Research map / Federated learning and privacy / Differential privacy
Privacy, Utility and Fairness: Navigating Trade-offs in Differentially Private Machine Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
A PhD research proposal (AAAI-25 doctoral consortium) on trade-offs between differential privacy, utility and fairness in DPSGD, plus planned work on privacy leakage from hyperparameter tuning.
Problem
Differential privacy in deep learning (DPSGD) often reduces utility and can cause disparate impact on accuracy across subgroups; it is unclear how disparities across metrics depend on hyperparameter tuning, and hyperparameter tuning itself leaks privacy.
Approach
Proposes an investigation of how DPSGD affects disparities across multiple performance and fairness metrics, and how hyperparameter choice (including for DPSGD-Global-Adapt) governs whether private models match non-private performance and fairness; plans to address privacy leakage via hyperparameter tuning.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply DPSGD instead of non-private training | Neural networks · Deep learning | accuracy | Hurt |
| apply DPSGD instead of non-private training | Neural networks · Subgroups | accuracy disparities | Mixed |
| optimize DP-model hyperparameters for performance rather than reusing non-private settings | DPSGD · Private neural networks | fairness | Helped |
| reuse non-private-model hyperparameters without further tuning | DPSGD · Private neural networks | fairness | Hurt |
Surprising findings
- Prior work (de Oliveira et al. 2023) suggests DPSGD does not necessarily harm fairness when DP hyperparameters are tuned, with disparate impact mainly arising when non-private hyperparameters are reused.
Limitations
- The described empirical study is reported as recently submitted; no results are presented in this abstract-style paper.
- Open questions remain on generality of findings across fairness and performance metrics.
Experimental setup
Builds on: DPSGD private hyperparameter tuning
Closest papers by meaning
- Differentially Private Fair Learning (ICML 2019 · similarity 0.85)
- Setting $\varepsilon$ is not the Issue in Differential Privacy (NeurIPS 2025 · similarity 0.85)
- Differentially Private Model Merging (NeurIPS 2026 · similarity 0.83)
- Beyond the Calibration Point: Mechanism Comparison in Differential Privacy (ICML 2024 · similarity 0.83)
- Private Hyperparameter Tuning with Ex-Post Guarantee (NeurIPS 2025 · similarity 0.82)
- Instance-Specific Asymmetric Sensitivity in Differential Privacy (NeurIPS 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.