Compression amplifies minority-class weakness in edge-deployed neural-network intrusion detection: a per-family analysis on NF-UQ-NIDS-v2

Companion archive for a study of what per-attack-family damage a chained knowledge distillation, structured pruning, and INT8 post-training quantization pipeline does to a NetFlow intrusion detector deployed on a Raspberry Pi 4. Contains the five-seed aggregation behind every reported number, the component-wise ablation across eight conditions and five seeds, cross-corpus transfer results on NF-CSE-CIC-IDS2018, NF-ToN-IoT and NF-UNSW-NB15, the family-aware pruning sweep, calibration results, the exported TensorFlow Lite models including the deployed INT8 artefact, and the full pipeline and analysis code. The NF-UQ-NIDS-v2 corpus itself is third-party data by Sarhan et al. and is not redistributed here.