Credit card fraud has grown increasingly common in today’s age, and with the rise in cybercrimes with fraud, several examples have been recorded in the past. The use of a distributed search plays a pivotal role in enhancing the performance of fraud detection systems. By enabling the aggregation and retrieval of data from multiple decentralized sources without compromising data privacy, it facilitates the efficient training of models on large, diverse datasets. Another technique used in controlling fraud losses is through applying Federated Learning (FL) for detecting fraudulent transactions. The models may gain the benefit of the dispersed data without actually sharing the data in this way. This paper focuses on the implementation of CNN with FL to improve a security and accuracy of fraud detection in financial transactions. The suggested model employs Kaggle credit card fraud dataset and uses enhanced techniques such as; SMOTE to work on class imbalance problem and one hot encoder to work on Categorical features. The proposed CNN-FL model surpassed other classifiers and yielded better accuracy, precision, and recall rates compared to traditional ML classifiers like NB, LR, and Gaussian Naive Bayes; accuracy $\mathbf{9 9. 8 6\%}$, precision 99.83%, recal 199.85%, and an F1-score of $\mathbf{9 9. 84\%}$. Thus, effectiveness of suggested CNN-based federated learning approach for enriching fraud detection systems is shown, with good generalisation and high accuracy on various types of transactions.