Skip to content
InfoResearchPreprintLLM-specific

HalluProbe: Exploiting VLM Hallucinations for Backdoor Detection

Published
Record updated
View JSON

Summary

Researchers show that backdoored Vision-Language Models hallucinate their backdoor target tokens more often across diverse inputs, while clean models produce random hallucinations without a stable target bias. They propose HalluProbe, which enhances hallucinations during output decoding and measures the distributional consistency of hallucinated tokens to flag backdoored models. Evaluated against six VLM backdoor attacks on three models, HalluProbe detects backdoored VLMs with over 90% accuracy, outperforming existing detection methods.