Investigating and developing efficient federated learning for air pollution monitoring

Abstract

Location-based data may be considered highly private; as such, handling location-based data requires that it cannot be used to track a user. In a network of multiple edge devices that each collect data, training a machine learning model would typically involve transmitting the data securely to a central server which requires strict privacy rules. Federated learning solves the privacy problem by not requiring data to be shared; instead, training of a machine learning model is performed on the device that gathered the data itself. Using federated learning with the Federated Stochastic Gradient Descent (fedsgd) algorithm, a similar training performance is expected as training a machine learning model on a single server with data transmitted to it. Overall less bandwidth may be used for communication between edge devices and the server. However, a higher computational cost is seen due to having to perform model training on the edge device, which lowers the potential data points that can be processed each day given the lower computational performance of an edge device versus a high power server. Whilst only a single edge device may train the model at a time, a different federated learning algorithm may be used on the server to enable multiple to train simultaneously