Differential Privacy Noise Allocation in Independent Media Impression Audits
Privacy noise allocation in media audits requires shifting epsilon from micro-slices to campaign aggregates to keep financial billing errors below two percent.
Contribution Bounding
Meaning
Constraint on data processing that limits the influence any single record or individual can exert on the output of a query. Data scientists apply contribution bounding to prevent outliers or heavy users from skewing results and increasing the risk of re-identification. This boundary is a prerequisite for calculating the sensitivity of a function, which in turn determines the amount of noise required to protect privacy.
It ensures that the impact of any one person is capped at a pre-defined limit.
Input Restriction
Limiting the count of records per user protects the statistical integrity of the data while maintaining compliance with privacy budgets. When contribution bounding is active, a user who appears a thousand times in a transaction log is reduced to a smaller, fixed number of entries. This modification allows the privacy loss to be bounded by a predictable constant.
Financial Risk
Incorrectly calibrated limits can lead to large data loss or inaccurate business intelligence, which affects the value of the insights being sold. Contracts for data access often specify the contribution bounding levels to ensure that the buyer knows exactly how much the data has been altered. A tighter bound provides better privacy but may reduce the utility of the data for identifying niche behaviors.
Sensitivity Control
The mechanism provides the foundation for more advanced privacy techniques by ensuring that the global sensitivity of the database is a known quantity.