CAREER: Empowering White-box Driven Analytics to Detect AI-synthesized Deceptive Content

该奖项全部或部分由《2021年美国救援计划法案》（公法117-2）资助。人工智能（AI）合成技术可以自动生成逼真的图像、视频和其他内容，在过去几年里有了显著的进步。尽管这些技术有很好的合法应用，但它们也带来了严重的信任和安全威胁。网络犯罪分子越来越多地使用人工智能合成技术来欺骗用户和操纵意见，而无需大量投入人工内容生成。例如，人工智能合成的个人资料照片被滥用来创建虚假账户，而模拟真人的深度假视频可以让网络犯罪分子有能力诽谤或冒充他人。现有的检测工作主要依赖于分析内容的“黑盒”方法，而没有考虑人工智能合成技术的工作方式。该项目的目标是使用考虑技术如何工作的“白盒”方法，既可以系统地检测人工智能合成的内容，又可以概述广泛的人工智能合成算法工作原理的一般原则，这将有助于检测算法适应新的合成技术的开发。这项研究的结果将加强用户对在线内容的信任，并帮助社交媒体网站和其他互联网平台减少人工智能合成内容的欺骗。项目团队将把本研究中开发的新数据集和技术整合到本科和研究生课程以及在线练习中，以培训未来的网络安全工作者。该小组还将支持研究的多样化参与，积极招募和指导妇女和其他代表性不足群体的人。本研究旨在提高人工智能综合检测的有效性、通用性和鲁棒性。这项工作的重点是检测人工智能合成的图像和视频，因为人类更容易被视觉内容吸引和欺骗。开发的分析原则有望激发这些领域的新工作，并扩展到检测其他类型的人工智能合成内容。该项目围绕三个研究重点组织。首先，该团队将开发一个统一的分析框架，系统地剖析人工智能合成模型，并深入了解模型中常见的合成模式。其次，基于这些发现，团队将设计基于频率和像素域的通用方法，以有效检测人工智能合成的图像和视频并进行大规模操作。第三，它将通过主动调查对抗性规避策略和优先考虑抵抗这些策略的检测技术来增强检测鲁棒性。该框架和开发的技术将通过大规模的真实世界数据进行全面评估。这项研究将有助于建立一个原则性的检测范式，并为战胜未来基于人工智能的欺骗和宣传形式提供见解。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This award is funded in whole or in part under the American Rescue Plan Act of 2021 (Public Law 117-2).Artificial intelligence (AI) synthesis techniques that automatically produce realistic images, videos, and other content have significantly improved over the past few years. Although there are promising legitimate applications of these techniques, they also raise serious trust and security threats. Cybercriminals increasingly weaponize AI synthesis techniques to deceive users and manipulate opinions without having to invest heavily in manual content generation. For instance, AI-synthesized profile photographs are abused to create fake accounts, while deepfake videos that simulate real people can give cybercriminals the ability to defame or impersonate others. Existing detection work mostly relies on "black-box" approaches that analyze content without considering the way the AI synthesis techniques work. This project's goal is to use "white-box" methods that consider how the techniques work, both to systematically detect AI-synthesized content, and to outline general principles that underlie how broad classes of AI synthesis algorithms work that will help detection algorithms adapt as new synthesis techniques are developed. The results of this research will reinforce user trust in online content and help social media sites and other Internet platforms mitigate deception through AI-synthesized content. The project team will integrate the new datasets and techniques developed in this research into undergraduate and graduate courses as well as online exercises to train future cybersecurity workers. The team will also support diverse participation in the research, actively recruiting and mentoring women and people from other under-represented groups.This research aims to advance AI synthesis detection in terms of efficacy, generalizability, and robustness. The work focuses on detecting AI-synthesized images and videos, as humans are more likely to be attracted to and deceived by visual content. The developed analytics principles are envisioned to inspire new work in these areas and expand to detection of other types of AI-synthesized content. The project is organized around three research thrusts. First, the team will develop a unified analytic framework to systematically dissect AI-synthesis models and gain deep understanding of synthesis patterns common across the models. Second, based on these findings, the team will design generalizable approaches based on the frequency and pixel domains to efficiently detect AI-synthesized images and videos and operate at scale. Third, it will enhance detection robustness by proactively investigating adversarial evasion strategies and prioritizing detection techniques resistant to those strategies. The framework and the developed techniques will be thoroughly evaluated with large-scale real-world data. This research will contribute to establishing a principled detection paradigm and provide insights to prevail over future forms of AI-based deception and propaganda.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.