EntUn: Mitigating the forget-retain dilemma in unlearning via entropy

초록

Advancements in natural language processing and computer vision have raised concerns about models inadvertently exposing private data and confidently misclassifying inputs. Machine unlearning has emerged as a solution, enabling the removal of specific data influences to meet privacy standards. This work focuses on unlearning in Instance-Removal (IR) and Class-Removal (CR) scenarios: IR targets the removal of individual data points, while CR eliminates all data related to a specific class. We propose EntUn, which maximizes entropy for the forget-set to reduce confidence in data to be forgotten and minimizes it for the retain-set to preserve discriminative power. An entropy-based intra-class mixup further stabilizes this process, using higher-entropy samples to guide controlled information removal. Experiments on CIFAR10, CIFAR100, and TinyImageNet show that EntUn outperforms state-of-the-art baselines, improving forgetting and enhancing privacy protection as confirmed by membership inference attack tests. This demonstrates entropy maximization as a robust strategy for effective unlearning.

키워드