Group Fair Guidance: Mitigating the Bias Amplification Phenomenon in Diffusion Model

Abstract

Recently, diffusion-based generative models have been widely used in various services, demonstrating significant impact and influence. However, these models exhibit a phenomenon known as bias amplification, where the bias level of the model worsens compared to the bias present in the training data. This issue significantly weakens group fairness, raising substantial concerns. In this thesis, we confirm that there is a quality- group fairness trade-off within the context of the quality-diversity trade-off, where group fairness deteriorates at higher guidance scales. To address this issue, we propose a method designed to mitigate such tendencies within Classifier Guidance (CG), called Group Fair Guidance. Group Fair Guidance reduces the variability in group fairness with changes in the guidance scale and improves the diversity of generated samples at the sample level.