SaTC: CORE: Small: Amplifying Deepfake Detection by Humans Using Cognitively-Inspired Interfaces

深度造假（使用深度学习工具制作的人为图像或视频）正变得越来越普遍，也越来越有说服力。它们可能很危险。它们可以更容易地制造虚假新闻、欺诈、勒索和未经同意的露骨图像。打击深度造假的一种方法可能是开发新的方法，让人类用户清楚地知道原始视频何时被操纵或完全被构建。这个项目探索了标记所谓深度伪造视频的不同方法，以了解哪种标签最明显，对人类用户最有说服力，最有可能帮助人们记住哪些视频是深度伪造的。这个项目团队正在评估三种不同的方法对深度伪造视频发出信号的有效性。两种使用文本或图标标记假视频的信号方法的灵感来自于事实核查界面的当前设计。项目团队开发了第三种视觉指示器，称为Deepfake Caricatures。这些漫画放大了操纵视频中的人工制品，这些人工制品是由深度伪造的方式产生的。这种放大会破坏深度伪造视频的视觉连贯性，使其看起来更明显是被操纵过的。项目团队正在进行实验，以确认Deepfake漫画比未触碰的深度伪造更容易被检测到，并验证这种方法是一种有用的视觉指标。项目团队还比较了不同的信号发送方法，以了解哪种方法更有助于提醒人类用户视频是深度伪造的。在一项这样的研究中，项目团队通过衡量每个视觉指标是否改变了用户对视频是深度伪造的信心，来比较它们是如何改变人们对视频的主观印象的。在另一项研究中，该项目团队通过测量用户在收到不同的视觉指示后对哪些视频是深度伪造的记忆程度，来比较这三种信号方法对用户记忆的影响。总的来说，这些研究将帮助研究人员了解哪些因素有助于在视频被篡改时警告人们，以及哪些因素有助于他们稍后记住这是一个深度伪造。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 (manipulated images or videos of people created using deep learning tools) are becoming more common and more convincing. They can be dangerous. They can make it easier to create false news, fraud, blackmail, and non-consensual explicit imagery. One way to combat deepfakes may be to develop new ways to make it clear to a human user when an original video has been manipulated or entirely constructed. This project explores different ways of labeling so-called deepfake videos to understand which kinds of labels are most visible, most convincing to a human user, and most likely to help people remember which videos are deepfakes.This project team is assessing the effectiveness of three different methods for signaling a deepfake video. Two signaling methods that use text or icons to flag a fake video are inspired by current designs in fact-checking interfaces. A third visual indicator, called Deepfake Caricatures, has been developed by the project team. These Caricatures magnify artifacts in the manipulated video, artifacts that result from the way deepfakes are created. This magnification disrupts the visual coherence of a deepfake video, making it look more obviously manipulated. The project team is performing experiments to confirm that Deepfake Caricatures are more detectable than untouched deepfakes and to validate that this method is a useful visual indicator. The project team also is comparing the different signaling methods to understand which is more helpful for alerting human users that a video is a deepfake. In one such study, the project team is comparing how each visual indicator changes people’s subjective impressions of videos by measuring whether they change the user’s confidence that a video is a deepfake. In another study, the project team is comparing the effect of the three signaling methods on users’ memories by measuring how well users remember which videos were deepfakes after being alerted with different visual indicators. Overall, these studies will help researchers understand what factors help warn people when a video has been manipulated, and what factors helps them remember that it was a deepfake later.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.