Federated learning algorithm for predicting error bit of NAND flash threshold voltage distribution

Abstract

The threshold voltage distribution (TVD) data of NAND flash affects the reliability and lifetime prediction of data. However, the TVD data varies depending on the user’s characteristics or usage frequency, and this changes the distribution of error bits. In this paper, we propose a federated learning algorithm to solve this problem. The federated learning algorithm trains the server model by aggregating the models learned by users without collecting their data. This method is advantageous for communication efficiency and privacy protection. The proposed algorithm makes the server model mimic the error bit estimation of user models. To do this, it calculates the reliability of each user model and generates pseudo error bits by weighted averaging the estimates of reliable user models. This way, the server model can cope with various user data distributions and improve the performance of error bit estimation. The experimental results show that the proposed algorithm performs more accurate and stable error bit estimation than existing algorithms.