Video Piracy Websites Detection Using Continual Learning with Elastic Weight Consolidation

Dianzhi Yu, Wentao Zhang, Zenglin Xu, Irwin King

Abstract: With the rapid development of the Internet and video streaming, video piracy websites (VPWs) pose a global challenge to the copyright protection of digital media as the black market for pirated videos continues to grow. In this paper, we design a Video Piracy Websites Detection (VPWD) architecture and implement the system that supports the end-to-end process of possible VPW collection to detection. When designing the architecture, we find that there are currently no publicly available datasets containing the contents of English and Chinese VPWs. Therefore, we build a real-world dataset of features targeting VPWs, containing 14,254 possible English and Chinese VPWs, with 5,759 of them labeled. To the best of our knowledge, it is the largest dataset for VPW detection. We aim to solve the challenges of detecting VPWs. Given the constant emergence of new VPWs with different VPW-related keywords and languages, a model trained once is not enough for detection. We demonstrate that such a fixed model will suffer from performance drops of over 8%, when possible VPWs with new keywords or languages are added to the database, resulting in feature distribution change. Our VPWD architecture is continual learning-based, which can continually learn features of VPWs. As far as we are aware, it is the first work to utilize continual learning for piracy website detection. We conduct experiments to compare continual learning algorithms in our VPWD architecture and achieve an accuracy of 96.4% on average, which is 0.5% higher than without continual learning algorithms.

External IDs:doi:10.1007/978-981-96-7008-6_21