Unveiling the Unseen: Video Recognition Attacks on Social Software

Video recognition attacks are a fine-grained privacy threat that can leak much private information. Moreover, due to the social nature of social software, their video preferences leak more personal information than traditional video software. Existing attack methods are unable to recognize rapidly updated videos on social software. In this paper, we propose an attack method to recognize videos watched on social software under different network environments and playback modes, and the key of the method is to restore the original length of video segments through the features of HTTP/2 protocol and TLS protocol. In addition, the method copes with the rapid update of social software videos through a video data rapid acquisition system. By evaluating our attack method’s performance under different network environments, playback modes, and operating systems, experimental results demonstrate that in a native network environment with sequential playback, the accuracy exceeds 98.70%, and the F1 score exceeds 98.51%. Even in a network environment with extra interference and users seeking playback, the accuracy exceeds 90.21%, and the F1 score exceeds 89.81%. In addition, the method shows good generalization that can be applied to software including Twitter, Facebook, and Instagram. These results demonstrate that video recognition attacks threaten user privacy and highlight the urgency of defending against such attacks.