Search papers, labs, and topics across Lattice.
This study investigates the effectiveness of black-box auditing for language model classifiers by introducing a novel metric called Groundedness Drift, which assesses whether the provided explanations remain relevant to the input. The method was tested on two 7B model architectures across five datasets and four attack families, demonstrating superior performance in terms of AUROC and reduced target attack success rates compared to existing detectors. Additionally, the research explores Unsupported Groundedness, which enhances detection signals but fails to fully mitigate adaptive attacks.
Groundedness Drift reveals that explanations can mislead in the presence of backdoor attacks, highlighting vulnerabilities in language model classifiers.
Language model classifiers with explanations are used for moderation, routing, topic triage, and low-resource annotation. We study black-box auditing when the defender has only clean calibration data without trigger information but can ask the classifier for a label plus a short rationale or quoted evidence. We introduce Groundedness Drift, a lightweight score measuring whether the answer summary remains grounded in the input. Across two 7B backbones, five datasets, and four common non-adaptive OpenBackdoor-style attack families, Groundedness Drift achieves higher AUROC and lower residual target ASR than every compared detector in all cases at a nominal 5\% clean-FPR budget. We then evaluate Unsupported Groundedness, a multi-probe escalation for explanation-camouflage stress cases. Unsupported Groundedness improves signals but does not close the adaptive gap.
Tsinghua AI