arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2401.12915cs.AIcs.CLcs.CV
收藏

对视觉语言模型进行红队测试

Red Teaming Visual Language Models

The University of Hong Kong(香港大学);Zhejiang University(浙江大学)

Mukai Li, Lei Li, Yuwei Yin, Masood Ahmed, Zhenguang Liu, Qi Liu

AI总结:

本文构建涵盖忠实性、隐私、安全性与公平性的红队测试数据集 RTVLM,系统评估 10 个开源 VLM,并证明基于该数据集的 SFT 对齐可显著提升 LLaVA-v1.5。

摘要:

VLM(视觉语言模型)扩展了 LLM(大语言模型)的能力,使其能够接受多模态输入。由于已经证实,LLM 可能通过特定测试用例(称为红队测试)被诱导生成有害或不准确内容,因此 VLM 在类似场景中表现如何,尤其是在文本与视觉输入相结合的情况下,仍是一个问题。为探究该问题,我们提出了一个新颖的红队测试数据集 RTVLM,它涵盖 4 个主要方面(忠实性、隐私、安全性、公平性)下的 10 个子任务(例如图像误导、多模态越狱、面部公平性等)。我们的 RTVLM 是首个从这 4 个不同方面对当前 VLM 进行基准测试的红队测试数据集。详细分析表明,10 个重要的开源 VLM 都在不同程度上难以应对红队测试,并且与 GPT-4V 之间存在最高达 31% 的性能差距。此外,我们直接使用 RTVLM 通过监督微调(SFT)对 LLaVA-v1.5 进行红队测试对齐,这使模型在 RTVLM 测试集上的性能提升 10%,在 MM-Hal 上提升 13%,并且在 MM-Bench 上没有明显下降,超过了其他使用常规对齐数据的基于 LLaVA 的模型。这表明当前开源 VLM 仍然缺乏红队测试对齐。我们的代码和数据集将开源。

英文摘要:

VLMs (Vision-Language Models) extend the capabilities of LLMs (Large Language Models) to accept multimodal inputs. Since it has been verified that LLMs can be induced to generate harmful or inaccurate content through specific test cases (termed as Red Teaming), how VLMs perform in similar scenarios, especially with their combination of textual and visual inputs, remains a question. To explore this problem, we present a novel red teaming dataset RTVLM, which encompasses 10 subtasks (e.g., image misleading, multi-modal jail-breaking, face fairness, etc) under 4 primary aspects (faithfulness, privacy, safety, fairness). Our RTVLM is the first red-teaming dataset to benchmark current VLMs in terms of these 4 different aspects. Detailed analysis shows that 10 prominent open-sourced VLMs struggle with the red teaming in different degrees and have up to 31% performance gap with GPT-4V. Additionally, we simply apply red teaming alignment to LLaVA-v1.5 with Supervised Fine-tuning (SFT) using RTVLM, and this bolsters the models' performance with 10% in RTVLM test set, 13% in MM-Hal, and without noticeable decline in MM-Bench, overpassing other LLaVA-based models with regular alignment data. This reveals that current open-sourced VLMs still lack red teaming alignment. Our code and datasets will be open-source.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑