Generation of Synthetic Tabular Data for Controlled Machine Unlearning

One of the most recent areas of research in the field of Artificial Intelligence concerns Machine Unlearning. Since research necessarily needs many tests and benchmarks to evaluate new proposals about methods or algorithms and compare them to the already existing ones, large datasets are requested. Complications about the difficulty to collect big amounts of data or privacy of individuals led to synthetic datasets. Not only they are faster to create than collecting real data, and don’t leak information about anyone, but they are also built ad hoc, so the user can decide,and tune characteristics based on the aim of the test. In this thesis both Machine Unlearning and Synthetic Data Generation topics will be analysed, and the focus will be about showing how with different synthetic datasets tuned on purpose, Machine Unlearning can become harder or easier, and how we can use these synthetic datasets to observe performances of Machine Unlearning methods and compare them. The core idea is that samples from more entangled datasets will be harder to be forgotten by the models, while datasets with more separated data will be easier. We will also observe what having entangled data means and how we can measure this phenomenon.