The Price of Forgetting: Data Redemption Mechanism Design for Machine Unlearning

Nowadays, technology companies spend efforts col-lecting datasets from massive users and training machine-learning models to enable innovative artificial intelligence (AI) applications. However, current data protection policies, such as General Data Protection Regulation (GDPR), enforce the right to be forgotten and require the server to perform machine unlearning and eliminate the effect of users' data on the trained model once receiving the data redemption requests. Such privacy regulations are unfair to the server, as it incurs extra costs for unlearning but suffers from a degraded model. In this paper, we propose the first incentive mechanism in machine unlearning to compensate for the server's cost to the best of our knowledge. We first characterize the accuracy degradation and consumed time as a function of the unlearning ratio by conducting experiments on three popular datasets and two widely used unlearning algorithms. Then, we model the interaction between the server and users as a two-stage Stackelberg game. We propose an iterative algorithm to optimize the unit price for data redemption by characterizing the convexity of server's profit maximization problem. The experimental results on real datasets show that our mechanism can achieve the largest server's profit and social welfare, compared with the GDPR and no redemption schemes.