This paper, titled "Fairness-aware Regression Robust to Adversarial Attacks," by Yulu Jin and Lifeng Lai, explores the intersection of two critical challenges in modern machine learning: ensuring models are fair to different demographic groups and making them resilient to malicious data manipulation.
TL;DR
The authors propose a minimax framework to design regression models that remain both accurate and fair even when an adversary poisons the training data. By decomposing complex, non-convex optimization problems into manageable sub-problems, they provide a way to build "adversarially robust fair" models that outperform standard fair models in hostile environments.
What Problem Does This Paper Address?
Machine learning models often face two distinct threats. First, they can inherit or amplify societal biases (e.g., discriminating based on race or gender). Second, they are vulnerable to adversarial attacks, where a malicious actor "poisons" the training data to degrade performance.
While researchers have studied fairness and robustness separately, they are rarely addressed together. This is a problem 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. The authors tackle the mathematical challenge of optimizing for both simultaneously, which results in a "nonsmooth nonconvex-nonconcave minimax problem"—essentially a worst-case scenario for standard optimization algorithms.
Methods
The authors frame the problem as a game between a defender (the model) and an attacker. The defender wants to minimize a combined objective of prediction error (MSE) and a group fairness measure (accuracy parity). The attacker wants to maximize this same objective.
They analyze two specific attack scenarios:
- Single Poisoned Data Point: The attacker adds one carefully designed data point to the dataset. The authors solve this by breaking the problem into four sub-problems, each being a non-convex quadratic minimization problem with multiple constraints (QCQP). They identify 8 specific cases to find a global minimizer.
- Rank-One Modification: A more powerful attacker modifies the entire feature matrix using a rank-one matrix. To solve this, the authors use variable augmentation to transform the "nasty" minimax problem into a "weakly-convex-weakly-concave" form, which can be solved using an inexact proximal point method.
Datasets
The authors evaluate their approach on three datasets:
- Synthetic Dataset (SD): 200 samples with 5 features, where bias is manually injected to test the model's limits.
- Law School Dataset (LSD): 1,823 records of law students. The goal is to predict undergraduate GPA using gender as the sensitive attribute.
- Medical Insurance Cost Dataset (MICD): 1,338 examples of medical expenses in the US, using gender as the sensitive attribute.
Key Insights and Ideas
- Fairness as a Vulnerability: The paper highlights that fairness isn't just a social goal; it's a security surface. An attacker can "weaponize" fairness constraints to make a model perform poorly for specific groups.
- The Three-Way Trade-off: There is an inherent tension between accuracy, fairness, and robustness. The authors use a trade-off parameter ($\lambda$) to allow practitioners to choose how much they value fairness versus raw accuracy in the face of an attack.
- Structural Exploitation: The most significant "aha moment" is that while the general optimization problem is NP-hard, the specific structure of regression allows for exact or near-exact solutions by carefully partitioning the problem into sub-cases.
Key Results
- Superiority under Attack: On the Law School Dataset, the proposed robust fair model maintained a significantly lower fairness gap than "unrobust" fair models when subjected to a poisoned data point.
- Resilience to Matrix Attacks: In the rank-one attack scenario on the MICD dataset, traditional fair models (like Fair Kernel Learning) saw their fairness gaps explode (e.g., increasing by nearly 7x), while the authors' robust model remained stable.
- Pareto Efficiency: Qualitative plots of "Error Gap vs. MSE" show that the robust model consistently occupies the "lower-left" corner—meaning it achieves a better balance of low error and low bias than traditional linear regression or standard fair models when data is poisoned.
Limitations and Future Work
- Linear Focus: The current work focuses primarily on linear regression. While they mention kernel methods, a full extension to deep neural networks remains a challenge due to the complexity of the minimax optimization.
- Attack Scale: The "one poisoned point" attack is a useful theoretical starting point, but real-world "clean-label" poisoning attacks often involve a larger percentage of the dataset.
- Computational Cost: Solving multiple sub-problems and using proximal point methods is more computationally expensive than standard training, which might limit scalability for massive datasets.
The authors suggest that future work could explore more complex, multi-point poisoning attacks and extend these robustness guarantees to classification tasks.
Summary updated