Explore our new Cybersecurity Terms Glossary, explained by our experts.
Cybersecurity Terms Glossary
Privacy-preserving machine learning is the set of techniques that let models be trained and used on sensitive data while limiting what anyone can learn about the individuals in it.

What Is Privacy-Preserving Machine Learning?
Privacy-preserving machine learning covers the techniques that allow a model to be built or used without exposing the data behind it. The need is practical rather than academic: a hospital trust wants to train on patient records, a bank wants to pool fraud signals with its peers, an insurer wants to analyse claims, and in each case the data is both essential and legally constrained. The techniques differ in what they protect against, which is the first thing to establish before choosing one.
| Technique | Protects against | Main cost |
|---|---|---|
| Differential privacy | Inferring any individual from outputs or a trained model | Accuracy, especially for small subgroups |
| Federated learning | Centralising raw data in one place | Complexity; updates still leak without added protection |
| Homomorphic encryption | The compute provider seeing the data at all | Severe performance overhead |
| Secure multi-party computation | Any single party seeing the joint input | Communication cost, protocol complexity |
| Trusted execution environments | The host operating system and operator | Hardware trust assumptions, side-channel history |
| Synthetic data | Direct exposure of real records | Fidelity, and residual privacy risk |
Differential privacy is the only approach offering a formal, measurable guarantee about individual exposure, applied in machine learning by clipping per-example gradients and adding calibrated noise during training. Federated learning avoids pooling data but exchanges model updates that can themselves leak, which is why it is normally combined with secure aggregation and differential privacy rather than used alone. Homomorphic encryption allows computation on encrypted values—the CKKS scheme supports the approximate arithmetic neural networks need—but remains orders of magnitude slower than plaintext computation, so it is used for narrow inference tasks rather than training. Secure multi-party computation splits a computation so that no participant sees the others’ inputs, which suits cross-institution analytics with a small number of parties. Trusted execution environments take a different approach entirely: the data is decrypted, but only inside hardware-isolated memory that the host cannot read, and modern GPU confidential computing has made this viable for model training at scale.
Deployments rarely rest on one technique. A cross-hospital research programme might federate training, aggregate updates securely, apply differential privacy to bound individual contribution, and run the aggregator inside an enclave so the coordinating institution cannot inspect anything either. Each layer addresses a different adversary: a curious coordinator, a colluding participant, a compromised host, an external analyst reading published results. Specify the adversary first; it is the only way to tell whether a proposed architecture is proportionate or theatre.
The ICO treats these methods as privacy-enhancing technologies and addresses them in its guidance on anonymisation, pseudonymisation, and PETs. Three points matter for compliance teams. First, using a PET is strong evidence of data protection by design under Article 25 and of data minimisation, which is exactly what a Data Protection Impact Assessment should record. Second, it does not automatically take data outside UK GDPR—identifiability remains the test, assessed on the parameters you chose and what else an attacker could reasonably obtain. Third, the model produced may itself be personal data if individuals can be inferred from it, which is the membership inference question and a live one for erasure requests. In health and public sector research the UK has pushed secure data environments, where analysts work inside a controlled setting rather than receiving extracts, and these techniques complement that model rather than replacing it. The ICO and the US authorities also ran a joint PETs prize challenge, which produced useful public reference implementations for financial crime and health use cases.
Start with the question the model must answer and the specific legal constraint you face, then choose the lightest technique that satisfies both. Overengineering is common and expensive; so is applying a technique with parameters so weak that the guarantee is nominal. Document the threat model, the parameters, and the residual risk, because that documentation is the deliverable a regulator or a partner institution will ask for. At Privia Security, our Cyber Risk Analysis Services assess whether a proposed privacy architecture addresses the adversary you actually have.
Identify the adversary first. If you are worried about what a published model or result reveals about individuals, differential privacy is the direct answer. If the obstacle is that data cannot leave each institution, federated learning addresses that. If the concern is the cloud provider seeing your data, confidential computing fits. Choosing before defining the threat usually produces an expensive architecture that misses the real risk.
For narrow inference tasks on small models, yes; for training large models, not at present. The performance overhead remains several orders of magnitude, and schemes supporting the operations neural networks need are the slowest. Most production systems that need a strong confidentiality guarantee use trusted execution environments instead, accepting a hardware trust assumption in exchange for usable performance.
No, though they can change which obligations bite. The ICO assesses identifiability on the facts, so a strong, well-parameterised approach may support an anonymisation argument while a weak one will not. Even where data remains in scope, applying a PET is persuasive evidence of data protection by design and of appropriate security measures under Article 32.
Mostly not. Differential privacy and federated learning determine how a model is trained, so they require retraining. Confidential computing and access controls can be applied to an existing model’s serving environment, and output minimisation and rate limiting reduce what an attacker can learn from it. For an existing high-risk model, test it for memorisation and plan a retrain if the results are poor.
You May Be Interested In These
We use a few cookies that are needed to make this website work, and they are always on. We would also like to set analytics cookies, to see which pages are useful; functional cookies, so that our live chat works; and advertising cookies, to measure our campaigns. We will not set any of those until you say yes, and you can change your mind at any time from the Cookie Settings link in the footer.
Choose the cookies you are happy for us to use. Strictly necessary cookies are always on; everything else is off until you turn it on.
These cookies keep the website working: they hold your security checks, your form submissions and the choice you make here. The site cannot run without them, so they are always on.
These let our live chat open and remember your conversation, and let the newsletter sign-up work. Switch them off and the chat window will not appear.
These tell us which pages are read and how people find us, so that we can improve the site. The reports are about groups of visitors, never about you personally.
These let us measure our advertising and show you our adverts on other websites, such as LinkedIn, Google and Meta. Switch them off and you will still see adverts, but they will be less relevant.