Communication-Efficient and Privacy-Preserving Aggregation in Federated Learning With Adaptability

Federated learning (FL) aims to protect data privacy while aggregating models. Existing works rarely focus simultaneously on the three issues of communication efficiency, privacy, and utility, which are the three main challenges facing FL. Specifically, sensitive information about the training data can still be inferred from the model parameters shared in FL. In recent years, differential privacy (DP) has been applied in FL to protect data privacy. The challenge of implementing DP in FL lies in the detrimental impact of DP noise on model accuracy. The DP noise affects the convergence of the model, leading to additional communication overhead. Moreover, considering the inherently high-communication costs of FL, FL process can be inefficient or even infeasible. In view of these, we propose a novel differentially private FL (DPFL) scheme named Adap-FedITK, which aims to achieve low-communication overhead and high-model accuracy while guaranteeing client-level DP. Specifically, we dynamically adjust the gradient clipping threshold for different clients in each round, based on the heterogeneity of gradients. This approach aims to mitigate the negative impact of DP and achieve a privacy-utility tradeoff. To alleviate the high-communication overhead problem in FL, we introduce an improved top-k algorithm, which utilizes sparsity and quantization to compress the model eliminates communication redundancy, and it also integrates coding techniques to further reduce communication. Extensive experimental results demonstrate that our method achieves the privacy-utility tradeoff and improves communication efficiency while ensuring client-level DPFL.