Exploring traffic pattern variability in vehicular federated learning

The emergence of software-defined vehicles has brought machine learning into the vehicular domain. To support these data-driven applications, techniques to incentivize users to share their vehicle data are crucial. Federated learning trains machine learning models in a distributed manner, leveraging client data without compromising its privacy. Nonetheless, in vehicular networks, the dynamic behavior of nodes affects client availability and the global model’s performance. Accordingly, this paper evaluates federated learning (FL) in a realistic vehicular network topology, accounting for real vehicle traffic in two Brazilian urban areas. The network simulation covers $3.7{\mathrm{km}}^2$ with 1290 vehicles per hour and road speeds, based on real data. Our paper provides a comprehensive analysis of the impact that different traffic behaviors can yield during the training phase of a federated learning model. We observe that there is a performance decay in urban areas with longer vehicle permanence. Interestingly, longer vehicle participation in FL training leads to a biased final model with reduced generalization. We propose a novel approach to verify vehicle variability over time, by using the Dice-Sørensen coefficient to compare the set of clients participating in different rounds of training. By maintaining the vehicle variability over the rounds we can reduce the effect of the bias on the model, and – with a 47% reduction of the communication overhead – achieve faster learning, higher convergence in the first 15 rounds, and an equivalent final accuracy. Additionally, we extend our analysis by conducting simulations under more extreme traffic scenarios across multiple datasets, using a MobileNetV3. The results confirm that sustaining high vehicle variability – in scenarios with a brief participation of vehicles in the training – yields comparable model performance while saving up to 83.5 GB in communication costs.