Coded Machine Unlearning using Machine Learning
Year: 2022
Type: article
Abstract: There are some applications that may necessitate removing the trace of a sample from the system, such as when a user requests that their data be deleted or when corrupted data is discovered. Simply removing a sample from storage does not necessarily remove its entire trace, because downstream machine learning models may store some information about the samples used to train them. If a sample is completely unlearned, it can be completely unlearned. We retrain all models that used it from scratch, removing that sample from their training dataset. When multiple such unlearning requests are anticipated, unlearning by retraining becomes prohibitively costly. The training data can be divided into smaller disjoint shards and assigned to non-communicating weak learners using ensemble learning. Each shard is used to create a faulty model. These models are then combined to form the final central model. In this paper, we propose a coded learning protocol in which the training data is encoded into shards prior to the learning phase using linear encoders. In addition, we present the corresponding unlearning protocol and demonstrate that it meets the perfect unlearning criterion.
Author: Tirumalasetty, Prasanth
Institution: University of Michigan–Dearborn
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7108339685