Privacy and differential privacy
Trained models can leak their training data — through memorised text an attacker can extract, and through membership inference. Differential privacy defines what it would mean not to; the Laplace mechanism and DP-SGD achieve it at a price; federated learning keeps data on devices but is not, by itself, a privacy guarantee.
50 min
After this lesson you can
- Describe training-data extraction and membership inference attacks, and why good generalisation doesn't prevent them
- State the (ε, δ) definition of differential privacy and explain each symbol
- Derive the Laplace mechanism and the attacker-accuracy bound that ε implies
- Implement DP-SGD (per-example clipping plus Gaussian noise) and read a privacy-accounting result
- Say what federated learning protects and what it does not
In this lesson
- How models leak
- Differential privacy: the definition
- The Laplace mechanism
- DP-SGD: private deep learning
- Federated learning: keeping data at home
This lesson is part of Descent Pro
Descent Pro unlocks every lesson on the route, from base camp to the summit. Every plan starts with a free trial.