Go to ICLR 2025 Conference homepageIs the Fairness Metric Truly Fair?
Image classification is a fundamental task in computer vision that has been widely adopted in critical applications such as face recognition and medical imaging, drawing considerable attention to its predictive fairness. Some researchers have proposed various fairness metrics and pipelines to enhance the fairness of deep learning models. However, recent studies indicate that existing fairness evaluation specifications and metrics have inherent flaws, as they focus on low-dimensional inputs, such as numerical data, and overlook partial correlations between target and sensitive attributes, leading to some degree of mutual exclusivity. This raises the question: Is the fairness metric truly fair? Through in-depth analysis, experiments conclude that the fairness of deep models is closely related to attribute sampling and the interdependencies among attributes. In this work, we address this challenge by introducing a new specification based on dynamic perturbation for image classification models. Specifically, we introduce an Attribute Projection Perturbation Strategy (APPS) that moves beyond the constraints of directly statistical discrete predictions by mapping sensitive attributes that may influence task attributes onto the same dimension for evaluation. Building on this, a Projection Fairness Metric System is proposed to quantifing the upper and lower bounds of fairness perturbations, examining and evaluating the impact of mapped sensitive attributes on the fairness of task predictions from different perspectives. Additionally, we conducted systematic evaluation experiments and extensive discussions, demonstrating that the proposed evaluation specification offers better objectivity and interpretability compared to existing metrics, in 24 image classification models including CNN and ViT architectures. It is hoped that this work will promote the standardization of fairness evaluation pipeline and metrics.