• class unlearning
• deep learning
• evaluation framework
• machine unlearning

Machine unlearning (MU), the process of selectively removing specific information from a trained AI model, is gaining significant attention to address critical issues such as privacy preservation and bias mitigation. Despite recent advancements, we observe a notable gap in understanding the unlearning process itself, as most current works focus mainly on assessing the predictive capabilities of unlearned models on retaining and forgetting sets. This narrow assessment neglects certain aspects of unlearning, such as how closely the behavior of an unlearned model approximates that of a retrained model, how the forgetting set is redistributed among the retaining classes, and whether unlearned classes are reassigned to semantically similar ones from a human perspective. This gab is partially due to the lack of a comprehensive evaluation framework for MU. To address this limitations, we propose a multi-dimensional evaluation framework aiming to move beyond the conventional approach by measuring the dissimilarity between an unlearned model and a ground truth retrained model, analyzes the redistribution of forgotten classes among retaining classes, and examines semantic consistency in class reassignment. Additionally, we present a new algorithm designed to perform MU without requiring access to the original training data. By leveraging semantic representations, our proposed method achieves performance comparable to or better than existing s-o-t-a methods. Importantly, this algorithm is designed to allow a human to guides the unlearning process by defining unlearning strategies for each class ensuring the resulting model not only forgets the targeted classes but also a coherence with the semantic of the classes.