Classifier-level long-tail corrections applied to a fine-tuned and frozen ET-BERT encoder on NUDT-Mobile—including logit adjustment, effective-number reweighting, class-balanced retraining (cRT), and feature mixup—improve Tail-F1 by at most 0.08 percentage points over the base ET-BERT (82.79%); Curv-Tail's end-to-end joint representation learning outperforms the best correction by 2.83 points in Tail-F1 and 3.05 points in Macro-F1.
From 2026-wang-curv-tail-lightweight-long-tailed — Curv-Tail: Lightweight Long-Tailed Encrypted Traffic Classification with Discrete Packet-Length Encoding and Lorentz Prototypes
· §3.2, Table 2
· 2026
· arXiv preprint
Implications
The bottleneck for ML traffic classification lies in feature extraction (packet lengths, inter-arrival times, byte patterns), not the classifier head; circumvention defenses must corrupt these low-level observable features rather than exploiting weaknesses in classifier post-processing pipelines.
End-to-end jointly trained classifiers are the correct threat model—frozen pre-trained encoders with corrected classifiers are far weaker, so defenses validated only against pre-trained models underestimate real adversary capability.