Strainer: Encrypted Video Traffic Identification for Mixed Segment Transmission Pattern

Identifying the video source of encrypted video traffic is a critical task of network regulation. Under Dynamic Adaptive Streaming over HTTP (DASH), the emergence of the mixed segment transmission pattern poses significant challenges for fingerprint matching, particularly in poor networks. To address these challenges, we propose Strainer, a DASH TLS-encrypted video traffic identification method tailored for the mixed segment transmission pattern. On the one hand, we observe that precise chunk sizes can be extracted from the TLS record layer of video traffic. On the other hand, by reversing the thinking, we develop Chunk Decomposition Model (CDM) via Imitation Learning (IL) localized preliminary training followed by Reinforcement Learning (RL) global optimization training. CDM is responsible for decomposing the traffic fingerprint into a segment sequence to enable distance computation with video fingerprints, thereby enhancing the efficiency of fingerprint matching. Strainer achieves a significant performance improvement compared with the other 6 SOTA methods across 6 scenarios.