Discrete packet-length tokenization is the single most influential component in Curv-Tail: replacing it with continuous length encoding reduces Tail-F1 by 11.04 percentage points on DataCon-Website (47.23% → 36.19%), a larger ablation gap than removing Lorentz geometry (−2.12 pp) or the length-retention auxiliary objective (−1.02 pp) combined.
From 2026-wang-curv-tail-lightweight-long-tailed — Curv-Tail: Lightweight Long-Tailed Encrypted Traffic Classification with Discrete Packet-Length Encoding and Lorentz Prototypes
· §3.3, Table 3
· 2026
· arXiv preprint
Implications
Traffic randomization must specifically target packet-length distributions: length-based discrete features dominate classifier accuracy more than timing jitter, payload byte patterns, or inter-arrival time, so padding/fragmenting to disguise characteristic length fingerprints yields the largest per-effort evasion gain.
Uniform-random length padding is insufficient; the classifier uses exact identity tokens for lengths ≤B and coarse bins above, so circumvention tools need to target both regimes—randomizing short packets (which retain exact tokens) matters as much as blurring large-packet histograms.