User Selection for Federated Learning in a Wireless Environment: A Process to Minimize the Negative Effect of Training Data Correlation and Improve Performance

This article describes our study about federated learning by using the front view-based training process in autonomous driving as an example. The overlap of camera views among users creates training data correlation. It is shown that this correlation undermines the optimization performance of the federated learning process. To reduce the negative effect of correlation within training data while limiting wireless channel resource utilization, a federated learning protocol with user selection is proposed. The protocol selects a limited number of users with sufficient spatial separation in each round of federated learning. An exclusion zone is applied to maintain separation during user selection. Experimental results show that in an example deployment scenario with a user density of 0.05, applying a discrete exclusion zone (DEZ) to prevent selecting the first three nearest users and applying a geographical exclusion zone (DEZ) to avoid selecting users within 70 m have equivalent effects on reducing training data. Furthermore, both methods can provide the performance of five-user federated learning approaches, an ideal case without correlation in the training data.

This article describes our study about federated learning by using the front view-based training process in autonomous driving as an example. The overlap of camera views among users creates training data correlation. It is shown that this correlation undermines the optimization performance of the federated learning process. To reduce the negative effect of correlation within training data while limiting wireless channel resource utilization, a federated learning protocol with user selection is proposed. The protocol selects a limited number of users with sufficient spatial separation in each round of federated learning. An exclusion zone is applied to maintain separation during user selection. Experimental results show that in an example deployment scenario with a user density of 0.05, applying a discrete exclusion zone (DEZ) to prevent selecting the first three nearest users and applying a geographical exclusion zone (DEZ) to avoid selecting users within 70 m have equivalent effects on reducing training data. Furthermore, both methods can provide the performance of five-user federated learning approaches, an ideal case without correlation in the training data.