Federated deep learning and explainable AI for real-time credit card fraud detection in highly imbalanced transaction streams

The growing complexities of credit card fraud, along with stringent data privacy laws and highly skewed streams of transactions, pose serious challenges to real-time fraud detection systems. This paper aims to solve these problems by suggesting a Federated Deep Learning (FDL) system with Explainable Artificial Intelligence (XAI) to ensure accurate, privacy-preserving, and interpretable fraud detection. The structure has a hybrid BiLSTM-CNN to efficiently describe temporal transaction behaviour and spatial feature associations. In order to reduce the severe class imbalance, a Generative Adversarial Network (GAN) -based augmentation method is proposed, which allows the model to learn realistic patterns of frauds not just in the traditional approaches of oversampling, such as SMOTE. The federated learning model enables a group of financial institutions to collectively train a world model without exchanging raw data containing sensitive information, in compliance with guidelines like GDPR and PCI DSS. Moreover, SHAP and LIME-based explainability modules are included to offer insights at the transaction level, enhancing auditability and trust among analysts. The experimental evaluation is carried out to make it clear and reproducible and with the Enhanced CreditCard₁0k dataset which is a controlled benchmark to test fraud detection under imbalanced conditions. The scenario about a large scale 2. 3 M + transaction is a simulated situation with many institutions to demonstrate scalability and applicability to real life. Nevertheless, the entire quantitative performance outcomes have been obtained in this study using only the Enhanced CreditCard₁0k dataset. Experimental findings show that the model has a high accuracy of 98. 7, a high recall rate of 95, and a low false positive rate of 0. 3, which is much better than the traditional and centralized base lines. Latency analysis: The system can ensure real-time inference performance (< 300 ms), including overhead of explainability. On the whole, the suggested solution can adequately balance accuracy, scalability, preserving privacy, and interpretability, which is why it is highly suitable to implement it in the contemporary distributed financial system.