A Machine Unlearning Method Based on a Forgetting Neural Network (FNN)

The study proposes a machine unlearning method based on a forgetting neural network (FNN), explicitly encoding forgetting with a multiplicative decay factor, and provides the first concrete implementation of an FNN. The method sets per-neuron forgetting factors, including allocation by ranking activation levels, and is evaluated on MNIST and Fashion-MNIST. Results show that it can systematically remove information from the forget set, maintain performance on retained data, and membership inference attacks also verify the unlearning effect.