1 Scope
This RFC is about research data that isn't public, such as licensed datasets, data under a data use agreement, and personal data. You classify the data first, and the class tells you which resource may hold it, which controls it needs, and who may access it. It builds on the Turing Institute's Data Safe Haven tiers and the SATRE specification.
2 Why
A data use agreement usually says things like "keep the data secure" and "don't share access". That's fairly clear on your own laptop. But what does it mean on a cluster, where many people share the same login nodes and file systems?
Experience with this varies a lot between researchers. Some have handled licensed and personal data for years, some are reading their first data use agreement, and many are somewhere in between. Should where your data ends up depend on that, or on who you happened to ask? Shared classes and requirements give researchers and providers the same starting point.
3 Classes
| Class | Data | Obligations |
|---|---|---|
| Public | Open data and published results. | None. |
| Restricted | Limited by a licence or data use agreement to named, approved people. Identifies nobody. | The agreement. |
| Personal | Can identify a person, directly or by linking. Includes pseudonymised data. | GDPR |
| Sensitive | Personal data in a special category, such as health or genetic data. | GDPR Art. 9, often an impact assessment (Art. 35). |
When more than one class applies, use the strictest. Licensed health records, for example, are both restricted and sensitive. If your university has its own classes, such as Aarhus University's four levels, map them to these.
4 From class to resource
- Classify the data. If it might identify a person, talk to your data protection officer.
- Pick a resource that meets the class (sections 5 and 6).
- Agree the terms: a data use agreement for restricted data, and a data processing agreement with the provider for personal data.
- Give access only to the people the agreement names.
5 Restricted data
| ID | Requirement |
|---|---|
| RD1 | The provider must give each person a personal account linked to an institutional identity, and must not permit shared logins. |
| RD2 | The provider must let a researcher keep data in a directory that no other user can read, encrypted with a key that the researcher holds. |
| RD3 | The provider must give each job temporary storage that no other job can read, and wipe it when the job ends. |
| RD4 | The provider must delete restricted data when a project ends, and must state how long it keeps logs, backups and snapshots. |
| RD5 | The provider must bind its administrators to confidentiality, and must tell the affected researchers about an exposure within 72 hours. |
| RD6 | The researcher must hold their own approval for each dataset, and register the dataset, purpose and end date with the provider. |
| RD7 | The researcher must keep the data encrypted in their private directory, and decrypt it only inside a job. |
| RD8 | The researcher must not try to identify anyone in the data, and must publish only aggregate results. |
| RD9 | The researcher must not send the data to a hosted AI service, and must treat model weights trained on it as restricted. A model on the resource is fine. |
| RD10 | The researcher must delete their copy when the project ends or the approval stops. |
| RD11 | The provider may offer a shared copy of a dataset, but only after the data owner confirms in writing that it fits the agreement. |
| RD12 | The provider must keep a shared copy read-only and readable only by approved people. It must check each approval before it adds someone, and remove them when it stops. |
Examples: WAYF or eduGAIN for accounts (RD1), gocryptfs or age for encryption (RD2), and a Slurm tmpfs job container for job storage (RD3).
6 Personal and sensitive data
There are no requirements here yet. What we know so far:
- The GDPR covers pseudonymised data, and a provider that processes personal data for a university needs a data processing agreement (Art. 28).
- Encryption is a security measure (Art. 32), but for whoever holds the key, encrypted data is still personal data.
- The P1 NGC HPC is one environment built for sensitive data. It's air-gapped, you reach it through a remote desktop, and each project gets its own data processing agreement (NGC).
7 Open questions
- Which requirements should personal and sensitive data have?
- When is encryption with a key the researcher holds enough, and when is a data processing agreement needed? Pseudonymised data may not be personal data for a party that cannot re-identify anyone (C-413/23 P).
- Which services should every provider offer, and which belong only in an environment such as the P1 NGC HPC?
- Should this RFC adopt an existing scheme, such as the Data Safe Haven tiers, instead of its own classes?