Jordan Wylie

Project Title: Evaluation of Federated Machine Unlearning using Membership Inference Attacks

The General Data Protection Regulation (GDPR) guarantees individuals the right to be forgotten, which involves the elimination of their data and any related knowledge from databases and associated Machine Learning (ML) models upon request.  Therefore, it is essential to make use of “machine unlearning” that accurately and efficiently removes knowledge from ML models.  This includes those trained using Federated Learning (FL), where multiple devices contribute to model training in a distributed way without exchanging data.  Despite the existence of federated machine unlearning solutions, there is limited research on methods that evaluate their effectiveness in erasing this knowledge.  This is vital as FL has shown to be beneficial in areas that manage sensitive data.

To evaluate these existing federated machine unlearning solutions, an assessment framework will be developed to verify that the specified data samples are correctly and efficiently forgotten in FL models.  To achieve this, membership-inference attacks will be used within this framework.  These attacks aim to determine if a given sample was part of the training data of a model.  After this, an evidence-based approach will be taken to design a new federated machine unlearning solution to address the major weaknesses identified in the current unlearning solutions.