Collaborative Research: SaTC: TTP: Small: DeFake: Deploying a Tool for Robust Deepfake Detection

深度造假——由人工智能生成或操纵的视频——对传播虚假信息、威胁勒索和新形式的网络钓鱼构成了重大威胁。它们已经被广泛用于制作未经双方同意的色情作品，并开始被用来破坏政府和选举。就连深度造假的威胁也让人们对新闻中视频的真实性产生了怀疑。记者在核实信息方面发挥着关键作用，他们尤其需要帮助来应对不断改进的深度造假技术。最近在检测深度伪造方面的结果很有希望，在实验室测试中接近100%的准确率，但很少有系统可用于实际应用。至关重要的是，要超越对精选数据集的准确性，并满足记者的需求，他们可以从这些进步中受益。这个过渡到实践项目的目标是开发DeFake工具，这是一个利用先进的机器学习来帮助记者以一种强大、直观的方式检测深度造假的系统，并提供向公众解释的结果。为了实现这一目标，项目团队从事四项主要任务：(1)使工具对新型深度伪造具有鲁棒性，并让它向用户展示为什么视频是假的；(2)保护工具免受对抗性示例的影响——对视频的微小扰动，这些干扰是专门为欺骗检测系统而制作的；(3)与记者合作，了解他们对该工具的需求，并建立一个在线社区，讨论深度造假及其检测；(4)将其他任务的进展整合成一个稳定、高效、有用的工具，并积极向记者传播。项目团队还利用视觉上有趣的深度伪造来开展引人入胜的教育和推广工作，例如针对所有年龄段的广大观众举办的深度伪造检测博物馆式展览。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。

Deepfakes – videos that are generated or manipulated by artificial intelligence – pose a major threat for spreading disinformation, threatening blackmail, and new forms of phishing. They are already widely used in creating non-consensual pornography, and have begun to be used to undermine governments and elections. Even the threat of deepfakes has cast doubts on the authenticity of videos in the news. Journalists, who have a key role in verifying information, especially need help to deal with ever-improving deepfake technology. Recent results on detecting deepfakes are promising, with close to 100% accuracy in lab tests, but few systems are available for real-world use. It is critical to move beyond accuracy on curated datasets and address the needs of journalists who could benefit from these advances.The objective of this transition-to-practice project is to develop the DeFake tool, a system that utilizes advanced machine learning to help journalists detect deepfakes in a way that is robust, intuitive, and provides results that are explainable to the general public. To meet this objective, the project team is engaged in four main tasks: (1) Making the tool robust to new types of deepfakes, and having it show users why a video is fake; (2) Protecting the tool from adversarial examples – small perturbations to a video that are specially crafted to fool detection systems; (3) Working with journalists to understand what they need from the tool, and building an online community to discuss deepfakes and their detection; and (4) Integrating advances from the other tasks into a stable, efficient, and useful tool, and actively disseminating this tool to journalists. The project team is also leveraging visually interesting deepfakes to develop engaging education and outreach efforts, such as a museum-style exhibit on deepfake detection meant for broad audiences of all ages.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.