Through a Fair Looking-Glass: Mitigating Bias in Image Datasets

TL;DR A U-Net learns small edits to each image that reduce the statistical dependence (HSIC) between the target attribute (Attractive) and gender. This de-biases the dataset itself, so classifiers trained on it are fairer.

Six CelebA test images — CelebA test images before and after the trained U-Net. The changes are subtle, mostly texture and facial detail, and they are enough to reduce the statistical dependence between *Attractive* and *Gender*. Drag across the image to compare.

The same six images after the de-biasing U-Net — CelebA test images before and after the trained U-Net. The changes are subtle, mostly texture and facial detail, and they are enough to reduce the statistical dependence between *Attractive* and *Gender*. Drag across the image to compare.

Abstract

With the recent growth in computer vision applications, the question of how fair and unbiased they are has yet to be explored. There is abundant evidence that the bias present in training data is reflected in the models, or even amplified. Many previous methods for image dataset de-biasing, including models based on augmenting datasets, are computationally expensive to implement. In this study, we present a fast and effective model to de-bias an image dataset through reconstruction and minimizing the statistical dependence between intended variables. Our architecture includes a U-net to reconstruct images, combined with a pre-trained classifier which penalizes the statistical dependence between target attribute and the protected attribute. We evaluate our proposed model on CelebA dataset, compare the results with a state-of-the-art de-biasing method, and show that the model achieves a promising fairness-accuracy combination.

Method

(a) Training: a U-Net reconstructs a batch of images. An MSE loss keeps the output close to the input, and a frozen two-headed ResNet predicts gender and the target attribute so that their statistical dependence (HSIC) can be penalized. (b) Use: the trained U-Net transforms the dataset, and downstream classifiers are trained on the result.

Results

Results on CelebA, using the attribute categories of Ramaswamy et al. (2021). Each value is the average over all attributes in that category.

	AP ↑			DP ↓			DEO ↓
	Incons.	G-dep	G-indep	Incons.	G-dep	G-indep	Incons.	G-dep	G-indep
Baseline	0.667	0.790	0.843	0.147	0.255	0.137	0.186	0.243	0.163
GAN debiasing	0.641	0.763	0.831	0.106	0.233	0.119	0.158	0.240	0.142
Adversarial debiasing	0.243	0.333	0.218	0.091	0.169	0.121	0.136	0.149	0.098
Ours	0.618	0.732	0.839	0.097	0.146	0.118	0.124	0.172	0.114

How much the model changes each attribute (red: change in demographic parity of that attribute's classifier) compared with how strongly the attribute depends on Attractive in the original data (blue: HSIC). The model mostly edits attributes that are entangled with the target. — How much the model changes each attribute (red: change in demographic parity of that attribute’s classifier) compared with how strongly the attribute depends on *Attractive* in the original data (blue: HSIC). The model mostly edits attributes that are entangled with the target.

Increasing the fairness weight λ trades accuracy for fairness. Each point is the mean of three training runs. Shaded bands show one standard deviation.

Citation

@inproceedings{rajabi2023looking,
  title     = {Through a Fair Looking-Glass: Mitigating Bias in Image Datasets},
  author    = {Rajabi, Amirarsalan and Yazdani-Jahromi, Mehdi and Ozmen Garibay, Ozlem and Sukthankar, Gita},
  booktitle = {Artificial Intelligence in HCI (HCII 2023)},
  series    = {Lecture Notes in Computer Science},
  pages     = {446--459},
  publisher = {Springer},
  year      = {2023},
  doi       = {10.1007/978-3-031-35891-3_27}
}

Amirarsalan Rajabi

Abstract

Method

Results

Citation