Research map / Image generation / Text-to-image diffusion
Exposing Implicit Vulnerabilities in Text-to-Image Models via Adversarial Agentic Probing
Chang Ma, Junlin Han, Shuo Chen, Runjia Li, Philip Torr, Jindong Gu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Cumulative Adversarial Decoding (ablation) | AdvPIE w/o CAD · safe-sd-v2-1 sexual category (Gemma) | ASR (%): 24.50 -> 37.67 | Improved |
| add Cumulative Adversarial Decoding (ablation) | AdvPIE w/o CAD · safe-sd-v2-1 sexual category (LLaVA Guard) | ASR (%): 25.17 -> 36.67 | Improved |
| add Cumulative Adversarial Decoding (ablation) | AdvPIE w/o CAD · safe-sd-v2-1 sexual category | bypass rate (%): 45.17 -> 54.83 | Improved |
| add global-relative structured feedback (ablation) | AdvPIE w/o feedback · safe-sd-v2-1 sexual category (Gemma) | ASR (%): 27.33 -> 37.67 | Improved |
| add global-relative structured feedback (ablation) | AdvPIE w/o feedback · safe-sd-v2-1 sexual category | bypass rate (%): 48.33 -> 54.83 | Improved |
| add safety reasoning rationale to feedback (ablation) | AdvPIE w/o reasoning · safe-sd-v2-1 sexual category (Gemma) | ASR (%): 30.50 -> 37.67 | Improved |
| add score-based guidance (top-k memory) to feedback (ablation) | AdvPIE w/o guidance · safe-sd-v2-1 sexual category (Gemma) | ASR (%): 21.83 -> 37.67 | Improved |
| add score-based guidance (top-k memory) to feedback (ablation) | AdvPIE w/o guidance · safe-sd-v2-1 sexual category | bypass rate (%): 37.83 -> 54.83 | Improved |
| add score-based guidance to feedback relative to no feedback at all (ablation) | AdvPIE w/o guidance · safe-sd-v2-1 sexual category (Gemma) | ASR (%): 27.33 -> 21.83 | Hurt |
| add safety reasoning rationale to feedback (ablation) | AdvPIE w/o reasoning · safe-sd-v2-1 sexual category | bypass rate (%): 47.17 -> 54.83 | Improved |
| use AdvPIE instead of ART | ART · Stable Diffusion API (Gemma judge) | ASR (%): 40.14 -> 50.45 | Improved |
| use AdvPIE instead of FLIRT | FLIRT · safe-sd-v1-5 (Gemma judge) | ASR (%): 33.26 -> 30.45 | Hurt |
Closest papers by meaning
- VERA-V: Variational Inference Framework for Jailbreaking Vision-Language Models (ICML 2026 · similarity 0.88)
- ART: Automatic Red-teaming for Text-to-Image Models to Protect Benign Users (NeurIPS 2024 · similarity 0.87)
- Prompting4Debugging: Red-Teaming Text-to-Image Diffusion Models by Finding Problematic Prompts (ICML 2024 · similarity 0.87)
- Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models (CVPR 2025 · similarity 0.87)
- UPAM: Unified Prompt Attack in Text-to-Image Generation Models Against Both Textual Filters and Visual Checkers (ICML 2024 · similarity 0.86)
- Discovering Failure Modes of Text-guided Diffusion Models via Adversarial Search (ICLR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
