Debiasing a face generator without its training data: a fairness method for GANs

Image generators inherit the imbalance of the data they learn from. The paper by Anubhav Jain, Nasir Memon and Julian Togelius (arXiv:2308.08638, 2023) shows a way to build an unbiased generative adversarial network (GAN) from an existing biased one, without the original training data.

The problem

Fairness in a generative model means equal representation across attributes such as ethnicity or gender: a fair model should be equally likely to produce a face of any group. Public models are not. Randomly drawing 10,000 samples from the StyleGAN model gives over 5,000 images of white people and fewer than 500 of black and Indian people combined. The super-resolution model PULSE generates only lighter-skinned faces whatever the input.

Earlier debiasing work assumed access to a smaller balanced dataset, a reference dataset or labeled data, which is unrealistic for models trained on billions of images, or for areas like medical imagery and biometrics. The authors assume none of that: no labels, no reference set and not even the original training data, which may be privacy-sensitive or gone. They assume only an existing auxiliary classifier.

The method

1. Generate balanced data. An evolutionary algorithm searches the biased model's latent space to extract a balanced set of synthetic images.
2. Retrain. A new model of the same architecture, or the same model retrained, is trained on that balanced set.
3. Add a fairness loss. A bias-mitigation loss function pushes the model's class distribution toward equiprobable.

The authors tested it on StyleGAN2 trained on the FFHQ face dataset, for racial fairness.

Results

• FFHQ: the fairness metric improved by almost 5 times while image quality stayed comparable on referenceless quality measures. The bias loss alone helped, and did better than a baseline importance reweighting technique.
• Imbalanced CIFAR10: with a long-tail version of the dataset holding 20,431 samples against 50,000 in the original, the method reached a fairness score of 0.0281 and an FID of 6.32, close to the 0.0286 and 4.36 of a model trained on fully balanced data. Its fairness value was about 10 times lower than from importance weighting or training directly on the imbalanced data.

The authors also argue that the popular Fréchet inception distance (FID) is sensitive to class balance in the data and should not be used to study bias when no balanced reference set exists.

Why it matters, and the caveat

Synthetic data can stand in for privacy-sensitive datasets, so the models that make it should not pass bias on to downstream tasks. Because the new model is trained once on generated data, with a focus on diversity, the authors say the risk of "model collapse" from repeated training on generated data is minimal. They suggest the approach applies to any GAN trained on any biased dataset.

Sources

• Anubhav Jain, Nasir Memon and Julian Togelius, "Fair GANs through model rebalancing for extremely imbalanced class distributions", arXiv:2308.08638 (2023) (https://arxiv.org/abs/2308.08638), licensed CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Changed here: summarised in hubnx's own words; figures, equations and appendices left out; rewritten in hubnx's own words.