开放式多模态AI安全评测基准 OpenMMSec

随着大模型（Deepseek、GPT、LLaMA等）的爆发式应用，AI技术已深度融入金融、医疗、智能终端语音交互场等核心领域，但用户所遭受的深度伪造身份攻击也随之浮出水面。为应对这一全球性挑战，由中国图象图形学学会、蚂蚁集团、云安全联盟CSA大中华区主办，广泛联合学界、机构共同组织发起2025全球AI攻防挑战赛。
本次大赛聚焦生活、金融、健康场景下数字身份AI大模型智能交互认证安全，共分为图、视、音三大场景及六个子赛题，并以攻防闭环的方式设置赛程突破静态攻防格局，通过赛事我们积累了百万量级的高价值样本数据，蚂蚁安全与智能实验室开源首个「开放式多模态AI安全基准数据集(OpenMMSec)」，希望对外打造国际领先的、覆盖全面的多模态AI模型安全对抗性攻防基准数据集，从而成为推动安全技术进化、筑牢可信AI生态的关键力量。

With the explosive proliferation of large language models (LLMs, including Deepseek, GPT, LLaMA, etc.), AI technology has been deeply integrated into core domains including finance, healthcare, and intelligent terminal voice interaction scenarios. Meanwhile, deepfake identity attacks targeting end users have emerged as a concomitant threat alongside this development.

To address this global challenge, the 2025 Global AI Attack and Defense Challenge was jointly organized and launched, hosted by the Chinese Society of Image and Graphics (CSIG), Ant Group, and Cloud Security Alliance (CSA) Greater China Chapter, in partnership with extensive academic circles and institutional stakeholders. This competition focuses on the security of intelligent interactive authentication for digital identities enabled by AI large models across daily life, finance, and healthcare scenarios. It encompasses three core scenarios (image, video, and audio) and six sub-challenges. The competition schedule is structured in a closed-loop attack and defense framework to break through the limitations of static attack and defense paradigms. Through this event, we have accumulated millions of high-value sample datasets. The Ant Security and Intelligence Lab has open-sourced the first "Open Multimodal AI Security Benchmark Dataset (OpenMMSec)". We aim to develop and release to the global community an internationally leading, comprehensively covered multimodal AI model security adversarial attack and defense benchmark dataset, thereby serving as a key driving force for the evolution of security technologies and consolidating a trustworthy AI ecosystem.

搜集汇总

背景与挑战

背景概述

OpenMMSec是一个开放式多模态AI安全评测基准数据集，包含图像、视频和音频三种模态的数据，用于评估AI模型在对抗性攻击下的安全性能。数据集来源于2025全球AI攻防挑战赛，覆盖生活、金融、健康场景下的数字身份认证安全，包含百万量级的高价值样本数据。