Research of Financial Transaction Model Based on Federated Learning Algorithm

As data storage across multiple devices becomes more complex, there is a growing need for efficient and secure ways to train models on decentralized personal data. Traditional centralized training methods, which require aggregating data from various sources, can compromise user privacy and data security. This report compares two federated learning algorithms: the widely-used FedAvg and the more recent pFedMe. FedAvg trains a global model by aggregating local updates from multiple clients. pFedMe improves on FedAvg by using Moreau envelopes for better personalized model training. Experiments with the MNIST dataset show that pFedMe outperforms FedAvg in accuracy and convergence speed.