A Federated Learning Algorithm Based on Combination of Prototype and Comparative Learning

Federated learning is a modern distributed machine learning paradigm that has garnered a lot of interest because of its benefits for collaborative model training and data privacy. However, federated learning is confronted with many challenges, including client data distribution heterogeneity and model drift. These factors place greater demands on model optimization and generalization capabilities. Traditional federated learning methods typically rely on gradient exchange and aggregation of global model parameters. This paper puts forward a privatized federated learning algorithm that combines prototype federated learning and model comparison federated learning, aiming to use prototype exchange and comparison learning techniques to increase the federated learning models' training effectiveness and classification accuracy. This algorithm significantly reduces the communication overhead by exchanging abstract class prototypes between the client and the server. Additionally, it uses comparison learning to enhance the loss algorithm, which improves the model's capacity for generalization. This, in turn, reduces the representation drift between the local and global models, thus improving the learning effectiveness of the model.