Skip to content

HalluProbe: using VLM hallucinations for backdoor detection

HalluProbe: Exploiting VLM Hallucinations for Backdoor Detection

arXiv: 其他 AI 安全论文· arXiv:2611.19868· Ji Guo, Xiaolong Qin, Jin Jiang, Yijing Liu, Pei Ke, Aiguo Chen, Wenbo JiangLaboratory of Intelligent Collaborative Computing, University of Electronic Science and Technology of China; School of Software Engineering, Chengdu University of Information Technology; National Key Laboratory of Wireless Communications, University of Electronic Science and Technology of China; School of Computer Science and Engineering, University of Electronic Science and Technology of ChinaIndexed by this site · 2 days ago· Originally publishedDefenses· Its place in the scan record →

Translation in progress — untranslated text is omitted.

Quick read

Defenses

Based on the paper PDF (AI-generated); refer to the original

AI summaryThe researchers propose a backdoor detection method, HalluProbe, that identifies vision-language models (VLMs) with backdoors implanted via data poisoning by amplifying VLM hallucinations during decoding and measuring the distributional consistency of hallucinated tokens. Its core finding is that hallucination is associated with backdoor behavior: poisoned training data imposes a strong language prior on the backdoor target token, and this prior is also a major source of hallucination, so amplifying hallucination can expose hidden backdoor-related priors. Backdoored VLMs hallucinate the backdoor target token more frequently across diverse inputs, whereas the hallucinated tokens of clean VLMs are largely random with no stable target bias. In experiments on three models and six representative VLM backdoor attacks, HalluProbe's detection accuracy exceeds 90%, which the authors state is better than existing backdoor detection methods.

Questions
VLMs are susceptible to backdoors implanted via data poisoning; the authors state that existing VLM backdoor detection mostly requires poisoned data, additional defense training, or trigger-containing inputs, limiting practical detection scenarios.
Methods
The authors consider that backdoor poisoning strengthens the language prior pointing to the target token, and that an overly strong language prior is exactly the source of hallucination. HalluProbe uses EHCD to amplify hallucination and VCD to suppress hallucination, takes the distribution of tokens differing between the two outputs, and uses the Top-1 vs. Top-2 frequency difference as the detection score to compare with a threshold.
Experiments and results
On InternVL2.5-8B, LLaVA-NeXT-8B, and Qwen3-VL-8B against 6 backdoor attacks, the authors report that HalluProbe's F1 is highest in all 18 settings; ablations show that removing EHCD or VCD lowers F1, and removing both brings it close to random.
Limitations and future work
The authors state that for clean-label backdoors they only discussed potential strategies and did not experimentally verify them (Appendix D.2); which objective dominates under multiple objectives and why it is inconsistent with the ASR ranking is an open question (Appendix D.1). The experiments cover only 3 8B open-source VLMs and 6 attacks, and the main experiments fix the poisoning rate at 5% and 3 training epochs.
Read originalarxiv.org