Clustered federated learning multi-classifier for non-IID scenario in drone devices
Jiaming Pei, Zhi Hong Yu, Wenxuan Liu, Weili Kong, Jiwei Zhang, Lukun Wang · 2022
Existing machine learning technology lacks effective support for practical drone devices in target recognition. The Clustered Federated Learning Algorithm (C-FLA) was proposed to enhance the joint training in drone groups. The algorithm considered the path fading and Rayleigh fading when the user side performs upstream data transmission. And it designed a federation learning method based on cluster-based training. The results showed that the traditional federal learning algorithm and C-FLA can converge to similar values under the same training conditions in the case of good channel state and small user transmit power constraints, and C-FLA converges faster. C-FLA can reduce the convergence value of the loss function by 10%-50% compared to the traditional centralized algorithm. It can be seen that C-FLA is more helpful for model training in multi-classifier scenarios.