Beyond siloed aggregation: An adaptive federated reinforcement learning model with multi-level knowledge distillation against evolving financial fraud

Credit card fraud detection (CCFD) is increasingly challenged by extreme class imbalance, non-IID data distributions across institutions, and rapidly evolving attack patterns. To address these issues, we present BSAFD, an adaptive federated reinforcement learning model with multi-level knowledge distillation in financial fraud detection, combining four synergistic components: a kernel-guided adversarial representation learning module that uses a compact encoder–decoder backbone with adaptive kernel sampling and adversarial augmentation to synthesize high-quality minority-class transactions and produce robust embeddings; hierarchical multi-level knowledge distillation that aligns each client's local classifier with the global model via logit-level soft labels and feature-relation alignment to transfer output confidence and preserve inter-sample geometry; PPO-based federated reinforcement learning that constrains local updates through clipped likelihood ratios to stabilize asynchronous gradient aggregation across heterogeneous participants; and label-driven federated fusion that groups clients by fraud-rate profiles and fuses their distilled feature representations into a unified classifier. Extensive experiments on six real-world fraud datasets demonstrate that BSAFD consistently outperforms ten state-of-the-art baselines in AUC, F1 score, and average precision.