<![CDATA[This study proposes a Hybrid Machine Unlearning framework that combines Reverse Gradient Descent (RGD) and Image Obfuscation to enable selective and privacy-preserving forgetting in deep neural networks without full retraining. The proposed method introduces a dual-layer mechanism that operates at both the model level, by reversing gradient updates to remove target class contribution, and the data level, by applying visual obfuscation to eliminate identifiable features prior to unlearning. Experiments were conducted on a subset of the Labeled Faces in the Wild (LFW) dataset containing 96 identity classes, evaluated under two unlearning scenarios with 5 and 10 target classes. Quantitative results show that pure RGD achieved values between 0.10–0.19 and 0.03–0.06 for the respective scenarios, while the hybrid RGD+Obfuscation configuration strengthened the forgetting effect up to 0.82 on MobileNetV2 and 0.26 on ResNet architectures. Model stability remained high, with below 0.27, indicating minimal degradation of non-target representations. The hybrid method achieved an unlearning time of 2–6s per epoch, with an additional 50–65% computational cost due to blurring and pixelation operations yet maintained an 80–90% reduction in total runtime compared to full retraining. These results demonstrate that the proposed hybrid approach effectively enhances forgetting strength while maintaining retention stability and computational efficiency. The method provides a scalable and resource-efficient solution for privacy-aware continual learning and AI lifecycle management, where fast and controlled machine unlearning is required without compromising model integrity.]]>
Copyrights © 2026
