Project: Machine Unlearning for Face Recognition
Winter Semester 2025/26
Face recognition is one of the most widely adopted biometrics across many domains, ranging from personal devices to security systems. However, its popularity makes it susceptible to many security and privacy issues. In practice, personal face data are often collected and stored without clear transparency or control, and once that data is embedded into a trained model, removing it is extremely difficult. This is especially problematic under regulations like the General Data Protection Regulation (GDPR), which explicitly requires that individuals have the right to request deletion of their data. Therefore, the concept of machine unlearning aims to address this issue by enabling selective removal of individuals’ data influence from trained models without the need for retraining.
Goal: The project aims to develop and implement efficient machine unlearning approaches capable of removing individual identities or data influences without retraining the model from scratch. Unlike conventional classification tasks, face recognition presents unique challenges for unlearning, as it operates on learned identity representations rather than fixed class labels, which requires adapting existing methods to this representation-based setting or developing new approaches suited to these challenges. Building on this, students will explore and implement state-of-the-art unlearning methods for face recognition, investigate research extensions such as backdoor forgetting, bias removal, and privacy risks in unlearning and design suitable, efficient verification metrics to evaluate unlearning effectiveness.
Resources:
For further information about the project structure and requirements, please check out the Project Group presentation (Presentation available here).
Also, these works for reference:
- Machine unlearning.
- Fast machine unlearning without retraining through selective synaptic dampening.
- A review on machine unlearning.