MultiTrust: A Comprehensive Benchmark Towards Trustworthy Multimodal Large Language Models

TL;DR

MultiTrust introduces the first comprehensive benchmark for evaluating trustworthiness in Multimodal Large Language Models, covering truthfulness, safety, robustness, fairness, and privacy via 32 diverse tasks and 21 models. It combines a two-level taxonomy with a scalable toolbox to assess multimodal risks and cross-modal impacts, revealing substantial gaps in open-source MLLMs and the persistent benefits of strong alignment in proprietary systems. Key findings include susceptibility to visual context, adversarial and jailbreaking attacks, and bias propagation, underscoring the need for improved multimodal alignment, dynamic evaluation, and principled safety mechanisms. The toolbox and large-scale analyses aim to standardize and accelerate future research toward trustworthy MLLMs in real-world, multimodal deployments.

Abstract

Despite the superior capabilities of Multimodal Large Language Models (MLLMs) across diverse tasks, they still face significant trustworthiness challenges. Yet, current literature on the assessment of trustworthy MLLMs remains limited, lacking a holistic evaluation to offer thorough insights into future improvements. In this work, we establish MultiTrust, the first comprehensive and unified benchmark on the trustworthiness of MLLMs across five primary aspects: truthfulness, safety, robustness, fairness, and privacy. Our benchmark employs a rigorous evaluation strategy that addresses both multimodal risks and cross-modal impacts, encompassing 32 diverse tasks with self-curated datasets. Extensive experiments with 21 modern MLLMs reveal some previously unexplored trustworthiness issues and risks, highlighting the complexities introduced by the multimodality and underscoring the necessity for advanced methodologies to enhance their reliability. For instance, typical proprietary models still struggle with the perception of visually confusing images and are vulnerable to multimodal jailbreaking and adversarial attacks; MLLMs are more inclined to disclose privacy in text and reveal ideological and cultural biases even when paired with irrelevant images in inference, indicating that the multimodality amplifies the internal risks from base LLMs. Additionally, we release a scalable toolbox for standardized trustworthiness research, aiming to facilitate future advancements in this important field. Code and resources are publicly available at: https://multi-trust.github.io/.

Figures (66)

• Figure 1: Framework of MultiTrust, including aspect division, evaluation strategy and design of the developed toolbox. Specifically, we study the trustworthiness by delving into the multimodal nature of MLLMs from a broader perspective, covering both multimodal risks and cross-modal impacts.
• Figure 2: Left: Rankings of MLLMs in each sub-aspect of MultiTrust. Right: (a) Correlation between the overall rankings of trustworthiness and those of general capabilities based on MMBench liu2023mmbench and MME fu2023mme. Top-8 are marked. (b) Pearson Correlation Coefficients between sub-aspects.
• Figure 3: ASRs (%, $\downarrow$) in Task \ref{['sec:robustness_untargeted']}/\ref{['sec:robustness_targeted']}.
• Figure 4: RtA rate (%, $\uparrow$) in Task \ref{['sec:vision_preference_selection']}.
• Figure B.1: The pool of irrelevant images for evaluating the cross-modal impacts.