CTS-DP: Publishing correlated time-series data via differential privacy

Analyzing and mining time-series data by taking advantage of the correlation between the data values can provide outstanding beneficial. But data owners may be unwilling to publish the data's true values due to privacy considerations. Recently, researchers have begun to leverage differential privacy to address this challenge. However, the Laplace noise series used in the current state-of-the-art approaches has a drawback in that it is independent and identically distributed. An adversary can remove the independent noise from the correlated time-series by utilizing a refinement method (e.g., filtering), resulting in a lesser than expected effective degree of privacy. To remedy this problem, we propose an effective correlated time-series data publication solution based on differential privacy by enforcing Series-Indistinguishability and designing a correlated Laplace mechanism. Based on the concept of indistinguishability from the unconditional security definition, Series-Indistinguishability guarantees that the correlation between the noise and original series is indistinguishable to an adversary. Furthermore, instead of using an independent Laplace mechanism, a correlated Laplace noise series is produced using four Gauss white noise series passed through a specific linear system, to satisfy Series-Indistinguishability. Experimental results demonstrate that our solution outperforms the state-of-the-art differential privacy mechanisms in terms of security and mean absolute error for large quantities of queries.