Oriel — AI & technology worth noticing, every day.

Not All Minorities Are Equal: Fairness Outcomes in Private Image Generation

Paper · https://arxiv.org/abs/2611.00059

arXiv:2611.00059v1 Announce Type: new Abstract: Differential privacy (DP) provides formal guarantees for protecting individuals in artificial intelligence systems, but its consequences for fairness in generative models remain poorly understood. We investigate how privacy-preserving training affects the representation of demographic groups and visual attributes in generative models of human faces. Across two facial image datasets and multiple privacy levels, we find that stronger privacy can disproportionately degrade the quality and utility of some underrepresented groups, suggesting that privacy can amplify existing disparities. However, representation alone does not explain these effects: several well-represented attributes also experience substantial degradation, while some underrepresented attributes remain comparatively robust. These findings show that the impact of privacy-preserving learning cannot be explained solely by data representation and may depend on the characteristics and interactions of the information being learned. Our results highlight the need to consider both privacy and fairness when evaluating generative models.

First seen by Oriel on .

No saved summary available.

Aliases

SourceExternal IDURL
arXiv AI Papersoai:arXiv.org:2611.00059v1https://arxiv.org/abs/2611.00059 Read discussion