Technologies tied to a privacy objective
Privacy-enhancing technologies reduce collection, association, disclosure, or inference while supporting a needed computation. Examples include local processing, purpose-specific identifiers, data minimization, aggregation, differential privacy, secure multiparty computation, homomorphic encryption, private set operations, trusted execution environments, query controls, and anonymous communication.
PET is not a quality label. Each technology protects a defined property against a defined observer and leaves other risks.
Select from the data flow
Start with parties, inputs, outputs, permitted computation, prohibited learning, collusion, endpoint trust, scale, latency, accuracy, and recovery. Remove data and parties before adding complex cryptography. Prefer local computation when the central service does not need raw input.
Choose a mechanism whose formal or operational claim matches the threat. A TEE changes host trust and retains hardware and code trust. Encryption in use can protect intermediate computation and still reveal output. Differential privacy protects release influence and not raw-data custody.
Composition and operation
Map keys, code, attestation, budgets, identifier scopes, access, updates, side channels, failure, logs, and recovery. Combine mechanisms only with a clear composed claim. Verify the complete implementation, not only the primitive.
Maintain privacy across export, support, debug, fallback, and degraded modes. A direct raw-data path can nullify a strong protected computation.
Failure and residual risk
Complex PETs add performance, correctness, key, hardware, and availability risk. Metadata and outputs can remain identifying. Colluding parties can change the threat. A central resolver can undo pseudonymization. A formal guarantee can be implemented with unsafe parameters or accounting.
Technology does not resolve unexpected purpose, coercion, unfair decision, or collection that should not occur.
Pomerium boundary
Pomerium can restrict which identities reach raw data, query services, key consoles, and confidential workloads. It is not itself a general PET for application data. The application and data owners must define the privacy objective, computation, observer, parameters, and output policy.
Evaluation checklist
- What exact collection, association, disclosure, inference, or observer does the technology reduce?
- Can the purpose be met by collecting less, processing locally, or using a less complex release model?
- Which raw data, metadata, outputs, endpoints, keys, hardware, code, collusion, and fallback remain trusted?
- Does the implemented and composed system preserve the claimed parameters and property?
- Which harmful use or decision remains possible even if the technology works correctly?
