PUBLISHING SENSITIVE TIME-SERIES DATA UNDER PRESERVATION OF PRIVACY AND DISTANCE ORDERS

摘要

In this paper, we address the problem of preserving mining accuracy as well as privacy in publishing sensitive time-series data. For example, people with heart disease do not want to disclose their ECG time-series, but they still allow mining some accurate patterns from their time-series. Our privacy model assumes that (1) data sources publish their time-series independently, and (2) all information used in publishing time-series can be publicly revealed. Based on this model, we introduce three assumptions: full disclosure, equi-uncertainty, and independency. We also derive two requirements: uncertainty preservation and distance order preservation. We show that only randomization methods satisfy all three assumptions, but even those methods do not satisfy both the requirements. Thus, we discuss the randomization-based solutions that satisfy all assumptions and requirements. For this purpose, we present a novel notion of the noise averaging effect of piecewise aggregate approximation (PAA), which is derived from a simple intuition that the summation of random noise converges to 0. This noise averaging effect can alleviate the problem of destroying distance orders in randomly perturbed time-series. Based on the noise averaging effect, we first propose two naive solutions that use the random data perturbation in publishing time-series while exploiting the PAA distance in computing distances. There is, however, a tradeoff between these two solutions with respect to uncertainty and distance orders. We thus propose three more advanced solutions that take advantages of both naive solutions. Experimental results show that our advanced solutions are superior to the naive solutions in the preservation of uncertainty, distance orders, and clustering accuracy.