Open Access

ARTICLE

EDESC-IDS: An Efficient Deep Embedded Subspace Clustering-Based Intrusion Detection System for the Internet of Vehicles

Lixing Tan^1,2, Liusiyu Chen¹, Yang Wang¹, Zhenyu Song^1,*, Zenan Lu^1,3,*

1 College of Information Engineering, Taizhou University, Taizhou, 225300, China
2 College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing, 210016, China
3 School of Computer Science, Fudan University, Shanghai, 200433, China

* Corresponding Authors: Zhenyu Song. Email: ; Zenan Lu. Email:

(This article belongs to the Special Issue: Advanced Networking Technologies for Intelligent Transportation and Connected Vehicles)

Computers, Materials & Continua 2026, 87(2), 42 https://doi.org/10.32604/cmc.2026.075959

Received 11 November 2025; Accepted 23 December 2025; Issue published 12 March 2026

Abstract

Anomaly detection is a vibrant research direction in controller area networks, which provides the fundamental real-time data transmission underpinning in-vehicle data interaction for the internet of vehicles. However, existing unsupervised learning methods suffer from insufficient temporal and spatial constraints on shallow features, resulting in fragmented feature representations that compromise model stability and accuracy. To improve the extraction of valuable features, this paper investigates the influence of clustering constraints on shallow feature convergence paths at the model level and further proposes an end-to-end intrusion detection system based on efficient deep embedded subspace clustering (EDESC-IDS). Following the standard learning approach, continuous messages are encoded into two-dimensional data frames via a frame builder, which are then input into an extended convolutional autoencoder for extracting shallow features from high-dimensional data. On this basis, the dual constraints of these output features and the embedding clustering module facilitate end-to-end training of the EDESC-IDS in various attack scenarios. Extensive experimental results show that such a system exhibits significant detection performance on four types of attack datasets, including DoS, Gear, Fuzzy, and RPM, with precision, recall, and F1 scores consistently above 97.79%, while maintaining a false negative rate (FNR) and an error rate (ER) below 2.22%.

---

Keywords

---