Title: Generative Adversarial Networks Applied for Privacy Preservation in Biometric-Based Authentication and Identification Show Chinese title

Title: 用于生物特征认证和识别中隐私保护的生成对抗网络

Abstract: Biometric-based authentication systems are getting broadly adopted in many areas. However, these systems do not allow participating users to influence the way their data is used. Furthermore, the data may leak and can be misused without the users' knowledge. In this paper, we propose a new authentication method that preserves the privacy of individuals and is based on a generative adversarial network (GAN). Concretely, we suggest using the GAN for translating images of faces to a visually private domain (e.g., flowers or shoes). Classifiers, which are used for authentication purposes, are then trained on the images from the visually private domain. Based on our experiments, the method is robust against attacks and still provides meaningful utility. Show Chinese abstract

Abstract: 基于生物特征的认证系统在许多领域得到了广泛采用。 然而，这些系统不允许参与用户影响其数据的使用方式。 此外，数据可能会泄露，并在用户不知情的情况下被滥用。 在本文中，我们提出了一种新的认证方法，该方法保护个人隐私，并基于生成对抗网络（GAN）。 具体而言，我们建议使用GAN将人脸图像转换到一个视觉隐私领域（例如，花朵或鞋子）。 然后，在视觉隐私领域的图像上训练用于认证的分类器。 根据我们的实验，该方法对攻击具有鲁棒性，并且仍然提供有意义的效用。