{
  "id": "2106.12576",
  "version": "v1",
  "published": "2021-06-22T20:37:12.000Z",
  "updated": "2021-06-22T20:37:12.000Z",
  "title": "DP-SGD vs PATE: Which Has Less Disparate Impact on Model Accuracy?",
  "authors": [
    "Archit Uniyal",
    "Rakshit Naidu",
    "Sasikanth Kotti",
    "Sahib Singh",
    "Patrik Joslin Kenfack",
    "Fatemehsadat Mireshghallah",
    "Andrew Trask"
  ],
  "comment": "4 pages, 3 images",
  "categories": [
    "cs.LG",
    "cs.AI",
    "cs.CR"
  ],
  "abstract": "Recent advances in differentially private deep learning have demonstrated that application of differential privacy, specifically the DP-SGD algorithm, has a disparate impact on different sub-groups in the population, which leads to a significantly high drop-in model utility for sub-populations that are under-represented (minorities), compared to well-represented ones. In this work, we aim to compare PATE, another mechanism for training deep learning models using differential privacy, with DP-SGD in terms of fairness. We show that PATE does have a disparate impact too, however, it is much less severe than DP-SGD. We draw insights from this observation on what might be promising directions in achieving better fairness-privacy trade-offs.",
  "revisions": [
    {
      "version": "v1",
      "updated": "2021-06-22T20:37:12.000Z"
    }
  ],
  "analyses": {
    "keywords": [
      "disparate impact",
      "model accuracy",
      "differential privacy",
      "significantly high drop-in model utility",
      "achieving better fairness-privacy trade-offs"
    ],
    "note": {
      "typesetting": "TeX",
      "pages": 4,
      "language": "en",
      "license": "arXiv",
      "status": "editable"
    }
  }
}