Zero-Shot Machine Unlearning

Modern privacy regulations grant citizens the right to be forgotten by products, services and companies. In case of machine learning (ML) applications, this necessitates deletion data not only from storage archives but also ML models. Due an increasing need for regulatory compliance required unlearning is becoming emerging research problem. The requests come in form removal a certain set or class already trained model. Practical considerations preclude retraining model scratch after discarding deleted data. few existing studies use either whole training data, subset some metadata stored during update weights unlearning. However, many cases, no related process samples may accessible purpose. We therefore ask question: it possible achieve with zero samples? paper, we introduce novel problem zero-shot that caters extreme practical scenario where original are available use. then propose two solutions based on (a) error minimizing-maximizing noise (b) gated knowledge transfer. These methods remove information forget while maintaining efficacy retain approach offers good protection against inversion attacks membership inference attacks. new evaluation metric, Anamnesis Index (AIN) effectively measure quality method. experiments show promising results deep models benchmark vision data-sets. source code here: https://github.com/ayu987/zero-shot-unlearning