Privacy-enhancing technologies in practice: the hype is ahead of the implementation

Synthetic data, federated learning, homomorphic encryption, differential privacy: the acronym count around PETs has grown faster than the production deployments. Regulators in the EU, UK, and Canada have published guidance endorsing PETs as a route to compliant data sharing. The UK Information Commissioner's Office and the US National Institute of Standards and Technology have both invested in PET-specific frameworks. A 2023 PricewaterhouseCoopers survey (conducted before the current period, so treat it as directional) found that roughly 60% of large enterprises planned to increase PET investment within two years. Analyst firms like Gartner have placed PETs on their privacy technology radars for several consecutive years. The consensus has formed, and it is coherent.

The consensus view, stated fairly

The dominant argument goes like this: data-sharing arrangements that were legally impossible or reputationally dangerous under GDPR and its equivalents can now proceed safely using PETs. Federated learning lets companies train shared models without moving raw data. Differential privacy adds calibrated statistical noise so that individual records cannot be reconstructed. Homomorphic encryption allows computation on encrypted data. Synthetic data generates statistically representative datasets that contain no real personal information. Together, these techniques open previously closed doors: cross-industry fraud detection, pharmaceutical research across hospital networks, financial risk modelling between competitors. The privacy regulator says yes, the CISO sleeps at night, and the data science team gets the signal it needs. Everyone wins.

Google's use of differential privacy in Chrome, Apple's use of it in iOS telemetry, and Google DeepMind's federated learning work on medical imaging are regularly cited as proof points. The technology works. The mathematics behind homomorphic encryption, developed by Craig Gentry at IBM Research and now commercialised by companies including Zama and IBM itself, is not contested.

Where the consensus goes wrong

The consensus makes two related errors. It conflates mathematical proof with operational readiness, and it treats early adopters as representative of average organisations.

Start with the computational cost problem. Fully homomorphic encryption remains expensive at scale. Processing encrypted data under HE can be orders of magnitude slower than processing plaintext, depending on the circuit depth and the library used. IBM and Zama (a vendor, whose figures should be cross-checked against independent benchmarks) have both published performance improvements, and the gap is narrowing. But for a mid-sized bank running nightly batch processes, "narrowing" is not the same as "solved". The production use cases that have actually shipped tend to be narrow in scope: a specific query type, a fixed model architecture, a well-bounded dataset. Generalising from these to enterprise-wide PET deployment is a category error.

Federated learning presents a different set of problems that receive less coverage. The technique assumes participants have comparable data qualitydata qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.View full definition →, compatible model architectures, and roughly symmetric contributions to the shared model. In practice, a hospital consortium where three institutions have electronic health records from 2015 onward and two have paper-only records until 2022 will produce a federated model that reflects those asymmetries in ways that are difficult to audit. Differential Privacy UK, the government initiative that ran trials on PETs for financial crime data sharing between 2022 and 2024, found that participant data heterogeneity was one of the primary barriers to usable outputs. That finding has not made it into most vendor pitches.

Synthetic data has its own failure mode that is gaining attention in 2026: membership inference attacks. Researchers at Google and several academic groups have demonstrated that synthetic datasets generated by generative models can leak information about the original training data if the generation model itself is not protected with differential privacy. Companies buying synthetic data from vendors like Mostly AI or Gretel (both commercial providers, whose security claims deserve independent validation) often do not ask whether the generation process was itself differentially private. The output looks clean; the generation pipelinepipelineAll active sales opportunities across the stages of the sales process, together with their combined potential value and probability of closing.View full definition → may not be.

The second error is governance theatre. A CDO can approve a PET pilot, brief the board on the compliance benefits, and file the regulator paperwork while the underlying data governancedata governanceData governance is the set of policies, roles, and processes that ensure data is accurate, secure, well-defined, and used responsibly across an organization.View full definition → infrastructure remains inadequate. PETs do not substitute for data minimisation, retention controls, or access management. They operate on top of those controls. An organisation that cannot accurately inventory its personal data holdings will not suddenly become compliant because it has deployed a federated learning wrapper around a poorly governed data lakedata lakeA data lake is a centralized repository that stores large volumes of raw data in its native format, from structured tables to unstructured files, until needed.View full definition →.

What a sharp operator should actually do

The first move is to stop treating PETs as a category and start treating them as a menu with different readiness levels for different use cases. Differential privacy for aggregate analytics reporting is deployable today by most data teams with moderate effort. Apple's open-source Swift Differential Privacy library and Google's open-source DP library are production-grade. That part of the consensus is correct. Homomorphic encryption for cross-institutional model inference is a different matter; it belongs in a 12-to-24-month research pipeline, not a Q3 roadmap.

The second move is to demand vendor-independent benchmarking before committing budget. When a PET vendor quotes latency or accuracy figures, those numbers are almost always generated on that vendor's preferred hardware configuration with their preferred workload. Ask for results on your architecture, with your data volume, and compare them against published academic benchmarks.

The third move is to use PET pilots as a governance audit. A federated learning pilot will immediately surface whether your data cataloguing is accurate enough to participate meaningfully. A synthetic data project will reveal whether your data scientists can distinguish a statistically valid synthetic dataset from one that simply looks plausible. These are valuable diagnostics independent of whether the PET pilot proceeds to production.

Regulatory pressure on cross-border data flows is not receding in 2026, and PETs will play a real role in enabling compliant data collaboration. The technology is not a mirage. But the CDOs who get ahead of this are the ones who read the ICO guidance and the academic benchmarks, not just the vendor white papers. A PET strategy built on pilots that cannot scale is still a compliance gap with a convincing slide deck.