Combining Pairwise Feature Matches from Device Trajectories for Biometric Authentication in Virtual Reality Environments
Combining Pairwise Feature Matches from Device Trajectories for Biometric Authentication in Virtual Reality Environments
复制标题
DOI：
10.1109/aivr46125.2019.00012
复制
发表时间：
2019-12
期刊：
影响因子：
--
通讯作者：
A. Ajit;N. Banerjee;Sean Banerjee
中科院分区：
文献类型：
--
作者：
A. Ajit;N. Banerjee;Sean Banerjee
本文提出了一种在虚拟现实(VR)环境中通过结合耳机、右手控制器和左手控制器的位置和方向特征来对用户进行无缝连续生物特征认证的方法。虚拟现实在军事训练、飞行模拟、治疗、制造和教育等关键任务应用中的快速增长，需要根据用户在虚拟现实空间中的行为对用户进行身份验证，而不是传统的基于PIN和密码的方法。为了模拟VR环境中可能发生的面向目标的交互，我们捕获了33名用户向虚拟目标投掷球的VR轨迹数据集，每个用户在训练日捕获10个样本，在测试日捕获10个样本。由于每个用户的训练样本数量的稀疏性，这是现实交互的典型情况，我们通过使用轨迹之间的成对关系来执行认证。我们的方法使用感知器分类器来学习来自耳机和手控制器的两个轨迹上位置和方向特征之间的匹配的权重，使得属于同一用户的轨迹获得较低的分类器分数，否则获得较高的分数。我们还对位置和方向特征的选择、设备组合、匹配度量和轨迹对准方法的选择进行了精度方面的广泛评估，并展示了使用右手控制器和耳机的方向匹配每个用户10个测试动作的最高准确率为93.03%。
In this paper we provide an approach to perform seamless continual biometric authentication of users in virtual reality (VR) environments by combining position and orientation features from the headset, right hand controller, and left hand controller of a VR system. The rapid growth of VR in mission critical applications in military training, flight simulation, therapy, manufacturing, and education necessitates authentication of users based on their actions within the VR space as opposed to traditional PIN and password based approaches. To mimic goal-oriented interactions as they may occur in VR environments, we capture a VR dataset of trajectories from 33 users throwing a ball at a virtual target with 10 samples per user captured on a training day, and 10 samples on a test day. Due to the sparseness in the number of training samples per user, typical of realistic interactions, we perform authentication by using pairwise relationships between trajectories. Our approach uses a perceptron classifier to learn weights on the matches between position and orientation features on two trajectories from the headset and the hand controllers, such that a low classifier score is obtained for trajectories belonging to the same user, and a high score is obtained otherwise. We also perform extensive evaluation on the choice of position and orientation features, combination of devices, and choice of match metrics and trajectory alignment method on the accuracy, and demonstrate a maximum accuracy of 93.03% for matching 10 test actions per user by using orientation from the right hand controller and headset.