Upon inspecting PATE-GAN, I noticed a couple of potential issues:
The model expects already processed/scaled data (the processing is done outside of the model, i.e., here), and as a consequence it doesn’t return synthetic data in the original scale.
The data bounds (min/max values) are directly extracted from the data in a non-DP way (here), which might lead to privacy vulnerabilities as shown in previous work.
Description
Upon inspecting PATE-GAN, I noticed a couple of potential issues: