Challenges in achieving privacy-preserving data valuation

The paper analyzes whether differentially private-compatible data valuation is feasible, pointing out that differential privacy requires outputs to be insensitive to a single record, whereas data valuation precisely aims to measure the influence of a single record, so there is a fundamental tension between the two. The authors identify the core algorithmic primitives in common valuation frameworks that lead to excessive sensitivity, and explain why directly applying DP mechanisms fails. The authors further propose design principles for valuation pipelines that more easily satisfy privacy constraints, and characterize on representative methods and datasets how privacy constraints reduce ranking fidelity.