MMDT: Decoding the Trustworthiness and Safety of Multimodal Foundation Models

2025-03-19

AI Summary

This paper addresses critical trustworthiness deficiencies—spanning safety, fairness, privacy, adversarial robustness, and out-of-distribution (OOD) generalization—in multimodal foundation models (MMFMs). To this end, we propose MMDT, the first unified evaluation framework for MMFM trustworthiness. MMDT systematically assesses six dimensions: safety, hallucination, fairness, privacy, adversarial robustness, and OOD generalization. It introduces novel methodologies including task-adaptive red-teaming attacks, construction of high-difficulty multi-scenario benchmarks, cross-modal bias quantification, and multimodal consistency analysis. Comprehensive evaluation using MMDT uncovers pervasive security vulnerabilities across state-of-the-art MMFMs. We provide a fully reproducible evaluation protocol and an open-source platform (mmdecodingtrust.github.io), thereby filling a fundamental gap in trustworthy MMFM assessment. MMDT advances multimodal AI toward verifiable, auditable, and reliable systems.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Philosophy and Ethics of AI: Safety, Robustness & Trustworthiness
• Computer Vision: Multi-modal Vision

Application Category

• Security and Privacy: Large-scale security measurements
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

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/.

Problem

• Evaluates safety and trustworthiness of multimodal foundation models
• Identifies vulnerabilities like unsafe content generation and biases
• Provides a unified platform for comprehensive model assessment

Innovation

• Unified platform for multimodal model evaluation
• Comprehensive safety and trustworthiness assessment
• Red teaming algorithms for vulnerability detection