Understanding and Improving Video Fingerprinting Attack Accuracy under Challenging Conditions

The threat of video fingerprinting attacks poses significant privacy concerns. These attacks can identify streamed videos with high accuracy despite the use of encryption, leveraging both heuristic-based and deep learning techniques. However, the real-world effectiveness of such attacks remains underexplored, as most research assumes ideal conditions. In this paper, we address the challenges posed by variable network conditions and live-streaming latency, which complicate the attacker's ability to collect useful training data. First, we evaluate several deep learning model architectures against video data under diverse network conditions, including two adaptations of existing website fingerprinting attacks tailored to video that we show boast notable improvements over the base attacks and previous state-of-the-art video fingerprinting attacks. Second, we introduce two augmentation techniques and demonstrate that they substantially enhance attack performance in suboptimal conditions, without knowledge of the victim's live latency. Finally, we analyze the effects of data limitations such as observation time, dataset size, and training time. Overall, our work provides new insights into the impact that several real-world challenges have on attack accuracy, presents new and improved attacks, and details two augmentation techniques that can further boost the performance of the new attacks. Combined, these significant advancements highlight the urgent need for effective defense mechanisms.