ForgetBase: a machine unlearning architecture for efficient privacy protection

With the rise of artificial intelligence, trained models pose significant privacy and security risks due to their tendency to memorize training data, prompting the need for machine unlearning technologies to comply with regulations. However, existing machine unlearning methods struggle with efficiency, especially in dynamic data environments, and often degrade model performance or fail to verify unlearning execution. To tackle these challenges, this paper presents ForgetBase, a novel machine unlearning framework that incorporates a dual-path filtering mechanism and a dynamic parameter adjustment strategy. By integrating these components, the framework is designed to efficiently address the limitations of existing methods in dynamic data environments. Analysis and discussions demonstrated that ForgetBase not only significantly reduces time and computational costs while maintaining high accuracy on test sets and effectively preventing unlearned data from being processed, but also enhances privacy protection by storing unlearned data in hash format. ForgetBase provides an efficient and practical solution for machine unlearning, balancing privacy and operational efficiency.