SafeBench

该数据集是根据OpenAI和Meta的使用政策精心策划的一个基准，旨在评估针对视觉-语言模型攻击的有效性。它包含了常见的安全主题和场景，在这些场景中，用户被禁止利用这些模型。数据集的规模涵盖了从三个视觉-语言模型家族生成的46,500个模型响应。其任务是评估针对视觉-语言模型的越狱攻击的有效性。This dataset is a carefully curated benchmark developed in accordance with the usage policies of OpenAI and Meta, with the general goal of evaluating the effectiveness of attacks against vision-language models (VLMs). It covers common security-related topics and scenarios where users are prohibited from exploiting these models. The dataset comprises 46,500 model responses generated from three distinct vision-language model families. Its specific task is to assess the effectiveness of jailbreak attacks against VLMs.

搜集汇总

背景与挑战

背景概述

SafeBench是一个针对视觉-语言模型的安全评估基准数据集，基于OpenAI和Meta的政策精心策划，涵盖常见安全主题和禁止场景。它包含从三个模型家族生成的46,500个响应，主要用于评估越狱攻击的有效性，帮助测试模型的安全性和抗攻击能力。