Algorithms that forget: Machine unlearning and the right to erasure

Abstract

Article 17 of the General Data Protection Regulation (GDPR) contains a right for the data subject to obtain the erasure of personal data. The right to erasure in the GDPR gives, however, little clear guidance on how controllers processing personal data should erase the personal data to meet the requirements set out in Article 17. Machine Learning (ML) models that have been trained on personal data are downstream derivatives of the personal data used in the training data set of the ML process. A characteristic of ML is the non-deterministic nature of the learning process. The non-deterministic nature of ML poses significant difficulties in determining whether the personal data in the training data set affects the internal weights and adjusted parameters of the ML model. As a result, invoking the right to erasure in ML and to erase personal data from a ML model is a challenging task.

This paper explores the complexities of enforcing and complying with the right to erasure in a ML context. It examines how novel developments in machine unlearning methods relate to Article 17 of the GDPR. Specifically, the paper delves into the intricacies of how personal data is processed in ML models and how the right to erasure could be implemented in such models. The paper also provides insights into how newly developed machine unlearning techniques could be applied to make ML models more GDPR compliant. The research aims to provide a functional understanding and contribute to a better comprehension of the applied challenges associated with the right to erasure in ML.