Studying Bias in Diffusion Models
Session Number
Project ID: CMPS 14
Advisor(s)
Dr. Yan Yan
Mr. Junyi Wu, Illinois Institute of Technology
Discipline
Computer Science
Start Date
17-4-2024 8:35 AM
End Date
17-4-2024 8:50 AM
Abstract
With text-to-image (“TTI”) models becoming increasingly popular, it is imperative that we ensure they are as unbiased as possible, an issue many text-to-image models are currently facing. For example, when we ran Google’s Image-FX text-to-image model to create an image of “a CEO from the Fortune 500 list” or “a CEO of the top 10 tech companies,” it always returned a white male. On the other hand, when the same Google text-to-image model was told to generate an image of a teacher, it almost always returned an image of a white woman. In this study, we also ran a white box Stablility Diffusion model to test and detect the biases within a text-to-image model. A key reason for bias in diffusion models is the limited diversity and existing bias in the training data, as bias enters the model through training data and transfer data. To mitigate bias in diffusion models, we need to be able to measure bias and get actionable signals that can help developers take corrective actions. For this purpose, we introduce a new bias factor that can be used to measure bias in a TTI model and other such models.
Studying Bias in Diffusion Models
With text-to-image (“TTI”) models becoming increasingly popular, it is imperative that we ensure they are as unbiased as possible, an issue many text-to-image models are currently facing. For example, when we ran Google’s Image-FX text-to-image model to create an image of “a CEO from the Fortune 500 list” or “a CEO of the top 10 tech companies,” it always returned a white male. On the other hand, when the same Google text-to-image model was told to generate an image of a teacher, it almost always returned an image of a white woman. In this study, we also ran a white box Stablility Diffusion model to test and detect the biases within a text-to-image model. A key reason for bias in diffusion models is the limited diversity and existing bias in the training data, as bias enters the model through training data and transfer data. To mitigate bias in diffusion models, we need to be able to measure bias and get actionable signals that can help developers take corrective actions. For this purpose, we introduce a new bias factor that can be used to measure bias in a TTI model and other such models.