Processed IoT/IIoT IDS Datasets: Stratified Splits and SMOTE–ENN–LOF Balanced NPZ Files for Lightweight Intrusion Detection Research

This Zenodo record provides processed versions of four publicly available IoT/IIoT intrusion detection datasets prepared for reproducible lightweight intrusion detection system (IDS) research. The released files include standardized train/test splits and balanced training datasets stored in NumPy .npz format. Both binary and multiclass variants are provided where applicable, enabling researchers to directly use the datasets for machine learning and deep learning experiments without repeating the complete preprocessing and balancing workflow. The original datasets were not created by the authors of this Zenodo record. Full credit belongs to the respective dataset creators and publishers. This release only provides processed and experiment-ready derivatives generated from the original public datasets. All datasets used in this study are publicly available and can be accessed through their respective repositories. The ACI-IoT 2023 dataset is hosted on IEEE Dataport and can be obtained at https://doi.org/10.21227/qacj-3x32. The UQ-IoT-IDS 2021 dataset is provided by the University of Queensland and is accessible at https://doi.org/10.48610/17b44bb. The Edge-IIoTset dataset is available on IEEE Dataport at https://doi.org/10.21227/mbc1-1h68. The WUSTL-IIoT 2021 dataset is publicly distributed by Washington University in St. Louis and can be accessed at http://www.cse.wustl.edu/~jain/iiot2/index.html. The processed datasets were generated using a consistent preprocessing pipeline followed in the associated lightweight IDS study. The workflow includes dataset-specific cleaning, removal of unsuitable or non-informative attributes, handling of invalid values, label mapping, encoding of categorical fields where required, feature scaling, and stratified train/test partitioning. For binary classification, benign traffic was encoded as class 0 and all attack traffic was encoded as class 1. For multiclass classification, the original or grouped attack categories were retained according to the experimental design of the study. Class balancing was applied only to the training data to avoid test-set leakage and to preserve realistic evaluation conditions. The balanced training files were generated using a hybrid SMOTE–ENN–LOF strategy. SMOTE was used to synthesize minority-class samples, ENN was used to remove ambiguous samples located near class boundaries, and LOF was used to eliminate potential outliers from the balanced training space. The test sets were not synthetically modified or balanced, so that model performance can be evaluated against the original class distribution. This processed dataset collection is intended to support reproducible research on lightweight IDS development, feature selection, class imbalance handling, and machine learning/deep learning model evaluation for IoT and IIoT security. Users of this Zenodo record must cite this processed dataset record as well as the original dataset sources listed above. Users are also responsible for ensuring that their use of each processed dataset complies with the license and access conditions of the corresponding original dataset provider.