Weight Coefficient Based Adaptive Federated Learning for Vehicular Data Transmission

With the ever-increasing amount of vehicle data being generated, the collection and transmission of this data-to-data processing centers is consuming significant amounts of communication resources. The traditional method of compressing and transmitting the vehicle data is not effective in addressing the issue of efficient utilization of this data. In order to overcome this challenge, we propose an adaptive federated learning approach that avoids the need for transmitting data per vehicle. Our approach leverages the vehicle as a distributed training device node and enables the training of vehicle data using the vehicle's own computing power, thereby eliminating the need to transmit the data over the network. To further enhance the efficiency of the federated learning aggregation calculation, we introduce the information entropy function and cosine similarity calculation. By computing the similarity between the model and the benchmark model, we are able to give a new round of model aggregation calculation weight. Finally, we validate the proposed algorithm using the actual MNIST dataset, demonstrating its high effectiveness.