Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MMDT: Decoding the Trustworthiness and Safety of Multimodal Foundation Models

Chejian Xu, Jiawei Zhang, Zhaorun Chen, Chulin Xie, Mintong Kang, Yujin Potter, Zhun Wang, Zhuowen Yuan, Alexander Xiong, Zidi Xiong, Chenhui Zhang, Lingzhi Yuan, Yi Zeng, Peiyang Xu, Chengquan Guo, Andy Zhou, Jeffrey Ziwei Tan, Xuandong Zhao, Francesco Pinto, Zhen Xiang, Yu Gai, Zinan Lin, Dan Hendrycks, Bo Li, Dawn Song

TL;DR

MMDT introduces a unified, modular platform to systematically evaluate multimodal foundation models across six trustworthiness dimensions: safety, hallucination, fairness, privacy, adversarial robustness, and out-of-distribution generalization. Through comprehensive red-teaming, scenario-based benchmarks, and cross-modal evaluations (T2I and I2T), the work reveals widespread vulnerabilities and varied strengths across leading MMFMs, with no model excelling uniformly across all perspectives. The platform provides actionable insights and mitigation strategies, from taxonomy-driven safety controls and bias-aware tuning to privacy-preserving training and robust data augmentation for OOD resilience. Overall, MMDT advances rigorous, multi-faceted assessment of MMFMs to guide safer deployment in safety-critical domains and to inform future mitigation research.

Abstract

Multimodal foundation models (MMFMs) play a crucial role in various applications, including autonomous driving, healthcare, and virtual assistants. However, several studies have revealed vulnerabilities in these models, such as generating unsafe content by text-to-image models. Existing benchmarks on multimodal models either predominantly assess the helpfulness of these models, or only focus on limited perspectives such as fairness and privacy. In this paper, we present the first unified platform, MMDT (Multimodal DecodingTrust), designed to provide a comprehensive safety and trustworthiness evaluation for MMFMs. Our platform assesses models from multiple perspectives, including safety, hallucination, fairness/bias, privacy, adversarial robustness, and out-of-distribution (OOD) generalization. We have designed various evaluation scenarios and red teaming algorithms under different tasks for each perspective to generate challenging data, forming a high-quality benchmark. We evaluate a range of multimodal models using MMDT, and our findings reveal a series of vulnerabilities and areas for improvement across these perspectives. This work introduces the first comprehensive and unique safety and trustworthiness evaluation platform for MMFMs, paving the way for developing safer and more reliable MMFMs and systems. Our platform and benchmark are available at https://mmdecodingtrust.github.io/.

MMDT: Decoding the Trustworthiness and Safety of Multimodal Foundation Models

TL;DR

MMDT introduces a unified, modular platform to systematically evaluate multimodal foundation models across six trustworthiness dimensions: safety, hallucination, fairness, privacy, adversarial robustness, and out-of-distribution generalization. Through comprehensive red-teaming, scenario-based benchmarks, and cross-modal evaluations (T2I and I2T), the work reveals widespread vulnerabilities and varied strengths across leading MMFMs, with no model excelling uniformly across all perspectives. The platform provides actionable insights and mitigation strategies, from taxonomy-driven safety controls and bias-aware tuning to privacy-preserving training and robust data augmentation for OOD resilience. Overall, MMDT advances rigorous, multi-faceted assessment of MMFMs to guide safer deployment in safety-critical domains and to inform future mitigation research.

Abstract

Multimodal foundation models (MMFMs) play a crucial role in various applications, including autonomous driving, healthcare, and virtual assistants. However, several studies have revealed vulnerabilities in these models, such as generating unsafe content by text-to-image models. Existing benchmarks on multimodal models either predominantly assess the helpfulness of these models, or only focus on limited perspectives such as fairness and privacy. In this paper, we present the first unified platform, MMDT (Multimodal DecodingTrust), designed to provide a comprehensive safety and trustworthiness evaluation for MMFMs. Our platform assesses models from multiple perspectives, including safety, hallucination, fairness/bias, privacy, adversarial robustness, and out-of-distribution (OOD) generalization. We have designed various evaluation scenarios and red teaming algorithms under different tasks for each perspective to generate challenging data, forming a high-quality benchmark. We evaluate a range of multimodal models using MMDT, and our findings reveal a series of vulnerabilities and areas for improvement across these perspectives. This work introduces the first comprehensive and unique safety and trustworthiness evaluation platform for MMFMs, paving the way for developing safer and more reliable MMFMs and systems. Our platform and benchmark are available at https://mmdecodingtrust.github.io/.

Paper Structure

This paper contains 86 sections, 4 equations, 18 figures, 54 tables.

Table of Contents

  1. Introduction
  2. MMDT-Safety: Safety
  3. MMDT-Hallucination: Hallucination
  4. MMDT-Fairness: Fairness
  5. MMDT-Privacy: Privacy
  6. MMDT-Adv: Adversarial Robustness
  7. MMDT-OOD: Out-of-Distribution Robustness
  8. Conclusion
  9. Preliminaries
  10. Introduction to multi-modal foundation models
  11. Multi-modal foundation models evaluated in this paper
  12. Text-to-image models
  13. Image-to-text models
  14. MMDT Platform Design
  15. Additional Discussion on Evaluation Results
  16. ...and 71 more sections

Figures (18)

  • Figure 1: Examples of unreliable responses of MMFMs on different trustworthiness perspectives.
  • Figure 2: Examples of harmful responses from MMFMs under different safety scenarios.
  • Figure 3: Examples of hallucinated responses from MMFMs under different scenarios. The examples are sampled from various models to demonstrate the prevalent hallucinations across different models.
  • Figure 4: Architecture of the MMDT Platform. The platform consists of modular components for benchmark orchestration, configuration, inference runtimes, and results analysis, ensuring scalability and extensibility.
  • Figure 5: A tree taxonomy of different perspectives of trustworthiness that our benchmark focuses on.
  • ...and 13 more figures