Privately Detecting Pairwise Correlations in Distributed Time Series

In this paper, we propose developing a generic framework for privately identifying similarities or correlations within and/or across basic statistics, e.g. mean, for independently owned, distributed participant data. To obscure the actual statistical values and improve the levels of privacy, we propose using scaled bin values instead of raw data. We find that while there is a natural trade off between privacy and accuracy, we can maintain reasonable correlation accuracy across different levels of privacy and different adversarial backgrounds for time series data with varying distributions.