From Internal Diagnosis to External Auditing: A VLM-Driven Paradigm for Online Test-Time Backdoor Defense

Binyan Xu, Fan Yang, Xilin Dai, Di Tang, Kehuan Zhang · 2026

Read From Internal Diagnosis to External Auditing: A VLM-Driven Paradigm for Online Test-Time Backdoor Defense by Binyan Xu, Fan Yang, Xilin Dai, Di Tang, Kehuan Zhang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-19448.txt.

Description

Binyan Xu and four co-authors defend neural networks at test time by auditing inputs from outside the model. Backdoor attacks plant hidden triggers that flip a model's answer on cue. Old test time defenses diagnose from inside, repairing the model or hardening inputs, but they stay tangled with the poisoned weights.

Their paradigm uses a vision language model as an external auditor of what comes in. An outside judge is not bound by the victim model's corruption. The shift from internal repair to external audit may hold up better as attacks grow more advanced.

Catalog record

CREATORBinyan Xu, Fan Yang, Xilin Dai, Di Tang, Kehuan Zhang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE85 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-19448

Browse the full library · About the Tower of Babel