Similarity Groups Scale Fairness Without Demographic Data

Dilermando Queiroz, João F. N. B. Cortese, Alceu Bissoto

Track: Tiny/Short Papers Track (up to 3 pages)

Keywords: Group Fairness, Individual Fairness, subgroup discovery, Scalable Fairness

TL;DR: Similarity Groups Fairness (SGF) uses representation similarity to define groups and jointly audit group and individual fairness, enabling scalable fairness evaluation and safeguards without demographic metadata in real-world medical imaging systems.

Abstract: Intelligent agent systems increasingly mediate interactions between humans and digital environments, making ethical guidance a critical component of their development. To achieve robust models capable of interaction with human, Self-Supervised Learning (SSL) has emerged as a dominant paradigm. This approach enables foundation models to be built only with the raw data, supporting reliable performance in real-world environments. However, ensuring fairness in these systems remains a major challenge, as existing methods for evaluation depend heavily on explicit metadata which are difficult to obtain at scale. Here, we introduce a unified framework for Similarity Group Fairness (SGF), a synthesis of group and individual fairness notions, that enables the auditing and safeguards of fairness without relying on demographic metadata.

Submission Number: 37