A correlation-based subspace analysis for data confidentiality and classification as utility in CPS

摘要

The concept of pairing confidential-relevant variables (connected variables) using ridge regression and bootstrap sampling has recently been proposed for developing perturbation models to data privacy in cyber-physical systems. In this approach, a single set of perturbation parameters for all the pairs of connected variables has been used to achieve trade-off between data confidentiality and classification as data utility. It has led to weaker confidentiality protection for some pairs of connected variables than the others. In this paper, we have determined that this discrepancy occurs due to varying correlation characteristics between the variables. The correlation between a connected variable and other confidential variables influences the correctness of the perturbation parameters of the ridge regression model studied for data privacy. In this paper, we have divided the feature space into correlated subspaces and studied the ridge regression-based perturbation model with bootstrap sampling in individual subspaces separately. Our experimental analysis with IRIS and NSL-KDD datasets has provided an interesting finding, indicating that the absolute Pearson correlation coefficient greater than 0.1, between the connected and confidential variables, can lead to strong confidentiality, as measured by signal-interference-ratio less than 20dB.