The enforcement of data privacy regulations, such as the GDPR and CCPA, has heightened the need for machine unlearning—the selective removal of specific training samples from machine learning models. Retraining large-scale deep neural networks from scratch is computationally expensive, motivating the exploration of more efficient strategies. This study investigates the application of the Sharded, Isolated, Sliced, and Aggregated (SISA) training framework to convolutional neural networks for privacy-preserving face recognition. The approach is evaluated on two widely used benchmark datasets, AR Face and ORL, by adapting SISA to AlexNet and VGG-16 architectures. Under the tested conditions, the SISA-based pipeline reduced retraining time by up to 62% for AlexNet and 58% for VGG-16, with a modest accuracy decline of about 0.4%. While these findings indicate potential for privacy-preserving applications, the evaluation remains limited to small datasets, lacks baseline comparisons, omits explicit privacy metrics, and does not yet analyze storage overhead. Future work will extend to larger and noisier datasets, incorporate alternative methods, and include privacy-specific and resource-related assessments.