Context-Shielded Backdoor Attack against Large Vision-Language Models

Summary

Researchers report that visual-token compression in large vision-language models can unmask a hidden backdoor. The authors identify context stripping, where pruning removes clean visual tokens that otherwise suppress an embedded trigger. They propose StripWake, a projector-only backdoor that stays at 0.4% attack success rate under full-token inference on LLaVA-1.5-7B, yet reaches 94.0%-99.7% across five primary compressors without CIDEr loss.

Topics

• Adversarial machine learning
• Multimodal models

Related items

• MediumCVE-2026-121936: easyVoice server-side request forgery through openaiBaseUrl in /api/v1/ttsSimilar attack · NVD/CVE Database
• MediumCVE-2026-118628: MLflow get-artifacts server-side request forgery information disclosureSimilar attack · NVD/CVE Database
• HighCVE-2026-118625: MLflow Tracking Server privilege escalation via get_endpointsSimilar attack · NVD/CVE Database
• InfoWhen trust becomes the payload: Inside the threats, etc quarterly reportSimilar attack · Elastic Security Labs
• MediumCritical Langflow flaws enable authentication bypass and RCESimilar attack · CSO Online