Enabling Pervasive Federated Learning using Vehicular Virtual Edge Servers

Recent works have proposed various distributed federated learning (FL) systems for the edge computing paradigm. These FL algorithms can assist pervasive applications in various aspects, e.g., decision making, pattern recognition, and behavior prediction. Existing solutions do not efficiently support the training based on the real-time location-specific data, because fundamentally, the “data collection” problem is rarely studied in the context of FL systems. To address this problem, we present a novel system, VC-SGD (Vehicular Clouds-Stochastic Gradient Descent), which seamlessly integrates the emerging concept of vehicular clouds with an edge-based FL. We show that by using vehicular clouds as virtual edge servers, VC-SGD is able to effectively support FL algorithms that use real-time location-specific data. We develop a general simulator that uses SUMO to simulate vehicle mobility and MXNet to perform real training. We use our simulator to verify the efficacy of VC-SGD. The experimental results demonstrate that VC-SGD improves over existing solutions.