A 2023 IIE/CAS master's thesis (author and advisor names redacted) develops a black-box adversarial-example attack against ML-based malicious-encrypted-traffic classifiers: transferable feature-space perturbations evade LSTM and FS-Net classifiers over 96.2% of the time (over 50.5% against SVM/LR/DNN), and a packet-length-and-timing perturbation algorithm that translates those perturbations into real traffic achieves an 82.1% evasion rate, 29.9 points higher than the prior state-of-the-art black-box method.
实验证明了本方法生成的对抗样本对于LSTM和FS-Net有超过96.2%的检测逃逸率，对于SVM、LR和DNN等模型也有超过50.5%的检测逃逸率...本文所提方法在更加实际的场景下生成的对抗样本能够取得82.1%的检测逃逸率，比目前最先进的黑盒对抗样本攻击方法高29.9%。
Defense implications
- Packet-length and inter-packet-timing perturbation alone, with no protocol change, achieved 82% evasion against ML-based traffic classifiers in this internal study — corroborates that traffic-shape/padding-and-delay defenses remain effective against MESA-lineage ML classifiers even without altering the underlying protocol.
Related findings
The same internal research note's second research point develops an ML-based detector for Geneva-style automated censorship-evasion traffic; simple flow-level features (flow size, max packet size, RST/SYN/FIN flag counts, forward init-window bytes, inter-arrival timing) achieve near-perfect (ROC-AUC ~1.00) classification of Geneva-generated evasion traffic against CICIDS2017 and MAWI backbone background traffic using decision trees, LightGBM, XGBoost and random forest, with abnormal flow size (~150 bytes vs. 1000-30000 bytes typical) identified as the single most discriminative feature.
A 2020 IIE CAS master's thesis (author and advisor names redacted) develops ML methods to identify audio/video traffic within encrypted flows without payload inspection, by quantifying on/off transmission burst patterns and binning long packet-length/timing sequences into short fixed-length feature vectors usable from as little as 5 seconds of traffic, plus a lightweight heuristic-threshold variant with a congestion-correction feature scheme built into a deployed identification system.
A 2020 MESA Lab monthly report describes building a sapp plugin that extracts packet-sequence features specifically "for DoH (DNS-over-HTTPS) service discovery," alongside a broader CSTNET DoH measurement-report effort and configuring a DNS-to-DoH gateway -- confirming sapp is used to fingerprint DoH traffic via statistical sequence features rather than plaintext DNS content.
A 2024 IIE/CAS PhD opening-report (author and advisor names redacted) lays out a four-part research plan for identifying specific encrypted video content transmitted over QUIC: (1) recovering per-video traffic fingerprints from downlink packet aggregation despite QUIC's full encryption, using linear-fit correction to reconstruct true video-chunk sizes; (2) matching randomized multi-segment video combinations via a random-walk segment model plus a traffic 'language model'; (3) a Siamese attention-CNN with contrastive learning to fingerprint-match mixed audio/video segment streams; and (4) a large-model-based classifier to keep fingerprint matching robust under network degradation and varied playback modes.
A 2020 IIE/CAS master's thesis (author and advisor names redacted; the advisor is a recurring MESA Lab supervisor named throughout this corpus), 'Research on Key Technologies of SSL Client Identification for Network Traffic,' proposes labeling training data by decrypting sessions through an SSL transparent-proxy MITM to auto-tag which client application generated each session, then builds a real-time classifier fusing SSL handshake (ClientHello) fingerprints with SSL 'Burst' packet-timing/size features to identify the client application without further decryption, validated against WhatsApp, WeChat, and Facebook traffic.
A 2024 MESA thesis proposal targets detecting "cross-border apps that illegally transmit personal information" inside encrypted traffic without decryption, by discovering app-specific tracking identifiers as <domain,parameter,value,interval> tuples and clustering flow structural similarity (Euclidean distance in a feature space) to separate device identifiers from background noise via a two-layer recognition model.