Computing frequency encoding of JA3, JA3S, and SNI across the entire dataset rather than within each cross-validation fold inflates the TLS-view F1 score by 0.28 — approximately ten times larger than any real detection effect measured in the study. When leakage is controlled, the largest genuine multi-view fusion gain over the best single view is only 0.022 F1.
From 2026-nguyen-huu-multi-view-fusion-encrypted — Multi-View Fusion for Encrypted C2 Detection: A Leakage-Controlled Measurement Study of Evaluation Pitfalls
· §V-A, Abstract
· 2026
· arXiv preprint
Implications
Treat published ML classifier accuracy claims for encrypted-traffic detection with extreme skepticism unless the evaluation protocol explicitly controls for cross-fold feature leakage; real gains are ~10× smaller than leaked scores suggest.
When benchmarking circumvention transports against ML detectors, require grouped cross-validation by capture or campaign — not random flow splits — to obtain realistic evasion estimates.