Coverage-Aware Federated Learning in Multihop Drone Networks via Distributed Participant Selection

Federated Learning (FL) can benefit applications using Multi-Hop Drone Networks (MDNs), but their distributed and dynamic connectivity challenges model training and aggregation. Traditional FL assumes stable topology and central server access, whereas MDNs rely on peer-to-peer exchanges without continuous connectivity, making it hard to ensure local model contributions are meaningful and non-redundant. To address this, we present a distributed participant selection algorithm in the context of MDNs. Specifically, at the beginning of each FL round, drones first identify their neighbors and measure the quality of communication links. With the identified neighbors, the drones select a subset of them to participate in the FL process based on their own and their neighbors’ contributions to the data needed for FL. This is done in a distributed manner, ensuring that only valuable data, which incorporates full sensor and area coverage, contributes to the FL model. For the subset of drones selected, we found the best path for these drones to transmit the local model via a distributed Steiner Tree problem formulation. Our simulation studies show that our distributed participant selection converges in 13 communication rounds for networks of up to 160 drones and achieves utility scores within 10% of a centralized baseline. Over 150 FL rounds with a real-world dataset, the trained global model using our approach yields an RMSE of 26.8% higher than the best centralized model. These results demonstrate the potential of our approach to enable resilient and scalable learning in dynamic drone networks, thereby advancing FL applications in distributed environments.