CLFLDP: Communication-efficient layer clipping federated learning with local differential privacy

Privacy preserving is a severe challenge in machine learning and artificial intelligence. Recently, many works have been devoted to solving this problem by proposing various federated learning frameworks and introducing local differential privacy. However, applying local differential privacy to federated learning has lower utility after perturbing the parameters. Therefore, to improve the accuracy and communication efficiency of the model while having strict privacy preserving, we propose CLFLDP, a communication-efficient layer clipping federated learning model with differential privacy preserving. First, a novel adaptive privacy budget allocation scheme is proposed to allocate the privacy budget based on the communication rounds and client relevance which can reduce the loss of privacy budget and the size of model noise. Second, a layer-based top k parameter selection method and aggregation scheme are proposed. The communication cost of the system is reduced by uploading layers with higher client-side relevance and excluding layers with lower relevance. Therefore, the proposed framework can achieve better balance between privacy preserving, communication efficiency and model accuracy of federated learning. Theoretical analysis and experiments on various commonly used image datasets demonstrate the superiority of our framework over the state-of-the-art federated learning frameworks.