Differentially Private Machine Learning

Sammanfattning

Sensitive information is constantly stored and processed for different reasons and bydifferent actors in this digital era. Differential Privacy (DP) is a notion of privacy achieved viarandomization in the data, thereby ensuring that privacy can be guaranteed. Three differentmethods of differential privacy were implemented on parameter estimation and numericallyanalyzed based on the trade-off between accuracy and privacy. The standard approach todifferential privacy is based on adding randomness to the sensitive data using the Laplacemechanism. This explicit approach was compared to an implicit randomization on Bayespoint estimate. In the third method, multiple data owners are considered who eachindependently add noise using the Laplace mechanism to the original data and only revealnoisy signals to a data acquisitor which estimates the unknown parameter. This reportexplores these three applications and discusses the advantages and disadvantages of each.The numerical results show that as noise is increased, privacy is ensured while the accuracydecreases and this is consistent throughout all three different methods.