Privacy-Enhancing Technologies (PETs)

Privacy-enhancing technologies (PETs) are a category of technical methods designed to let data be processed, analyzed, or shared while limiting exposure of the underlying sensitive information — including anonymization and pseudonymization, differential privacy, synthetic data generation, federated learning, and cryptographic techniques like homomorphic encryption and secure multi-party computation. Each technique protects data differently and trades off usability, accuracy, and computational cost in different ways, which is why PETs are better understood as a toolkit to select from than a single solution to adopt.

This matters for AI specifically because different AI use cases call for different privacy guarantees. Anonymizing a transcript before it reaches a summarization model is a very different problem from training a model across multiple hospitals' patient data without any single hospital's raw records ever leaving its own servers — the first is well served by entity-level anonymization, while the second typically requires a technique like federated learning or secure multi-party computation instead. Treating "PETs" as interchangeable with anonymization alone can leave an organization reaching for the wrong tool for a given risk.

A consortium of banks wanting to jointly train a fraud-detection model without sharing their raw customer transaction data with each other is a clear example of PETs beyond simple anonymization. Federated learning allows each bank to train a shared model locally, on its own data, and only share model updates rather than the underlying records — meaning no bank ever sees another's actual customer data, while all of them benefit from a model trained on a far larger pattern set than any single institution's data alone would support.

A healthcare research project analyzing patient outcomes across multiple hospitals without physically centralizing patient records is a similar case, often using a combination of federated learning and differential privacy to add statistical noise that prevents any individual patient from being re-identified from the aggregate results. A software vendor wanting to let customers test a product against realistic data without exposing real customer records might instead use synthetic data generation, creating artificial records that preserve the statistical patterns of the original data without containing any actual individual's information at all. In each case, the right PET depends on the specific problem — sharing a model without sharing data, publishing aggregate statistics safely, or testing against realistic-but-fake data — not on a single default technique applied regardless of fit.

Organizations that treat anonymization as a complete substitute for the full range of PETs can end up under-protected in use cases anonymization wasn't designed to solve. Anonymization works well for masking identifiers in a document or transcript before it reaches a model, but it doesn't address a scenario like multiple organizations needing to collaboratively train a model without any of them seeing each other's raw data, or a need to publish aggregate statistics with a mathematical guarantee that no individual can be re-identified from them, the way differential privacy specifically provides.

Treating anonymization as the one PET to reach for every time is a mismatch — different privacy problems call for different tools, and picking the wrong one can leave the actual risk unaddressed even while something labeled "privacy protection" is technically running.

Questa AI is built around anonymization and re-identification control as its core privacy-enhancing technique, specifically suited to the most common AI risk scenario: sensitive data reaching a model or a log in the course of everyday AI use. This is a deliberate focus rather than an attempt to cover every PET — anonymization is the right tool for masking identifiers in documents, prompts, and transcripts before they reach an AI model, which is the scenario most organizations' everyday AI workflows actually involve.

Questa's governance dashboard and Blackbox recording give organizations visibility into where this specific technique is applied and how effectively, which matters for demonstrating proportionate data protection to regulators and auditors. For use cases requiring a different PET entirely — federated learning across organizations, or differential privacy for published aggregate statistics — those typically call for a purpose-built approach alongside, rather than instead of, the anonymization Questa applies to everyday AI data flows.