Federated Learning in Drone-based Systems

Drone-based systems are extremely dynamic as these drones move independent of one another depending on the demand source for the services they provide. To effectively learn concepts in such systems, we need sensor-based data from all the drones. However, accumulation and processing of such data at a central location is subject to vulnerabilities such as exposure of private drone-based information. Federated Learning (FL) is a natural choice for such applications as each drone only shares its updated model and not its data with the central entity. However, learning in such a dynamic setup is rendered difficult due to the challenges in maintaining consistent wireless link with necessary bandwidth between all the drones and the base station. Given these constraints, we study the tradeoff between the number of drones selected for each FL learning round and the amount of information transferred from each drone to the server. We use an approximate solution to the multiple-choice online algorithm to select a subset of the drones for each learning round and random selection of the amount of information sent from each drone during each learning round. We observe that the number of drones that participate in any given FL training round and the amount of information shared by each participating drone are complementary. Our preliminary results show that the lightweight multiple-choice online algorithm for drone selection can be effectively used along with an appropriate information selection algorithm for improved system performance.