Reproduce the label-smoothing geometry result on encrypted-traffic data

Reproduce the geometric analysis of prior label-smoothing studies on the CESNET-TLS-Year22 encrypted-traffic classification setting to determine whether label smoothing’s effect on out-of-distribution detection depends on the dataset, class structure, smoothing strength, or distance of unknown services from the training distribution.

Background

Prior studies cited by the paper report that label smoothing can degrade out-of-distribution detection by collapsing the geometry of the penultimate representation. In contrast, the experiments on 102 fine-grained TLS services do not reproduce that degradation, and the label-smoothed student performs particularly well with feature-space detectors. The authors suggest that differences in modality, class count, smoothing strength, and near-versus-far unknown structure may explain the discrepancy, but they do not resolve it.

References

Reproducing the geometric analysis of those papers on this dataset would settle the point and is left for future work.

— Unknown-Traffic Detection, Calibration and Shortcut Reliance in Distilled Encrypted-Traffic Classifiers over One Year  (2609.31141 - Abbasi, 25 Sep 2026) in Section 6, “Where we disagree with the literature,” subsection “Label smoothing”