Abstract:
|
The Bayesian MCMC bootstrapping approach provides a viable alternative to the classic Differential Privacy protection approach (Dwork and Roth (2014)) by random sampling of the individual models’ tuples of their coefficients, standard errors, the training sample sizes and aggregating them into one federated model. In this way, this method provides the data source privacy protection without hurting the model’s predictive power that is unavoidable result of adding Differential Privacy random noise to the model’s training data set.
|