Federated Vehicular Networks – Opportunities and Challenges

Federated Learning (FL) is a promising Machine Learning (ML) paradigm for training models in a privacy-aware manner, while still reaching model convergence. In this setting, users receive the specifications of the ML models, which are then trained over their local datasets, so that the trained model is distributed back to the network to be aggregated with the models from other clients.

In this context, the network plays a crucial role in the FL process, constituting the biggest bottleneck in the system. Techniques for the quantization of neural network weights for sending over a wireless channel, scheduling of the model aggregation from different users, and the selection of which users in the network participate in the process; are all attempts to mitigate the network overhead in the FL process.

Our research investigates mitigating such impact by performing FL model aggregations in a D2D manner, such that contributions to the trained model are not all performed through cellular network uplink and downlink. In this manner, we aim to achieve better convergence rates and greater model robustness by allowing more participants in the FL process.

Related Literature:
[1] Ye, Dongdong, et al. "Federated learning in vehicular edge computing: A selective model aggregation approach." IEEE Access 8 (2020): 23920-23935.
[2] Du, Zhaoyang, et al. "Federated learning for vehicular internet of things: Recent advances and open issues." IEEE Open Journal of the Computer Society 1 (2020): 45-61.
[3] Posner, Jason, et al. "Federated learning in vehicular networks: opportunities and solutions." IEEE Network 35.2 (2021): 152-159.