Abstract
Bank card fraud is one of the biggest challenges in digital finance space, which needs detection models to address class imbalance, interpretability, and adaptability to changing tactics of fraud. The paper proposes a neuro-adaptive architecture established on a highly structured preprocessing pipeline with stratified splitting, feature normalisation, and representation learning via a Denoising Autoencoder. At the core of this framework lays an Artificial Neural Network optimised by the Firefly Algorithm for fast hyperparameter tuning facilitated by Elastic Weight Consolidation that promotes continual learning without sacrificing past performance. The proposed Adaptive ANN + FA outperforms baseline ANN, CNN, and LSTM models mainly in F1-Score, precision, and recall-the main metrics in fraud detection. Also, SHAP breaks out feature contribution and prediction reasonability making the results very transparent. Optimised adaptive and explainable models are positioned here as strong enablers of real-world fraud discovery and subsequent robustness in the financial systems.