← News

ResearchSources: arXiv · 18 Nov 2026

On-policy distillation can pass visual backdoors from teacher to student models, study finds

A new arXiv paper argues that a distillation method once thought resistant to backdoor propagation can still transmit them, driven by the teacher's probability distributions rather than malicious responses.

In plain words

Distillation is a way to train a smaller 'student' AI model by copying a larger 'teacher' model. A backdoor is a hidden flaw that makes a model misbehave on a specific trigger. The paper reports that in on-policy distillation for vision-language models, backdoors can spread even when the teacher never supplies malicious example responses.

What was announced

On-policy distillation (OPD) supervises responses the student generates itself using the teacher's probability distributions, rather than training on teacher-written responses. Because a benign student receives no explicit supervision from malicious target responses, the authors say OPD appears to offer inherent resistance to backdoor propagation. Their paper challenges that intuition, reporting experiments across multiple model families and datasets plus theoretical analysis.

The authors report three findings. Backdoors rarely appear at low poisoning ratios, but the attack success rate rapidly approaches 100% once a critical threshold is exceeded, unlike the gradual propagation they describe in traditional offline distillation. Propagation does not require explicit malicious responses; it is driven by continuous accumulation of anomalous probability mass assigned by the teacher. Building on this, they propose a trigger design enabling visual backdoor propagation with zero explicit trigger injection. The work is presented as revealing an overlooked propagation mechanism and potential security risk in OPD.

What it does not show

  • The evidence is the paper's abstract only; the experiments, threshold values and trigger design are not independently verified in the supplied material.
  • The abstract does not state which model families, datasets or poisoning thresholds were used, so the findings' precise scope cannot be assessed.

For security and IT teams

Scope
The finding concerns visual backdoor propagation in on-policy distillation for vision-language models, as described in the paper's abstract; it does not cover other training paradigms or modalities.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.