PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards causal benchmarking of bias in face analysis algorithms

Guha Balakrishnan, Yuanjun Xiong, Wei Xia, Pietro Perona

arXiv:2007.06570v1cs.CVcs.CYcs.LG

TL;DR

Measuring bias in computer vision requires assessing whether attribute changes affect algorithm performance, but controlling other variables in image data is difficult. The authors use human-guided synthetic transects to compare experimental and observational bias analyses, which reach different conclusions about the sources of gender-classifier errors.

  • Problem

    Bias measurement asks whether changing a protected attribute while holding others fixed changes algorithm performance, but image data make this experimental control difficult.

    Source: §1 · p. 3

  • Method

    The authors generate synthetic matched-image transects that vary selected attributes, use human annotations to quantify perceived manipulations, and compare those annotations with algorithm outputs.

    Source: §1 · p. 4

  • Results

    The observational analysis flagged skin color while the experimental analysis did not, and the experimental analysis revealed age, hair-length, and facial-hair biases; both methods agreed on gender.

    Source: §6.1 · p. 27

  • Takeaways & Limitations

    The experimental findings point to adding training images of long-haired men and short-haired women across races, and of children, teens, and possibly seniors, to address identified biases.

    Source: §6.1 · p. 27

    The method reduces unwanted correlations with annotated variables but does not eliminate them or account for hidden confounders.

    Source: §6.2 · p. 27

Abstract

from arXiv · show

Measuring algorithmic bias is crucial both to assess algorithmic fairness, and to guide the improvement of algorithms. Current methods to measure algorithmic bias in computer vision, which are based on observational datasets, are inadequate for this task because they conflate algorithmic bias with dataset bias. To address this problem we develop an experimental method for measuring algorithmic bias of face analysis algorithms, which manipulates directly the attributes of interest, e.g., gender and skin tone, in order to reveal causal links between attribute variation and performance change. Our proposed method is based on generating synthetic ``transects'' of matched sample images that are designed to differ along specific attributes while leaving other attributes constant. A crucial aspect of our approach is relying on the perception of human observers, both to guide manipulations, and to measure algorithmic bias. Besides allowing the measurement of algorithmic bias, synthetic transects have other advantages with respect to observational datasets: they sample attributes more evenly allowing for more straightforward bias analysis on minority and intersectional groups, they enable prediction of bias in new scenarios, they greatly reduce ethical and legal challenges, and they are economical and fast to obtain, helping make bias testing affordable and widely available. We validate our method by comparing it to a study that employs the traditional observational method for analyzing bias in gender classification algorithms. The two methods reach different conclusions. While the observational method reports gender and skin color biases, the experimental method reveals biases due to gender, hair length, age, and facial hair.

1 Introduction

Observational bias studies can confound attribute effects with correlated features and test-set composition, limiting causal conclusions. The paper proposes controlled synthetic transects with human annotations and evaluates them on gender classifiers, finding a different bias picture from observational analysis.

1 Introduction

  • Bias measurement is intended to reveal whether changing a protected attribute while holding other factors equal changes algorithm output, and to guide corrective actions.The stated goal is to reveal causal connections between attributes and algorithmic performance.Source: §1 · p. 3
  • Observational studies cannot reliably attribute performance differences to a variable of interest when other variables correlate with it.Wild datasets may also underrepresent combinations of attributes and fail to represent the target population, making generalization hit-or-miss.Source: §1 · pp. 2–3
  • The proposed method synthesizes matched image transects that vary selected attributes and compares algorithm predictions with human annotations.Human annotations provide generator-independent ground truth; the method can vary multiple attributes in grid-like samples.Source: §1 · pp. 3–4, Figure 2 · p. 4
  • The experimental analysis and PPB observational analysis diverged most on skin color; the experimental method also identified age, hair length, and facial hair as bias sources, while both agreed on gender.The authors suggest correlations among hair length, skin color, and gender in PPB may explain the differing conclusions.Source: §1 · p. 4
  • The authors evaluate the method on two gender classifiers and report more balanced coverage of key attributes than in wild face datasets.They describe the work as a first step; further design and experimentation are needed for broadly applicable and reliable techniques.Source: §1 · p. 4

2 Related Work

Prior face-analysis bias studies often use observational data, raising the question of whether measured disparities reflect algorithms or test-data correlations. This work distinguishes its controlled measurement goal from related dataset-bias, model-training, generative-model, and network-interpretation research.

2 Related Work

  • The paper treats measuring bias in fixed, pre-trained black-box systems as complementary to work that trains models for subgroup-equal performance.Its focus is reliable measurement, rather than mitigating bias through model training.Source: §2 · p. 5
  • Prior face-analysis and face-recognition bias evaluations use observational rather than interventional techniques, so their conclusions warrant caution.The observational setup leaves open whether measured biases depend on algorithmic bias or correlations in test data.Source: §2 · pp. 4–5
  • Earlier experimental work altered skin color in photographs, but extending that approach to attributes such as hair length was described as onerous.The authors aim for a practical method that can study any attribute independently.Source: §2 · p. 5
  • Some generative-model studies assume one attribute can be changed at a time; this framework instead uses human annotations to account for residual attribute correlations.The cited assumption depends on unbiased generative-model training data, which the passage says is almost always impractical.Source: §2 · p. 5
  • Unlike saliency, network dissection, and TCAV approaches to interpreting networks, this method synthesizes modified inputs and quantifies them with human annotations.The passage describes those interpretation approaches as examining salient input regions, neuron effects, or conceptual sensitivity.Source: §2 · p. 5

3 Face Attribute Annotation in Synthetic Images

For synthetic faces, the paper treats attributes as perceived outward appearances rather than intrinsic states of real people. Human observers supply those judgments, with attributes discretized to fit the study and observers’ reporting abilities.

3 Face Attribute Annotation in Synthetic Images

  • Synthetic images have no person or intrinsic gender, ethnicity, age, or emotion; the study concerns observers’ perceptions of displayed attributes.The paper uses the perceived gender of idealized statues as an analogy for perception without a real person behind the image.Source: §3 · p. 6
  • An attribute measured from an image means an algorithm’s prediction of a casual observer’s report of perceived outward display.The authors report that observer measurements are consistent and reproducible, and therefore treat their statistics as objective quantities.Source: §3 · p. 6
  • The study discretizes continuous attributes into classes, without claiming its chosen discretization is ideal for every context.The choices accommodate casual observers and conform to existing conventions such as the Fitzpatrick skin-tone scale.Source: §3 · p. 6
  • Gender is measured as perceived gender, not gender identity, and is represented on a five-step one-dimensional scale.Observers could not reliably report beyond the traditional M/F dimension; the five steps accommodate differing confidence and ambiguity.Source: §3 · p. 6

4 Method

The method creates synthetic, human-annotated image transects to vary semantic face attributes in matched samples, then analyzes classifier error across subgroups and estimates covariate-adjusted effects. Its design combines attribute-controlled synthesis, orthogonalized traversal directions, and checks for remaining confounding and image realism.

4 Method

  • The framework pairs semantic-attribute-controlled image synthesis with human annotations to analyze recognition-system errors and estimate covariate-adjusted effects.Transects are grid-like matched samples whose axes manipulate different attributes.Source: §4 · pp. 6–7

4.1 Transects: A Walk in Face Space

  • A face generator maps sampled latent vectors to images, and the method learns attribute directions to create grids spanning multiple attributes.The authors use StyleGAN2 and generalize latent-space attribute traversals to transects.Source: §4.1 · p. 7

4.1.1 Estimating Latents-to-Attributes Linear Models

  • Human labels guide attribute modeling rather than classifier-generated labels, reducing the risk that classifier bias from attribute correlations enters the synthesis method.The authors note that native StyleGAN2 sampling favors Caucasian-looking faces because FFHQ is demographically skewed; human-guided sampling partially mitigates this.Source: §4.1.1 · p. 8

4.1.2 Multi-attribute Transect Generation

  • Multidimensional transects project a latent point onto the intersection of attribute hyperplanes, then sample a grid along the selected attribute directions.The iterative projection converges to the intersection; the authors found convergence in fewer than 50 iterations.Source: §4.1.2 · p. 9

4.1.3 Orthogonalization of Traversal Directions

  • Because attribute directions can be correlated, traversal vectors are adjusted to be perpendicular to other attribute normals, reducing unintended attribute changes.Orthogonalization removes the unwanted masculine appearance shift in shorter-hair transects, while skin-color transects show little difference.Source: §4.1.3 · p. 10

4.1.4 Setting Step Sizes and Transect Dimensions

  • Transect dimensions are kept small to limit annotation costs, with attribute ranges set manually and interpolated across each dimension.The examples use five points for 1D and 2D transects and two points for 3D transects.Source: §4.1.4 · p. 10

4.2 Analyses Using Transects

  • Analysis compares classifier error across attribute subgroups and estimates average treatment effects, using regression adjustment when annotated covariates may still confound comparisons.For bounded error values, the experiments use logistic regression; its coefficient represents change in log odds per unit attribute change.Source: §4.2 · pp. 10, 12

4.3 Human Annotation

  • Human annotators label eight attributes on discrete scales, with five annotations per image and attribute; images judged too fake are excluded from analysis.The annotation effort totaled 0.52 million judgments and took about 10 hours; annotators generally showed good agreement.Source: §4.3 · p. 13

5 Experiments

The experiments compare synthetic, human-annotated transects with wild-collected face datasets to examine gender-classifier bias. Transects expose attribute effects that differ from the pattern suggested by PPB, especially the limited role of skin tone relative to other attributes.

5 Experiments

  • PPB reported higher error on dark-skinned females, and the experiments use transects to test whether changing gender and skin color while holding other attributes constant changes classifier performance.The comparison is framed as a test of the hypothesis prompted by PPB's observational findings.Source: §5 · pp. 13–15

5.1 Gender Classifiers

  • Two ResNet-50 gender classifiers were trained on CelebA and FairFace; their biases were reported as comparable to those in the original PPB study.The authors chose research-grade models rather than commercial systems to support reproducibility.Source: §5.1 · pp. 15–17

5.2 Transect Data

  • The team generated 1,000 2×2×2 transects varying skin color, hair length, and gender while holding other face characteristics constant.The StyleGAN2-based method used a style space selected for semantic-attribute disentanglement.Source: Figure 9 · p. 17, §5.2 · p. 18

5.2.1 Dataset Pruning

  • Pruning removed faces rated at least “Likely fake” and ambiguous mid-range gender, skin-color, or hair-length images, leaving 5,713 faces.The attribute ranges removed were [0.4, 0.6] for skin color and gender and [0.3, 0.5] for hair length.Source: §5.2.1 · p. 18, Figure 10 · p. 19

5.3 Comparison of Transects to Real Face Datasets

  • Wild-collected datasets show gender imbalances in attributes, including hair length, and hair length correlates with skin color; transects provide better balance across these attributes.Transects were designed to mirror PPB's skin-color and age distributions while mitigating hair-length imbalance.Source: §5.3 · pp. 18, 21, Figure 11 · p. 20, Figure 12 · p. 20

5.4 Analysis of Bias

  • Transect errors broadly resemble PPB's female-higher pattern, but shorter hair raises women's errors and skin tone shows no consistent effect within homogeneous groups.Transect errors are comparable to or higher than PPB errors; matched samples also make relative attribute comparisons possible where PPB has sparse groups.Source: §5.4 · p. 21, Figure 14 · p. 22, Figure 13 · p. 22

5.5 Regression Analysis

  • Facial hair, gender, makeup, hair length, and age significantly affect classifier error, while skin color has a negligible effect in the transect regression analysis.The authors conclude that observational PPB analysis and the experimental transect method point to different causes of gender-classification bias.Source: Figure 15 · p. 23, §5.5 · p. 23

5.5.1 Joint Effects of Attributes on Classification Error

  • Age and gender have joint effects on errors, with gender differences varying for children, teenagers, and seniors compared with adults.ResNet-CelebA particularly misclassifies young male children and babies as female, consistent with the age-pattern analysis.Source: Figure 16 · p. 24, §5.5.1 · p. 24

6 Discussion and Conclusions

The experimental approach produces conclusions that differ from observational analyses and supports causal, intersectional bias analysis. Its findings also depend on careful control of residual correlations and synthesis limitations.

6.1 Summary

  • Human-calibrated synthesis creates matched transects that can explore any human-annotatable attribute without requiring perfect synthesis.Matched samples support counterfactual analysis; approximately uniform attribute sampling also makes underrepresented and intersectional groups easier to assess.Source: §6.1 · pp. 25–26
  • Bias measurements may generalize across populations and support predictions of disparate treatment and disparate impact for a specified population.The paper attributes this to identifying causally linked attributes independently of their prevalence.Source: §6.1 · p. 26
  • Synthetic datasets reduce ethical and legal risks because they do not depict real people, and can be produced and annotated relatively quickly and affordably.The authors say this may make bias testing more widely accessible.Source: §6.1 · p. 27
  • Transects and PPB diverge on skin-color bias: PPB flags it, while the experimental analysis does not; both methods agree on gender bias.The experimental analysis additionally identifies age, hair length, and facial hair as bias sources, and attributes the divergence likely to correlated hair length, skin color, and gender in PPB.Source: §6.1 · p. 27
  • The differing diagnoses point to different training priorities: long-haired men, short-haired women across races, and child-teen and possibly senior age groups.The PPB-based analysis could instead suggest adding more dark-skinned women; these are recommendations inferred from the respective bias measurements.Source: §6.1 · p. 27
  • Bias analysis should include attributes beyond those of immediate interest because hidden confounders and nonlinear attribute interactions can matter.The paper points to the age–gender interaction and argues for intersectional analysis and multidisciplinary judgment when selecting attributes.Source: §6.1 · p. 27

6.2 Limitations and Future Work

  • Transects can retain unwanted correlations: hair-length edits may add facial hair, and gender edits may add earrings that remain unannotated.Researchers must watch for residual correlations and hidden confounders; discovered factors can be annotated, compensated for in analysis, and mitigated in synthesis.Source: §6.2 · pp. 27–28
  • Face synthesis still has visible artifacts, limited coverage of producible physiognomies and attribute combinations, and a StyleGAN2 tendency to generate Caucasian faces from latent sampling.The authors also call for improved synthesis controls and note that extending the method beyond gender classification to tasks such as face recognition is not straightforward and requires further study.Source: §6.2 · p. 28

Related papers

Loading 2007.06570v1…