Towards Reducing Bias in Gender Classification

Komal K. Teru, Aishik ChakrabortyarXiv 2019 · arXiv

Introduction

Machine learning models are fundamentally constrained by the quality and representativeness of their training data. In facial analysis, datasets often reflect societal prejudices or significant demographic imbalances. A seminal 2018 study by Buolamwini and Gebru demonstrated that commercial gender classification systems frequently exhibit lower performance on individuals with darker skin tones. Such biases in AI pipelines—whether used for security or recruitment—can lead to systemic unfairness. This research addresses racial bias by employing adversarial training to "strip" race-specific information from image representations, aiming for equitable performance across all ethnic groups.

Methodology

The proposed approach utilizes Fader Networks, an architecture designed to disentangle specific attributes from underlying data representations through adversarial training. The system comprises three primary components:

  1. The Encoder: Compresses a face image into a latent representation.
  2. The Discriminator (Adversary): Attempts to predict the individual's race from the latent representation.
  3. The Decoder: Reconstructs the original image using the latent representation and a target race label.
The Fader Network architecture consists of an inference model and an adversarial component designed to filter out race-specific information.

The core mechanism involves a training loop where the Encoder is penalized if the Discriminator successfully identifies the race. This pressure forces the Encoder to learn a "race-invariant" representation that retains gender-relevant features while discarding racial markers. A final Gender Classifier (FaderCNN) is then trained on these filtered representations.

Experimental Setup

The researchers evaluated the model using the UTKFace dataset, which contains over 23,000 images across five ethnic categories: White (Class 0), Black (Class 1), Asian (Class 2), Indian (Class 3), and Others (Class 4). The dataset is highly imbalanced, with Class 0 representing approximately 88% of the training data. To ensure rigorous evaluation, a stratified test set was created with 474 images from each racial category.

The FaderCNN was compared against two baselines:

  • SimpleCNN: A standard convolutional neural network.
  • SimpleCNN-WL: A CNN utilizing weighted loss to mitigate data imbalance.

Results and Discussion

The primary metric for fairness in this study is the variance in accuracy across different racial groups. A lower variance indicates more consistent and equitable performance.

Quantitative comparison of accuracy across racial classes and overall model variance.

As shown in the results, the FaderCNN achieved a significant reduction in bias. The variance dropped from 11.26 in the standard SimpleCNN to 6.66 in the FaderCNN—a reduction of approximately 41%. While the FaderCNN saw a slight decrease in overall accuracy (84.83% vs 89.51%), it provided much more consistent results across minority groups, particularly Class 2 (Asian), where standard models often struggle. This trade-off suggests that standard models may inadvertently "cheat" by using racial features to optimize for the majority demographic.

Conclusion and Future Work

This work serves as a proof of concept for using adversarial filters to enhance AI fairness. While the current implementation faced computational constraints—preventing the discriminator from reaching true random-chance performance—the results clearly demonstrate the potential of race-invariant representations. Future research will focus on more sophisticated data augmentation, hyperparameter optimization for the adversarial trade-off, and extending this framework to other sensitive attributes like age or socioeconomic status.

Summary updated

Abstract

Societal bias towards certain communities is a big problem that affects a lot of machine learning systems. This work aims at addressing the racial bias present in many modern gender recognition systems. We learn race invariant representations of human faces with an adversarially trained autoencoder model. We show that such representations help us achieve less biased performance in gender classification. We use variance in classification accuracy across different races as a surrogate for the racial bias of the model and achieve a drop of over 40% in variance with race invariant representations.
BibTeX
@misc{teru2019towards,
  author = {Komal K. Teru and Aishik Chakraborty},
  title = {{Towards Reducing Bias in Gender Classification}},
  year = {2019},
  eprint = {1911.08556},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1911.08556},
}
Figure