Latent Imitator: Generating Natural Individual Discriminatory Instances for Black-Box Fairness Testing

TL;DR

This paper tackles the problem of fair ML testing by generating natural individual discriminatory instances that reflect real data distributions. It introduces Latent Imitator (LIMI), which builds a surrogate decision boundary in GAN latent space and uses one-step latent probing with two proximal candidates to locate discriminatory examples near the true boundary without iterative search. Empirical results across tabular datasets and models show LIMI outperforms state-of-the-art baselines in effectiveness, efficiency, and naturalness, and retraining with LIMI-sourced data yields meaningful improvements in both individual and group fairness, with minimal accuracy loss. The work demonstrates LIMI’s potential to rapidly and reliably reveal fairness violations in black-box settings and suggests promising extensions to imaging and multi-attribute scenarios.

Abstract

Machine learning (ML) systems have achieved remarkable performance across a wide area of applications. However, they frequently exhibit unfair behaviors in sensitive application domains, raising severe fairness concerns. To evaluate and test fairness, engineers often generate individual discriminatory instances to expose unfair behaviors before model deployment. However, existing baselines ignore the naturalness of generation and produce instances that deviate from the real data distribution, which may fail to reveal the actual model fairness since these unnatural discriminatory instances are unlikely to appear in practice. To address the problem, this paper proposes a framework named Latent Imitator (LIMI) to generate more natural individual discriminatory instances with the help of a generative adversarial network (GAN), where we imitate the decision boundary of the target model in the semantic latent space of GAN and further samples latent instances on it. Specifically, we first derive a surrogate linear boundary to coarsely approximate the decision boundary of the target model, which reflects the nature of the original data distribution. Subsequently, to obtain more natural instances, we manipulate random latent vectors to the surrogate boundary with a one-step movement, and further conduct vector calculation to probe two potential discriminatory candidates that may be more closely located in the real decision boundary. Extensive experiments on various datasets demonstrate that our LIMI outperforms other baselines largely in effectiveness ($\times$9.42 instances), efficiency ($\times$8.71 speeds), and naturalness (+19.65%) on average. In addition, we empirically demonstrate that retraining on test samples generated by our approach can lead to improvements in both individual fairness (45.67% on $IF_r$ and 32.81% on $IF_o$) and group fairness (9.86% on $SPD$ and 28.38% on $AOD$}).

Figures (4)

• Figure 1: Overview of LIMI framework. LIMI first imitates the target model's decision boundary by deriving a surrogate boundary based on $\mathcal{D}_{aux}$, which coarsely depicts a region with better naturalness. Later, to obtain more natural samples, LIMI manipulates random latent vectors to the surrogate boundary with a one-step movement and further probes two potential discriminatory candidates that may be closer to the real boundary.
• Figure 2: Visualization of the data distribution of generated instances. The distribution of our LIMI is the closest to the original one, which indicates the better naturalness of generated instances.
• Figure 3: Scalability study. Comparison between baselines and combined methods. Results are shown in $\#\mathcal{D}_{idi}$.
• Figure 4: Discriminatory images generated by LIMI. Men are predicted with not smile (Top), while women are predicted with smile (Bottom).