- https://publications.jrc.ec.europa.eu/repository/handle/JRC134350
- https://www.sciencedirect.com/science/article/pii/S2352340924005274?via%3Dihub
- Source: jrc
TL;DR
This EU Joint Research Centre report surveys six privacy-enhancing technology categories — differential privacy, federated learning, SMPC, homomorphic encryption, synthetic data, and TEEs — and provides a decision framework for selecting among them based on data sensitivity, trust model, and GDPR requirements.
Summary
No single privacy technology suits all data-sharing scenarios. This JRC report surveys six PET categories: differential privacy, federated learning, secure multi-party computation, homomorphic encryption, synthetic data, and trusted execution environments. Each is assessed on maturity, use cases, limitations, and GDPR alignment. A decision framework guides practitioners in selecting the appropriate PET based on data sensitivity, trust model, and performance constraints — bridging technical capabilities with regulatory obligations for European practitioners.
Key contributions
- Provides a structured survey of six PET categories with consistent assessment criteria across maturity, use cases, and limitations.
- Introduces a decision framework mapping data sensitivity and trust model to appropriate PET selection.
- Assesses GDPR alignment for each technology, connecting technical controls to legal compliance obligations.
- Identifies federated learning and differential privacy as the most mature PETs for production data-sharing deployments.
- Demonstrates that hybrid PET architectures can cover individual weaknesses at the cost of increased complexity.
When to cite
- When selecting among PETs (differential privacy, FL, SMPC, HE, synthetic data, TEEs) for a data-sharing use case.
- When justifying a privacy technology choice against GDPR requirements in a European regulatory context.
- When comparing the maturity and practical limitations of homomorphic encryption vs. federated learning for production use.
- When arguing that no single PET is sufficient and hybrid privacy architectures should be considered.