Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Privacy-Preserving Debiasing using Data Augmentation and Machine Unlearning

Zhixin Pan, Emma Andrews, Laura Chang, Prabhat Mishra

TL;DR

This work tackles the dual challenges of data bias and privacy leakage in machine learning by marrying guided diffusion-based data augmentation with step-wise, distributed machine unlearning. The approach iteratively generates bias-aware synthetic samples and recollects the model to forget original data, aiming to reduce bias without sacrificing performance while defending against membership inference attacks. Key contributions include the first integrated framework for data augmentation and unlearning, a bias-aware guided-diffusion objective using KL-divergence as a bias metric, and a distributed unlearning scheme with a deletion-capacity bound. Empirical results on CIFAR-10 and CelebA demonstrate substantial bias reduction and improved resilience to MIAs, with high-quality synthetic data and minimal degradation in accuracy, underscoring the practical potential of privacy-preserving debiasing in real-world deployments.

Abstract

Data augmentation is widely used to mitigate data bias in the training dataset. However, data augmentation exposes machine learning models to privacy attacks, such as membership inference attacks. In this paper, we propose an effective combination of data augmentation and machine unlearning, which can reduce data bias while providing a provable defense against known attacks. Specifically, we maintain the fairness of the trained model with diffusion-based data augmentation, and then utilize multi-shard unlearning to remove identifying information of original data from the ML model for protection against privacy attacks. Experimental evaluation across diverse datasets demonstrates that our approach can achieve significant improvements in bias reduction as well as robustness against state-of-the-art privacy attacks.

Privacy-Preserving Debiasing using Data Augmentation and Machine Unlearning

TL;DR

This work tackles the dual challenges of data bias and privacy leakage in machine learning by marrying guided diffusion-based data augmentation with step-wise, distributed machine unlearning. The approach iteratively generates bias-aware synthetic samples and recollects the model to forget original data, aiming to reduce bias without sacrificing performance while defending against membership inference attacks. Key contributions include the first integrated framework for data augmentation and unlearning, a bias-aware guided-diffusion objective using KL-divergence as a bias metric, and a distributed unlearning scheme with a deletion-capacity bound. Empirical results on CIFAR-10 and CelebA demonstrate substantial bias reduction and improved resilience to MIAs, with high-quality synthetic data and minimal degradation in accuracy, underscoring the practical potential of privacy-preserving debiasing in real-world deployments.

Abstract

Data augmentation is widely used to mitigate data bias in the training dataset. However, data augmentation exposes machine learning models to privacy attacks, such as membership inference attacks. In this paper, we propose an effective combination of data augmentation and machine unlearning, which can reduce data bias while providing a provable defense against known attacks. Specifically, we maintain the fairness of the trained model with diffusion-based data augmentation, and then utilize multi-shard unlearning to remove identifying information of original data from the ML model for protection against privacy attacks. Experimental evaluation across diverse datasets demonstrates that our approach can achieve significant improvements in bias reduction as well as robustness against state-of-the-art privacy attacks.
Paper Structure (18 sections, 1 theorem, 11 equations, 12 figures, 1 table, 2 algorithms)

This paper contains 18 sections, 1 theorem, 11 equations, 12 figures, 1 table, 2 algorithms.

Table of Contents

  1. Introduction
  2. Background and Preliminaries
  3. Data Bias and Countermeasures
  4. Membership Inference Attack
  5. Notations and Preliminaries
  6. Proposed Method
  7. Overview
  8. Guided Diffusion Model based Augmentation
  9. Machine Unlearning
  10. Experimental Evaluation
  11. Experimental Setup
  12. Reduction of Data Bias
  13. Class-Imbalance Reduction:
  14. Preservation of Data Quality
  15. Performance of Model Unlearning
  16. ...and 3 more sections

Key Result

Theorem 1

Let $\epsilon = 1$ and $\delta < 0.005$. There exists a 4-Lipschitz, 1-strongly convex function $f$ and distribution $D$, such that for any learning algorithm $A$ and unlearning algorithm $\hat{A}$, the deletion capacity where $c$ depends on the properties of function $f$ and is strictly less than1.

Figures (12)

  • Figure 1: Integration of data augmentation and machine unlearning can reduce both data bias and privacy attacks.
  • Figure 2: An ML model trained with a biased feature distribution (e.g., no black cat) can lead to mispredictions.
  • Figure 3: Proposed framework with data augmentation and machine unlearning can achieve both privacy and fairness.
  • Figure 4: Our ML framework can compute both learning and unlearning in a parallel and distributed fashion.
  • Figure 5: Distribution of individual classes in randomly selected subset of CIFAR-10 with 5000 samples
  • ...and 7 more figures

Theorems & Definitions (4)

  • Definition 1
  • Definition 2
  • Definition 3
  • Theorem 1