PaperlayerView exact revision on arXiv

Source-linked AI summary

Fair Attribute Classification through Latent Space De-biasing

Vikram V. Ramaswamy, Sunnie S. Y. Kim, Olga Russakovsky

arXiv:2012.01469v3cs.CV

TL;DR

The paper addresses bias in visual classifiers when target labels correlate with protected attributes. It uses GANs and latent-space perturbations to augment datasets with balanced examples, and reports improved fairness metrics across many attributes while maintaining comparable performance. The study evaluates this approach on CelebA and analyzes its effectiveness and limitations.

  • Problem

    Visual classifiers can exploit correlations between target labels and protected attributes, while balanced data collection is often infeasible.

  • Method

    The method perturbs GAN latent vectors to preserve target attributes while changing protected attributes, then augments real data with the resulting generated images.

  • Results

    The augmented classifiers achieve quantitative improvements across several fairness metrics over baselines and prior work while maintaining comparable performance.

  • Takeaways & Limitations

    GAN latent-space data augmentation shows promise for de-biasing attribute classifiers across many attributes and metrics.

  • Takeaways & Limitations

    The evaluation uses a binary Male label as perceived gender expression, although assigned sex, gender identity, and gender expression are not binary.

Abstract

from arXiv · show

Fairness in visual recognition is becoming a prominent and critical topic of discussion as recognition systems are deployed at scale in the real world. Models trained from data in which target labels are correlated with protected attributes (e.g., gender, race) are known to learn and exploit those correlations. In this work, we introduce a method for training accurate target classifiers while mitigating biases that stem from these correlations. We use GANs to generate realistic-looking images, and perturb these images in the underlying latent space to generate training data that is balanced for each protected attribute. We augment the original dataset with this perturbed generated data, and empirically demonstrate that target classifiers trained on the augmented dataset exhibit a number of both quantitative and qualitative benefits. We conduct a thorough evaluation across multiple target labels and protected attributes in the CelebA dataset, and provide an in-depth analysis and comparison to existing literature in the space.

1. Introduction

Visual classifiers can exploit correlations between target and protected attributes in training data, causing failures when correlated cues are absent. The paper proposes GAN-based latent-space augmentation to create balanced data while preserving the target attribute.

  • Motivation: Visual recognition models may exploit spurious correlations in data, contributing to disparate impact across groups.Examples include underrepresented racial groups receiving lower face-recognition accuracy and activity models amplifying gender correlations.
  • Motivation: A hat classifier trained on correlated hat-and-glasses data may fail to recognize hats without glasses, and vice versa.
  • Approach: The proposed augmentation generates images with hats but not glasses and glasses but not hats, decorrelating WearingHat from Glasses.The method makes these changes through perturbations in the GAN latent space.
  • Approach: The improved latent perturbation method preserves WearingHat while changing Glasses, using GAN-generated images to augment real-world training data.The paper illustrates this behavior with latent-space manipulation.

2. Related Work

Prior work documents bias in visual recognition and develops tools and training-based methods to detect or mitigate it. This work instead focuses on using GAN-generated, de-biased data to train classifiers.

  • Bias in gender and racial attributes is well documented, and models can amplify biases present in training data.
  • Dataset-bias mitigation includes oversampling, adversarial learning, separate protected-attribute classifiers, constraints, and regularization.
  • AI Fairness 360 and REVISE surface biases in large-scale datasets for preemptive analysis.
  • CelebA contains discrepancies in attribute labels categorized as inconsistent, gender-dependent, or gender-independent.
  • GANs have been used for low-shot and long-tail dataset augmentation, while this work targets classifiers affected by dataset bias.
  • Related GAN-based approaches generate less biased data, including methods that require small unbiased datasets or train generators for particular settings.

3. Method

The method learns semantic target and protected-attribute directions in a GAN’s latent space, then generates complementary image pairs that preserve target labels while reversing protected attributes. These generated samples augment the original data to reduce target–protected-attribute correlations across multiple attributes.

  • Problem formulation: The goal is to generate synthetic data in which target and protected attributes are de-correlated, satisfying P[ft(x)=1|fg(x)=1] = P[ft(x)=1].
  • Training and scope: The augmented training set combines real data with generated de-biased data; one GAN can support multiple target and protected attributes, including simultaneous augmentation of two protected attributes.
  • Synthetic data generation: A GAN generates images from latent vectors, while learned classifiers ft and fg assign target and protected-attribute labels to those images.
  • Latent-space modeling: Latent-space classifiers ht and hg approximate ft∘G and fg∘G using sampled latent vectors inherited from the original data distribution.
  • Latent perturbation: For each latent vector z, the method constructs z′ with ht(z′)=ht(z) and hg(z′)=−hg(z), producing a complementary pair.
  • Assumptions and implementation: The approach assumes latent spaces are approximately linearly separable in semantic attributes and uses a closed-form perturbation when classifiers are linear hyperplanes.
  • Labeling: The generated images x=G(z) and x′=G(z′) receive the same target label ft(G(z)) while their protected-attribute labels are opposite.
  • Labeling: Using ft(x) for both generated images is intended to improve target-label capture, though alternative labeling choices are possible.

4. Experiments

Experiments on CelebA evaluate the augmentation method across attribute-label categories, fairness metrics, perturbation choices, and attribute discriminability. The method generally improves fairness while retaining comparable accuracy, with stronger gains for harder target attributes and balanced latent perturbations.

  • Dataset and attributes: CelebA provides 2,022,599 face images with 40 binary attributes; experiments use 26 attributes categorized as inconsistently labeled, gender-dependent, or gender-independent.Male is treated as a binary perceived-gender-expression protected attribute, while the authors acknowledge that gender-related concepts are not binary.
  • Dataset and attributes: Inconsistently labeled attributes may yield erratic classifiers, whereas gender-dependent labels vary with perceived gender and gender-independent labels appear reasonably consistent.The gender-dependent category includes Young, while gender-independent attributes include Glasses and WearingHat.
  • Comparison with the baseline: For gender-independent attributes, AP changes from 83.9 to 83.0, while DEO changes from 16.7 to 13.9, BA from 0.3 to 0.0, and KL from 1.1 to 0.9.The augmentation method improves all three fairness metrics while maintaining comparable AP; gains are smaller for gender-dependent attributes.
  • Comparison with the baseline: Score changes under protected-attribute perturbation decrease from 0.12 to 0.09 for inconsistently labeled, 0.11 to 0.07 for gender-dependent, and 0.09 to 0.06 for gender-independent attributes.For ArchedBrows, the baseline score change is 0.41 for an underrepresented combination but 0.094 for a more prevalent one.
  • Attribute discriminability and skew: Fairness gains are greater when the target attribute is harder to learn than the protected attribute; for Young, DEO improvement is -0.2 for easy targets versus 2.1 for hard targets.Gender expression is among the easiest CelebA attributes to learn, whereas Young is harder than all but four other attributes.
  • Ablation studies: Adding GAN-generated images without latent perturbation raises AP to 82.9 versus 82.6 for the proposed model but worsens fairness, whereas adding perturbed G(z′) improves fairness.The reported fairness values are DEO 19.7 versus 16.1, BA 1.1 versus 0.5, and KL 1.6 versus 1.3; the authors attribute the result to synthetic noise and skew.

5. Extensions of our method

The paper extends latent-space de-biasing with domain-dependent target hyperplanes and tests GAN inversion for augmenting real images. Domain-dependent hyperplanes improve fairness for gender-dependent attributes, whereas GAN inversion does not appreciably improve results.

  • Domain-dependent hyperplanes: Domain-dependent hyperplanes model target attributes separately within each protected-attribute domain.The method learns separate target hyperplanes for protected labels g=1 and g=−1.
  • Domain-dependent hyperplanes: The constrained latent-space solution minimizes ||z − z′||2 by alternating gradient descent with projection onto the feasible region.The feasible region is the intersection of linear hyperplanes encoding the constraints.
  • Domain-dependent hyperplanes: For gender-dependent attributes, domain-dependent hyperplanes improve fairness without sacrificing accuracy.On validation data, DEO decreases from 21.4 to 17.2, BA from 1.5 to 0.4, and KL from 1.2 to 1.0.
  • Domain-dependent hyperplanes: For gender-independent attributes, domain-dependent hyperplanes provide no significant improvement because the target hyperplanes are similar across domains.When wt1 and wt−1 are similar, the constraints collapse to the original method.
  • GAN inversion: GAN inversion enables perturbing real images in latent space, but this more complex augmentation procedure shows no appreciable improvement.The procedure inverts real images to latent vectors, perturbs them, and generates images with the same target and opposite protected labels.

6. Conclusions

The paper introduces GAN-based latent-space augmentation to train fairer attribute classifiers when target and protected attributes are correlated. Across multiple attributes and metrics, the authors analyze when the method is effective and compare it with existing techniques.

  • The proposed method uses GAN-based data augmentation to mitigate bias from correlations between target and protected attributes.
  • The evaluation covers many attributes and fairness metrics and includes comparisons with existing techniques.
  • The authors analyze the settings in which latent-space data augmentation is most effective.

Appendix

The supplementary document adds methodological derivations, attribute-level analyses, effectiveness studies, ablation details, and an investigation of protected-attribute label requirements.

  • Section A: Section A derives a closed-form solution for z′ to manipulate latent vectors.
  • Section B: Section B reports attribute-level results and further analysis of the main experiments.
  • Section C: Section C discusses factors that influence, or do not influence, the method’s effectiveness.
  • Section D: Section D provides additional details on the ablation studies.
  • Section E: Section E investigates how many protected-attribute labels are required to achieve the desired performance.

A. Derivation

The method computes a latent-space perturbation that changes the protected-attribute score while preserving the target-attribute score, selecting the minimum-distance solution under a linear-separability assumption.

  • The perturbation seeks z′ whose protected-attribute score changes while its target-attribute score remains unchanged.The target and protected scores are approximated by functions h_t and h_g.
  • The derivation assumes the latent space is approximately linearly separable in semantic attributes, represented by normalized weight vectors and intercepts.The target and protected attribute models use w_t, w_g, b_t, and b_g.
  • Because the score constraints have infinitely many solutions, the method chooses the perturbation minimizing the distance between z and z′.The displacement z′ − z is restricted to the span of the target and protected attribute directions.
  • The derivation produces a closed-form solution for the perturbed latent vector z′.The closed-form expression is introduced after reducing the constraint equations.
  • The resulting perturbation is verified to change the protected-attribute score while maintaining the target-attribute score.

B. Attribute-level results

The attribute-level analysis tests the linear-separability assumption using synthetic samples and relates hyperplane quality, dataset skew, and downstream classifier improvement.

  • Attribute hyperplanes are estimated with 10,000 samples using linear SVM.
  • Table 8 reports hyperplane accuracy, average precision, positive-sample percentage, training-set skew, and improvement over the baseline across attributes.
  • Most attributes are well separated by the estimated hyperplanes, except highly skewed attributes with too few underrepresented-subgroup examples.The skew compares positive-sample counts across protected-attribute classes.
  • The authors report improvement over the baseline on four evaluation metrics but find no immediate correlation between hyperplane quality and downstream performance.

B.2 Changes in baseline score

The analysis evaluates whether latent perturbations preserve target scores and examines how manipulating one attribute affects others. Target-score changes are larger for less frequent label combinations, while some attributes are especially sensitive to manipulation.

  • Changes in baseline score: The target-score preservation analysis uses the absolute change in baseline classifier score over 5,000 images as a proxy.The authors explicitly note that the proxy is flawed because the baseline classifier performs worse on minority examples.
  • Changes in baseline score: A strong negative correlation links baseline-score change with the real-world fraction of images having the expected target and protected labels.The authors suggest either poorer target preservation or minority-sample misclassification as possible explanations.
  • Attribute interactions: Some attributes change drastically when balanced datasets are created for other attributes, indicating greater sensitivity to latent-space manipulations.Attractive changes substantially regardless of which target attribute is preserved.
  • Changes in baseline score: Figure 8 shows larger score changes when generated images have minority labels.The figure separates changes according to the new ground-truth attribute values.
  • Data skew: The analysis also investigates performance under differently skewed training sets for attributes such as HighCheeks.The experiments start from the 162,770-image CelebA training set.

C.2 Discriminability of attributes

The discriminability experiment tests whether one attribute is easier to learn than another under controlled correlations. Gender expression is among the easiest attributes, whereas Young is relatively difficult, offering a possible explanation for differences in bias influence.

  • Experimental design: The experiment creates fully skewed datasets and evaluates classifiers on balanced CelebA validation subsets containing all four attribute combinations.Average precision is computed separately for each attribute.
  • Discriminability criterion: An attribute is considered easier to learn when it achieves higher average precision than its counterpart in both created datasets.
  • Results: Table 11 compares attributes in rows and columns, marking whether the row attribute is easier to learn than the column attribute.The table uses y for easier and n for harder.
  • Results: Gender expression is one of the easiest attributes to learn, outperforming every tested attribute except WearingHat and Glasses.The authors connect this discriminability pattern with the prevalence of gender bias in many models.
  • Results: Young is relatively hard to learn, ranking harder than all but four other tested attributes.Its correlations with other attributes may therefore be less influential.

D. Ablation studies

The ablations test whether hyperplane balance, synthetic-image labeling, and the amount of protected-attribute supervision materially affect downstream accuracy and fairness. Results indicate limited sensitivity to hyperplane sample balance, modest disadvantages for alternative labeling schemes, and gradual fairness gains from more labeled examples.

  • Hyperplane balance: Hyperplane estimation used a fixed total of 12,000 samples while varying positive and negative fractions across gender expression.The resulting classifiers were evaluated on CelebA validation data after augmenting real training images with synthetic pairs.
  • Hyperplane balance: Underrepresentation in hyperplane-estimation samples does not appear to substantially affect downstream classification performance.This conclusion is summarized by the ablation table for HighCheeks.
  • Synthetic labels: Alternative synthetic-image selection and labeling schemes remain comparable but lose AP and generally worsen fairness metrics relative to the proposed method.AP is 79.8 and 82.1 versus 82.6; average DEO is 18.1 and 17.4 versus 16.1, while BA is 0.9 and 0.7 versus 0.5.
  • Protected-attribute supervision: More labeled examples for gender-expression classifiers yield gradual fairness improvements that remain within error bars.Across four attributes, DEO changes from 11.1 with 10 samples to 9.6 with all 162k training examples, while BA changes from 0.6 to 0.4 and KL from 0.6 to 0.5.
Loading 2012.01469v3…