private-data-generation
private-data-generation copied to clipboard
PATE-GAN: Processing and Metadata
Description
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.