An intelligent cybersecurity framework for banking systems: integrating neural networks, large language models, and federated learning for anti-money laundering and fraud detection

Abstract The fraud of money laundering costs the global financial system USD 800 billion to USD 2 trillion annually, while digital banking contributes to the increasing number and complexity of money laundering transactions. If there are adversarial forces that are constantly adapting their approach to avoid complying with a conventional rule-based compliance program, that isn’t going to work. In this paper, we present a comprehensive, layered cybersecurity framework for banking systems, combining graph neural networks (GNNs), long short-term memory (LSTM) architectures, large language models (LLMs) and federated learning, aimed at tackling anti-money laundering (AML) and transaction fraud detection on a massive scale. The framework is based on the theoretical constructs of agency theory, information asymmetry theory, and routine activity theory and is situated in the context of a formal mathematical model that encodes the temporal dependencies of transactions, the topology of the relationships between transactions, and the semantic context of their behavior. This multistage detection pipeline involves supervised classification, unsupervised anomaly detection, regulatory narration with the assistance of LLM and an explainable-AI (XAI) layer for regulatory transparency. It has been developed and designed to operate in a federated manner, which means that the data can be shared among different institutions without the risk of it being shared with the raw data, thus ensuring data privacy in line with GDPR provisions. Benchmarks with publicly available AML datasets show that the proposed hybrid model outperforms the standalone rule-based model and single-model baseline with an area under the ROC curve (AUROC) of 0.961. The framework gives a hands-on and theory-based guide for intelligent, trustworthy financial crime detection to address regulators, compliance officers and data scientists’ evolving needs.