Differential privacy (DP)
DP is a strong approach to protecting data privacy with a specific objective; that is, it requires that the result of a query or an algorithm on a dataset be insensitive to removing or adding any single record in the dataset. Using aggregates of individual data points and adding noise to outputs to avoid personal identification is at the core of DP.
As a result, it protects against leaking information about individual records, such as the training data that goes into an ML model. DP can be applied to various stages of ML, such as data preprocessing, model training, and prediction serving.
DP relies on strong mathematical foundations to make it virtually impossible for an attacker to infer any individual’s data from the output. This makes it the gold standard in privacy but also hard to adopt. For a thorough treatment of DP, please see the extensive paper How to DP-fy ML: A Practical Guide to Machine Learning with Differential Privacy, published...
