PUB - Publikationen an der Universität Bielefeld

Federated Bayesian optimization (FBO) has been introduced in recent years to avoid privacy leakage when multiple clients involve in finishing a global optimization task. Parameter-sharing-based FBOs, as one branch of FBOs, however, compromise the optimization efficacy due to the reduced fitting ability of parameterized Gaussian processes (GPs). In this work, we propose a data-sharing federated framework based on compressed sensing by directly sharing perturbed decision variables of raw data. Specifically, decision variables of raw data in each client are perturbed to a certain level by controlling the reconstruction rate to make the reconstructed data similar to but indistinguishable from the real raw data. By doing this, the reconstructed data can be used as the perturbed raw data to be shared with other clients. Then, a curator is randomly selected at each round of surrogate updates to train a global GP model using the union of perturbed and real data set, helping explore the whole search space. Additionally, we put forward a novel standard for evaluating privacy levels in BO algorithms, promoting fair performance benchmarks. Compared to other FBOs, the proposed algorithm has demonstrated to be very effective without compromising privacy, as evidenced by experimental results on the CEC2005 benchmark.