Rui Zeng (Zhejiang University), Xi Chen (Zhejiang University), Yuwen Pu (Zhejiang University), Xuhong Zhang (Zhejiang University), Tianyu Du (Zhejiang University), Shouling Ji (Zhejiang University)

Backdoors can be injected into NLP models to induce misbehavior when the input text contains a specific feature, known as a trigger, which the attacker secretly selects. Unlike fixed tokens, words, phrases, or sentences used in the textit{static} text trigger, textit{dynamic} backdoor attacks on NLP models design triggers associated with abstract and latent text features (e.g., style), making them considerably stealthier than traditional static backdoor attacks. However, existing research on NLP backdoor detection primarily focuses on defending against static backdoor attacks, while research on detecting dynamic backdoors in NLP models remains largely unexplored.

This paper presents CLIBE, the first framework to detect dynamic backdoors in Transformer-based NLP models. At a high level, CLIBE injects a textit{"few-shot perturbation"} into the suspect Transformer model by crafting an optimized weight perturbation in the attention layers to make the perturbed model classify a limited number of reference samples as a target label. Subsequently, CLIBE leverages the textit{generalization} capability of this "few-shot perturbation" to determine whether the original suspect model contains a dynamic backdoor. Extensive evaluation on three advanced NLP dynamic backdoor attacks, two widely-used Transformer frameworks, and four real-world classification tasks strongly validates the effectiveness and generality of CLIBE. We also demonstrate the robustness of CLIBE against various adaptive attacks. Furthermore, we employ CLIBE to scrutinize 49 popular Transformer models on Hugging Face and discover one model exhibiting a high probability of containing a dynamic backdoor. We have contacted Hugging Face and provided detailed evidence of the backdoor behavior of this model. Moreover, we show that CLIBE can be easily extended to detect backdoor text generation models (e.g., GPT-Neo-1.3B) that are modified to exhibit toxic behavior. To the best of our knowledge, CLIBE is the first framework capable of detecting backdoors in text generation models without requiring access to trigger input test samples. The code is available at https://github.com/Raytsang123/CLIBE.

View More Papers

Victim-Centred Abuse Investigations and Defenses for Social Media Platforms

Zaid Hakami (Florida International University and Jazan University), Ashfaq Ali Shafin (Florida International University), Peter J. Clarke (Florida International University), Niki Pissinou (Florida International University), and Bogdan Carbunar (Florida International University)

Read More

VoiceRadar: Voice Deepfake Detection using Micro-Frequency and Compositional Analysis

Kavita Kumari (Technical University of Darmstadt), Maryam Abbasihafshejani (University of Texas at San Antonio), Alessandro Pegoraro (Technical University of Darmstadt), Phillip Rieger (Technical University of Darmstadt), Kamyar Arshi (Technical University of Darmstadt), Murtuza Jadliwala (University of Texas at San Antonio), Ahmad-Reza Sadeghi (Technical University of Darmstadt)

Read More

Cascading Spy Sheets: Exploiting the Complexity of Modern CSS...

Leon Trampert (CISPA Helmholtz Center for Information Security), Daniel Weber (CISPA Helmholtz Center for Information Security), Lukas Gerlach (CISPA Helmholtz Center for Information Security), Christian Rossow (CISPA Helmholtz Center for Information Security), Michael Schwarz (CISPA Helmholtz Center for Information Security)

Read More

Panel on “Security and Privacy Issues in New 5G...

Moderator: Arupjyoti (Arup) Bhuyan, Ph.D. Director, Wireless Security Institute, Idaho National Laboratory Panelists: Ted K. Woodward, Ph.D. Technical Director for FutureG, OUSD (R&E) Phillip Porras, Program Director, Internet Security Research, SRI Donald McBride, Senior Security Researcher, Bell Laboratories, Nokia

Read More