Knowledge Distillation with BERT for Image Tag-Based Privacy Prediction

Proceedings of Recent Advances in Natural Language Processing, pages 1616–1625 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_181 1616 Knowledge Distillation with BERT for Image Tag-Based Privacy Prediction Chenye Zhao University of Illinois at Chicago Chicago, USA czhao43@uic.edu Cornelia Caragea University of Illinois at Chicago Chicago, USA cornelia@uic.edu Abstract Text in the form of tags associated with online images is often informative for predicting pri- vate or sensitive content from images. When using privacy prediction systems running on social networking sites that decide whether each uploaded image should get posted or be protected, users may be reluctant to share real images that may reveal their identity, but may share image tags. In such cases, privacy-aware tags become good indicators of image privacy and can be utilized to generate privacy de- cisions. In this paper, our aim is to learn tag representations for images to improve tag- based image privacy prediction. To achieve this, we explore self-distillation with BERT, in which we utilize knowledge in the form of soft probability distributions (soft labels) from the teacher model to help with the train- ing of the student model. Our approach effec- tively learns better tag representations with im- proved performance on private image identifi- cation and outperforms state-of-the-art models for this task. Moreover, we utilize the idea of knowledge distillation to improve tag represen- tations in a semi-supervised learning task. Our semi-supervised approach with only 20% of annotated data achieves similar performance compared with its supervised learning counter- part. Last, we provide a comprehensive anal- ysis to get a better understanding of our ap- proach. 1 Introduction With the rapid growth of the number of users on online social networking sites, image privacy has become a major concern (Ahern et al., 2007; Squic- ciarini et al., 2017). Users may accidentally dis- close their sensitive information such as locations, habits or personal relationships from images that they post to their social networking sites (Squic- ciarini et al., 2017), which could be used in the detriment of the users (Tonge and Caragea, 2020). Zerr et al. (2012a) defines private images as ones that belong to the private sphere (e.g., portraits, family, home) or capture sensitive contents that can not be shared with everyone on the Internet. The remaining images are considered to be pub- lic. Binary image privacy classifiers are developed (Tonge and Caragea, 2018; Yang et al., 2020; Zerr et al., 2012a) aiming to identify whether images belong to the public class or the private class. How- ever, the access to the image content is not always allowed since users may be reluctant to share the real images (revealing user’s identity through the face, and friends, etc.) for visual content analysis. In such cases, tags attached by users to describe their images are found to be informative about the image contents and are good indicators of the pri- vacy settings and improve the privacy prediction methods (Tonge et al., 2018). Privacy prediction models trained with image tags achieve compet- itive results compared with vision-based privacy prediction models …

Content truncated to 500 words