Robustifying Vision Transformer without Retraining from Scratch by Test-Time Class-Conditional Feature Alignment

Takeshi Kojima, Yutaka Matsuo, Yusuke Iwasawa