ISSN 1000-1239 CN 11-1777/TP

Journal of Computer Research and Development ›› 2022, Vol. 59 ›› Issue (2): 430-439.doi: 10.7544/issn1000-1239.20200717

Previous Articles     Next Articles

A Perturbation Mechanism for Classified Transformation Satisfying Local Differential Privacy

Zhu Suxia, Wang Lei, Sun Guanglu   

  1. (School of Computer Science and Technology, Harbin University of Science and Technology, Harbin 150080) (Research Center of Information Security and Intelligent Technology, Harbin University of Science and Technology, Harbin 150080)
  • Online:2022-02-01
  • Supported by: 
    This work was supported by the National Natural Science Foundation of China (61502123), the Science Foundation for Returned Overseas Students of Heilongjiang Province (LC2018030), the Heilongjiang University Special Foundation for Basic Scientific Research (JMRH2018XM04), and the Natural Science Foundation of Heilongjiang Province (LH2021F032).

Abstract: As the state-of-the-art privacy protection technology, local differential privacy is widely used to compute the mean value of continuous numerical data. The perturbation mechanism will directly affect the accuracy of the mean value. In order to further improve the accuracy of mean value estimation, a perturbation mechanism for classified transformation satisfying differential privacy is proposed. In this mechanism, continuous numerical data is divided into transformation range, which is then segmented. What’s more, it transforms the segmentation into one-dimensional binary category data. After transformation, the mechanism of random response is used to perturb the data. More importantly, it extracts the value randomly as well as uniformly from the numerical segment identified by the perturbation data as the perturbed value. The experimental results of mean value estimation in both real data and synthetic data show that the mechanism proposed in the paper greatly improves the accuracy. In addition, this perturbation mechanism is used to build a mini-batch gradient descent algorithm satisfying local differential privacy and the linear regression learning task is completed successfully. The experimental results show that this method not only is superior to other existing mechanisms but also can obtain a smaller mean square error at the same time.

Key words: local differential privacy, data transformation, mean value estimation, mini-batch gradient descent, random response

CLC Number: