Privacy Filtering Using Word Embedding for 3D Point Cloud Based Spatial Sharing Systems

With the advancement of information technology and high-performance input/output interfaces, cyber-physical space sharing systems (hereafter CPSSS) have been proposed to enable highly immersive interactions by sharing spatial data in a virtual space. The main advantage of CPSSS is that users sharing a space can have highly immersive interactions by sharing threedimensional images in multiple spaces. However, sharing all the information obtained from each space could violate the privacy of the participating users, thus necessitating the implementation of filtering, such as making certain aspects visible or invisible. On the other hand, making all images invisible compromises the sense of immersion; therefore, developing technologies for appropriate filtering according to various contexts is essential. This paper aims to develop privacy filtering technologies considering context and individual user characteristics by investigating privacy requirements for all possible objects to be shared in CPSSS. In preliminary research, we collected and analyzed data on different personality traits and privacy awareness for different objects through crowdsourcing with an online cooking class scenario. In this paper, we develop a method to infer privacy requirement levels using personality traits and object names as features. In the proposed method, first, we apply a text encoding using the distributed representation of object names to infer the privacy requirement level for unknown objects not included in the dataset. Second, to improve the inference accuracy of the privacy requirement level, we apply transfer learning to the requirement inference model for each user. As a result, we found that the method with word embedding improved the accuracy of recognizing the privacy requirement level by approximately 9% compared to the baseline. In addition, implementing transfer learning improved accuracy by approximately 14%.