Complex credit card stealing schemes are further posing an uphill challenge to financial institutions since they are no longer restricted to organizational boundaries. Fraud detection is an efficient process which needs various transaction data access; legal, ethical and regulatory standards do not allow data sharing between the banks. The proposed study presents a federated learning (FL) model of credit card fraud detection across several banks without the aggregation of data in a centralized location. The framework allows individual banks to train a global model of detecting fraud collectively without sharing details about the data. The key tools applied in the system are federated aver-aging, secure parameter exchange, and local model updates. The elements enable learning across institutions without breaching privacy conventions. These elements manage to stop the exposure of the raw data, but maximize the level of detection leading to a comparative assessment of this method and a traditional centralized deep neural network in a paper that establishes elevated privacy standards, and a non-existent performance sacrifice. Experiments simulate the pattern of transactions using synthetic, but representative data which has been constructed with embedded anonymization methods. Through eight key measures, the accuracy, the false positive rates, and computational efficiency are explored and compared. The model provided results in a significant boost in collaborative fraud detection capacity without violating legal constraints and consumer trust. It solves the urgent need of privacy-preserving, scalable and high-accuracy credit card fraud detection in a distributed setting where it is either not feasible or even illegal to share data.