In recent years, credit card transaction fraud has inflicted significant losses on both consumers and financial institutions. To address this critical issue, we propose an optimized framework for fraud detection. This study deals with non-identically independent distributions (IIDs) involving different numbers of clients. The proposed framework empowers banks to construct robust fraud detection models using their internal training data. Specifically, by optimizing the initial global model before to the federated learning phase, the suggested optimization technique accelerates convergence speed by reducing communication costs when moving forward with federal training. The optimization techniques using the three most recent metaheuristic Optimizers, namely: An improved gorilla troops optimizer (AGTO), Coati Optimization Algorithm (CoatiOA), Coati Optimization Algorithm (COA). Furthermore, credit card data is highly skewed, which makes it challenging to predict fraudulent transactions. The resampling strategy is used as a preprocessing step to improve the outcomes of unbalanced or skewed data. The performance of these algorithms is documented and compared. Computation time, accuracy, precision, recall, F-measure, loss, and computation time are used to assess the algorithms' performance. The experimental results show that AGTO and (CoatiOA) exhibit higher accuracy, precision, recall, and F1 scores compared to the baseline FL Model. Additionally, they achieve lower loss values.