Lazy Machine Unlearning Strategy for Random Forests

Abstract

Removing the impact of some revoked training data from the machine learning models, i.e., machine unlearning, is a non-trivial task, which plays a pivotal role in fortifying the privacy and security of ML-based applications. This paper focuses on the problem of machine unlearning for random forests efficiently, with the streaming setting of revocation requests. The existing works are all devoted to speeding up the unlearning process of a single revocation request, none of works target at the streaming scenario. A straightforward solution is to carry out the unlearning technique tailored for a single request immediately when a new revocation request arrives. Undoubtedly, that is time-inefficient, since the time cost is proportional to the number of requests involved in the streaming. To solve this problem, this paper proposes a lazy unlearning strategy to carry out the unlearning operations involved in different revocation requests in a single batch, so as to avoid redundant computations and implement computation sharing. In particular, we adopt a node level unlearning policy, which carries out the unlearning operations on demand of the testing request by checking whether the revocation requests on this node can affect the inference of the testing request. Experiments on several real datasets show that compared to the baseline, lazy unlearning strategy can improve the unlearning efficiency by 1.1X-4X on different datasets and the number of retraining times is reduced to 1/4 on average.