Image Counterfactual Sensitivity Analysis for Detecting Unintended Bias

By Emily L. Denton, Ben Hutchinson, Margaret Mitchell, Timnit Gebru. 2019

129 citations recorded by the source corpus.

Facial analysis models are increasingly used in applications that have serious impacts on people's lives, ranging from authentication to surveillance tracking. It is therefore critical to develop techniques that can reveal unintended biases in facial classifiers to help guide the ethical use of facial analysis technology. This work proposes a framework called \textit{image counterfactual sensitivity analysis}, which we explore as a proof-of-concept in analyzing a smiling attribute classifier trained on faces of celebrities. The framework utilizes counterfactuals to examine how a classifier's prediction changes if a face characteristic slightly changes. We leverage recent advances in generative adversarial networks to build a realistic generative model of face images that affords controlled manipulation of specific image characteristics. We then introduce a set of metrics that measure the effect of manipulating a specific property on the output of the trained classifier. Empirically, we find several different factors of variation that affect the predictions of the smiling classifier. This proof-of-concept demonstrates potential ways generative models can be leveraged for fine-grained analysis of bias and fairness.

Subfields: Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

Citations in the corpus

Generative Models for Effective ML on Private, Decentralized Datasets (2019, background): “ factor in prediction if the model has erroneously ﬁt an assumption about them during training (e.g., a smiling face detector that judges smiling vs. non-smiling based on gender or gender expression (Denton et al., 2019)). When entire groups are unor underrepresented in training, the result can be b”
Designing an encoder for StyleGAN image manipulation (2021, background): “Early works [10, 11, 34] use a fully-supervised approach and find latent directions for binary-labeled attributes such as young ↔ old or smile ↔ no-smile.”
Encoding in Style: a StyleGAN Encoder for Image-to-Image Translation (2020, background): “One popular approach is to ﬁnd linear directions that correspond to changes in a given binary labeled attribute, such as young ↔ old, or no-smile ↔ smile [35, 11, 10, 3].”
Disentangling in Latent Space by Harnessing a Pretrained Generator (2020, background): “A popular approach is to find linear directions that correspond to changes in a given binary labeled attribute, such as young↔ old, or no-smile↔ smile [15, 51, 20].”
Only a matter of style: age transformation using a style-based regression model (2021, background): “…either by directly learning a disentangled mapping of an attribute of interest [Nitzan et al. 2020] or by a latent space traversal [Abdal et al. 2020b; Denton et al. 2019; Goetschalckx et al. 2019; Härkönen et al. 2020; Shen et al. 2020; Shen and Zhou 2020; Voynov and Babenko 2020; Wu et al. 2020].”
HyperStyle: StyleGAN Inversion with HyperNetworks for Real Image Editing (2021, background): “These range from full-supervision in the form of semantic labels [3, 16, 19, 56] and facial priors [59,60] to unsupervised approaches [24,57,63,64].”
Pivotal Tuning for Latent-based Editing of Real Images (2021, methodology): “Some using fullsupervision in the form of semantic labels [10, 11, 34], others [15, 30, 35] find meaningful directions in a selfsupervised fashion, and finally recent works present unsupervised methods to achieve the same goal [14, 39, 40], requiring no manual annotations.”
Towards causal benchmarking of bias in face analysis algorithms (2020, background): “CelebA is the most popular public face dataset due to its size and rich attribute annotations, but is known to have severe imbalances [19].”
Explaining Visual Models by Causal Attribution (2019, background): “[4] conceives counterfactuals in the sense of altering just one of the latent factors and estimating its effect.”
Taxonomizing and Measuring Representational Harms: A Look at Image Tagging (2023, methodology): “…example, researchers have re-taken photos using props, such as wigs (Rodden 2017; Algorithmic Justice League 2021), or used GANs to generate faces with different characteristics (Denton et al. 2021; McDuff et al. 2019), testing whether there are any differences in the tags applied to these images.”
Fair Attribute Classification through Latent Space De-biasing (2020, methodology): “[14] and measure the change in target attribute score as we perturb the protected attribute in images.”