What do privacy-enhancing technologies actually enable, and what do they cost in practice?

The main families and what each is for

• Differential privacy: calibrated noise added to results so that no individual's presence materially changes the output. The only approach that gives a quantified, composable guarantee, and it is the one that addresses repeated querying rather than single-query thresholds.
• Federated learning: models trained on-device with only updates sent centrally, so raw data never leaves. Genuinely reduces collection, and the updates themselves can leak information without additional protection.
• Secure multi-party computation: several parties compute a joint function without revealing inputs. Powerful for cross-organisation analysis and computationally expensive.
• Homomorphic encryption: computation on encrypted data. Strongest guarantee, and the performance cost confines it to narrow high-value operations.
• Trusted execution environments: hardware-isolated processing. Practical and fast, and the trust rests on the hardware vendor and its attestation chain.

The costs to state plainly

Accuracy — differential privacy trades precision for the guarantee, and the privacy budget is finite, so the analysis programme is bounded rather than open-ended. Performance — the cryptographic approaches range from expensive to impractical for general workloads. Complexity — correct implementation is specialist work, and a subtly wrong implementation provides a guarantee that does not hold, which is worse than none because it is relied upon.

Where they earn their cost

Cross-organisation analysis that would otherwise not be permitted at all. The right framing for the business is that the alternative is not less-protected analysis, it is a data-sharing agreement nobody will sign — constrained measurement that both legal teams approve delivers more value than an unconstrained design that never launches.

The trap to avoid

Treating a privacy technology as a substitute for governance. Aggregate-only outputs are defeated by differencing across queries; federated updates can be inverted; a trusted execution environment protects the data and not the analyst's intent. The technology bounds the disclosure; policy still has to bound the purpose.