Data Collaborative Federated Learning for Non-i.i.d Data in Wireless Networks

Federated Learning (FL) has been advocated to solve the problem of data privacy. Nonetheless, the model’s performance suffers degradation when it comes to the non-independent and identically distributed (non-i.i.d) data in wireless networks. This poses a great challenge for deploying FL to wireless networks. To overcome the obstacle of non-i.i.d data in FL, we propose a novel framework in this paper named Data Collaborative Federated Learning (DCFL). The DCFL differentiates from FL that it utilizes dimensional reduced data to fine-tune the trained model by data collaboration, which significantly promotes the model performance. Moreover, a cluster-DCFL framework that divides users into multiple clusters is further proposed to alleviate the excessive communication burden in DCFL when users are scattered over the wireless networks. The experimental results on the MNIST, CIFAR10, and RadioML datasets show that DCFL outperforms several FL baselines including FedAvg, FedProx and a finetuned FL (FLFT). Meanwhile, the cluster-DCFL further reduces the communication burden of DCFL in the specific experiment.