Federated Learning Powered Semantic Communication for UAV Swarm Cooperation

In recent years, the unmanned aerial vehicle (UAV) swarm has garnered significant attention from both industry and academia. The UAV swarm possesses the capacity to achieve sophisticated and collaborative operations, which might result in a large volume of data exchange. Nevertheless, the current data-oriented communication scheme may not adequately guarantee such data exchange, thus impeding the progression of UAV swarm intelligence. In this article, we introduce the semantic communication paradigm into the UAV swarm in which UAVs can transmit semantics extracted from raw data instead of from the entire data set. This approach minimizes the transmitted data size and enhances communication efficiency. However, training precise semantic models in resource-limited UAVs presents a formidable challenge. In particular, the dynamic topology and volatile communication links within the UAV swarm may deteriorate training efficiency. To solve the aforementioned challenges, we adopt the hierarchical federated learning (HFL) framework wherein the UAV swarm is divided into clusters led by cluster heads. In this framework, each UAV trains its semantic model according to local data, and subsequently uploads the model parameters to the associated cluster head for intermediate aggregation. Then, decentralized aggregation is conducted among cluster heads to achieve global parameter consensus. To incentivize UAVs' participation in federated learning (FL) tasks, cluster formation in intra-cluster training is modeled as an evolutionary game. We further design a novel cluster management mechanism to handle the UAVs dropout issue due to the dynamics of the UAV swarm.