Detecting credit-card fraud is still a challenge because transactions are heavily imbalanced and fraud patterns evolve very quickly. In this paper, an attention-enhanced neural network-based scheme for credit card fraud detectionCredit card fraud detection is proposed, which consists of adding two small lightweightLightweight multihead attention blocks to a standard fraud detection multilayer perceptronMultilayer perceptron (MLP) (MLP) (referred to as Baseline MLP Scheme). In our approach, a reproducible pipeline is used to download the data from the public IEEE-CIS Fraud Detection datasetDataset, apply label encoding, standardization, and Synthetic Minority Oversampling Technique (SMOTESynthetic Minority Over-sampling Technique (SMOTE)) rebalancing, then train both the baseline scheme and the proposed one on identical splits with early stopping. The attention-enhanced neural network-based scheme reaches an AUC of 0.986, an average precision of 0.985, and a recall of 0.91, surpassing the baseline scheme (with an AUC of 0.971, an average precision of 0.970, and a recall of 0.81). Training time is found to double on a single CPU core, but memory increases only by \(11\%\) , making our proposed method practical for real-time risk engines.