Examining Pathological Bias in a Generative Adversarial Network Discriminator: A Case Study on a StyleGAN3 Model

TL;DR

This paper uncovers internal biases in a StyleGAN3 discriminator that are not explained by training data, revealing systematic stratification affecting various demographic categories.

Contribution

It identifies and analyzes internal color and luminance biases in a pre-trained GAN discriminator, highlighting issues beyond training data biases.

Findings

Discriminator exhibits internal color and luminance biases.

Biases are not explained by training data.

Scores are stratified by demographic categories.

Abstract

Generative adversarial networks (GANs) generate photorealistic faces that are often indistinguishable by humans from real faces. While biases in machine learning models are often assumed to be due to biases in training data, we find pathological internal color and luminance biases in the discriminator of a pre-trained StyleGAN3-r model that are not explicable by the training data. We also find that the discriminator systematically stratifies scores by both image- and face-level qualities and that this disproportionately affects images across gender, race, and other categories. We examine axes common in research on stereotyping in social psychology.