AI systems learn to treat individuals fairly, not just demographic groups
Researchers have developed a method to reduce bias in AI vision models by ensuring similar people receive similar treatment, rather than just balancing outcomes across demographic groups. The technique could matter for companies deploying AI in hiring, lending, and healthcare, where individual fairness claims are increasingly common in litigation and regulation.
Originaltitel: Individually fair representation learning for DINOv2
<p>Deep learning models frequently inherit biases from training data, leading to unfair outcomes. Fair representation learning aims to mitigate these biases by reducing sensitivity to protected attributes while preserving task-relevant information. However, existing methods primarily target group fairness and neglect individual fairness, which requires similar individuals to receive similar outcomes. We introduce an individual fairness framework applied to modern Vision Transformer (ViT) architectures, specifically DINOv2 and domain-specific RAD-DINO. This approach adapts representations leveraging adversarial perturbations along sensitive attribute(s), minimizing differences between representations of similar individuals while preserving information for downstream classification tasks. In addition, we propose a task-agnostic fair representation strategy combining adversarial perturbations with a reconstruction loss that ensures information preservation of the adapted representations. The reconstruction loss operates directly in the representation space, avoiding the need for a high-quality image decoder while still being effective. We then certify fairness and robustness in the learned representations using center and randomized smoothing. Evaluations on the CelebA and NIH Chest X-ray datasets across various tasks show that incorporating individual fairness constraints boosts fairness in most cases with only a minor drop in accuracy for different downstream tasks. Compared to the general-domain DINOv2, RAD-DINO achieves superior baseline accuracy, while Fair-RAD-DINO delivers a more balanced and superior fairness–accuracy trade-off compared to Fair-DINOv2. The task-agnostic training shows comparable effectiveness in balancing accuracy and fairness trade-offs on downstream classification tasks with negligible accuracy drops ( for CelebA, 1% for NIH) relative to representations learned with adversarial perturbation plus auxiliary task-specific strategies.</p>