Communication-Efficient Federated Learning with DRL-Based Parameter Quantization

In internet of vehicles (IoVs), data privacy has received more attention recently and should be protected based on the user requirement. Federated learning (FL), as an effective solution, has been proposed by sharing the parameters of local models to finish the whole data training process. However, its significant communication overhead brings out lots of new challenges for the resource-constrained vehicles. In this paper, we propose a hierarchical FL framework which involves twice aggregations to improve the energy efficiency of FL in IoVs. we also present a deep reinforcement learning (DRL) algorithm to dynamically adjust the upload quantization precision based on training progress and current model accuracy. Experimental results show that our proposed algorithm reduces the energy consumption by 84.9 % and latency by 61.6 % without compromising model accuracy and stability compared to the traditional methods.