Defense against Backdoor Attack on Pre-trained Language Models via Head Pruning and Attention Normalization.

Xingyi Zhao, Depeng Xu, Shuhan Yuan

Browse the full ICML paper archive.