The Cost of Privacy in Asynchronous Differentially-Private Machine Learning

We consider training machine learning models using data located on multiple private and geographically-scattered servers with different privacy settings. Due to the distributed nature of data, communicating all collaborating owners simultaneously may prove challenging or altogether impossible. differentially-private asynchronous algorithms for collaboratively machine-learning datasets. The implies that a central learner interacts one-on-one whenever they are available communication without needing aggregate query responses construct gradients entire fitness function. Therefore, algorithm efficiently scales many owners. define cost as difference between privacy-preserving model trained in absence concerns. demonstrate has an upper bound is inversely proportional combined size datasets squared sum budgets squared. validate theoretical results experiments financial medical illustrate collaboration among more than 10 at least 10,000 records greater equal 1 superior comparison isolation only one datasets, illustrating value privacy. number can be lowered if budget higher.