Machine Unlearning by Reversing the Continual Learning

Recent legislations, such as the European General Data Protection Regulation (GDPR), require user data holders to guarantee individual’s right be forgotten. This means that must completely delete upon request. However, in field of machine learning, it is not possible simply remove these from back-end database wherein training dataset stored, because learning model still retains this information. Retraining using a with removed can overcome problem; however, lead expensive computational overheads. In order remedy shortcoming, we propose two effective methods help owners or private trained model. The first method uses an elastic weight consolidation (EWC) constraint term and modified loss function neutralize removed. second approximates posterior distribution Gaussian distribution, after unlearning computed by decreasingly matching moment (DMM) neural network on all Finally, conducted experiments three standard datasets backdoor attacks evaluation metric. results show both are removing triggers deep models. Specifically, EWC reduce success rate 0. IMM ensure prediction accuracy higher than 80% keep below 10%.