EAGER: DCL: SaTC: Enabling Interdisciplinary Collaboration: Evaluating Bias In The Creation and Perception of GAN-Generated Faces

不良行为者经常使用机器人和虚假资料攻击个人或团体，破坏社会和谐和集体运动。这些虚假的个人资料可能会使用人脸图像来表明人类的真实性。直到最近，还可以通过反向图片搜索来识别恶意演员，因为许多虚假的个人资料使用的是库存照片。基于机器学习的通用对抗网络（GANs）的最新进展使得创造不存在且无法识别的人的超现实面孔成为可能。这些脸可以做成动画，用来造成伤害。为了帮助开发更安全、更值得信赖的网络空间，了解人类感知者（单独或借助计算辅助）是否以及如何检测真实面孔与人造面孔，以及他们的检测策略和结果在不同群体之间的差异是至关重要的。本项目研究生成面孔的gan是否存在种族偏见，以及这种偏见是否表现在群体内与群体外面孔的可检测性差异上。该项目测试了gan存在种族偏见的假设，因为训练数据集本身就存在偏见，白人面孔（尤其是白人女性面孔）被过度代表。此外，当创建工具来控制生成什么样的面孔时，这些工具也可能具有种族偏见，因为它们提取的是有偏见的参数。这些有偏见的过程可能导致gan生成的面孔对少数种族个体比多数种族个体更容易被检测到。为了验证这些假设，该项目正在开发一个不同面孔的训练数据集，并对肤色和性别等感兴趣的维度进行注释。这些注释可用于训练具有任意数量检查点的GAN，以检查GAN生成的面部在创建的不同阶段如何出现。该项目正在研究人们如何在GAN的每个阶段感知生成的面孔。该项目有助于激发对机器学习如何工作的理论见解，并为不同群体的本科生研究人员提供计算机科学和社会心理学方面的培训。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。

Bad actors often use bots and fake profiles to attack individuals or groups and to undermine social harmony and collective movements. These fake profiles may use face images to signal human authenticity. Until recently it was possible to identify bad-faith actors via reverse image searches because many fake profiles used stock photos. Recent advances in machine learning-enabled general adversarial networks (GANs) have made it possible to create hyper-realistic faces of people who do not exist and cannot be identified. These faces can be animated and used to cause harm. To help develop more secure and trustworthy cyberspaces, it is critical to understand whether and how human perceivers (alone or with computational aids) can detect real vs. artificial faces, and how their detection strategies and outcomes differ across groups. This project investigates whether the GANs that generate faces are racially biased and whether this bias is manifested in differential detectability of ingroup vs. outgroup faces. The project tests the hypothesis that GANs are racially biased because the training dataset is itself biased, with White faces (especially White female faces) overrepresented. Furthermore, when tools are created to control what kind of face is generated, these tools may be racially biased as well because they are extracting biased parameters. These biased processes may result in GAN-generated faces that are more detectable to racial minority individuals vs. racial majority individuals. To test these hypotheses, the project is developing a training dataset of diverse faces, with annotations for dimensions of interest such as skin tone and gender. These annotations can be used to train a GAN with any number of checkpoints to examine how GAN-generated faces appear at different stages of creation. The project is examining how people perceive the generated faces at each stage of the GAN. This project is helping spur theoretical insights into how machine-learning works, and provides training in computer science and social psychology for a diverse group of undergraduate researchers.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.