Zhaohan Xi

backdoor attack (2) adversarial defense (2) few-shot learning (2) self-supervised learning (1) harmful content (1) safety alignment (1) model alignment (1) pre-trained language model (1) data curation (1) model customization (1) trigger inversion (1) large language model (1) representation invariance (1) harmful content mitigation (1) backdoor removal (1) adversarial prompt tuning (1) soft token (1) poisoning sample (1) safety compromise (1) representation learning (1)