When machine unlearning meets script identification

Abstract

Machine Unlearning (MU) has emerged as a new paradigm for forgetting data samples from a given model. However, existing MU methods have focused on popular classification problems, leaving the landscape of unlearning for script identification and document analysis relatively unexplored. This paper addresses this gap by proposing an MU framework for script identification in scene text scenarios, utilizing deep learning networks. We conducted extensive experiments to assess the impact of data removal on different combinations of classes, including single and multiple classes, along with varying percentages of the forget set. We implemented two unlearning strategies: retraining from scratch (US) and fine-tuning (UF) for efficient forgetting manipulation. We evaluated our approach using a tiny vision transformer variant and ConvNeXt pre-trained models for scene text script identification on the SIW-13 dataset. Our results demonstrate that fine-tuning minimizes performance degradation.