Fairness-aware Regression Robust to Adversarial Attacks

Yulu Jin, Lifeng Lai

Abstract: In this paper, we take a first step towards answering the question of how to design fair machine learning algorithms that are robust to adversarial attacks. Using a minimax framework, we aim to design an adversarially robust fair regression model that achieves optimal performance in the presence of an attacker who is able to add a carefully designed adversarial data point to the dataset or perform a rank-one attack on the dataset. By solving the proposed nonsmooth nonconvex-nonconcave minimax problem, the optimal adversary as well as the robust fairness-aware regression model are obtained. For both synthetic data and real-world datasets, numerical results illustrate that the proposed adversarially robust fair models have better performance on poisoned datasets than other fair machine learning models in both prediction accuracy and group-based fairness measure.

Introduction

Machine learning models are increasingly deployed in sensitive domains, yet they face two distinct threats: societal bias and adversarial vulnerability. While researchers have studied fairness and robustness separately, they are rarely addressed together. This is problematic because adversarial attacks can specifically target a model's fairness, widening the "fairness gap" even if the model was designed to be fair on clean data. This paper explores the intersection of these challenges, proposing a framework for regression models that remain both accurate and fair in hostile environments.

Methodology

The authors frame the problem as a minimax game between a defender (the model) and an attacker. The defender seeks to minimize a combined objective of mean squared error (MSE) and a group fairness measure (accuracy parity), while the attacker seeks to maximize it.

The study analyzes two specific attack scenarios:

  1. Single Poisoned Data Point: The attacker adds one carefully designed data point. This is solved by partitioning the problem into four non-convex quadratic minimization sub-problems.
  2. Rank-One Modification: A more powerful attacker modifies the entire feature matrix. The authors use variable augmentation and an inexact proximal point method to solve this complex minimax problem.

Experimental Evaluation

The approach was evaluated on three datasets: a Synthetic Dataset (SD), the Law School Dataset (LSD), and the Medical Insurance Cost Dataset (MICD). The results demonstrate that the robust fair model consistently maintains lower fairness gaps and higher accuracy under attack compared to standard fair models.

Performance comparison on LSD and MICD datasets under single-point poisoning attacks.

As shown in the figures above, the proposed robust models (solid lines) maintain lower MSE and significantly smaller fairness gaps across varying trade-off parameters ($\lambda$) compared to their unrobust counterparts (markers). Specifically, on the MICD dataset, traditional fair models saw their fairness gaps explode under attack, while the robust model remained stable.

Pareto Efficiency

The robust model consistently occupies the "lower-left" corner of trade-off plots, indicating a superior balance of low error and low bias.

Pareto efficiency plots showing the trade-off between error gap and $R^2$ for different models.

Conclusion

This work highlights that fairness is not just a social goal but a security surface that can be weaponized. By exploiting the structural properties of regression, the authors provide a method to build models resilient to such manipulation. Future work will look to extend these guarantees to deep neural networks and classification tasks, as well as more complex multi-point poisoning attacks.

Summary updated

Abstract

In this paper, we take a first step towards answering the question of how to design fair machine learning algorithms that are robust to adversarial attacks. Using a minimax framework, we aim to design an adversarially robust fair regression model that achieves optimal performance in the presence of an attacker who is able to add a carefully designed adversarial data point to the dataset or perform a rank-one attack on the dataset. By solving the proposed nonsmooth nonconvex-nonconcave minimax problem, the optimal adversary as well as the robust fairness-aware regression model are obtained. For both synthetic data and real-world datasets, numerical results illustrate that the proposed adversarially robust fair models have better performance on poisoned datasets than other fair machine learning models in both prediction accuracy and group-based fairness measure.
BibTeX
@article{jin2023fairness,
  author = {Yulu Jin and Lifeng Lai},
  title = {{Fairness-aware Regression Robust to Adversarial Attacks}},
  journal = {IEEE Trans. Signal Process.},
  year = {2023},
  doi = {10.1109/tsp.2023.3328111},
}
Figure