Hermes: Rule-Driven Adaptive Gradient Level Quantization in Federated Learning for Internet of Vehicles

Abstract

Federated learning can collaboratively train machine learning models by utilizing model parameters, ensuring that vehicle nodes do not share original data, thereby enhancing data interaction safety in complex traffic environments. However, it encounters two key challenges: high communication cost of federation aggregation due to frequent updates of model parameters in Internet of vehicles; and low model aggregation accuracy due to the non-adaptive gradient quantization strategy. To this end, we propose Hermes, which first calculates the gradient variance of vehicle nodes to dynamically adjust the quantization level, and then utilizes the scalar of the quantization level to alleviate the ineffective quantization phenomenon of gradient quantization to zero, followed by employing an error feedback mechanism to guarantee the integrity of the model. Experimental results show that, compared to the other baselines, Hermes reduces upload traffic to approximately 2.94% of the full-precision transmission volume; meanwhile, the average accuracy of the aggregate model is improved by 2.49%. Besides, Hermes can control the quantization relative error within 0.05%.