A Network Intrusion Detection Approach Based on Asymmetric Convolutional Autoencoder

Shujian Ji, Kejiang Ye, Chengzhong XuLecture Notes in Computer Science 2020 · DOI · OpenAlex

Abstract

Modern Network Intrusion Detection Systems (NIDS) struggle with high-dimensional data and the detection of infrequent but critical "small-sample" attacks. This paper proposes ACAE-RF, a hybrid model combining an Asymmetric Convolutional Autoencoder (ACAE) for unsupervised feature extraction and a Random Forest (RF) classifier. By reshaping 1D network traffic into 2D matrices, the ACAE captures spatial feature correlations, while the RF ensemble provides robust classification. Evaluated on KDD99 and NSL-KDD datasets, the model achieves 98.39% accuracy and significantly improves detection rates for rare User-to-Root (U2R) and Remote-to-Local (R2L) attacks.

Introduction

As digital infrastructure expands, the volume and complexity of network traffic grow exponentially. Traditional NIDS face two primary hurdles: the labor-intensive nature of manual feature engineering and the "needle in a haystack" problem, where dangerous but rare attacks (like U2R or R2L) are often ignored by models optimized for overall accuracy. This research addresses these challenges by bridging the gap between deep learning’s feature extraction capabilities and shallow learning’s classification efficiency.

Methodology

The proposed ACAE-RF architecture follows a two-stage process: unsupervised feature learning followed by supervised classification.

1. Data Preprocessing

Network traffic vectors (typically 41 features) are normalized and converted into 8x8 2D matrices using zero-padding. This transformation allows the model to treat network data like a low-resolution image, enabling the use of Convolutional Neural Networks (CNNs) to uncover spatial relationships between features.

2. Feature Extraction (The ACAE)

Unlike standard symmetric autoencoders, the ACAE focuses on the encoding phase. By using a multi-layer stacked encoder, the model reduces computational overhead while producing highly abstract data representations. It utilizes convolutional layers for "local perception" and parameter sharing.

The complete pipeline of the ACAE-RF model, illustrating the flow from data preprocessing through feature extraction to final classification.

3. Classification (The Random Forest)

Instead of a standard Softmax layer, the learned features are fed into a Random Forest classifier. RF is an ensemble of decision trees known for its resistance to overfitting and its ability to handle high-dimensional feature spaces effectively.

The convolutional and deconvolutional layers used to generate feature maps and reconstruct inputs during the training phase.

Experimental Evaluation

The model was evaluated on two industry-standard benchmarks:

  • KDD99: A classic, large-scale dataset for intrusion detection.
  • NSL-KDD: An improved version that removes redundant records, making it significantly harder for models to achieve high accuracy through simple memorization.

Key Results and Insights

The ACAE-RF model outperformed existing methods like Deep Belief Networks (DBN) and Stacked Non-symmetric Deep Autoencoders (S-NDAE), particularly in difficult categories:

  • Overall Accuracy: Reached 98.39% on both datasets.
  • Small-Sample Breakthrough: In the U2R category, the model achieved a precision of 41.32%, whereas previous S-NDAE models failed to detect them entirely (0%).
  • Rare Attack Recall: For R2L attacks, the model achieved a recall of 88.36%, a massive improvement over the 9.31% seen in earlier non-symmetric autoencoder implementations.
  • Optimal Architecture: Performance peaked at 7 layers. Beyond 9 layers, accuracy dropped due to overfitting. Additionally, the model reached peak performance at just 20 epochs; training longer led to an increase in false alarms.

Conclusion

The ACAE-RF approach demonstrates that an asymmetric architecture focusing on convolutional feature extraction can significantly improve the detection of rare network threats. While R2L accuracy remains an area for further refinement, the model provides a robust framework for handling the imbalances found in real-world network traffic. Future work will explore the model's performance against more diverse, modern network event categories.

Summary updated

BibTeX
@inproceedings{ji2020network,
  author = {Shujian Ji and Kejiang Ye and Chengzhong Xu and Cheng-zhong Xu},
  title = {{A Network Intrusion Detection Approach Based on Asymmetric Convolutional Autoencoder}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2020},
  doi = {10.1007/978-3-030-59635-4\_9},
}
Figure