A Framework for Machine Unlearning Using Selective Knowledge Distillation into Soft Decision Tree

With growing privacy regulations, removing user-related information from machine learning models has become essential.Machine unlearning addresses this by enabling selective removal of learned information, but most existing methods rely on deep learning models, which are computationally expensive and lack interpretability.To overcome these limitations, we propose a novel machine unlearning framework using selective knowledge distillation into a Soft Decision Tree (SDT).A convolutional neural network (ConvNet) is first trained to generate soft labels and intermediate features, which are transferred to the SDT.During distillation, an unlearning algorithm adjusts specific leaf node distributions and routing weights using soft redistribution and path pruning.This enables class-specific forgetting without retraining and preserves accuracy on non-target classes.Experiments on MNIST and CIFAR-10 demonstrate that our framework effectively removes class-specific knowledge while maintaining overall model performance.The interpretable SDT structure also allows for clear visualization of model changes before and after unlearning.