Researching the Performance of AutoML Platforms in Confidential Computing

Abstract

The paper is dedicated to testing the performance indicators of automatic machine learning platforms when they function in standard and confidential modes using the example of a nonlinear multidimensional regression. A general protocol of distributed machine learning trusted in the sense of security is proposed. It is shown that within the framework of confidential virtualization, when optimizing the architecture of machine learning pipelines and hyperparameters, the best quality indicators of generated pipelines for multidimensional regressors and speed characteristics are demonstrated by solutions based on Auto Sklearn compared with Azure AutoML, which is explained by different learning strategies. The results of the experiments are presented.