Bivariate Correlation-Based Attack: A Challenge for Privacy Preservation of Location Sequence Data

Abstract. Data correlation is a critical issue for location sequence data privacy protection. Series-indistinguishability provides a theoretical basis for the differential privacy protection of temporally correlated location data and is implemented by the correlated Laplace mechanism (CLM), which has become a novel privacy protection method. However, location sequence data is essentially a multivariate time series with data correlation not only within each dimension but also between different dimensions. The CLM-based location privacy scheme only adapts to the auto-correlation in each dimension and ignores cross-correlation between different dimensions, which compromises its privacy performance. To evaluate its actual privacy protection performance, we propose a bivariate correlation-based attack (BCA) utilizing a filtering method and theoretically derive the optimal filter parameters. Based on simulations and real data experiments, the results show that the privacy performance of the CLM-based privacy scheme is significantly reduced under BCA, confirming that this scheme cannot achieve complete series-indistinguishability for bivariate sequences. Furthermore, the results suggest that BCA is an effective tool for privacy performance evaluation.