1: \begin{abstract}

2: Federated learning (FL) has been proposed to enable distributed learning on Artificial Intelligence Internet of Things

3: (AIoT) devices with guarantees of high-level data privacy.

4: Since random initial models in FL can easily result in unregulated  Stochastic Gradient Descent (SGD) processes, existing

5: FL methods greatly suffer from  both slow convergence and poor

6: accuracy, especially in non-IID scenarios. 

7: To address this problem, we propose a novel method named CyclicFL,  which can quickly derive effective initial models to

8: guide the SGD  processes, thus

9: improving the overall FL training performance. 

10: We formally

11: analyze the significance of data

12: consistency between the pre-training and training stages of CyclicFL, showing 

13: the limited Lipschitzness of loss for the pre-trained models by CyclicFL.

14: Moreover, we systematically prove that our method can achieve faster convergence speed under various convexity assumptions.

15: Unlike traditional centralized pre-training methods 

16: that require public proxy data,  CyclicFL pre-trains initial models 

17: on selected AIoT devices cyclically without exposing their local data. 

18: Therefore, they can be easily integrated into any security-critical FL methods. 

19: Comprehensive experimental results show that CyclicFL can

20: not only improve the maximum classification 

21: accuracy by up to $14.11\%$ but also significantly accelerate the overall FL 

22: training process.

23: \end{abstract}

24: