A Framework with Randomized Encoding for a Fast Privacy Preserving Calculation of Non-linear Kernels for Machine Learning Applications in Precision Medicine

摘要

For many diseases it is necessary to gather large cohorts of patients with the disease in order to have enough power to discover the important factors. In this setting, it is very important to preserve the privacy of each patient and ideally remove the necessity to gather all data in one place. Examples include genomic research of cancer, infectious diseases or Alzheimer's. This problem leads us to develop privacy preserving machine learning algorithms. So far in the literature there are studies addressing the calculation of a specific function privately with lack of generality or utilizing computationally expensive encryption to preserve the privacy, which slows down the computation significantly. In this study, we propose a framework utilizing randomized encoding in which four basic arithmetic operations (addition, subtraction, multiplication and division) can be performed, in order to allow the calculation of machine learning algorithms involving one type of these operations privately. Among the suitable machine learning algorithms, we apply the oligo kernel and the radial basis function kernel to the coreceptor usage prediction problem of HIV by employing the framework to calculate the kernel functions. The results show that we do not sacrifice the performance of the algorithms for privacy in terms of F1-score and AUROC. Furthermore, the execution time of the framework in the experiments of the oligo kernel is comparable with the non-private version of the computation. Our framework in the experiments of radial basis function kernel is also way faster than the existing approaches utilizing integer vector homomorphic encryption and consequently homomorphic encryption based solutions, which indicates that our approach has a potential for application to many other diseases and data types.