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.
Sensitivity Truncation
Meaning
Threshold based transformation that restricts the range of input values to manage query sensitivity by capping individual data points. Engineers apply sensitivity truncation to ensure that extreme values or outliers do not disproportionately affect the results of an analysis. By setting a maximum and minimum allowable value, they can pre-calculate the sensitivity of the query, which is a necessary step for adding the correct amount of noise.
This technique is common in financial and healthcare data analysis.
Range Limitation
Defining the truncation range requires a deep understanding of the data and the business context. If the range is too narrow, valuable information may be lost, but if it is too wide, the required noise will be too high, making the results useless. This balance is critical for maintaining the utility of the data.
Contractual Agreement
Specific parameters used for sensitivity truncation are often documented in the data processing agreement between the parties. This ensures that both the data provider and the recipient understand how the data is being modified. It also provides a clear baseline for auditing the privacy and accuracy of the system.
Capping Procedure
The capping of extreme values prevents individual records from dominating the final results of the query.