New metric PCER audits fairness in differentially private ML

Researchers have introduced a new group fairness criterion called the Privacy-Cost Equity Ratio (PCER) for differentially private machine learning systems. PCER addresses the issue that differential privacy mechanisms like DP-SGD can disproportionately increase privacy costs for certain demographic groups, leading to wider accuracy disparities. By framing fairness as compensatory, PCER argues that groups bearing higher privacy exposure should receive greater system benefits. The metric, which requires only per-group accuracy data, can be used as a post-hoc audit tool to uncover fairness issues missed by outcome-based metrics, as demonstrated on the COMPAS dataset where it revealed a double disadvantage for a protected group. AI

IMPACT Introduces a new metric to audit fairness in privacy-preserving machine learning, potentially improving equitable outcomes.

RANK_REASON Academic paper introducing a new metric for differentially private machine learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• COMPAS
• differential privacy
• DP SGD
• machine learning
• natural language processing
• Privacy Cost Equity Ratio

• paper
• safety

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