One of the most promising applications of machine learning is the ability to generate large amounts of data, dubbed synthetic data, from a small set of original input data. The synthetic data resembles the properties of the original data, but being completely fictious, it does not constitute personal data in the traditional sense. Hence, it can be assumed that synthetic data would be a valid way of generating anonymous data. Unfortunately, this assumption turns out to be wrong in several respects. For one, personal information and properties of human individuals that were part of the original data can still be extrapolated from the synthetic data (see the paper on "Anonymization Groundhog Day"!). Second, synthetic data is likely to increase biases in the dataset generated, thereby leading to severe, often systematic discrimination of people not (or noch sufficiently) represented in the original input dataset.

These topics are touched upon in this lecture.

slides: pdf/pptx

You can find additional readings and literature of relevance in the course module.