A Hard-Voting Ensemble of DNN, LSTM, And CNN For Network Intrusion Detection on the UNSW-NB15 Dataset

Abstract

The increasing sophistication of cyber threats demands intrusion detection systems that are both accurate and robust. While deep learning models have shown promise in network intrusion detection, individual architectures often fail to capture the full complexity of network traffic patterns. This paper presents a hard-voting ensemble approach that combines three distinct deep learning architectures—Dense Neural Network (DNN), Long Short-Term Memory (LSTM), and Convolutional Neural Network (CNN)—to improve intrusion detection performance on the UNSW-NB15 dataset. Each base model is designed to capture complementary characteristics of network traffic: the DNN learns complex feature interactions, the LSTM captures temporal dependencies, and the CNN extracts spatial patterns. The ensemble combines their predictions through majority voting. Experimental results demonstrate that the proposed ensemble achieves an accuracy of 85.01%, outperforming individual models (DNN: 84.54%, CNN: 84.55%, LSTM: 82.95%). The ensemble also achieves precision of 84.38%, recall of 85.01%, and an F1-score of 81.67%. A confusion matrix analysis reveals balanced performance across nine attack categories. These findings establish that ensemble methods, particularly hard voting, offer a practical pathway toward more reliable network intrusion detection systems.