New Study: Counterfactual Fairness Doesn't Imply Group Fairness in Image Classification

A new study published on arXiv investigates the relationship between counterfactual fairness (CF) and group fairness (GF) in image classification. Researchers constructed new datasets, \oursceleb and \ourslfw, to evaluate both fairness metrics simultaneously. Their findings indicate that CF does not necessarily imply GF in image classification, a divergence from previous observations in tabular datasets. This discrepancy is attributed to the presence of latent attributes correlated with sensitive attributes, such as secondary sex characteristics and hair length. To address this, the study proposes Counterfactual Knowledge Distillation (CKD) as a method to reduce reliance on these latent attributes, thereby enabling CF-achieving models to satisfy GF. AI

IMPACT This research highlights potential disparities in fairness metrics for image classification models, suggesting a need for new methods to ensure equitable outcomes.

RANK_REASON Academic paper on AI fairness metrics. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Counterfactual Fairness Circularity
• Counterfactual Knowledge Distillation
• Group Fairness
• image classification
• \oursceleb
• \ourslfw

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →