Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Leveraging Diffusion Perturbations for Measuring Fairness in Computer Vision

Nicholas Lui, Bryan Chia, William Berrios, Candace Ross, Douwe Kiela

TL;DR

This work tackles the challenge of evaluating downstream fairness in computer vision by introducing diffusion perturbations to create a race-balanced occupation dataset. It combines base image generation, race-guided inpainting, and filtering with FairFace to produce 24k images across five occupations, accompanied by a novel fairness metric that measures sensitivity of true-label probabilities to demographic perturbations. The core contribution is a diffusion-based pipeline and a robust evaluation framework that reveals model disparities across vision-language systems, particularly for non-Caucasian identities, and demonstrates that diffusion methods can meaningfully aid fairness assessment. By releasing the dataset and code, the authors enable broader reuse and comparative studies, advocating diffusion-based fairness analyses as a scalable, automated auditing tool for vision models.

Abstract

Computer vision models have been known to encode harmful biases, leading to the potentially unfair treatment of historically marginalized groups, such as people of color. However, there remains a lack of datasets balanced along demographic traits that can be used to evaluate the downstream fairness of these models. In this work, we demonstrate that diffusion models can be leveraged to create such a dataset. We first use a diffusion model to generate a large set of images depicting various occupations. Subsequently, each image is edited using inpainting to generate multiple variants, where each variant refers to a different perceived race. Using this dataset, we benchmark several vision-language models on a multi-class occupation classification task. We find that images generated with non-Caucasian labels have a significantly higher occupation misclassification rate than images generated with Caucasian labels, and that several misclassifications are suggestive of racial biases. We measure a model's downstream fairness by computing the standard deviation in the probability of predicting the true occupation label across the different perceived identity groups. Using this fairness metric, we find significant disparities between the evaluated vision-and-language models. We hope that our work demonstrates the potential value of diffusion methods for fairness evaluations.

Leveraging Diffusion Perturbations for Measuring Fairness in Computer Vision

TL;DR

This work tackles the challenge of evaluating downstream fairness in computer vision by introducing diffusion perturbations to create a race-balanced occupation dataset. It combines base image generation, race-guided inpainting, and filtering with FairFace to produce 24k images across five occupations, accompanied by a novel fairness metric that measures sensitivity of true-label probabilities to demographic perturbations. The core contribution is a diffusion-based pipeline and a robust evaluation framework that reveals model disparities across vision-language systems, particularly for non-Caucasian identities, and demonstrates that diffusion methods can meaningfully aid fairness assessment. By releasing the dataset and code, the authors enable broader reuse and comparative studies, advocating diffusion-based fairness analyses as a scalable, automated auditing tool for vision models.

Abstract

Computer vision models have been known to encode harmful biases, leading to the potentially unfair treatment of historically marginalized groups, such as people of color. However, there remains a lack of datasets balanced along demographic traits that can be used to evaluate the downstream fairness of these models. In this work, we demonstrate that diffusion models can be leveraged to create such a dataset. We first use a diffusion model to generate a large set of images depicting various occupations. Subsequently, each image is edited using inpainting to generate multiple variants, where each variant refers to a different perceived race. Using this dataset, we benchmark several vision-language models on a multi-class occupation classification task. We find that images generated with non-Caucasian labels have a significantly higher occupation misclassification rate than images generated with Caucasian labels, and that several misclassifications are suggestive of racial biases. We measure a model's downstream fairness by computing the standard deviation in the probability of predicting the true occupation label across the different perceived identity groups. Using this fairness metric, we find significant disparities between the evaluated vision-and-language models. We hope that our work demonstrates the potential value of diffusion methods for fairness evaluations.
Paper Structure (36 sections, 1 equation, 2 figures, 8 tables)

This paper contains 36 sections, 1 equation, 2 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Demographic perturbations of images.
  4. Datasets balanced along demographic traits.
  5. Method
  6. Prompt Creation
  7. Text-to-Image Generation
  8. Automated Filtering of Base Image Using VQA Model
  9. Text-to-image faithfulness.
  10. Limb realism.
  11. Overall realism.
  12. Mask Generation
  13. Perturbation Using Inpainting
  14. Filtering Perturbed Sets Using FairFace
  15. Limitations of Image Generation
  16. ...and 21 more sections

Figures (2)

  • Figure 1: Our main contributions. We propose (1) a novel diffusion-based approach to generate a dataset balanced along demographic traits, and (2) a fairness metric to measure a model's robustness to demographic perturbations. We apply these techniques to (3) the creation of an occupations dataset and (4) produce fairness insights.
  • Figure 2: Samples of images generated for each occupation. The base image is generated using the text prompt "A photo of the face of a $<$occupation$>$". We then generate a mask over the person(s) in the image. The original prompt is perturbed 4 times to include 4 different race identifiers: "A photo of the face of a [Black$|$Caucasian$|$Asian$|$Indian] $<$occupation$>$". Base image-mask pairs are passed into the inpainting pipeline which produces four variants of the base image.