The Trade-off Between Privacy and Utility in Local Differential Privacy

In statistical queries work, such as frequency estimation, the untrusted data collector could as an honest-but-curious (HbC) or malicious adversary to learn true values. Local differential privacy(LDP) protocols have been applied against the untrusted third party in data collecting. Nevertheless, excessive noise of LDP will reduce data utility, thus affecting the results of statistical queries. Therefore, it is significant to research the trade-off between privacy and utility. In this paper, we first measure the privacy loss by observing the maximum posterior confidence of the adversary (data collector). Then, through theoretical analysis and comparison we obtain the most suitable utility measure that is Wasserstein distance. Based on these, we introduce an originality framework for privacy-utility tradeoff framework, finding that this system conforms to the Pareto optimality state and formalizing a payoff function to find optimal equilibrium point under Pareto efficiency. Finally, we illustrate the efficacy of our system model by the Adult dataset from the UCI machine learning repository.