Research on Machine Unlearning Verification Based on Predictive Cross-Entropy

The General Data Protection Regulation (GDPR), implemented by the European Union in 2018, set a global standard for personal data privacy protection. Among its requirements is the “right to be forgotten” which mandates that service providers must delete a user's personal data upon the user's request to withdraw data authorization. However, deep learning networks tend to retain learned information, making them susceptible to privacy attacks. While several methods exist for achieving machine unlearning to enforce the right to be forgotten, verifying the effectiveness of machine unlearning remains challenging. In this paper, we propose the metric “Percentage of Unlearning” (PoU), which measures the extent of machine unlearning through the change in predictive cross-entropy of the data intended to be forgotten during the unlearning process. We introduce a machine unlearning verification framework based on predictive cross-entropy, enabling users to self-verify whether their data has been deleted without solely relying on the service providers' assurances. This framework also assists service providers in ensuring complete machine unlearning while minimizing time and computational costs. We compared the PoU metric with existing verification metrics such as ACC and MIA recall. Experimental results demonstrate that, compared to existing verification methods, the PoU metric not only confirms the effectiveness of machine unlearning but also addresses the shortcomings of existing metrics by quantifying the degree of unlearning.