Privacy-enhancing technologies (PETs) are technical methods that allow data to be collected, analyzed, shared or computed on while minimizing the exposure of personal information to the parties involved. In Privacy engineering, PETs are the concrete toolbox that turns principles such as data minimization and privacy by design into enforceable technical guarantees, rather than relying only on contracts and access policies.
The category is broad. It ranges from long-established measures such as pseudonymization and encryption to advanced cryptographic and statistical methods such as homomorphic encryption, secure multi-party computation and differential privacy. What unites them is a shared goal: reduce what any single party can learn while preserving useful results.
How privacy-enhancing technologies protect data
PETs protect data by changing what is revealed, to whom and at what stage. Some transform the data itself so that individuals are harder to identify; others keep data encrypted or isolated during computation; still others add calibrated noise so that outputs cannot be traced to individuals.
A useful way to organize the field is by the point in the data lifecycle where protection applies. The table shows the main groups and the guarantee each provides.
| Family | Representative techniques | What it protects |
|---|---|---|
| Data transformation | Pseudonymization, masking, generalization, k-anonymity | Identity in stored or shared datasets |
| Cryptographic computation | Homomorphic encryption, secure multi-party computation | Inputs while they are being processed |
| Statistical output protection | Differential privacy | Individuals in published results |
| Hardware isolation | Trusted execution environments | Data in use on untrusted infrastructure |
| Distributed processing | Federated learning, on-device processing | Raw data that never leaves the device |
| Network privacy | Onion routing, Oblivious HTTP | Link between user identity and request |
| Synthetic data | Generative models trained on real data | Real records replaced by artificial ones |
Most production systems combine several families. A data clean room, for example, may join pseudonymized datasets inside an isolated environment and release only aggregates protected by thresholds or differential privacy.
Key techniques and their trade-offs
Every PET trades privacy strength against utility, performance and complexity. Choosing well requires understanding these costs, which are often underestimated in early project planning.
| Technique | Privacy strength | Performance cost | Main limitation |
|---|---|---|---|
| Pseudonymization | Low to medium | Negligible | Data remains personal data and linkable |
| k-anonymity | Medium | Low | Vulnerable to homogeneity and background knowledge attacks |
| Differential privacy | High, quantified by epsilon | Low to moderate | Noise reduces accuracy for small groups |
| Fully homomorphic encryption | Very high | Very high, often orders of magnitude | Slow and limited for complex workloads |
| Secure multi-party computation | High | High network and compute overhead | Requires coordination among parties |
| Trusted execution environments | Medium to high | Low to moderate | Trust shifts to hardware vendor; side channels |
| Federated learning | Medium | Moderate | Model updates can still leak information |
In differential privacy, the privacy budget epsilon expresses the guarantee: smaller values give stronger protection and noisier results. Values below 1 are considered strong, while deployed systems often use larger budgets, so the chosen epsilon should always be disclosed and justified.
Privacy-enhancing technologies and related concepts
PETs are often confused with neighboring ideas in the privacy engineering vocabulary. The distinction matters because a principle is not a technology, and a legal status is not a technical property.
| Concept | Nature | Relationship to PETs |
|---|---|---|
| Privacy by design | Engineering and legal principle | PETs are one means of fulfilling it |
| Anonymization | Outcome in which data no longer identifies people | Some PETs aim at it; few achieve it alone |
| Data security controls | Protection against unauthorized access | PETs also limit what authorized parties learn |
| LINDDUN | Privacy threat modeling method | Identifies threats that PETs then mitigate |
| Data clean room | Collaboration environment | A product that packages several PETs |
The difference from classic security is essential. Encryption at rest stops outsiders, but a PET restricts the legitimate processor as well, for example by letting an analyst compute an average without ever seeing individual salaries.
Choosing a PET for a use case
Selection starts with the threat model: who must be prevented from learning what, and what output is genuinely needed. The decision table below maps common situations to typical choices.
| Situation | Typical choice | Reason |
|---|---|---|
| Publishing population statistics | Differential privacy | Bounds what any release reveals about one person |
| Joint analysis by competitors | Secure multi-party computation or clean room | No party sees another’s raw data |
| Outsourced computation on sensitive data | Trusted execution or homomorphic encryption | Cloud operator cannot read inputs |
| Training models on phones | Federated learning with secure aggregation | Raw data stays on devices |
| Test data for developers | Masking or synthetic data | Realistic structure without real customers |
| Hiding who requests a resource | Oblivious HTTP or onion routing | Separates identity from content |
Maturity also counts. Pseudonymization and masking are routine and inexpensive, while fully homomorphic encryption remains specialized. A sound plan applies the simplest technique that meets the threat model, then adds stronger methods only where residual risk justifies the cost.
Limitations and common misconceptions
PETs reduce risk but do not make processing automatically lawful or harmless. Several recurring misconceptions lead to weak deployments:
- Pseudonymized data is still personal data under the GDPR, because re-identification remains possible with additional information.
- Removing names is not anonymization; combinations of attributes such as postcode, birth date and gender can single people out.
- Encryption alone does not hide metadata such as timing, volume and the identity of the parties.
- PETs protect inputs, not purposes: a privacy-preserving computation can still produce a discriminatory or unexpected inference.
Governance therefore remains necessary. Documentation of parameters, independent review of implementations and testing against realistic re-identification attacks separate genuine protection from privacy claims that exist only on paper.
A plain-language summary
Privacy-enhancing technologies are tools that let organizations work with data without seeing more than they need. Imagine several people wanting to know their average salary without anyone revealing their own number. With the right method, they can get the correct average while every individual salary stays secret.
Some of these tools blur details, some keep data locked even while it is being used, and some add a little random noise to results. The idea is always the same: get the useful answer, expose as little as possible.
An example from practice
A group of five regional hospitals wants to study how a treatment affects recovery times, but regulations and patient trust prevent them from pooling raw records. Their privacy engineers design a combined approach instead of a central database.
Each hospital keeps its records locally and pseudonymizes patient identifiers with a shared keyed hash so that the same patient can be matched across sites. A secure multi-party computation protocol then calculates the joint statistics without any hospital seeing another’s data. Before publication, the results pass through a differential privacy mechanism that adds calibrated noise, and groups smaller than a defined threshold are suppressed.
The study reports recovery differences with confidence intervals that account for the added noise. No hospital transferred raw patient data, and the published tables cannot reasonably be used to infer whether a specific patient took part.