When analyzing data privately (), you must limit each person's contribution size (clipping bound). For data with very large outliers (heavy-tailed data), picking this limit is tough. Existing methods require prior knowledge or use percentiles that fail.
The authors introduce a fully private method that automatically chooses the clipping bound. It tests a series of possible limits and uses residual coherence, a measure of how aligned the large contributions are, to pick the best one. They prove this relates to reducing error from clipping.
Across several datasets, their method outperforms existing baselines. This removes a major hurdle to deploying differential privacy, especially for heavy-tailed data common in web activity.