DP-SGD vs PATE: Which Has Less Disparate Impact on Model Accuracy?
Authors:
- Archit UniyalAUArchit Uniyal
- Rakshit NaiduRNRakshit Naidu
- Sasikanth KottiSKSasikanth Kotti
- Sahib SinghSSSahib Singh
- Patrik Joslin KenfackPKPatrik Joslin Kenfack
- Fatemehsadat MireshghallahFMFatemehsadat Mireshghallah
- Andrew TraskATAndrew Trask
Preprint
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.Keywords
All Related Versions
This publication has 8 references indexed in Scilit:
- Decision Making with Differential Privacy under a Fairness Lens2021
- A Survey on Bias and Fairness in Machine LearningACM Computing Surveys, 2021
- Differential privacy of hierarchical Census data: An optimization approachArtificial Intelligence, 2021
- Revisiting Membership Inference Under Realistic AssumptionsProceedings on Privacy Enhancing Technologies, 2021
- Deep Learning With Gaussian Differential PrivacyHarvard Data Science Review, 2020
- Fair decision making using privacy-protected data2020
- Membership Inference Attacks Against Machine Learning Models2017
- Deep Residual Learning for Image Recognition2016