What are Privacy-Enhancing Technologies (PET)?
Privacy-enhancing technologies, told as a story about a toolbox with several different ways to protect a guest.
Hide the whole list, and there's nothing left to analyze.The guests are safe, but the list is useless for analysis.
Show the whole list, and guests are in danger.The analysis works great — but names, ages, and home cities are exposed.
PET is a toolbox holding several different tools, each protecting the guest its own way.There isn't just one way to protect a guest — you pick the tool that fits the situation.
Differential privacy, homomorphic encryption, federated learning, and synthetic data each fit a different situation.Some, like differential privacy, are fast; others, like homomorphic encryption, are slow and costly.
Perfect privacy and perfect accuracy can't both happen at once.Just like pseudonymization, how much to hide is always a trade-off between safety and accuracy.
PET (privacy-enhancing technologies) = a set of tools — differential privacy, homomorphic encryption, federated learning, synthetic data — that protect data while keeping it usable.
PET is an umbrella term for technologies that let data be analyzed or used for training while still protected or encrypted. Differential privacy adds statistical noise to results so no single individual's presence can be detected; homomorphic encryption lets a party compute directly on ciphertext without decrypting it (the key holder decrypts the result); federated learning gathers only trained models from each device, never the raw data itself; synthetic data generates fake records that mimic the real distribution. No single technique perfectly satisfies both privacy and accuracy or efficiency at once, so they are mixed and matched to fit the situation.
When grown-ups say it
- Differential privacy
- Mixing a little blur into the result. Adds statistical noise so no one can tell whether a single person changed the result.
- Homomorphic encryption
- Computing while still locked. Computes without opening the lock; the key holder opens the answer. → the letter nobody can read without the key
- Federated learning
- Learning separately, at each house. The data never leaves home — only what was learned gets collected.
- Synthetic data
- A look-alike fake list. Fake guests with similar traits stand in for real ones, in analysis or testing.
- Zero-knowledge proof
- Proving something without showing it. Lets you prove you know a password without ever saying it.
- Privacy-accuracy trade-off
- A seesaw: one side rises as the other falls. Hide more, and it's safer but less accurate.
- Pseudonymization
- Swapping in a number tag. The most basic way of hiding a name. → swap in a number tag, lock the real name away
- Data linkage/combination
- Combining ledgers from several castles. One way to do it is inside a controlled special room. → the special reading room you can't carry data out of