Model
Differential-privacy noise and budget
DP is nearly free on large aggregates and very costly on small ones; and some budgets are depleted by reading.
- Privacy parameter
- budget spent by one release (smaller is more private)
- Sensitivity
- the most that one individual can change the answer (1 for a count)
How it behaves
One input on the horizontal axis, the others held fixed.
ε = 0.5ε = 2
Data table
| True count | series | Relative error |
|---|---|---|
| 1 | ε = 0.5 | 200% |
| 1.26 | ε = 0.5 | 159% |
| 1.58 | ε = 0.5 | 126% |
| 2 | ε = 0.5 | 100% |
| 2.51 | ε = 0.5 | 79.6% |
| 3.16 | ε = 0.5 | 63.2% |
| 3.98 | ε = 0.5 | 50.2% |
| 5.01 | ε = 0.5 | 39.9% |
| 6.31 | ε = 0.5 | 31.7% |
| 7.94 | ε = 0.5 | 25.2% |
| 10 | ε = 0.5 | 20% |
| 12.6 | ε = 0.5 | 15.9% |
| 15.8 | ε = 0.5 | 12.6% |
| 20 | ε = 0.5 | 10% |
| 25.1 | ε = 0.5 | 7.96% |
| 31.6 | ε = 0.5 | 6.32% |
| 39.8 | ε = 0.5 | 5.02% |
| 50.1 | ε = 0.5 | 3.99% |
| 63.1 | ε = 0.5 | 3.17% |
| 79.4 | ε = 0.5 | 2.52% |
| 100 | ε = 0.5 | 2% |
| 126 | ε = 0.5 | 1.59% |
| 158 | ε = 0.5 | 1.26% |
| 200 | ε = 0.5 | 1% |
| 251 | ε = 0.5 | 0.796% |
| 316 | ε = 0.5 | 0.632% |
| 398 | ε = 0.5 | 0.502% |
| 501 | ε = 0.5 | 0.399% |
| 631 | ε = 0.5 | 0.317% |
| 794 | ε = 0.5 | 0.252% |
| 1K | ε = 0.5 | 0.2% |
| 1.26K | ε = 0.5 | 0.159% |
| 1.58K | ε = 0.5 | 0.126% |
| 2K | ε = 0.5 | 0.1% |
| 2.51K | ε = 0.5 | 0.08% |
| 3.16K | ε = 0.5 | 0.063% |
| 3.98K | ε = 0.5 | 0.05% |
| 5.01K | ε = 0.5 | 0.04% |
| 6.31K | ε = 0.5 | 0.032% |
| 7.94K | ε = 0.5 | 0.025% |
| 10K | ε = 0.5 | 0.02% |
| 12.6K | ε = 0.5 | 0.016% |
| 15.8K | ε = 0.5 | 0.013% |
| 20K | ε = 0.5 | 0.01% |
| 25.1K | ε = 0.5 | 0.008% |
| 31.6K | ε = 0.5 | 0.006% |
| 39.8K | ε = 0.5 | 0.005% |
| 50.1K | ε = 0.5 | 0.004% |
| 63.1K | ε = 0.5 | 0.003% |
| 79.4K | ε = 0.5 | 0.003% |
| 100K | ε = 0.5 | 0.002% |
| 126K | ε = 0.5 | 0.002% |
| 158K | ε = 0.5 | 0.001% |
| 200K | ε = 0.5 | 0.001% |
| 251K | ε = 0.5 | 0.001% |
| 316K | ε = 0.5 | 0.001% |
| 398K | ε = 0.5 | 0.001% |
| 501K | ε = 0.5 | 0% |
| 631K | ε = 0.5 | 0% |
| 794K | ε = 0.5 | 0% |
| 1M | ε = 0.5 | 0% |
| 1.26M | ε = 0.5 | 0% |
| 1.58M | ε = 0.5 | 0% |
| 2M | ε = 0.5 | 0% |
| 2.51M | ε = 0.5 | 0% |
| 3.16M | ε = 0.5 | 0% |
| 3.98M | ε = 0.5 | 0% |
| 5.01M | ε = 0.5 | 0% |
| 6.31M | ε = 0.5 | 0% |
| 7.94M | ε = 0.5 | 0% |
| 10M | ε = 0.5 | 0% |
| 12.6M | ε = 0.5 | 0% |
| 15.8M | ε = 0.5 | 0% |
| 20M | ε = 0.5 | 0% |
| 25.1M | ε = 0.5 | 0% |
| 31.6M | ε = 0.5 | 0% |
| 39.8M | ε = 0.5 | 0% |
| 50.1M | ε = 0.5 | 0% |
| 63.1M | ε = 0.5 | 0% |
| 79.4M | ε = 0.5 | 0% |
| 100M | ε = 0.5 | 0% |
| 1 | ε = 2 | 50% |
| 1.26 | ε = 2 | 39.7% |
| 1.58 | ε = 2 | 31.5% |
| 2 | ε = 2 | 25.1% |
| 2.51 | ε = 2 | 19.9% |
| 3.16 | ε = 2 | 15.8% |
| 3.98 | ε = 2 | 12.6% |
| 5.01 | ε = 2 | 9.98% |
| 6.31 | ε = 2 | 7.92% |
| 7.94 | ε = 2 | 6.29% |
| 10 | ε = 2 | 5% |
| 12.6 | ε = 2 | 3.97% |
| 15.8 | ε = 2 | 3.15% |
| 20 | ε = 2 | 2.51% |
| 25.1 | ε = 2 | 1.99% |
| 31.6 | ε = 2 | 1.58% |
| 39.8 | ε = 2 | 1.26% |
| 50.1 | ε = 2 | 0.998% |
| 63.1 | ε = 2 | 0.792% |
| 79.4 | ε = 2 | 0.629% |
| 100 | ε = 2 | 0.5% |
| 126 | ε = 2 | 0.397% |
| 158 | ε = 2 | 0.315% |
| 200 | ε = 2 | 0.251% |
| 251 | ε = 2 | 0.199% |
| 316 | ε = 2 | 0.158% |
| 398 | ε = 2 | 0.126% |
| 501 | ε = 2 | 0.1% |
| 631 | ε = 2 | 0.079% |
| 794 | ε = 2 | 0.063% |
| 1K | ε = 2 | 0.05% |
| 1.26K | ε = 2 | 0.04% |
| 1.58K | ε = 2 | 0.032% |
| 2K | ε = 2 | 0.025% |
| 2.51K | ε = 2 | 0.02% |
| 3.16K | ε = 2 | 0.016% |
| 3.98K | ε = 2 | 0.013% |
| 5.01K | ε = 2 | 0.01% |
| 6.31K | ε = 2 | 0.008% |
| 7.94K | ε = 2 | 0.006% |
| 10K | ε = 2 | 0.005% |
| 12.6K | ε = 2 | 0.004% |
| 15.8K | ε = 2 | 0.003% |
| 20K | ε = 2 | 0.003% |
| 25.1K | ε = 2 | 0.002% |
| 31.6K | ε = 2 | 0.002% |
| 39.8K | ε = 2 | 0.001% |
| 50.1K | ε = 2 | 0.001% |
| 63.1K | ε = 2 | 0.001% |
| 79.4K | ε = 2 | 0.001% |
| 100K | ε = 2 | 0.001% |
| 126K | ε = 2 | 0% |
| 158K | ε = 2 | 0% |
| 200K | ε = 2 | 0% |
| 251K | ε = 2 | 0% |
| 316K | ε = 2 | 0% |
| 398K | ε = 2 | 0% |
| 501K | ε = 2 | 0% |
| 631K | ε = 2 | 0% |
| 794K | ε = 2 | 0% |
| 1M | ε = 2 | 0% |
| 1.26M | ε = 2 | 0% |
| 1.58M | ε = 2 | 0% |
| 2M | ε = 2 | 0% |
| 2.51M | ε = 2 | 0% |
| 3.16M | ε = 2 | 0% |
| 3.98M | ε = 2 | 0% |
| 5.01M | ε = 2 | 0% |
| 6.31M | ε = 2 | 0% |
| 7.94M | ε = 2 | 0% |
| 10M | ε = 2 | 0% |
| 12.6M | ε = 2 | 0% |
| 15.8M | ε = 2 | 0% |
| 20M | ε = 2 | 0% |
| 25.1M | ε = 2 | 0% |
| 31.6M | ε = 2 | 0% |
| 39.8M | ε = 2 | 0% |
| 50.1M | ε = 2 | 0% |
| 63.1M | ε = 2 | 0% |
| 79.4M | ε = 2 | 0% |
| 100M | ε = 2 | 0% |
Worked examples, recomputed
Every number the mechanisms quote for this relationship is recomputed from it, so none of them disagrees with it.
- ε = 0.5: a count of 100 → about 2% error
- a count of 5 → about 40%
- budget 1.0 at 0.1 per query → 10 queries, ever
Where it appears
Each mechanism states it in its own names. The general symbol each name stands for is shown with it.
- Security & Privacy
How do you protect data from someone who watches your design and then chooses how to attack it?