Federated Learning

Traditional ML has been operating in its current form for decades. Naturally, it has some pros and cons—with the pros typically maximizing performance and predictive accuracy. There are many alternatives to traditional ML that provide flexibility these pros and cons, particularly with respect to privacy, explainability, and function. In this lesson, we cover FL, the distributed ML approach that allows models to train on highly sensitive, non-centralized datasets.

For additional context, imagine we have several hospitals with highly sensitive diagnosis information for patients with a very rare disease. In these cases, differential privacy alone is not enough to guarantee an individual’s privacy due to the rarity of the disease and the sparsity of the data. Synthetic data can’t be generated because of the small sample size (GANs and other generators are limited to the variance of their dataset and can generate lots of similar examples but no dramatically different ones). How can this data be used?

Perhaps the simplest answer is: it can’t. Taking the data out of a hospitals’ environment would be tantamount to data privacy violation.