Abstract
As machine learning models are increasingly deployed in high-stakes domains, ensuring both fairness and adversarial robustness has become paramount. However, these two objectives are often studied in isolation. This paper investigates the relationship between them, uncovering a fundamental conflict: enforcing fairness can significantly increase a model's vulnerability to adversarial attacks. Through theoretical analysis and empirical evaluation on datasets such as UTKFace, Fashion-MNIST, and CIFAR-10, we demonstrate that fairness constraints tend to pull decision boundaries closer to data points, thereby reducing the safety margin required for robustness. We propose the use of a bounded loss function, specifically Ramp loss, to mitigate this trade-off and achieve a more balanced performance across accuracy, fairness, and security.
Introduction
The rapid adoption of AI in sensitive applications like hiring and legal sentencing has brought fairness and robustness to the forefront of research. Fairness aims to prevent discrimination against protected groups, while robustness ensures that models remain reliable under small, malicious input perturbations. This study addresses the "dichotomy" between these two properties, asking whether making a model fairer inherently makes it less secure.
The Conflict Between Fairness and Robustness
The core finding of this research is that fairness and robustness frequently pull a model's decision boundary in opposite directions. Fairness constraints often act as a magnet, drawing the decision boundary toward data points to equalize error rates across different groups. In contrast, robustness requires maximizing the distance between the data and the decision boundary to create a safety margin.
This "Proximity Problem" is most severe when groups have different levels of data variance. Fairness attempts to compensate for high variance in one group by moving the boundary closer to it, which inadvertently makes that group more susceptible to tiny amounts of noise that a regular model would ignore.
Methodology and Decision Boundary Analysis
The authors utilize a theoretical framework based on binary classification with Gaussian mixtures to mathematically prove how optimal fair classifiers differ from robust ones. They identify the average distance to the decision boundary as the critical metric linking these two properties.
To address the sensitivity of standard unbounded loss functions (like cross-entropy) to outliers—which can warp the decision boundary when combined with fairness constraints—the paper introduces a mitigation strategy using Ramp loss. This bounded loss function limits the influence of misclassified samples, helping to find a better balance between competing objectives.
Experimental Results
Empirical validation across various deep learning architectures (CNNs, ResNet-50, VGG-13) confirms the theoretical predictions. Fair models were found to have boundary errors up to 9% higher than standard models.
As shown in the analysis of the fairness penalty ($\lambda$), increasing the weight of fairness constraints leads to a sharp drop in the distance to the decision boundary, directly correlating with increased adversarial vulnerability.
Conclusion
This work highlights a critical challenge in building trustworthy AI: the trade-off between equity and security. While enforcing fairness is essential, it must be done with an awareness of the potential impact on robustness. The proposed Ramp loss strategy offers a path forward, but the authors emphasize that this is a call for further research into training paradigms that can simultaneously satisfy both requirements without compromise.
Summary updated