Lion Cipher Lab
Privacy-preserving machine learning: the basics
- AI Security
- Beginner
- ~40 min
- Free · no account required
Why it matters
It is tempting to think that once data is 'in the model,' it is safely blended away -- averaged into billions of numbers no one could pull back out. That intuition is wrong, and the gap between it and reality is where a lot of AI privacy harm lives. Models can and do memorize specific training examples, and researchers have shown you can often tell whether a particular person's record was in the training set just by probing the finished model. That matters enormously when the training data is medical records, location histories, or private messages. The good news is that this is not a…
What you'll learn
- understanding how models leak training data
- membership-inference at a conceptual level
- what differential privacy is and what the privacy budget epsilon means
- the accuracy-vs-privacy trade-off
- federated learning at a high level
- data minimization as a first principle
- reading privacy claims critically
Steps in this lab
- The uncomfortable truth: models remember
- The cheapest defense is the data you never collect
- Differential privacy: a promise you can measure
- See it: add Laplace noise to a count
- Turn the dial: what epsilon actually costs
- Federated learning: train without gathering the data
- Putting it together (and the honest limits)
- What you built here, and the instinct to keep