PRESERVING FLEXIBLE PERSONALIZED FEDERATED LEARNING ENVIRONMENT WITH LOCAL DIFFERENTIAL PRIVACY

Abstract

Privacy leakage and non-IID data distributions have been a major concern in personalized federated learning. With the advancement in technology, the heterogeneity of devices and user data is increasing. It causes serious privacy issues in federated learning models. One of the major current problems of personalized federated learning is the need for more flexibility in personalization and local data privacy. To ensure privacy at the local level, differential privacy techniques have been used previously. These techniques added noise to all local model parameters to ensure privacy. Adding noise ensures privacy but may cause bad performance and slow convergence issues in federated learning models. Research resolved inflexible personalization concerns by using Fisher's information matrix. The research also solved the model’s slow convergence in a personalized Federated learning Environment by adding noise only to less informative parameters. For this purpose, local differential privacy was applied to preserve local parameters while improving the performance of the FL models. Python and PyTorch were used to do the implementation of the research. The accuracy of this method was evaluated with the help of two different data sets: SVHN and MNIST. According to the findings of the experiments on SVHN, the performance has undergone an excellent improvement, while maintain utility and privacy. On MNIST dataset, the mean accuracy improved and showed excellent results while maintaining privacy of data.