The paper introduces a novel approach to evaluate machine unlearning (MU) by identifying the worst-case forget set, which is the subset of data points that poses the greatest challenge for unlearning. This approach addresses the limitations of existing methods that rely on random data forgetting, which may not accurately reflect the true performance of unlearning algorithms. By adopting a bi-level optimization (BLO) framework, the authors aim to identify the most difficult data subset for unlearning, thereby providing a more reliable and comprehensive evaluation of MU. The worst-case forget set is identified through a two-level optimization process. At the upper level, the goal is to maximize the difficulty of unlearning by selecting the most challenging data subset. At the lower level, the unlearning process is performed to ensure the model's utility on data not targeted for unlearning. This approach allows for a balance between data influence erasure and model utility. The authors demonstrate the effectiveness of their method through extensive experiments on various datasets and models, including image classification and text-to-image generation. The results show that the worst-case forget set evaluation reveals critical strengths and weaknesses of existing unlearning strategies, highlighting the complex challenges of MU in practice. The method is also shown to be compatible with other metrics for evaluating unlearning effectiveness and utility. The study also explores the application of the worst-case forget set in different unlearning scenarios, including class-wise forgetting and prompt-wise forgetting. The results indicate that the worst-case forget set can be effectively used to identify the most challenging data points for unlearning, leading to more accurate and robust unlearning algorithms. The paper concludes that the proposed approach provides a more reliable and comprehensive evaluation of MU, addressing the limitations of existing methods and offering a new perspective on the challenge of unlearning. The results suggest that incorporating curriculum learning into MU could significantly improve unlearning effectiveness, and the process of identifying the worst-case forget set offers a way to attribute data influence based on their unlearning difficulty.The paper introduces a novel approach to evaluate machine unlearning (MU) by identifying the worst-case forget set, which is the subset of data points that poses the greatest challenge for unlearning. This approach addresses the limitations of existing methods that rely on random data forgetting, which may not accurately reflect the true performance of unlearning algorithms. By adopting a bi-level optimization (BLO) framework, the authors aim to identify the most difficult data subset for unlearning, thereby providing a more reliable and comprehensive evaluation of MU. The worst-case forget set is identified through a two-level optimization process. At the upper level, the goal is to maximize the difficulty of unlearning by selecting the most challenging data subset. At the lower level, the unlearning process is performed to ensure the model's utility on data not targeted for unlearning. This approach allows for a balance between data influence erasure and model utility. The authors demonstrate the effectiveness of their method through extensive experiments on various datasets and models, including image classification and text-to-image generation. The results show that the worst-case forget set evaluation reveals critical strengths and weaknesses of existing unlearning strategies, highlighting the complex challenges of MU in practice. The method is also shown to be compatible with other metrics for evaluating unlearning effectiveness and utility. The study also explores the application of the worst-case forget set in different unlearning scenarios, including class-wise forgetting and prompt-wise forgetting. The results indicate that the worst-case forget set can be effectively used to identify the most challenging data points for unlearning, leading to more accurate and robust unlearning algorithms. The paper concludes that the proposed approach provides a more reliable and comprehensive evaluation of MU, addressing the limitations of existing methods and offering a new perspective on the challenge of unlearning. The results suggest that incorporating curriculum learning into MU could significantly improve unlearning effectiveness, and the process of identifying the worst-case forget set offers a way to attribute data influence based on their unlearning difficulty.