Synthetic Data: Security and Data Protection

How synthetic data is created

The starting point is almost always a real source dataset. A generator learns its patterns and links. Then it creates new records. Common methods are:

• Statistical methods: Distributions and correlations are modelled and then sampled again, for example with Bayesian networks or copulas. They are easy to understand, but weak at complex links.
• Generative models: Generative adversarial networks (GANs), variational autoencoders and diffusion models create realistic tables, images or time series.
• Language models: Large language models write synthetic text, dialogues or sample documents, for instance to train classifiers.
• Rule-based test data: For software tests, data is often built from format rules alone, with no source data at all. This is harmless, but not very realistic.

Typical use cases

For mid-sized companies, three cases stand out:

• test and development systems that today hold copies of live data
• AI projects where real data is scarce, unbalanced or sensitive
• sharing data with service providers or research partners

The first case matters most for security. Test systems are usually less well protected than live systems.

Security and privacy risks

Inferences about real people

Generative models can memorise training records and repeat them almost word for word. Outliers are most at risk. An employee with an unusual salary or a patient with a rare diagnosis can often still be recognised in synthetic data. Membership inference attacks can also show whether a person was in the source data. Attribute inference can work out missing traits of a known person.

Quality and bias risks

Synthetic data takes on the biases of its source and can make them worse. Rare but important cases, such as fraud patterns or defect types, may not show up correctly. Models trained mostly on synthetic data can do worse in the real world. If generative models are trained again and again on their own output, quality can slowly drop. This is called model collapse.

Manipulating the data pipeline

The generator and its settings are a target for attacks. Anyone who tampers with it can feed skewed data into many models further down the line. This is a form of data poisoning. It is hard to spot, because people expect synthetic data to look "artificial" anyway.

Data protection perspective

Creating synthetic data from personal source data is itself processing under the GDPR. It needs a legal basis under Art. 6. The output falls outside the GDPR only if no one can identify a person with means reasonably likely to be used (Recital 26). You must justify and document this. Calling data "synthetic" is not enough.

The EU AI Act (Regulation (EU) 2024/1689) names synthetic and anonymised data as the preferred option. Only if they are not enough may special categories of personal data be processed to detect and correct bias. The Digital Omnibus has been in force since 27 July 2026. Since then, this legal basis no longer covers only high-risk AI. Under strict necessity rules, it covers all AI systems and general-purpose AI models.

Practical steps

1. Define the purpose: Functional tests often need no more than rule-based data with no source. You only need realistic synthesis when the statistical traits really matter.
2. Cut down source data: Remove direct identifiers and any traits you do not need before you start.
3. Build privacy into the method: Generators with differential privacy limit, in a provable way, how much one record can shape the output.
4. Measure privacy: Check how far each synthetic record is from the closest real one. Look for duplicates and run membership inference tests.
5. Measure utility: Compare distributions. Train models on synthetic data, test them on real data and compare how well they do.
6. Secure the generator: Limit access to the source data, the generator model and its settings, and log all changes. The generator model itself may hold personal data.

Conclusion

Synthetic data can greatly lower the risk of data leaks, above all in test and development systems. But it is not anonymous by default. You need a documented check of both privacy and quality. SecTepe supports risk assessments for such projects as part of ISMS work and data protection impact assessments.

Frequently asked questions

Can I share synthetic data freely?

Only if it has been shown that no one can draw conclusions about real people or trade secrets. Without that check, treat it like the source data.

Does synthetic data replace anonymisation?

It is one possible way to anonymise data, not a guarantee. It is judged by the same criteria as any other anonymisation method.