Machine unlearning refers to the process of removing the influence of specific training data from a machine learning model, thereby supporting privacy compliance and data governance. In this study, we extend prior work on weight-resetting unlearning methods by investigating the impact of selective layer-wise freezing on unlearning performance. Using the CIFAR-100 dataset and the ResNet-50 architecture as a testbed, we design a series of experiments that freeze different hierarchical layers during unlearning to assess their contribution to forgetting effectiveness and model recovery. We employ six comprehensive evaluation metrics, including accuracy on forget/retain sets, membership inference attacks (MIA), activation distance, Jensen-Shannon divergence, and Zero Retrain Forgetting (ZRF), to quantify the behavioral shift of the model during unlearning. Our results show that unlearning primarily relies on adjusting high-level features, with deeper layers being more influential in eliminating class-specific knowledge. Additionally, t-SNE visualizations reveal that forgotten samples tend to be reassigned to semantically similar categories, emulating a form of natural forgetting. These findings provide actionable insights into the internal dynamics of unlearning and suggest that targeted manipulation of higher-level features can significantly enhance unlearning effectiveness while preserving model utility.