Approximating Robust Linear Regression With An Integral Privacy Guarantee

Most of the privacy-preserving techniques suffer from an inevitable utility loss due to different perturbations carried out on the input data or the models in order to gain privacy. When it comes to machine learning (ML) based prediction models, accuracy is the key criterion for model selection. Thus, an accuracy loss due to privacy implementations is undesirable.The motivation of this work, is to implement the privacy model “integral privacy” and to evaluate its eligibility as a technique for machine learning model selection while preserving model utility. In this paper, a linear regression approximation method is implemented based on integral privacy which ensures high accuracy and robustness while maintaining a degree of privacy for ML models. The proposed method uses a re-sampling based estimator to construct linear regression model which is coupled with a rounding based data discretization method to support integral privacy principles. The implementation is evaluated in comparison with differential privacy in terms of privacy, accuracy and robustness of the output ML models. In comparison, integral privacy based solution provides a better solution with respect to the above criteria.