Privacy, Utility and Fairness: Navigating Trade-offs in Differentially Private Machine Learning

Lea DemeliusAAAI 2025 · DOI

TL;DR

This paper explores the complex interplay between privacy, utility, and fairness in differentially private machine learning (DPML). It highlights that while DPML protects individual data, it can inadvertently lead to disparate impacts on model accuracy across different subgroups, raising fairness concerns. The research investigates how hyperparameter tuning can mitigate these trade-offs, aiming to make DPML more practical and trustworthy for real-world applications.

What Problem Does This Paper Address?

As Artificial Intelligence (AI) becomes increasingly integrated into our daily lives, ensuring its trustworthiness is paramount. This includes guarantees around privacy, fairness, and utility. A significant challenge arises from the potential for AI models to leak sensitive information from the data they are trained on. Traditional anonymization techniques have proven insufficient, leading to a growing interest in Differential Privacy (DP). DP offers a mathematical framework to protect individual data while still allowing for aggregate analysis.

However, implementing DP, particularly in complex machine learning models like neural networks using methods like Differentially Private Stochastic Gradient Descent (DPSGD), often introduces trade-offs. Improving privacy can negatively impact model performance (utility) and, crucially, can lead to unfair outcomes. Specifically, research has shown that DPSGD can disproportionately affect the accuracy of certain subgroups within the data, creating a fairness concern. This paper addresses the critical question of how to navigate these trade-offs, particularly focusing on the role of hyperparameter tuning in achieving both privacy and fairness without sacrificing too much utility.

Methods

The paper's core contribution lies in its investigation of the relationships between differential privacy, utility, and fairness in machine learning. The authors focus on Differentially Private Stochastic Gradient Descent (DPSGD), a standard method for achieving DP in neural networks. Their research involves conducting in-depth experiments to understand how DPSGD influences disparities across various metrics and how these disparities are affected by hyperparameter choices. They also examine the effectiveness of hyperparameter tuning in developing private models that maintain comparable performance and fairness to non-private models. A specific variant, DPSGD-Global-Adapt, designed to mitigate disparate impact, is also explored.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • DPML's Fairness Blind Spot: A key insight is that while DP aims to protect individual privacy, it can inadvertently create or exacerbate fairness issues by disproportionately impacting the accuracy of certain data subgroups. This is a critical challenge for deploying AI responsibly.
  • Hyperparameter Tuning as a Lever: The research strongly suggests that careful hyperparameter tuning is not just about optimizing for accuracy but is crucial for balancing privacy, utility, and fairness. Reusing hyperparameters from non-private models for DP models without adjustment is likely to lead to fairness issues.
  • Interconnectedness of Trade-offs: The paper emphasizes that privacy, utility, and fairness are not independent. Improvements in one often come at the cost of another, and understanding these intricate interdependencies is vital for developing trustworthy AI.
  • The Need for Specialized Tuning: The findings imply that standard hyperparameter tuning strategies may not be sufficient for DPML. New approaches or careful adaptation of existing ones are needed to account for the unique challenges introduced by differential privacy.

Key Results

The paper reports on experimental findings that were recently submitted for publication. While specific quantitative results are not detailed in this summary, the key findings indicate:

  • DPSGD can indeed lead to disparate impacts on model accuracy across different subgroups.
  • The extent of these disparities is significantly influenced by the choice of hyperparameters.
  • Effective hyperparameter tuning can help mitigate these disparate impacts, potentially leading to private models with comparable or even better performance and fairness than their non-private counterparts.
  • The effectiveness and reliability of hyperparameter tuning strategies for achieving this balance are under investigation.

Limitations and Future Work

The authors acknowledge that while their research suggests hyperparameter tuning can mitigate disparate impacts, there are still open questions regarding the general applicability and the impact of different performance and fairness metrics.

Future work planned includes addressing a research gap concerning privacy leakage through hyperparameter tuning in differentially private machine learning. Even when individual training runs are differentially private, the process of hyperparameter tuning itself can potentially leak information. While composition can be used to account for this, it often leads to overly loose privacy guarantees. The authors plan to explore alternative methods for obtaining improved privacy guarantees during hyperparameter tuning, particularly in complex scenarios involving cross-validation and early stopping. They are also interested in investigating practical attacks that exploit information leakage through non-private hyperparameter tuning.

Summary updated

BibTeX
@inproceedings{demelius2025privacy,
  author = {Lea Demelius},
  title = {{Privacy, Utility and Fairness: Navigating Trade-offs in Differentially Private Machine Learning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i28.35205},
}
Figure