Learn to Forget: Machine Unlearning Via Neuron Masking

Nowadays, machine learning models, especially neural networks, have became prevalent in many real-world applications. These models are trained based on a one-way trip from user data: as long users contribute their data, there is no way to withdraw. To this end, machine unlearning becomes popular research topic, which allows the model trainer unlearn unexpected data model. In paper, we propose first uniform metric called forgetting rate measure effectiveness of unlearning method. It concept membership inference and describes transformation eliminated “memorized” “unknown” after conducting unlearning. We also novel method Forsaken . superior previous work either utility or efficiency (when achieving same rate). benchmark with eight standard datasets evaluate its performance. The experimental results show that it can achieve more than 90% average only causeless 5% accuracy loss.