AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Images are Achilles' Heel of Alignment: Exploiting Visual Vulnerabilities for Jailbreaking Multimodal Large Language Models

Yifan Li, hangyu guo, Kun Zhou, Wayne Xin Zhao, Ji-Rong Wen

ECCV 2024 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 115 citations (OpenAlex)

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 changedWhereMeasuredEffect
add a blank image to a harmful text instructionLLaVA-1.5 · Harmful-instruction datasetaverage ASR (%): +25.20Hurt
replace blank image by a harmful (toxic) imageLLaVA-1.5 · Harmful-instruction datasetaverage ASR (%): 54.13 -> 68.13Hurt
fine-tune more backbone parameters (full vs LoRA)LLaVA-1.5 · Toxic settingaverage ASR (%): 62.80 -> 68.13Hurt
add text-to-image pointer to typography image (ablation)HADES · HADES datasetaverage ASR (%): LLaVA-1.5 -7.47, LLaVA-1.5 LoRA -14.13, LLaVA +6.53, Gemini +16.93, GPT-4V +1.47Mixed
add optimized harmful image generated by diffusion with LLM prompt refinement (ablation)HADES · HADES datasetaverage ASR (%): LLaVA +33.86, Gemini Pro Vision +37.20, GPT-4V +13.34Improved
add gradient-optimized adversarial image on top (ablation)HADES on LLaVA-1.5 · HADES datasetaverage ASR (%): 78.53 -> 90.26Improved
more LLM-driven prompt optimization steps for harmful image generation (ablation)HADES · HADES datasetASRImproved
HADES instead of adversarial-image or compositional jailbreak baselinesLLaVA-1.5 · Violence/Animal/Self-Harm categoriesASR (%): Animal 74.67 -> 83.33; Self-Harm 80.67 -> 89.33Improved
fine-tune with contrastive harmlessness LoRA on OCR and captioning refusalsLLaVA-1.5 · HADES test splitaverage ASR (%): 89.53 -> 5.07Improved
add contrastive harmlessness LoRALLaVA-1.5 · LLaVA-Benchscore: 63.40 -> 57.90Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy