The communication efficiency evaluation of unmanned aerial vehicles (UAVs) is of great significance to the completion of their missions. With the increase in the number of devices, the amount of data for communication interaction has grown significantly. However, the existing evaluation systems have insufficient computational resources and data transmission resources, and the demand for communication security has also increased significantly. Therefore, this paper proposes an improved Stacking ensemble learning method. This method uses federated learning as the underlying framework and applies Kernel Independent Component Analysis (KICA) to preprocess the data within local computing nodes. Three base learners are deployed inside each federated node, which complete the training, prediction, and meta-feature generation of the base learners locally, and finally realize the Stacking integration of the prediction results within the node through logistic regression. On the one hand, this method can improve the overall performance by combining the prediction results of multiple models. On the other hand, it takes advantage of federated learning to only transmit model prediction results without exchanging raw data, which greatly reduces the risk of privacy data leakage and ensures data security to a certain extent. Experiments show that the improved Stacking ensemble learning method achieves a mean prediction accuracy of 79.25% (95% CI: [76.32%, 82.13%], outperforming single models and traditional Stacking (73.92% ± 2.8%). Additionally, KICA data preprocessing reduces training time by 50.9%–58.4%.