Chengkun Wei (Zhejiang University), Wenlong Meng (Zhejiang University), Zhikun Zhang (CISPA Helmholtz Center for Information Security and Stanford University), Min Chen (CISPA Helmholtz Center for Information Security), Minghu Zhao (Zhejiang University), Wenjing Fang (Ant Group), Lei Wang (Ant Group), Zihui Zhang (Zhejiang University), Wenzhi Chen (Zhejiang University)

*Prompt-tuning* has emerged as an attractive paradigm for deploying large-scale language models due to its strong downstream task performance and efficient multitask serving ability. Despite its wide adoption, we empirically show that prompt-tuning is vulnerable to downstream task-agnostic backdoors, which reside in the pretrained models and can affect arbitrary downstream tasks. The state-of-the-art backdoor detection approaches cannot defend against task-agnostic backdoors since they hardly converge in reversing the backdoor triggers. To address this issue, we propose LMSanitator, a novel approach for detecting and removing task-agnostic backdoors on Transformer models. Instead of directly inverting the triggers, LMSanitator aims to invert the *predefined attack vectors* (pretrained models' output when the input is embedded with triggers) of the task-agnostic backdoors, which achieves much better convergence performance and backdoor detection accuracy. LMSanitator further leverages prompt-tuning’s property of freezing the pretrained model to perform accurate and fast output monitoring and input purging during the inference phase. Extensive experiments on multiple language models and NLP tasks illustrate the effectiveness of LMSanitator. For instance, LMSanitator achieves 92.8% backdoor detection accuracy on 960 models and decreases the attack success rate to less than 1% in most scenarios.

View More Papers

Automatic Adversarial Adaption for Stealthy Poisoning Attacks in Federated...

Torsten Krauß (University of Würzburg), Jan König (University of Würzburg), Alexandra Dmitrienko (University of Wuerzburg), Christian Kanzow (University of Würzburg)

Read More

ORL-AUDITOR: Dataset Auditing in Offline Deep Reinforcement Learning

Linkang Du (Zhejiang University), Min Chen (CISPA Helmholtz Center for Information Security), Mingyang Sun (Zhejiang University), Shouling Ji (Zhejiang University), Peng Cheng (Zhejiang University), Jiming Chen (Zhejiang University), Zhikun Zhang (CISPA Helmholtz Center for Information Security and Stanford University)

Read More

On the Vulnerability of Traffic Light Recognition Systems to...

Sri Hrushikesh Varma Bhupathiraju (University of Florida), Takami Sato (University of California, Irvine), Michael Clifford (Toyota Info Labs), Takeshi Sugawara (The University of Electro-Communications), Qi Alfred Chen (University of California, Irvine), Sara Rampazzi (University of Florida)

Read More

SOC Service Areas: Identification, Prioritization, and Implementation

Christopher Rodman, Breanna Kraus, Justin Novak (SEI/CERT)

Read More