Open Access

ARTICLE

ATC-FusionNet: A Hybrid Deep Learning Ensemble for Network Intrusion Detection Systems

Liping Wang¹, Jiang Wu^1,2,*, Liang Wang³

1 School of Information Science and Engineering (School of Cyberspace Security), Zhejiang Sci-Tech University, Hangzhou, China
2 Key Laboratory of Intelligent Textile and Flexible Interconnection of Zhejiang Province, Hangzhou, China
3 Zhejiang Shannon Communication Technology Company Ltd., Hangzhou, China

* Corresponding Author: Jiang Wu. Email:

Computers, Materials & Continua 2026, 88(1), 42 https://doi.org/10.32604/cmc.2026.078591

Received 04 January 2026; Accepted 13 March 2026; Issue published 08 May 2026

Abstract

The rapid growth of networked systems and the increasing diversity of cyberattack behaviors have posed significant challenges to intrusion detection, particularly in scenarios characterized by high-dimensional features and severe class imbalance. Conventional detection approaches based on handcrafted rules or shallow representations often exhibit limited robustness under such conditions. To address these issues, this paper presents a hybrid deep learning framework for network intrusion detection that integrates complementary feature learning mechanisms within a dual-branch architecture. Specifically, a Transformer branch is employed to model long-range temporal dependencies in network traffic, while a convolutional neural network branch (CNN) is used to capture localized and fine-grained feature patterns. An attention-based fusion strategy is further introduced to adaptively aggregate branch-specific representations and enhance intrusion-sensitive features. In addition, an autoencoder-based feature reconstruction module is incorporated before the dual-branch network to compress and reconstruct input features through an encoder–decoder structure, thereby preserving essential behavioral characteristics of network traffic and improving feature discriminability. To mitigate the impact of class imbalance, a Dynamic Weighted Logit-adjusted Focal Loss (DWLF) is introduced to reduce the bias toward majority classes during model optimization. Extensive experiments conducted on two public benchmark datasets demonstrate the effectiveness of the proposed approach. The proposed model achieves an overall accuracy of 90.01% on the UNSW-NB15 dataset and 97.82% on the NF-CSE-CIC-IDS2018 dataset. Experimental results indicate improved robustness under highly imbalanced data distributions, demonstrating stable performance across datasets with different traffic characteristics.

---

Keywords

---