Mirage adversarial image perturbation achieves 100% bypass of CLIP safety checker on Stable Diffusion 1.5
Auto-published from arXiv:2610.30348 by the mine-arxiv pipeline. Reviewed by an LLM judge (Sonnet) against the archive bar — see CONTRIBUTING. Notes: cleared review (confidence 0.83, flags: [possible-zero-day, no-prompt-excerpt]) ⚠ needsDisclosureReview = true — review for undisclosed vulnerability before any public submission.
Category
multimodal
Model
Stable Diffusion 1.5
Surface
image-to-image generation pipeline with post-generation safety checker
Setup
Attacker holds a reference NSFW image as an adversarial target. A norm-bounded perturbation (L∞, ε=8/255) is optimized on a benign input image using a contrastive loss: ℒ_contras = -cos(y_adv, y_t) + λ∑cos(y_adv, y_NSFW^j), which simultaneously aligns the perturbed image embedding with the NSFW target and repels it from NSFW concept embeddings used by the safety checker. The perturbed benign image is then submitted to SD 1.5's image-to-image pipeline. Surrogate encoders ViT-H/14 and ViT-bigG/14 (LAION-2B pretrained) are used for optimization; the target safety checker (CLIP, ViT, or NudeNet) is treated as a black box.
Reproduction anchor
benchmark: NSFW Image Dataset (referenced from HuggingFace); I2P dataset (Inappropriate Image Prompts, top-100 most toxic descriptions). Paper states implementation code will be released but no URL provided as of publication.
Observed behavior
The CLIP safety checker passes all 100 perturbed generated images (100% bypass rate), and 84% of them are semantically aligned with the NSFW target (84% ASR). Average of 2.41 black-box queries needed. ViT Checker: 90% bypass, 80% ASR, 3.58 queries. NudeNet: 96% bypass, 84% ASR, 3.82 queries.
Expected behavior
Post-generation safety checkers should detect and withhold NSFW-content images before they reach the user, regardless of whether the input image is adversarially perturbed.
Reproducibility
medium
Threat model
A malicious user of a diffusion model API (e.g., a platform wrapping SD 1.5 with safety filtering) crafts a visually innocuous input image that causes the pipeline to generate and deliver targeted NSFW content. The platform's post-generation safety layer — the intended last line of defense — is fully bypassed without any access to its internals.
Novelty
First attack to exploit the shared CLIP-like representation space to jointly optimize for NSFW target alignment and NSFW concept repulsion, achieving near-perfect bypass of heterogeneous post-generation safety checkers with under 4 black-box queries.
Source
- arXiv: 2610.30348
- PDF: https://arxiv.org/pdf/2610.30348
- Categories: cs.CR
- Authors: Shanghao Shi, Chaoyu Zhang, Heng Jin, Hao Li, Ning Zhang, Y. Thomas Hou, Wenjing Lou
Triage notes (auto)
- paperType:
red-team-vuln - estimatedCaseCount: 3
- triage reason: Demonstrates reproducible jailbreak attacks against post-generation safety checkers in deployed diffusion models (SD 1.5, SDXL, Kandinsky 2.2) using imperceptible perturbations. Concrete attack methodology with evaluation across multiple models and datasets constitutes case-rich red-team research.
0 comments
─────────────────────────────────────────────────────────────────────
// no comments yet