Privacy-enhancing technologies

Privacy-enhancing technologies reduce the amount of personal information exposed during collection, storage, processing, or disclosure. The term covers measures such as data minimization, pseudonymization, anonymization, encryption, access controls, and computation on protected data. The United States National Institute of Standards and Technology groups PETs around minimizing data, obscuring identities or values, separating processing environments, and aggregating information.¹

PETs do not make data automatically private or eliminate legal responsibilities. A dataset can remain identifying when supposedly anonymous records are combined with outside information, while encrypted data may become exposed through keys, endpoints, metadata, or implementation errors. Effective privacy protection therefore combines technology with retention limits, purpose specification, governance, auditing, and user rights. The European Union's General Data Protection Regulation treats privacy by design and by default as organizational and technical obligations rather than as a single product feature.²

Different PETs protect different points in the data lifecycle. Differential privacy adds carefully calibrated randomness to queries or outputs so that the presence or absence of one person has limited influence on the result; its privacy parameter, epsilon, helps quantify the privacy–utility trade-off.³ Statistical agencies and researchers use it for aggregate releases, although repeated queries consume a limited privacy budget.

Encryption protects data in transit and at rest, while homomorphic encryption permits selected computations on ciphertexts. Secure multiparty computation lets several parties jointly calculate a result without revealing their private inputs to one another. Federated learning keeps training data on participating devices or servers, but model updates can still leak information and may require secure aggregation or differential privacy.⁴ Zero-knowledge proofs allow one party to demonstrate that a statement is true without revealing the underlying secret, a technique useful in authentication and verifiable credentials.

PETs are valuable when organizations need useful results but should not receive identifiable records. Health institutions can use secure computation for collaborative studies; governments can publish differentially private statistics; financial institutions can compare fraud signals without exchanging complete customer files; and messaging services can combine end-to-end encryption with minimized metadata. These arrangements can support research and service delivery while reducing unnecessary disclosure.

Protection is nevertheless conditional. Pseudonyms remain linkable, aggregation can fail when groups are small, and anonymization may be undermined by auxiliary datasets. Differential privacy can reduce statistical accuracy, particularly for rare events or detailed subgroups. Cryptographic systems introduce computational cost, key-management requirements, and risks from compromised devices or malicious participants. PET selection therefore depends on the threat model, the sensitivity and distribution of data, the intended output, and who controls the infrastructure—not simply on whether a method is labeled “privacy-preserving.”⁵

Privacy protection can concern metadata and inferences as much as the original records. Even when message content is encrypted, communication patterns, locations, timing, device identifiers, and social relationships may reveal sensitive facts. Private information retrieval can let a user obtain an item from a database without revealing which item was requested, while private set intersection enables parties to discover common entries without exposing their nonmatching records. These techniques have specialized uses in contact discovery, intelligence, and cross-institutional analytics.

PET research also examines data trusts, personal data stores, synthetic data, and “compute-to-data” models in which approved code runs near protected records and only vetted outputs leave the environment. Synthetic data can lower direct exposure but may reproduce confidential patterns or permit membership inference if generated carelessly. The most durable approach is layered: collect less, separate identities from attributes, restrict computation, measure disclosure risk, and test systems against realistic attacks.⁶

Differential privacy

A formal privacy framework that limits how much the output of an analysis can change when one person's data is added or removed.

Pseudonymization

Replacing direct identifiers with artificial identifiers while retaining a separate means of re-identification.

Homomorphic encryption

Encryption that supports specified computations on ciphertexts, producing an encrypted result that can later be decrypted.

Secure multiparty computation

Cryptographic protocols that let parties compute a joint function while keeping their individual inputs concealed.

Federated learning

A machine-learning arrangement in which models are trained across local datasets and updates, rather than centralizing all raw training data.