Is synthetic data always anonymous? about ctgan HOT 1 CLOSED

As far as I understand, yes this is possible.

We're trying to run similar techniques in live systems now and there we do a check on the output rows to verify no identical rows are present in the synthesized data. However, the probability of a sample exactly the same as from the original dataset is very small, especially when there are continuous columns present. However, given the dataset you are describing (limited nr. of columns, limited options per column), the probability of these values occuring rises.

The probability of a datapoint of the real data being in the synthetic data is (assuming that CTGAN can fit the original distribution quite well) very close to that point being sampled from the joint probability distribution.

If you have a specific subset of columns that cannot occur identically together, make sure to check on them if those will release private information.

from ctgan.