Per-flow unknown-scores and model checkpoints for "What Does the Student Inherit? Unknown-Traffic Detection, Calibration and Shortcuts in Distilled Traffic Classifiers over Time"

Derived data for a pre-registered study of knowledge distillation in encrypted-traffic classification. It contains the per-flow unknown-traffic scores of every teacher and student, for every evaluation window, and the trained model weights: 335 files, 8.5 GB in total. The models are trained on CESNET-TLS-Year22 at three start dates and evaluated on 18 test windows spanning 35 weeks. Scores are given under the two pre-registered logit-based rules (energy, maximum softmax probability) and, for the checkpoints scored in the revision, under two feature-space rules (Mahalanobis distance and feature-space k-nearest-neighbour distance). This deposit does not redistribute CESNET-TLS-Year22. The scores are derived quantities and the checkpoints are trained weights; neither contains the captured flows. Use of the dataset itself remains subject to its own terms. Code and analysis: https://github.com/Mahmoud-Abbasi-svg/kd-encrypted-traffic-inheritancePre-registration: https://osf.io/rts6n