Abstract
Machine learning fairness concerns about the biases towards certain protected or sensitive group of people when addressing the target tasks. This paper studies the debiasing problem in the context of image classification tasks. Our data analysis on facial attribute recognition demonstrates (1) the attribution of model bias from imbalanced training data distribution and (2) the potential of adversarial examples in balancing data distribution. We are thus motivated to employ adversarial example to augment the training data for visual debiasing. Specifically, to ensure the adversarial generalization as well as cross-task transferability, we propose to couple the operations of target task classifier training, bias task classifier training, and adversarial example generation. The generated adversarial examples supplement the target task training dataset via balancing the distribution over bias variables in an online fashion. Results on simulated and real-world debiasing experiments demonstrate the effectiveness of the proposed solution in simultaneously improving model accuracy and fairness. Preliminary experiment on few-shot learning further shows the potential of adversarial attack-based pseudo sample generation as alternative solution to make up for the training data lackage.
The Accuracy-Fairness Paradox
In machine learning, researchers often encounter a frustrating trade-off: making a model "fairer" (reducing bias against protected groups) usually results in lower overall accuracy. This phenomenon is known as the Accuracy-Fairness Paradox. Existing debiasing techniques typically fall into three categories: pre-processing (re-sampling data), in-processing (adding constraints during training), or post-processing (adjusting outputs). However, these methods often lose information or restrict the model's representative power.
The authors argue that the root cause of bias is imbalanced data distribution. As shown in the figure above, there is a clear correlation between the ratio of female images in the training set and the resulting gender bias for specific attributes like "arched eyebrows." When a dataset contains many more "females with arched eyebrows" than "males with arched eyebrows," a model learns to use gender as a shortcut for the facial feature. Collecting perfectly balanced real-world data is often prohibitively expensive.
Adversarial Example-based Data Augmentation (AEDA)
To address this, the researchers propose Adversarial Example-based Data Augmentation (AEDA). Instead of searching for rare real-world images, they use adversarial perturbations to "create" them.
The Core Logic
If a dataset lacks "males with arched eyebrows," AEDA takes an image of a "female with arched eyebrows" and applies a slight adversarial perturbation. This perturbation is designed to make the model's bias classifier see the image as "male," while the target feature (arched eyebrows) remains visually and structurally intact. These generated "pseudo-samples" serve as synthetic training data to fill gaps in the distribution, effectively balancing the training set without requiring new raw data.
Implementation Flavors
- AEDA_pre: Generates adversarial examples once before training begins.
- AEDA_online: Couples target classifier training with adversarial generation. As the model's feature extractor evolves, the adversarial examples are updated to ensure they remain informative and maintain cross-task transferability.
- AEDA_robust: The most advanced version, which adds an adversarial training module to the bias classifier. This ensures the generated examples contain generalized, high-level features of the minority group rather than mere pixel noise.
Experimental Results
The approach was evaluated on C-MNIST (simulated color bias) and CelebA (real-world facial attribute bias).
- Breaking the Paradox: Unlike traditional methods like down-sampling, which reduces bias at the cost of accuracy, AEDA acts as a supplement. By providing more data, the model learns the correct features more effectively, leading to higher accuracy alongside better fairness.
- Performance: On CelebA, AEDA_robust achieved the highest balanced average accuracy (74.30% bACC) and the lowest model bias (3.27) compared to standard debiasing methods. On C-MNIST, it boosted accuracy from 55.62% to 91.80% while slashing bias.
- Few-Shot Potential: Preliminary tests showed a 27% improvement in few-shot learning scenarios, suggesting AEDA is a viable alternative to traditional data collection when samples are scarce.
Conclusion
AEDA demonstrates that adversarial perturbations—often viewed as model vulnerabilities—can be harnessed as a powerful tool for data augmentation. By decoupling target features from biased attributes in the feature space, AEDA successfully mitigates model bias while simultaneously improving predictive performance, offering a robust path forward for fair and accurate computer vision systems.
Summary updated