_| | | | |_

Colored Metric (CM) : An Evaluation Metric for Machine Unlearning Applied to Generative Adversarial Networks

ABSTRACT

As artificial intelligence advances, potential risks and ethical issues are emerging, and in particular, interest in the security of user data and the right to be forgotten is increasing. Machine unlearning is gaining attention as a method to delete unwanted data with minimal cost. Generative models generate various types of data, but the quality evaluation of such data mostly relies on subjective judgment. The current evaluation of machine unlearning is the same. Generative model evaluation indices such as CID index, IS (Inception Score), and FID (Fréchet Inception Distance) have limitations in evaluating the quality of machine unlearning. Therefore, this study proposed a Colored Metric (CM) that evaluates the influence of a class to be machine unlearned on surrounding classes, and verified its usefulness by comparing it with FID.

Machine Unlearning, Generative Adversarial Networks(GAN), Evaluation Metric , Fine Tuning, Privacy