DP-SGD vs PATE: Which Has Less Disparate Impact on Model Accuracy?

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…

Patrik Kenfack

PhD Candidate,
Computer Science

• Montréal QC, Canada
• ÉTS Montréal & Mila
• Google Scholar
• LinkedIn
• GitHub
• Email

DP-SGD vs PATE: Which Has Less Disparate Impact on Model Accuracy?

Archit Uniyal, Rakshit Naidu, Sasikanth Kotti, Sahib Singh, Patrik Kenfack, Fatemehsadat Mireshghallah, Andrew Trask

Published in ICML Workshop on Machine Learning for Data: Automated Creation, Privacy, Bias,, 2021

Workshop

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.

PaperPDF