Abstract The Internet of Things (IoT) generates high-volume network traffic that is heterogeneous and highly imbalanced, creating a critical need for intrusion detection systems (IDS) that remain accurate under resource-constrained edge gateways. This study presents an optimized gradient-boosting IDS framework evaluated on the CICIoT2023 benchmark. The pipeline integrates stratified undersampling to mitigate class imbalance and gain-based feature selection to remove redundant flow attributes. We evaluate the framework across three tasks: binary detection, 8-class attack-family classification, and 34-class fine-grained classification. Across all tasks, the evaluated models achieved strong detection performance with low inference latency; specifically, XGBoost attained 99.61% accuracy and a 0.995 macro-F1 score in binary detection with 0.355 µs/sample inference time. Feature selection reduced the input space from 46 to 23 dimensions, and an ablation study showed that this reduction decreased inference latency by 30–51% and training time by 33–40% while keeping accuracy variation within ± 0.014% points. These results indicate an accuracy–efficiency trade-off suitable for real-time deployment. All datasets, code, and results are publicly available to facilitate reproducibility. Data availability The raw CICIoT2023 dataset analyzed in the current study is available in the Canadian Institute for Cybersecurity repository at [https://www.unb.ca/cic/datasets/iotdataset-2023.html](.)30. Researchers intending to replicate the full data ingestion and feature selection pipeline described in Algorithm 1 are advised to download the raw dataset from the original source.**Extended data31: ****Zenodo Record: ** *Supplementary material - An Optimized Gradient Boosting Framework for IoT Intrusion Detection: A Comprehensive Evaluation on the CICIoT2023 Dataset (Version 3.0.0).* This repository contains: 1. **Source Code: ** Python scripts implementing the complete experimental pipeline, including stratified undersampling, gain-based feature selection (Algorithm 1), and model evaluation for Binary, 8-class, and 34-class tasks. 2. **Processed Datasets: ** Stratified subsets of the training and testing data (with the optimized 23-feature set) used for the reported ablation studies. References Asharf, J.