Comparison of Ensemble and Federal Learning for Secure Data Collaboration in Satellite Networks
Zhen Gao, Wen Sun, Yuqi Zhang · 2022
Satellite network is becoming an important component of 6G and future networks for seamless coverage. Each satellite is generating a large amount of data based on their rich sensors, e.g. images, but the data from different satellite may not be gathered in some cases for performance improvement of deep learning task due to processing capability and privacy problems. Federal learning (FL) is a decentralized data collaboration solution without data exposure. But exchange of model parameters poses heavy network traffic between satellites. Ensemble learning (EL) is a popular method to improve task performance by collaborating diverse base leaners, and it has been initially proved to be effective for data collaboration. Since only the inference result is exchanged between clients, EL enables easier deployment of data collaboration than FL with dramatic decreasing of network traffic. At present, the performance of EL based data collaboration has not been well studied. In this paper, we compare the performance of FL and EL based data collaboration systems for different data division scenarios. Experiment results show that EL based scheme can achieve close or higher performance than FL based scheme for small number of clients with some data sharing.