Visual backdoors grow as student model rollouts continue
As Student Rollouts Continue, Visual Backdoors Keep Growing
The paper finds that in On-Policy Distillation (OPD) post-training of vision-language models (VLMs), visual backdoors can propagate without explicit malicious responses, with attack success rate rapidly approaching 100% once the poisoning ratio exceeds a critical threshold. OPD requires no teacher-generated responses and supervises only the student's online-generated responses with the teacher's probability distribution, and thus appears naturally resistant to backdoor propagation; the paper challenges this intuition through experiments across multiple model families and datasets and through theoretical analysis. Backdoor propagation is driven by the continuous accumulation of anomalous probability mass assigned by the teacher; unlike the more gradual propagation in offline distillation, OPD rarely exhibits backdoors at low poisoning ratios. Based on these findings, the authors propose a new trigger design that achieves visual backdoor propagation with zero explicit trigger injection. The authors say these findings reveal a neglected backdoor propagation mechanism in OPD and highlight its potential security risks.