Abstract
The fairness of AI decision-making has garnered increasing attention, leading to the proposal of numerous fairness algorithms. In this paper, we aim not to address this issue by directly introducing fair learning algorithms, but rather by generating entirely new, fair synthetic data from biased datasets for use in any downstream tasks. Additionally, the distribution of test data may differ from that of the training set, potentially impacting the performance of the generated synthetic data in downstream tasks. To address these two challenges, we propose a diffusion model-based framework, FADE (Fairness-Aware Diffusion with Meta-training). FADE introduces two types of gradient induction during the sampling phase of the diffusion model: one to ensure that the generated samples belong to the desired target categories, and another to make the sensitive attributes of the generated samples difficult to classify. To overcome data distribution shifts, we train the diffusion model and the classifiers used for induction within a meta-learning framework. Experiments on real datasets demonstrate that FADE achieves better accuracy and optimal fairness in downstream tasks.
Introduction
In real-world applications, machine learning models frequently encounter distribution shift, where the test data distribution differs from the training data. A critical challenge in this context is Fairness-aware Domain Generalization (FairDG). Traditional fairness techniques often degrade when applied to unseen domains (e.g., a model trained on loan data from one city exhibiting bias when applied to another).
Existing solutions typically attempt to "disentangle" sensitive features from useful ones, but these methods often rely on rigid mathematical assumptions. FADE addresses this by shifting the focus from model-correction to data-generation, creating a diverse, unbiased synthetic dataset that can be used to train any downstream classifier.
The FADE Framework
FADE operates through a three-stage pipeline designed to produce data that is both fair and domain-invariant.
1. Domain-Invariant Pre-training
Using a meta-learning approach (MAML-style), FADE pre-trains three components: a Score-based Diffusion Model (SDM), a label classifier, and a sensitive attribute classifier. By splitting source data into "support" and "query" sets across different domains, the framework forces these models to learn features that are consistent across all environments, rather than domain-specific quirks.
2. Classifier-Guided Fair Sampling
This stage is the core innovation of the framework. To generate a new data point, FADE starts with random noise and uses the pre-trained classifiers to guide the diffusion process:
- Fair Control (Entropy as a Shield): The process guides generation to maximize the entropy of the sensitive attribute classifier. This effectively "bleaches" bias by forcing the generator to produce data where a classifier cannot distinguish sensitive attributes (e.g., race or gender).
- Label Guidance: Simultaneously, the label classifier ensures the generated data corresponds to the desired target class.
3. Downstream Training
A standard classifier is trained from scratch using only this "cleaned" synthetic data. Because the synthetic data is already domain-invariant and fair, the resulting model inherits these properties without requiring complex in-processing constraints.
Experimental Evaluation
The researchers evaluated FADE on three real-world datasets: Adult (Census Income), Bank Marketing, and New-York-Stop-and-Frisk (NYSF). These datasets represent various challenges in fairness and domain shift.
Key Results
FADE consistently outperformed state-of-the-art baselines, such as FGAN and DECAF, in the accuracy-fairness trade-off.
- Superior Fairness: On the Adult dataset, FADE achieved an average Demographic Parity difference ($\Delta_{DP}$) of 0.09, significantly better than the VAE (0.21) or standard DDPM (0.17).
- Maintained Accuracy: FADE maintained high average accuracy (84.55% on Adult; 80.84% on Bank), often surpassing other fair-generation methods by several percentage points.
- Robustness to Shift: In the NYSF dataset, which features high variance between boroughs, FADE demonstrated stable performance, proving its ability to generalize to unseen target domains.
Conclusion
FADE demonstrates that treating fairness as a data-generation problem via score-based diffusion models is a highly effective strategy for FairDG. By decoupling generation from training and using entropy-based guidance, FADE produces synthetic data that is robust to distribution shifts while maintaining high utility and fairness. Future work will explore extending this framework to high-dimensional image data and addressing complex concept drift.
Summary updated