Consumers and financial institutions all across the world struggle with financial fraud on a regular basis. Every year, they lose billions of monies. As a result, it's critical to have an efficient Fraud Detection System (FDS) to reduce loss for both customers(users) and financial institutions. Utilizing machine learning algorithms, which aid in future prediction and pattern recognition by analyzing vast amounts of data, is a typical method of detecting fraud. A big dataset is necessary to obtain a well-performing model, yet datasets have the drawback of being skewed. i.e., samples of fraudulent transactions are far less common than samples of honest transactions. Furthermore, banks and other financial organizations often are not permitted to disclose their transaction data due to the data privacy and security connected with transaction datasets. When these issues are combined, it is challenging for the centralized FDS to identify fraud tendencies. In this thesis, present a framework for federated learning, a machine learning environment where numerous entities cooperate to solve a machine learning issue under the supervision of a central server or service provider, to train a fraud detection model. With this strategy, financial institutions may profit from a common model that has witnessed more fraud than each bank individually while avoiding sharing the dataset. As a result, the user’s sensitive information is safe guarded. Thesis's findings suggest that when it comes to identifying financial fraud, the federated model (Federated Averaging) may match or even exceed the central model (Multi-Layer Perceptron).

Fraud detection, credit card, financial transaction, cashless transaction, fraudulent transaction, machine learning, deep learning