Privacy Enhancing Technology
also called PETs
Techniques that allow useful computation over data without exposing the underlying records — differential privacy, secure multi-party computation, homomorphic encryption, federated learning.
These moved from research to procurement conversations quickly, and an architect should be able to say which problem each actually solves rather than treating them as one category.
Differential privacy adds calibrated noise so that no individual's presence measurably changes any published result, with a quantified privacy budget. It is the mature one, used in census publication and telemetry at scale, and its cost is accuracy — small populations and rare values suffer most, which is exactly where analysts often want detail.
Federated learning trains a model across devices or institutions without centralising the data, sending updates rather than records. It fits health and mobile keyboard use cases well, and it is not private by itself: model updates leak information, so it is normally combined with differential privacy or secure aggregation.
Secure multi-party computation lets parties jointly compute a function over their private inputs with nobody seeing the others' data. Practical today for constrained problems such as set intersection, which is what many clean-room overlap analyses actually need.
Homomorphic encryption allows computation on ciphertext. Genuinely powerful, still slow enough that it is confined to narrow high-value cases.
The advice worth giving when one is proposed: name the specific threat being mitigated. These are expensive and constraining, and they are frequently proposed where access control and minimisation would do the job.