Table of Links

Abstract

1 Introduction

2 Background

3 Approach and Implementation

3.1 Seed Image Collection and 3.2 Neutral Prompt List Collection

3.3 Image Generation and 3.4 Properties Assessment

3.5 Bias Evaluation

4 Evaluation

4.1 Experimental Setup

4.2 RQ1: Effectiveness of BiasPainter

4.3 RQ2 - Validity of Identified Biases

4.4 RQ3 - Bias Mitigation

5 Threats to Validity

6 Related Work

7 Conclusion, Data Availability, and References

3 APPROACH AND IMPLEMENTATION

In this section, we present BiasPainter, a metamorphic testing framework designed for measuring the social bias in image generation models. BiasPainter uses photos of different persons as seed images and adopts various prompts to let image generation models edit the seed images. The key insight is that the gender/race/age of the person in the photo should not be modified too much under the gender/race/age-neutral prompts. Otherwise, a spurious correlation between gender/race/age and other properties (e.g. career) exists in the model. Namely, a suspicious bias is detected. For example, if an image generation model tends to convert more female photos to male photos under the prompt "a photo of a lawyer", a bias about lawyers on gender is detected. Figure 2 depicts the framework of BiasPainter, which consists of five stages:

(1) Seed Image Collection: collect photos of people across different races, genders and ages as seed images.

(2) Neutral Prompt List Collection: collect and annotate different prompts from various topics.

(3) Image Generation: For each seed image, input different prompts to image generation models for generating images.

(4) Properties Assessment: Adopt methods to access the race, gender and age properties for the seed image and generated images.

(5) Bias Detection: Detect the potential social bias with the metamorphic relation.