AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

"Eyes Closed, Safety On: Protecting Multimodal LLMs via Image-to-Text Transformation"

Yunhao Gou, Kai Chen, Zhili LIU, Lanqing Hong, Hang Xu, Zhenguo Li, Dit-Yan Yeung, James Kwok, Yu Zhang

ECCV 2024 · Vision-language and multimodal › Multimodal LLMs · approach: Benchmarks and analysis of MLLMs · 63 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
remove images when querying MLLMs with malicious text (ablation)five MLLMs · VLSafeharmless rate: about 20% -> nearly 100%Improved
apply ECSO (self-check, query-aware caption, answer without image)LLaVA-1.5-7B · MM-SafetyBench SD+OCRharmless rate: 32.1 -> 86.4Improved
apply ECSO on SD images where Direct is already fairly safeLLaVA-1.5-7B · MM-SafetyBench SDharmless rate: 85.0 -> 95.4Improved
apply ECSOfive MLLMs · MME, MM-Vet, MMBench utilityutility: e.g. LLaVA-1.5-7B MMBench 64.6 -> 64.2; ShareGPT4V MME-C 376.4 -> 380.7No clear effect
keep image features in addition to the query-aware caption at step 4 (ablation)ECSO · MM-SafetyBench SD+OCR and VLSafeharmless rateHurt
use generic caption instead of query-aware caption (ablation)ECSO on ShareGPT4V-7B · MME-P and MMBenchutility: MME-P 1567.1 -> 1562.7; MMBench 66.1 -> 65.8Hurt
add HARMLESS and ETHICAL safety instruction to ECSO final step (ablation)ECSO · VLSafeharmless rate: 71.3 -> 91.8 for LLaVA-1.5-7BImproved
detect harm from the model response instead of image and query (ablation)ECSO step 2 · 1000 labelled responsesdetection accuracy: 75.6 -> 95.0 for LLaVA-1.5-7BImproved
use original LLM instead of finetuned LLM in step 4 (ablation)ECSO · VLSafeharmless rate: 91.8 -> 93.3 for LLaVA-1.5-7BImproved
SFT on ECSO-generated responses instead of ground-truth VLGuard responses (ablation)LLaVA-1.5-7B · MM-SafetyBench SD+OCRharmless rate / utility: utility 1679.7 -> 1727.2Improved
combine finetuning with ECSO at inference (ablation)LLaVA-1.5-7B finetuned on VLGuard · MM-SafetyBench SD+OCRharmless rate: 73.68 -> 89.18Improved

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