Standard preprocessing defenses (JPEG, Gaussian blur, median filter, random masking) leave SISS ASR statistically unchanged at ~68%
Auto-published from arXiv:2611.16744 by the mine-arxiv pipeline. Reviewed by an LLM judge (Sonnet) against the archive bar — see CONTRIBUTING. Notes: cleared review (confidence 0.77, flags: [no-prompt-excerpt, possible-zero-day])
Category
multimodal
Model
Qwen2.5-VL-7B-Instruct (defense evaluation surrogate)
Surface
API (VLM image input with upstream preprocessing)
Setup
A fixed 100-image defense set (10 payloads × 10 background prompts, step-600 LoRA checkpoint, control scale 1.5, seed 42) is evaluated under five generic input-level preprocessing defenses applied before VLM inference: (1) area resampling, (2) 3×3 median reconstruction, (3) JPEG quality-75 + area resampling, (4) Gaussian low-pass σ=1, and (5) randomized masking with 5 replicas + majority aggregation. Evaluation uses Qwen2.5-VL-7B-Instruct at 512×512 and after Lanczos→64×64 downsampling. No verbatim prompt excerpt for the defense eval is in the body; judging prompt is in Supplementary.
Reproduction anchor
Defense protocol: 100-image split (§IV-E); primary judge: Qwen2.5-32B three-level rubric (scores ≥0.5 count as payload detection); exact McNemar paired tests reported; judging prompt in Supplementary. No public code URL found in body.
Observed behavior
All five generic defenses leave ASR statistically unchanged: defended ASR ranges 67–69% vs. 68% undefended; all paired McNemar tests yield p=1.0. ScaleGuard (multi-scale consistency detector, 9 low-resolution probes across 3 sizes and 3 kernels) is the only evaluated defense that achieves a significant reduction, lowering non-adaptive ASR from 61.4% to 24.3% (p=2.98×10⁻⁸) with zero clean false positives among 70 controls.
Expected behavior
Preprocessing transformations (compression, blurring, random masking) should disrupt the payload encoding and reduce attack success; defenders should be able to apply standard image hardening to mitigate visual prompt injection.
Reproducibility
medium
Threat model
Security engineers deploying VLM pipelines who rely on standard image preprocessing (JPEG re-encode, resizing, noise addition) as a mitigation for visual prompt injection will find these controls entirely ineffective against SISS-style attacks, leaving production systems silently vulnerable.
Novelty
Empirically demonstrates that intuitive input-level mitigations for visual prompt injection provide no measurable protection against cross-resolution semantic shift attacks, while a purpose-built multi-scale consistency check (ScaleGuard) achieves significant reduction — motivating a new class of scale-aware defenses.
Source
- arXiv: 2611.16744
- PDF: https://arxiv.org/pdf/2611.16744
- Categories: cs.AI
- Authors: Yiheng Zhao, Runze Ji, Chenyu Yi, Yi Yu, Bingquan Shen, Alex Kot, Xudong Jiang
Triage notes (auto)
- paperType:
red-team-vuln - estimatedCaseCount: 3
- triage reason: Demonstrates a concrete visual prompt injection vulnerability in VLMs where hidden instructions become recoverable when images are downsampled—a reproducible failure across 4 closed-source models with 300-case benchmark and quantified attack success rates (58.75% ASR). No explicit disclosure to vendors mentioned.
0 comments
─────────────────────────────────────────────────────────────────────
// no comments yet