Privacy Protection
Add noise to protect individuals.
Definition
ε-differential privacy. Adding noise to query results. Guarantees individual-level privacy.
Mechanisms
Laplace mechanism for numeric queries. Exponential mechanism for general. Gaussian mechanism for approximate.
In ML
Differential privacy SGD: clip gradients, add noise. DP-SGD provides privacy guarantees.
Key Takeaways
- Add noise for privacy
- ε controls privacy-utility tradeoff
- DP-SGD for private training