Gendered Double Standards in CelebA: A Multidimensional Audit
arXiv:2605.15312
Abstract
Media scholarship has long documented gendered double standards of ageing and beauty, yet how these standards appear in combinations of visual attributes and their evaluation in computer vision datasets remains underexplored. We examine how gendered appearance norms are reflected in CelebA's subjective "Attractive" annotation and in models trained to predict it. Drawing on a multidimensional approach to representational bias, we audit 202,599 images across dataset structure, learned feature associations, and spatial model sensitivity. Hierarchical clustering identifies seven attribute bundles, including groupings associated with youth, makeup, and adornment, and with ageing, facial hair, and formal attire. Logistic regression with cluster-by-perceived-sex interactions reveals an asymmetry obscured by overall label rates. Although female-perceived faces receive positive attractiveness annotations more often overall, an ageing- and masculine-coded bundle has a substantially stronger negative association with their attractiveness odds than with those of male-perceived faces, relative to a youthful, feminine-coded reference bundle. SHAP analysis of XGBoost predictions shows related differences: body-size attributes contribute negatively for female-perceived faces but are near neutral or slightly positive for male-perceived faces. Exploratory Grad-CAM analysis contrasts central facial sensitivity among the highest-scoring young female-perceived faces with upper-forehead sensitivity among older male-perceived faces. Together, these findings provide evidence consistent with gendered appearance double standards, particularly a narrower association of female attractiveness with youthful, feminine-coded presentation. Multidimensional audits can thus reveal evaluative patterns that overall label rates and predictive accuracy leave unresolved.
arXiv comment25 pages, 8 figures