As Student Rollouts Continue, Visual Backdoors Keep Growing
Organizations: Alibaba Group · Guangxi University · Wuhan University of Technology · Institute of Automation, Chinese Academy of Sciences ·
Abstract
Conventional knowledge distillation can facilitate the propagation of backdoors from teacher models to student models. However, how this risk manifests in On-Policy Distillation (OPD) remains insufficiently studied. OPD has recently emerged as an important post-training paradigm for Vision-Language Models (VLMs). Unlike traditional offline distillation, OPD does not require teacher-generated responses. Instead, it uses the teacher's probability distributions to supervise response trajectories generated online by the student. Since the initial student model typically generates benign responses and receives no explicit supervision from malicious target responses, OPD appears to offer some inherent resistance to backdoor propagation. In this paper, we challenge this intuition and reveal a more subtle risk of visual backdoor propagation in OPD. Through experiments across multiple model families and datasets, together with theoretical analysis, we identify three main findings. First, unlike the relatively gradual propagation observed in traditional offline distillation, backdoors rarely emerge in OPD at low poisoning ratios. However, the attack success rate rapidly approaches 100% once a critical threshold is exceeded. Second, backdoor propagation does not depend on explicit malicious responses. Instead, it is driven by the continuous accumulation of anomalous probability mass assigned by the teacher. Finally, based on these findings, we propose a novel trigger design that enables visual backdoor propagation with zero explicit trigger injection. These findings further reveal an overlooked mechanism of backdoor propagation in OPD and highlight its potential security risks.
Figures & tables
| Attack | Clean quality | Attack | Clean quality | ||||||
| Ratio (%) | ASR | FTR | BLEU-4 | CIDEr | Ratio (%) | ASR | FTR | BLEU-4 | CIDEr |
| LLaVA-1.5 (Checkpoint-400) | LLaVA-1.5 (Checkpoint-400) | ||||||||
| 0–5 | 0–10 | ||||||||
| 7.5 | 12.5 | ||||||||
| 10 | 15 | ||||||||
| Qwen2.5-VL (Checkpoint-400) | Qwen2.5-VL (Checkpoint-400) | ||||||||
| Ratio (%) | First CP | ASR (%) | FTR (%) | |
|---|---|---|---|---|
| Teacher Top- | ||||
| 10 | 2.5 | 100 | {\color[rgb]{0.125,0.1563,0.1875}100.00}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 0.00} | {\color[rgb]{0.125,0.1563,0.1875}100.00}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 0.00} |
| 50 | 2.5 | 100 | {\color[rgb]{0.125,0.1563,0.1875}99.97}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 0.06} | {\color[rgb]{0.125,0.1563,0.1875}99.43}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 0.64} |
| 100 | 5 | 100 | {\color[rgb]{0.125,0.1563,0.1875}97.33}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 1.57} | {\color[rgb]{0.125,0.1563,0.1875}57.70}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 15.56} |
| Student Top- | ||||
| 10 | 12.5 | 400 | {\color[rgb]{0.125,0.1563,0.1875}96.40}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 3.00} | {\color[rgb]{0.125,0.1563,0.1875}0.90}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 0.95} |
| ASR (%) | FTR (%) | |||
|---|---|---|---|---|
| Teacher Student | Unfiltered | Filtered | Unfiltered | Filtered |
| LLaVA-1.5 | {\color[rgb]{0.125,0.1563,0.1875}99.07}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 0.25} | {\color[rgb]{0.125,0.1563,0.1875}97.50}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 1.97} | {\color[rgb]{0.125,0.1563,0.1875}0.07}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 0.12} | {\color[rgb]{0.125,0.1563,0.1875}1.50}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 2.42} |
| LLaVA-1.5 | {\color[rgb]{0.125,0.1563,0.1875}99.80}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 0.35} | {\color[rgb]{0.125,0.1563,0.1875}98.77}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 1.57} | {\color[rgb]{0.125,0.1563,0.1875}0.27}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 0.23} | {\color[rgb]{0.125,0.1563,0.1875}0.17}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 0.29} |
| Qwen2.5-VL | {\color[rgb]{0.125,0.1563,0.1875}99.73}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 0.31} | {\color[rgb]{0.125,0.1563,0.1875}99.67}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 0.15} | {\color[rgb]{0.125,0.1563,0.1875}0.13}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 0.15} | {\color[rgb]{0.125,0.1563,0.1875}0.07}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 0.12} |
| Qwen2.5-VL | {\color[rgb]{0.125,0.1563,0.1875}72.80}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 7.37} | {\color[rgb]{0.125,0.1563,0.1875}78.07}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 7.71} | {\color[rgb]{0.125,0.1563,0.1875}0.37}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 0.25} | {\color[rgb]{0.125,0.1563,0.1875}1.27}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 1.02} |
| InternVL2.5 | {\color[rgb]{0.125,0.1563,0.1875}97.57}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 0.75} | {\color[rgb]{0.125,0.1563,0.1875}95.13}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 6.28} | {\color[rgb]{0.125,0.1563,0.1875}0.27}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 0.15} | {\color[rgb]{0.125,0.1563,0.1875}0.43}\,{\color[rgb]{0.5742,0.6094,0.6484}\scriptstyle\pm 0.32} |
Backdoor Attacks on VLMs
Backdoor attacks on vision-language models study hidden triggers that make multimodal systems behave maliciously while appearing normal.
See all 40 papers