ETD

• computer vision
• deep learning
• fine tuning
• machine unlearning
• noisy data
• object detection
• transformers
• user privacy

Machine Unlearning (MU) is a rapidly growing research field that aims to remove specific infor-
mation from AI models. While existing MU approaches in computer vision mainly focus on image
classification, their applicability to spatial localization tasks, such as object detection (OD), remains
underexplored. This thesis establishes the first framework for unlearning in OD. Specifically, this
thesis designs two main setups for this problem: User Privacy (UP) to comply with privacy regula-
tions such as the GDPR and CCPA, and Resolving Confusion (RC), which aims to correct the
model from noisy label training. Its theoretical foundations and evaluation include proposing novel
metrics, such as adapting the Membership Inference Attack (MIA) to assess unlearning in OD
or Interclass Confusion. Additionally, this thesis extends state-of-the-art MU techniques, originally
developed for image classification, to OD, demonstrating their applicability to both CNN and
Transformer-based architectures across diverse unlearning scenarios. This thesis provides a
starting point in the field of machine unlearning in OD, leading future research for privacy-compliant
AI and removing influences of noisy labels from DL models.