achieves validation F1 of 0.6962, XGBoost 0.6887, and the hybrid meta-model 0.7760 on the Elliptic Bitcoin dataset with a strict out-of-time split, surpassing both models. Operational AML systems require failure-aware, production-ready, reliability-focused design, according to this study. Researchers believe this is the first failure-aware AML framework to incorporate representation, monitoring, and dependability. The hybrid architecture indicates that real-world deployment prioritizes durability, calibration, and interpretability over prediction performance. Per-timestep analysis shows that performance decline reflects distributional drift, validating PSI as an early warning method. The reliability-aware meta-model provides selective automation and human review for risk-aware decision-making by distinguishing high-confidence predictions from uncertain scenarios. This shows that a unified system with structural learning, drift monitoring, and reliability estimates is more resilient and operationally viable than standalone anti-money laundering models.

Keywords: Anti-Money Laundering (AML); Concept Drift; Population Stability Index; Elliptic Dataset; Adversarial Machine Learning; Failure-Aware Systems; Hybrid Meta-Learning.

Received on: 02/10/2025, Revised on: 25/11/2025, Accepted on: 16/01/2026, Published on: 19/08/2026

DOI: 10.69888/FTSFDS.2026.000741

FMDB Transactions on Sustainable Finance and Data Science, 2026 Vol. 1 No. 3, Pages: 145-161

• Views : 33
• Downloads : 6

Download PDF