Bias in Differential Privacy

Check our AAAI 2021 paper on Bias and Variance of Post-processing in Differential Privacy.  It studies the effects of post-processing differentially private outputs (e.g., to restore feasibility of some constraints) on the noise distribution. The paper takes a first step towards understanding the properties of post-processing and quantifies the bias and variance introduced by a wide class of post-processing algorithms based on projection. It includes an analysis of the release of important quantities based on census data. The most interesting result is Theorem 5.