To ensure privacy compliance, systems are often required to remove the personal identities of specific individuals they were trained to recognize, leading to the emergence of machine unlearning methods. In this paper, we aim to achieve task-agnostic instance unlearning to forget specific individuals from facial images while maintaining high recognition accuracy for the model. Our approach is evaluated on the basis of the model’s accuracy after the unlearning process, the forgetting score, and the final score. Specifically, we employ the ResNet-18 model to support the facial recognition task, with its weights adjusted using our proposed method, Personal Identity Unlearning for Facial Recognition (PIU-FR). Our method, named PIU-FR, combines two key techniques: a new machine unlearning approach that integrates the widely used NegGrad and SCRUB methods, and an innovative data partitioning strategy that generates an appropriate unseen dataset for each facial dataset to support the unlearning process. The results demonstrated that the ResNet-18 model retains high task accuracy after undergoing the unlearning process and meets the set objectives. Furthermore, PIU-FR outperforms previous state-of-the-art (SOTA) approaches and even surpasses retraining from scratch. This highlights that our approach allows machine learning models to effectively remove specific data samples used during training without the need for complete retraining. Data availability and access dataset’s samples, experimental results, and proposed architectures are available publicly at https://github.com/Unlearn-PIU-FR .