Why You Should Never Use Conventional Rounding for Protecting Privacy
Conventional (“deterministic”) rounding takes confidential statistics for aggregate tables and rounds them to a fixed base (such as the nearest 5 or 10) prior to publishing. While widely used for its simplicity, deterministic rounding creates notable privacy weaknesses and adds unnecessary error compared to random noise addition using a framework such as Differential Privacy (DP).
Read the full brief
Key Findings
- Higher Accuracy: Conventional rounding adds uniform noise, making a large error just as likely as a small error. Random noise added through DP can be calibrated to add the same level of error as rounding, but where larger values of noise are less likely to occur. This results in narrower confidence intervals and less uncertainty around the confidential value, providing more accurate statistics.
- Elimination of Boundary Vulnerabilities: Rounding fails when confidential values fall on rounding interval boundaries (e.g., 14 rounded down versus 15 rounded up). An attacker with external background information can infer true values with 100% certainty across rounding boundaries. DP maintains the same protections for all confidential values.
- Release of Multiple Statistics: Deterministic rounding privacy protections decline rapidly as more statistics are released, leading to a high risk of disclosures. Random noise added through DP allows data curators to maintain stronger protections as more statistics are published.