NIDS Backdoor Benchmark: cleaned NIDS datasets and backdoor-poisoned training sets for XGBoost intrusion detectors

Data release for the paper Read Early, Remove Better: Early-Training Dynamics against Dirty-Label Backdoors in XGBoost Intrusion Detectors (Gia-Hien Tran, Duy-Truong Nguyen and Thanh-Quan Nguyen, University of Science, Vietnam National University Ho Chi Minh City; Phuoc-Hung Truong, Ho Chi Minh City Open University; submitted to ICCIES 2027). Code: https://github.com/VISTA-OUS/TrainingDynamicTree. The release holds four network intrusion detection datasets, cleaned, deduplicated and split 70/15/15 with the same procedure, and every poisoned training set used in the paper. TON_IoT, UNSW-NB15 and Edge-IIoTset are the evaluation datasets; CIC-IDS2017 is the development dataset, on which the reading point of Early-Training Dynamics (ETD) and the orientation of every screen were fixed, and it enters no evaluation number. Each poisoned set is stored as the rows it changes in the clean training split, and rebuild_poisoned_set.py rebuilds the full poisoned training set and its triggered test set exactly. Attack grid: seven attack families (Representative, Distribution-Matching, Influence-Aware, Density-Aware, Spectral-Aware, ETD-Aware, Strip-Aware) x trigger sizes k in {1, 3, 5} x poisoning rates {0.5, 1, 2, 5, 10}% of the source class: 105 dirty-label sets per dataset, plus 75 clean-label sets per dataset that the paper does not evaluate. 232 of the 315 dirty-label sets of the evaluation datasets implant a backdoor (net attack success rate of at least 0.70 over five evaluation runs); they are the sets evaluated in the paper (column in_detector_comparison of sets.csv). Stress tests of Sect. 5.4: 60 white-box sets aimed at ETD (ETD value-aware) and Confident Learning (19 implanted), and 10 sets of the joint ETD/Isolation Forest attack (5 implanted). Version 2.0.0 adds the joint-attack sets, rewrites the README for the current paper, and removes the code archive, the supplementary tables and the redrawn-set manifests of an earlier manuscript (still available in version 1.3.0). All clean splits and poisoned sets of version 1.3.0 are unchanged byte for byte. The data derive from CIC-IDS2017, TON_IoT, UNSW-NB15 and Edge-IIoTset, whose terms apply; use is limited to non-commercial academic research (CC BY-NC 4.0). See README.md for the layout, column definitions and source citations.