Machine unlearning aims to safeguard data privacy by mitigating the data’s impact on machine learning models. Nonetheless, machine unlearning practices can introduce new privacy vulnerabilities, leaving models susceptible to various forms of attack, such as confidence attack and label-only attacks. Existing defense methods encounter challenges in striking a balance between defending against attacks and sustaining model performance. In this paper, we propose a posterior perturbation method to defend against membership inference attacks by randomizing the model’s outputs without requiring adjustments. To prevent confidence attacks, we employ optimization algorithms to generate adversarial noise that disrupts the model outputs’ confidence scores, thereby obscuring the output differences between the original and unlearned models. We also propose a label perturbation method to defend against label-only attacks by randomizing the model’s output labels through high-dimensional sphere sampling. The experimental results demonstrate that our proposed method effectively defends against membership inference attacks while maintaining the model’s performance.