A model-based filter to improve local differential privacy:

Local differential privacy has gained popularity in academia and industry as an effective mechanism for computing descriptive statistics while guaranteeing privacy for end users. In the proposed local differential privacy mechanisms, the introduction of noise is key to preserving privacy, but may considerably limit the estimation power of the provider. In the paper we propose to include a machine learning model as a pre-filter to help discard observations that are so noisy that they do not add value to the estimation process. For this we validate our approach with synthetic and real databases, and identify an average reduction of up to 31% in the mean squared error (MSE).