Pre-training has been a necessary phase for deploying pre-trained language models (PLMs) to achieve remarkable performance in downstream tasks. However, we empirically show that backdoor attacks exploit such a phase as a vulnerable entry point for task-agnostic. In this paper, we first propose $\mathtt{maxEntropy}$, an entropy-based poisoning filtering defense, to prove that existing task-agnostic backdoors are easily exposed, due to explicit triggers used. Then, we present $\mathtt{SynGhost}$, an imperceptible and universal task-agnostic backdoor attack in PLMs. Specifically, $\mathtt{SynGhost}$ hostilely manipulates clean samples through different syntactic and then maps the backdoor to representation space without disturbing the primitive representation. $\mathtt{SynGhost}$ further leverages contrastive learning to achieve universal, which performs a uniform distribution of backdoors in the representation space. In light of the syntactic properties, we also introduce an awareness module to alleviate the interference between different syntactic. Experiments show that $\mathtt{SynGhost}$ holds more serious threats. Not only do severe harmfulness to various downstream tasks on two tuning paradigms but also to any PLMs. Meanwhile, $\mathtt{SynGhost}$ is imperceptible against three countermeasures based on perplexity, fine-pruning, and the proposed $\mathtt{maxEntropy}$.
翻译:暂无翻译