Explore our new Cybersecurity Terms Glossary, explained by our experts.

View the Glossary View the Glossary
PRIVIA

Cybersecurity Terms Glossary

What Is Differential Privacy?

Differential privacy adds calibrated mathematical noise to results so that no individual’s presence in a dataset can be inferred, giving a measurable rather than assumed privacy guarantee.

Illustration for the cybersecurity term Differential Privacy
Home / Blog / Cybersecurity Terms / What Is Differential Privacy?
17 November 2026

What Is Differential Privacy?

What Is Differential Privacy?

Differential privacy is a mathematical definition of privacy loss, introduced by Cynthia Dwork and colleagues in 2006. A process satisfies it if the result is almost exactly the same whether or not any single individual’s record is in the dataset. “Almost exactly” is where the mathematics lives: the permitted difference is bounded by a parameter, so for the first time privacy protection became something you could measure and budget rather than assert.

The Guarantee in Plain Terms

Imagine an analyst who already knows everything about a dataset except whether you are in it. Differential privacy ensures that looking at the published result tells them almost nothing new. The bound is written as epsilon: a smaller epsilon means tighter protection and more noise, a larger epsilon means more useful numbers and weaker protection. A second parameter, delta, allows a small probability of the guarantee failing in practical implementations. The property that makes this usable at scale is composition—privacy loss from repeated queries adds up in a quantifiable way, so an organisation can set an overall budget for a dataset and track consumption instead of guessing when it has released too much.

Two Places to Add the Noise

ModelWhere noise is addedTrade-off
CentralBy a trusted curator holding the raw data, before publishing resultsBetter accuracy, but the curator sees everything
LocalOn each person’s device, before data is sentNo trusted curator needed, but far more noise for the same utility
ShuffleLocally, then anonymised and mixed by an intermediaryA middle position used in modern telemetry designs

Where It Is Already Running

The US Census Bureau applied differential privacy to the 2020 census as its disclosure avoidance system, the largest public deployment to date and the one that generated the most debate about utility. Apple has used local differential privacy for usage telemetry since iOS 10 in 2016, and Google has deployed it in RAPPOR and in mobility reporting. In machine learning, differentially private stochastic gradient descent clips per-example gradients and adds noise during training, producing models with a formal bound on what any single training record can contribute. That directly limits membership inference and memorisation, which is why it appears throughout privacy-preserving machine learning.

The Costs Nobody Mentions First

Noise reduces accuracy, and the effect falls hardest on small subgroups—exactly the populations that research and public policy often care about most. Training with differential privacy increases compute cost and usually lowers model quality. Choosing epsilon is a policy decision dressed as a technical one: values below 1 are considered strong, values in the low single digits are common in practice, and some deployments have used figures high enough that the formal guarantee carries little practical meaning. Implementation also matters, since floating-point noise generation and timing have both produced real vulnerabilities. NIST SP 800-226, published in final form in 2025, exists precisely to help evaluate whether a claimed guarantee is credible.

The UK Data Protection View

The ICO addresses differential privacy in its guidance on anonymisation, pseudonymisation, and privacy-enhancing technologies. The important point for compliance teams is that applying the technique does not automatically place data outside UK GDPR. Whether output is anonymous depends on the epsilon chosen, the composition across all releases, and what else is reasonably available to an attacker—the same identifiability test that applies to any anonymisation claim. Used well, it is strong evidence of data protection by design under Article 25 and a credible answer to the data minimisation principle. Used with a large epsilon and no budget tracking, it is a label rather than a control. Document the parameters, the budget, and the reasoning, because that documentation is what an auditor or the ICO will ask to see. At Privia Security, our Regulation and Compliance Services help organisations evaluate whether a privacy technique genuinely meets the obligation it is being used to satisfy.

FAQ – Frequently Asked Questions

What Value of Epsilon Should We Use?

There is no universal answer, which is why the choice must be documented. Values below 1 are generally regarded as strong, single digits are common in production, and anything in the tens offers limited practical protection. Decide based on the sensitivity of the data, how many releases you plan, and who might attack the output, then record the reasoning alongside the budget you set.

Does Differential Privacy Make Data Anonymous Under UK GDPR?

Not automatically. The ICO assesses identifiability in the round: the parameters you chose, the cumulative effect of every release from the same data, and what other information an attacker could reasonably combine with it. A strong epsilon with a tracked budget supports an anonymity argument; a weak one with repeated queries does not. Treat it as evidence, not as a conclusion.

Can We Apply It to an Existing Model?

Not retrospectively in any meaningful sense. The guarantee comes from how the model was trained, so you would need to retrain using differentially private optimisation. For an existing model, the practical alternatives are output controls, rate limiting, and testing for memorisation and membership inference, which reduce exposure without providing a formal bound.

How Much Accuracy Will We Lose?

It depends on dataset size, the strength you choose, and how skewed the data is. Large datasets with common patterns absorb noise well; small subgroups and rare categories suffer most. Expect a measurable drop in model quality under tight budgets, and test on the subpopulations that matter rather than on aggregate accuracy, which can hide the worst effects.

Related Terms

You May Be Interested In These

What Is Dirty COW?

Cybersecurity Terms

17 November 2026

What Is Dirty COW?

Read More
What Is ISO/IEC 42001?

Cybersecurity Terms

17 November 2026

What Is ISO/IEC 42001?

Read More