Recent advances in deep learning have raised growing concerns about privacy leakage, as models can memorize sensitive training data. To address this issue, machine unlearning aims to remove the influence of specific data samples (the forget set) while...
Recent advances in deep learning have raised growing concerns about privacy leakage, as models can memorize sensitive training data. To address this issue, machine unlearning aims to remove the influence of specific data samples (the forget set) while preserving the knowledge learned from the remaining data (the retain set). However, existing unlearning approaches often fail to achieve a balanced trade-off between forgetting effectiveness, model utility, computational efficiency and task robustness. In this work, we formulate machine unlearning as a feature distribution alignment problem and propose an Optimal Transport–based framework that aligns the feature representations of the forget set with the retain feature manifold. By modeling unlearning as a geometric transformation in the feature space, our approach effectively eliminates structural traces of forget samples while maintaining performance on the retain set. To ensure computational scalability and effective unlearning process, we introduce a prototype guided alignment mechanism using class prototypes that represent the retain distribution, along with a class-aware cost function to encourage semantic dispersion in sample-level unlearning. Extensive experiments on various datasets and models demonstrate that our proposed method achieves superior performance against state-of-the-art methods.