Federated Unlearning via Class-Discriminative Pruning

Model pruning schemes require far-reaching modifications of the existing architecture of the model in the unlearning process, which could affect the performance of the unlearned models.

“Model pruning schemes require far-reaching modifications of the existing architecture of the model in the unlearning process [55, 56], which could affect the performance of the unlearned models.”

Wang et al. [55] evaluate verifiability based on model accuracy.

“Wang et al. [55] evaluate verifiability based on model accuracy.”

In Wang et al.'s [55] scheme, the unlearned model is only available when the fined-tuned training process is complete.

“Therefore, the unlearned model is only available when the fined-tuned training process is complete.”

In Wang et al.'s [55] unlearning scheme, while the parameters associated with the class that needs to be unlearned are pruned, the parameters with other classes also become incomplete, which will affect the model performance.

“In their unlearning scheme, however, while the parameters associated with the class that needs to be unlearned are pruned, the parameters with other classes also become incomplete, which will affect the model performance.”

Wang et al. [55] regarded the output of a channel as a word and the feature map of a category as a document.

“Similar to analyzing how relevant a word is to a document in a set of documents, they regarded the output of a channel as a word and the feature map of a category as a document.”

Wang et al. [55] provided an unlearning procedure via channel pruning [97], followed by a fine-tuning process to recover the performance of the pruned model.

“An unlearning procedure via channel pruning [97] was also provided, followed by a fine-tuning process to recover the performance of the pruned model.”

Wang et al. [55] introduced the concept of term frequency-inverse document frequency (TF-IDF) [96] to quantify the class discrimination of the channels.

“They introduced the concept of term frequency-inverse document frequency (TF-IDF) [96] to quantify the class discrimination of the channels.”

Wang et al. [55] analyzed the problem of selectively unlearning classes in a federated learning setting.

“Inspired by the observation that different channels have a varying contribution to different classes in trained CNN models, Wang et al. [55] analyzed the problem of selectively unlearning classes in a federated learning setting.”

Channels with high TF-IDF scores have more discriminatory power in the target categories and thus need to be pruned.

“Channels with high TF-IDF scores have more discriminatory power in the target categories and thus need to be pruned.”

Wang et al.'s [55] scheme pruned specific parameters in the original models, which will also reduce the accuracy of the model due to the lack of some model prediction information.

“Second, Wang et al.'s [55] scheme pruned specific parameters in the original models, which will also reduce the accuracy of the model due to the lack of some model prediction information.”