Who May Use the Model? Policy-Aware Selection of Data Contributors in Federated Learning

The project

Federated Learning (FL) lets multiple organisations train a shared machine learning model without exchanging their raw data. In emerging data ecosystems, such as European data spaces, data is however rarely offered unconditionally. Data owners attach access and usage policies to their datasets, increasingly in machine-readable form using the W3C Open Digital Rights Language (ODRL): the data may be used for research but not for commercial purposes, only within the EU, only until a certain date, only with attribution, not for redistribution, and so on.

When such datasets are used to train a federated model, a natural question arises: which usage rights does the resulting model carry? A reasonable starting assumption is that the model inherits the restrictions of every dataset it was trained on – permissions narrow down to what all contributors allow, while prohibitions and obligations accumulate. This creates a fundamental trade-off. Including more contributors typically means more and more diverse training data, and therefore a better model – but also a model that is more restricted in how it can be used, or even unusable because of conflicting policies. Including fewer contributors keeps the model permissive, but may leave it under-trained.

Research focus

The goal of this MSc project is to find combinations of data contributors that are not too restrictive for the intended practical use of the model, while still providing a sufficient amount of high-quality data to train a good model. The thesis may address one or several of the following questions:

• Policy composition: how can the ODRL (or other) usage policies of individual datasets be combined into a policy for the resulting model, and how can conflicts between them be detected automatically?
•  Quantifying both sides of the trade-off: how can the permissiveness of a combined policy be measured (e.g. against a target usage profile such as “commercial use within the EEA”), and how can the value of each contributor's data be estimated before or during training (quantity, quality, diversity, contribution scores such as Shapley values)?
•  Contributor selection: formulating the choice of contributors as a constrained or multi-objective optimisation problem, and solving it using e.g. integer programming, greedy heuristics or evolutionary algorithms, producing Pareto-optimal trade-offs between model performance and freedom of use.
•  Dynamics: handling contributors who join, leave, or change or revoke their policies during or after training.

The approach will be implemented and evaluated in a simulated FL environment (e.g. Flower), using benchmark datasets distributed among clients and annotated with realistic ODRL policies. The topic is connected to ongoing SINTEF research on federated learning and data sharing.

Expected results and learning outcome

• A formal model of how usage policies propagate from individual datasets to a federated model.
•  A policy-aware contributor selection method, implemented as a prototype on top of an FL framework.
•  An experimental evaluation illustrating the trade-off between model quality and usage rights.
•  Experience with FL, data valuation, policy languages and optimisation, at the intersection of machine learning and data governance – a field of growing importance with the European data strategy and the EU Data Act and AI Act. Results of sufficient quality can be published at a scientific workshop or conference together with the supervisors.

Desired qualifications

• Experience with machine learning and common frameworks (PyTorch, TensorFlow, scikit-learn, Hugging Face, etc.)
• Good programming skills in Python
• Interest in data governance and knowledge representation; familiarity with semantic web technologies (RDF, ODRL, SPARQL, SHACL) is an advantage
• Basic knowledge of optimisation (e.g. integer or multi-objective optimisation) is an advantage
• Experience with a Federated Learning framework such as Flower, TensorFlow Federated or FATE (desirable)

References

• W3C. ODRL Information Model 2.2. W3C Recommendation, 2018. https://www.w3.org/TR/odrl-model/
• T. Li, A. K. Sahu, A. Talwalkar, V. Smith. Federated learning: Challenges, methods, and future directions. IEEE Signal Processing Magazine, 2020.
• A. Ghorbani, J. Zou. Data Shapley: Equitable valuation of data for machine learning. ICML, 2019.
• L. Fu, H. Zhang, G. Gao, M. Zhang, X. Liu. Client selection in federated learning: Principles, challenges, and opportunities. IEEE Internet of Things Journal, 2023.