SHAP-based Feature Stability Analysis for High-dimensional IoT Intrusion Detection Using Deep Neural Network
In today's world, the Internet of Things (IoT) has emerged as a vital network infrastructure in the process of digitalization.However, existing works struggle with high-dimensional and redundant features that increase computation and reduce efficiency.Furthermore, they fail to distinguish critical features for specific attack types, making their security decisions difficult to interpret.This manuscript presents a SHapley Additive exPlanations (SHAP)-based Feature Stability Analysis Network (SFSA-Net) for an Intrusion Detection System (IDS) in IoT.The proposed SFSA-Net uses SHAP to extract consistent stable features from a baseline Deep Neural Network (DNN), which then optimizes the final network for all attack types.To evaluate feature consistency, a stability-weighted feature ranking index, the Refined Explainable Index (REI), is established by conducting a three-stage stability analysis.The proposed approach achieves an accuracy of 98.75%±0.23%on UNSW-NB15 and 98.85%±0.28%on CICIoT-2023 datasets, respectively.Moreover, the proposed SFSA-Net achieves an optimal feature reduction of 78.7% on UNSW-NB15 and 97.4% on CICIoT-2023, minimizes memory consumption by 83.63% on UNSW-NB15 and 89.03% on CICIoT-2023, while simultaneously improving the detection of minority classes, specifically increasing the recall for worm attacks by 39.25 percentage points.
Authors
- Priti Mishra
- Praveen Shivanna
Institutions
- Visvesvaraya Technological University (IN)
Publication Details
- Journal
- International journal of intelligent engineering and systems
- Published
- 2026-11-17
- DOI
- https://doi.org/10.22266/ijies2026.1231.04
- Primary Topic
- Network Security and Intrusion Detection
- Type
- article
- Field-Weighted Citation Impact
- 0.00