In recent years, with the development of software and hardware, migrating databases to the cloud has become an emerging development trend and can reduce database operation and maintenance costs for small and medium-sized enterprises and individual users. Furthermore, the development of cloud databases has led to a massive market demand for database operation and maintenance. Researchers have proposed many database self-tuning technologies to support automatic optimization of database knobs. To improve tuning efficiency, existing technologies have shifted from focusing solely on the tuning problem itself to focusing on how to reuse historical experience to find the optimal parameter configuration for the current database instance. However, with the development of cloud databases, users have gradually increased their requirements for privacy protection, hoping to avoid privacy leakage while having efficient data access efficiency. Existing methods do not consider protecting the privacy of users’ historical tuning experience, which may cause user load characteristics to be perceived, causing economic losses. This study analyzes the characteristics of cloud database tuning tasks in detail, organically combines the server side and the user side, and proposes a cloud database knob tuning technology based on federated learning. First, to solve data heterogeneity in federated learning, this study proposes an experience screening method based on meta-feature matching to eliminate historical experiences with large differences in data distribution in advance to improve the efficiency of federated learning. To protect user privacy, this study organically combines the characteristics of cloud database services and proposes a federated Bayesian optimization algorithm with the node end as the training center. Through random Fourier features, it achieves user privacy protection without distorting the tuning experience. The results on extensive public benchmarks present that the proposed method could achieve competitive tuning performance compared with existing tuning methods. Moreover, due to the reuse of historical experience, it can greatly improve tuning efficiency.