Another issue I have with data synthetization is that by trying to reproduce the original data as faithfully as possible you waste a lot of your privacy budget on meaningless features.
That said I'm happy to be convinced otherwise. Do you have a publicly available demo dataset that I could look at?
Tangent: I suppose humans choice in models and statistics to use can be argued to be a similar thing, but I'm still dubious of how this concept is being communicated.
What we do is extract features in an unsupervised way and build a model to recreate the dataset from said feature. When the synthesized dataset is indistinguishable from the real thing. You should be able to build a model that gets you 90%-100% of the way there with synthetic data, then tune / retrain on the real data.
The model that does decisioning itself doesn’t matter. If you needed to adjust the synthetic dataset, you’d have to either build a new generative model OR otherwise weight the dataset.
https://medium.com/capital-one-tech/why-you-dont-necessarily...
It’s even good enough in many cases to build models off of.
However, it’s important to not that numbers such as SSNs are always “real”, in the sense they link to someone. A random SSN has with near certainty been used before and belonged to someone (perhaps multiple) people. The trick is ensuring the rest of the attributes don’t match the individual.