arXiv releases CoRTeC for differentially private synthetic tabular data
Original title: CoRTeC: A Differentially Private Release for Synthetic Tabular Data, Decodable by Any Generator
CoRTeC releases synthetic data under differential privacy using noisy counts, decodable without language models or real records.
CoRTeC releases synthetic versions of records under differential privacy by spending the privacy budget once on a small file of noisy counts. For each group, the file gives size, balance of outcomes, histogram of every column for each outcome and the outcome rate in non-overlapping cells. A reviewer can check every number before the file leaves the institution. Anyone can then turn the counts into synthetic rows, as often as they like and at no further privacy cost. A maximum-entropy decoder can do this without any language model. At ε=2, the automatically configured release trains classifiers to within 0.02 AUC of a real sample on the Adult dataset, across three data splits.
- Privacy budget spent once on noisy counts file.
- Maximum-entropy decoder works without language model.
- Without selection step, stays within 0.05 AUC on NHANES and bank marketing.
- Strongest membership attack stays far below the budget bound.
- Selection step matches rows to released counts.
From the original article
1 min read at arxiv.orgThe records that could train the most useful models, such as a hospital's patient files, are often the ones an institution cannot share. We propose CoRTeC, a method that releases synthetic versions of such records under differential privacy. Differential privacy places a proven limit on what the synthetic data can reveal about any person, and in particular on membership inference.
By Calister Nnona · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.