Analyzing VGG-19’s Bias in Facial Beauty Prediction: Preference for Feminine Features

From an evolutionary perspective, sexual dimorphism has been linked to perceived attractiveness, with masculine traits preferred in men and feminine traits in women. Moreover, symmetry is a strong predictor of facial attractiveness across both sexes. Recent advancements in the field of artificial intelligence have enabled algorithms to accurately predict facial attractiveness. This study aims to investigate whether these algorithms accurately replicate human judgments of attractiveness. We hypothesized that sexually dimorphic manipulations (masculinized men and feminized women) (H1), as well as symmetrized versions (H2), would elicit higher attractiveness ratings from a facial beauty prediction algorithm. Employing transfer learning, we trained six deep-learning models using four facial databases with attractiveness ratings (n = 6848). The top-performing model, VGG-19, demonstrated a high prediction correlation of .86 on the test set. Surprisingly, our findings revealed an interaction effect between sex and sexual dimorphism. Feminized versions of both men’s and women’s faces obtained higher attractiveness ratings than their masculinized counterparts. For symmetry, our results indicated that symmetrized faces were perceived as more attractive, albeit exclusively among women. These findings offer novel insights into the understanding of facial attractiveness from both algorithmic and human behavioral perspectives.

@article{Fernandes2024,
  title     = {Analyzing VGG-19’s Bias in Facial Beauty Prediction: Preference for Feminine Features},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {15},
  number    = {10},
  year      = {2024},
  publisher = {The Science and Information Organization},
  author    = {Nuno Fernandes and Sandra Soares and Joana Arantes},
  doi       = {10.14569/IJACSA.2024.0151089},
  url       = {https://doi.org/10.14569/IJACSA.2024.0151089}
}