Research map / Vision-language and multimodal / Multimodal LLMs
MultiTrust: A Comprehensive Benchmark Towards Trustworthy Multimodal Large Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces MultiTrust, a unified benchmark of 32 tasks across five trustworthiness aspects (truthfulness, safety, robustness, fairness, privacy) that evaluates 21 MLLMs including cross-modal impact tests and a modular toolbox.
Problem
Existing evaluations of MLLM trustworthiness are narrow, covering only one or a few aspects and phenomena, and largely ignore cross-modal interactions and lack standardized, scalable tooling.
Approach
A holistic benchmark organized by 5 aspects and 10 sub-aspects that tests both multimodal risks (new image-based threats) and cross-modal impacts (how adding relevant or irrelevant images changes text-only task behavior), backed by a unified, modular evaluation toolbox.
Key results
- Proprietary GPT-4V and Claude3 show consistently top trustworthiness; GPT-4V and Claude3 refuse 69.5% of NSFW images on average while most open-source models reject none.
- Under untargeted adversarial attack, image captioning accuracy of most models drops below 20% from more than 90%.
- Correlation of 0.60 between general capability and trustworthiness rankings across 21 MLLMs; GPT-4 GPT-Score correlates 0.91 with human ratings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Pair text-only PII leakage prompts with relevant images | LLaVA-NeXT · PII Leakage in Conversations | RtA (%): 51.00 -> 1.50 | Hurt |
| Pair text-only PII leakage prompts with irrelevant images | LLaVA-NeXT · PII Leakage in Conversations | RtA (%): 51.00 -> 25.67 | Hurt |
| Pair text-only PII leakage prompts with relevant images | MiniGPT-4-L2 · PII Leakage in Conversations | RtA (%): 100.00 -> 24.00 | Hurt |
| Pair text-only PII leakage prompts with irrelevant images | MiniGPT-4-L2 · PII Leakage in Conversations | RtA (%): 100.00 -> 66.78 | Hurt |
| Fine-tune the aligned Llama2 base LLM during multimodal training rather than keeping it frozen (ablation) | Llama2-based MLLMs · Trustworthiness tasks | trustworthiness | Hurt |
| Use a vision encoder with more parameters | MLLMs · Image Captioning under Adversarial Attack | attack success rate | Helped |
Surprising findings
- Only placing harmful behaviors in images without deliberate prompts jailbreaks many MLLMs, e.g., 71% for InternLM-XC2 and 80% for LLaVA-1.5.
- Multimodal training of Llama2-based MiniGPT-4 (frozen LLM) preserves privacy protection better than mPLUG-Owl2 which fine-tunes its LLM.
Limitations
- Machine ethics excluded due to cultural and subjective variability.
- Limited analysis of closed-source models due to unknown training details.
- Some dataset parts could be maliciously misused.
Experimental setup
Datasets: AMBER MME Ferret RTVLM SVT ImageNet AdvGLUE AdvGLUE++ SST-2 COCO-O RealToxicityPrompts POPE
Benchmarks: MultiTrust
Metrics: accuracy ASR RtA rate toxicity score Pearson correlation p-value GPT-Score Leakage Rate
Baselines: GPT-4V Claude3 Gemini-Pro Qwen-VL-Plus LLaVA-1.5-7B LLaVA-1.5-13B MiniGPT-4-L2 mPLUG-Owl2 InternLM-XC2 InternVL-Chat
Builds on: DecodingTrust TrustLLM POPE MM-SafetyBench AdvGLUE COCO-O
Closest papers by meaning
- MLLMGuard: A Multi-dimensional Safety Evaluation Suite for Multimodal Large Language Models (NeurIPS 2024 · similarity 0.91)
- MoHoBench: Assessing Honesty of Multimodal Large Language Models via Unanswerable Visual Questions (AAAI 2026 · similarity 0.90)
- VMDT: Decoding the Trustworthiness of Video Foundation Models (NeurIPS 2025 · similarity 0.90)
- Position: TrustLLM: Trustworthiness in Large Language Models (ICML 2024 · similarity 0.89)
- When Understanding Becomes a Risk: Authenticity and Safety Risks in the Emerging Image Generation Paradigm (CVPR 2026 · similarity 0.88)
- Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models? (CVPR 2025 · similarity 0.88)
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.