FADE: Towards Fairness-aware Generation for Domain Generalization via Classifier-Guided Score-based Diffusion Models

TL;DR

FADE addresses fairness-aware domain generalization under covariate shift by generating unbiased, domain-invariant data with a classifier-guided score-based diffusion model. It pre-trains an SDM and two classifiers across multiple domains, then derives a debiased generator by guiding diffusion with classifier signals and entropy-driven adjustments to obscure sensitive information. The generated data train a downstream fair, domain-invariant classifier, yielding superior accuracy-fairness trade-offs across three real-world datasets. This approach demonstrates robust performance under distribution shifts and offers a practical, training-efficient path to fair generalization in real-world deployments.

Abstract

Fairness-aware domain generalization (FairDG) has emerged as a critical challenge for deploying trustworthy AI systems, particularly in scenarios involving distribution shifts. Traditional methods for addressing fairness have failed in domain generalization due to their lack of consideration for distribution shifts. Although disentanglement has been used to tackle FairDG, it is limited by its strong assumptions. To overcome these limitations, we propose Fairness-aware Classifier-Guided Score-based Diffusion Models (FADE) as a novel approach to effectively address the FairDG issue. Specifically, we first pre-train a score-based diffusion model (SDM) and two classifiers to equip the model with strong generalization capabilities across different domains. Then, we guide the SDM using these pre-trained classifiers to effectively eliminate sensitive information from the generated data. Finally, the generated fair data is used to train downstream classifiers, ensuring robust performance under new data distributions. Extensive experiments on three real-world datasets demonstrate that FADE not only enhances fairness but also improves accuracy in the presence of distribution shifts. Additionally, FADE outperforms existing methods in achieving the best accuracy-fairness trade-offs.

Figures (4)

• Figure 1: An overview of FADE. Supposing there are 4 domains in the training set $\mathcal{D}^{src}$. For the $k$-th iteration, we randomly select a batch as $\mathcal{T}_k^{sup}$, and sample from the remaining three domains to compose $\mathcal{T}_k^{qry}$. We perform gradient descent using the loss $\mathcal{L}^{sup}_k(\Phi)$ obtained on $\mathcal{T}_k^{sup}$ to obtain a temporary set of parameters $\Phi'$, then use the model with $\Phi'$ to obtain loss $\mathcal{L}^{qry}_k(\Phi'_k)$ values on $\mathcal{T}_k^{qry}$. We use the weighted sum of two losses to update $\Phi=\{\boldsymbol{\mathbf{\theta}}_s,\boldsymbol{\mathbf{\theta}}_y,\boldsymbol{\mathbf{\theta}}_z\}$. When using the score network $s_{\boldsymbol{\theta}_s}$ to generate data from noise, we can guide the generation process with $\phi_{\boldsymbol{\theta}_y}$ and $\psi_{\boldsymbol{\theta}_z}$ to obtain unbiased data. Finally, these data are used to train downstream tasks to achieve the final results.
• Figure 2: (a) LDA visualization of the original dataset and the dataset generated by FADE. The orange, green, and blue curves represent the probability densities of the original data for two sensitive groups and the generated data, respectively. The three dashed lines indicate the means. The markers below the x-axis show the one-dimensional data, with red points representing the mean points. (b) Average $D_{fair}$ results for all fairness-aware methods across three datasets.
• Figure 3: Accuracy-fairness trade-offs of FADE by various $\lambda_z \in \{0.1, 1, 10, 50, 100\}$ on (a) Adult and (b) Bank datasets over different baselines. The upper left indicates a better trade-off performance. Results are averaged across all domains.
• Figure : FADE