[Paper Review] Fairness for Image Generation with Uncertain Sensitive Attributes

Ajil Jalal, Sushrut Karmalkar+3 more|arXiv (Cornell University)|Jun 23, 2021

Generative Adversarial Networks and Image Synthesis5 citations

TL;DR

This paper introduces Conditional Proportional Representation (CPR), a fairness definition for image generation that ensures reconstructed images proportionally represent protected groups without requiring predefined group identities. It proves that Posterior Sampling via Langevin dynamics achieves CPR obliviously—without relying on group labels—making it robust to ambiguous or contested sensitive attributes like race.

ABSTRACT

This work tackles the issue of fairness in the context of generative procedures, such as image super-resolution, which entail different definitions from the standard classification setting. Moreover, while traditional group fairness definitions are typically defined with respect to specified protected groups -- camouflaging the fact that these groupings are artificial and carry historical and political motivations -- we emphasize that there are no ground truth identities. For instance, should South and East Asians be viewed as a single group or separate groups? Should we consider one race as a whole or further split by gender? Choosing which groups are valid and who belongs in them is an impossible dilemma and being "fair" with respect to Asians may require being "unfair" with respect to South Asians. This motivates the introduction of definitions that allow algorithms to be \emph{oblivious} to the relevant groupings. We define several intuitive notions of group fairness and study their incompatibilities and trade-offs. We show that the natural extension of demographic parity is strongly dependent on the grouping, and \emph{impossible} to achieve obliviously. On the other hand, the conceptually new definition we introduce, Conditional Proportional Representation, can be achieved obliviously through Posterior Sampling. Our experiments validate our theoretical results and achieve fair image reconstruction using state-of-the-art generative models.

Motivation & Objective

• Address fairness in image generation where traditional group fairness definitions fail due to ambiguous or contested sensitive attributes like race.
• Recognize that defining protected groups (e.g., Asian vs. South Asian) is inherently political and context-dependent, undermining standard fairness definitions.
• Propose a fairness framework that does not depend on predefined groupings, enabling fairness across all possible groupings simultaneously.
• Develop and validate a method that achieves fairness without access to sensitive attribute labels, ensuring robustness to group ambiguity.
• Demonstrate that existing methods like PULSE exacerbate bias in image reconstruction due to model priors and data imbalance, even when trained on diverse data.

Proposed method

• Introduce Conditional Proportional Representation (CPR), a fairness notion ensuring that the probability of generating a sample from a group is proportional to its prior prevalence in the data.
• Prove that CPR can be achieved obliviously—without knowledge of group identities—using Posterior Sampling via Langevin dynamics.
• Formalize the incompatibility between CPR and Representation Demographic Parity (RDP), showing RDP's strong dependence on arbitrary group definitions.
• Use Bayesian inference to model the posterior distribution over image reconstructions, ensuring fairness through stochastic sampling.
• Implement the method using deep generative models and Langevin dynamics to sample from the posterior, avoiding deterministic reconstruction biases.
• Validate the approach empirically on image super-resolution tasks using the FFHQ and AFHQ datasets, comparing with PULSE and other baselines.

Experimental results

Research questions

• RQ1 Can fairness in image generation be defined in a way that is independent of arbitrary or contested group definitions such as race or ethnicity?
• RQ2 What fairness definitions are compatible with obliviousness—i.e., fairness across all possible groupings without requiring group labels?
• RQ3 Why do existing generative models like PULSE produce racially biased reconstructions even when trained on diverse data?
• RQ4 Can a single algorithm achieve fairness across all possible groupings simultaneously, without prior knowledge of sensitive attributes?
• RQ5 Is there a fairness notion in image generation that ensures proportional representation of groups without relying on demographic parity or independence constraints?

Key findings

• Representation Demographic Parity (RDP) is strongly dependent on the choice of protected groups and cannot be achieved obliviously, making it unsuitable for ambiguous attributes like race.
• Conditional Proportional Representation (CPR) is a conceptually new fairness definition that ensures proportional representation of groups without requiring group labels.
• Posterior Sampling via Langevin dynamics is the only known method that can achieve CPR fully obliviously, making it robust to group ambiguity.
• Experiments show that PULSE reconstructs 80% of cats as dogs when dogs are the 80% majority, demonstrating severe bias due to model priors and data imbalance.
• In balanced settings, Posterior Sampling satisfies RDP, SPE, and PR, while PULSE fails to maintain fairness, confirming theoretical claims.
• Even when fairness in representation is satisfied, poor reconstruction quality for underrepresented groups can still undermine fairness—highlighting the need for quality-aware fairness metrics.