Information Leakage Through Packet Lengths in RTC Traffic

Modern real-time communication (RTC) depends on efficient audio and video compression to minimize bandwidth requirements. While the codecs are certainly up to the task, the combined effect of live data, lossy compression, and length-preserving encryption leads to the possibility of leakage: the lengths of encrypted packet sequences can leak information about the underlying, unencrypted data. In this work, we measure leakage in real-world RTC platforms and analyze whether information is recoverable. We survey a range of real-world RTC platforms (along with various configurations of each platform) for their propensity to leak data through encrypted packet lengths. We then conduct an in-depth study on Zoom and collect 268,392 network traces amounting to over 786.8 h of recorded media. We use this data to train models to automatically recover information from packet lengths. This is achieved by leveraging existing action recognition datasets, along with a new dataset of virtual meeting recordings. We train neural networks for action recognition over encrypted packet lengths rather than the source data itself. We find that under certain conditions, our models significantly outperform random selection, in effect quantifying leakage for the given learning problem.