Dynamic Semi-Synchronous Federated Learning for Connected Autonomous Vehicles

Due to the increased computational capacity of Connected and Autonomous Vehicles (CAVs) and concerns about transferring private information, storing data locally and moving network computing to the edge is becoming increasingly appealing. This makes Federated Learning (FL) appealing for CAV applications. However, the synchronous protocols used in FL have several limitations, such as low round efficiency. In this context, this work presents FALCON, a semi-synchronous protocol for FL based on the link duration. FALCON leverages data periodically transmitted by CAVs to compute link duration and establish a dynamic temporal synchronization point. Additionally, FALCON includes a client selection mechanism that considers the local model versions and models with higher local loss. FALCON reduces the communication rounds and the number of selected clients while maintaining the same level of accuracy for FL applications.