Federated Bayesian optimization via compressed sensing

Qiqi Liu, Leming Wu, Yaochu Jin

Abstract: Highlights•This work explores the combination of compressed sensing and BO to achieve privacy-preserving federated optimization via data sharing. Different from other data-sharing strategies in other privacy-preserving BO work, the proposed compressed sensing-based data sharing can not only preserve the relative distance among solutions in one single client but also can ensure the relative distances of solutions in different clients, which is the key to utilizing the raw data in different clients effectively.•This work strikes a balance between exploration and exploitation over the search process through a global GP trained using perturbed data in a curator, and a local GP model trained using local raw data in the curator. Additionally, at each round of surrogate updating, a curator is randomly selected from the set of clients. Over successive rounds, every client will have the opportunity to serve as a curator. This approach, coupled with the use of both global and local GPs, allows the proposed algorithm to effectively balance exploration and exploitation.•This work presents a privacy indicator for measuring the level of privacy preservation, ensuring the fairness of performance among privacy-preserving BO algorithms. Current data-perturbation-based BO either perturb objective values or decision variables, or both, using different strategies such as differential privacy. However, this is a lack of a standard measuring the privacy protection level. In this work, we propose measuring the privacy protection level by calculating the differences between the landscape fitted by the perturbed solutions and the original landscape fitted by original solutions.

External IDs:dblp:journals/isci/LiuWJ24