Research project REFLECT-ML: Right to be Forgotten through Non-Redundant Information Control in ML
Flagship project within Digital Futures that examines how the Right to be Forgotten applies when personal data has been used to train AI systems.
Photo: Unsplash
REFLECT-ML is an interdisciplinary research project on the Right to be Forgotten in machine learning. The project investigates what personal information may remain in AI models, what residual traces of such information may be legally acceptable, and how effective methods for data deletion and machine unlearning can be developed. Its overall aim is to bridge the gap between legal requirements and technical possibilities.
Objective
- Careful assessment of what new information due to personal data of an individual has been memorised;
- Guidelines outlining what residuals of personal information in a ML model should be legally acceptable;
- Explore how the RTBF can be effectively implemented, ultimately in the context of ML systems.
Background
In the context of machine learning (ML), there is an urgent need to clarify what “deletion” under the Right to be Forgotten (RTBF) truly entails. As ML models generalize and internalize patterns from data, achieving complete removal of an individual data point remains a major technical challenge. We argue, however, that such full deletion often exceeds what the law actually requires. REFLECT-ML aims to explore how the RTBF can be effectively implemented, ultimately in the context of ML systems. We will address the disconnect between the abstract legal language of regulators and operationally founded technical theoretical measures that can be used to quantify the new information associated with individual data as well as the practical complexities in estimating those. Further, we will develop different technical approaches to control its memorization as well as exploring machine unlearning attempts.
Cross-disciplinary collaboration
The team consists of PI Oechtering (researcher in information theory, relevant for information quantification), PI Colonna (researcher in law, relevant knowledge related to the GDPR and AI Act), and PI Johansson (researcher in optimization, relevant for the design of efficient learning algorithms). The outlined work is a cross-disciplinary effort since information measures need to be legal compliant, legal requests need to be algorithmically feasible, and algorithms need to aim for the right objective.