Sequential Model Based Intrusion Detection System for IoT Servers Using Deep Learning Methods

Ming Zhong, Yajin Zhou, Gang ChenSensors 2021 · DOI

TL;DR

This paper proposes a novel Intrusion Detection System (IDS) for IoT servers that leverages sequential deep learning models, specifically GRU and Text-CNN, to analyze network traffic and system routines. By treating sequential data as a language, these models achieve higher accuracy and F1-scores compared to traditional machine learning methods, offering a more robust defense against evolving cyber threats in the IoT landscape.

What Problem Does This Paper Address?

The Internet of Things (IoT) is rapidly expanding, connecting a vast array of devices and generating massive amounts of data. While this connectivity offers immense benefits, it also creates a significantly larger attack surface, making IoT servers prime targets for cyber threats. Traditional network security measures are struggling to keep pace with the sophistication and volume of these attacks. Existing Intrusion Detection Systems (IDS) often rely on older machine learning techniques that may not effectively capture the complex, sequential nature of modern network traffic and system interactions, leading to lower accuracy and performance. This paper addresses the critical need for more effective and accurate IDS for IoT servers by exploring the potential of deep learning, particularly sequential models, to better understand and defend against these evolving threats.

Methods

The core of this research lies in the application of sequential deep learning models for intrusion detection. The proposed system collects features from both the network layer (using tcpdump packets) and the application layer (via system routines). The key innovation is the use of models that can treat this sequential data as a language:

  • Text-CNN (Convolutional Neural Network): Adapted from natural language processing, Text-CNN is used to extract local features from sequences. It employs convolutional filters of varying widths to capture different n-gram patterns within the system call sequences.
  • GRU (Gated Recurrent Unit): An advanced recurrent neural network (RNN) architecture, GRU is chosen for its ability to handle sequential data and long-term dependencies. GRU uses gating mechanisms (reset and update gates) to selectively remember or forget information, making it more efficient than traditional LSTMs for certain tasks.

These models are integrated into a deep neural network architecture that includes a preprocessing layer, a feature extraction module (where GRU or Text-CNN are applied), a classification decision module (an MLP), and an output module (softmax regression). The system is trained using backpropagation and stochastic gradient descent with momentum, with cross-entropy as the loss function.

Datasets

The paper evaluates the proposed system using two well-established datasets:

  • KDD99: A widely used dataset for network intrusion detection, containing simulated network traffic with various attack types (e.g., Denial of Service, User to Root, Remote to Local). The authors note that KDD99 suffers from data imbalance and the need for oversampling techniques.
  • ADFA-LD: A host-based intrusion detection dataset that records Linux system calls. This dataset is characterized by its focus on system call sequences and includes common Linux server functions, with attacks like brute-force password attacks and web shells.

Key Insights and Ideas

  • Sequential Data as Language: The paper's central idea is treating network traffic and system routines as a form of language. This allows powerful NLP techniques like Text-CNN and GRU to be applied effectively for anomaly detection.
  • Deep Learning's Advantage: Deep learning models, by their nature, can learn more complex and abstract features from data compared to traditional machine learning methods. This is crucial for identifying subtle anomalies that might indicate an intrusion.
  • GRU and Text-CNN Synergy: The choice of GRU and Text-CNN highlights a strategic approach: GRU excels at capturing temporal dependencies and memory, while Text-CNN is adept at identifying local patterns and features within sequences.
  • Importance of Feature Extraction: The study emphasizes that the feature extraction module has a significant impact on the overall performance of the IDS.

Key Results

The paper demonstrates that the proposed deep learning-based IDS, utilizing GRU and Text-CNN, consistently outperforms traditional machine learning methods (C4.5 Decision Tree, Naïve Bayes, SVM, and SVM-RBM) across various attack types and datasets.

  • Higher F1-Scores: The experiments show that the deep learning models achieve significantly higher F1-scores, indicating a better balance between precision and recall in detecting intrusions. While specific percentage improvements are not explicitly stated in the abstract, the figures (Figures 7-14) clearly illustrate that GRU and Text-CNN consistently achieve the highest F1-scores across different attack categories and datasets, often reaching scores above 0.9.
  • Improved Accuracy and Performance: The abstract states that the deep learning methods have higher F1-scores, implying improved accuracy and performance.
  • Effectiveness on KDD99 and ADFA-LD: The models proved effective on both the KDD99 and ADFA-LD datasets, showcasing their generalizability.

Limitations and Future Work

  • Data Integrity: The paper acknowledges that deep learning systems have limitations, including concerns about the integrity of the input data. This can affect the reliability of the generated input and output.
  • Computational Cost: While GRU is presented as an evolution of RNN, the paper notes that GRU can be computationally expensive due to its memory function.
  • Need for Faster Evaluation: The authors mention the increasing demand for faster and more useful data evaluation, suggesting that further optimization for speed is a potential area for future work.
  • Variational Contexts: The paper suggests that new deep learning methods are becoming popular because they can fit "variational contexts related to data," implying that adapting models to diverse and evolving IoT environments is an ongoing challenge and opportunity.

Summary updated

Abstract

IoT plays an important role in daily life; commands and data transfer rapidly between the servers and objects to provide services. However, cyber threats have become a critical factor, especially for IoT servers. There should be a vigorous way to protect the network infrastructures from various attacks. IDS (Intrusion Detection System) is the invisible guardian for IoT servers. Many machine learning methods have been applied in IDS. However, there is a need to improve the IDS system for both accuracy and performance. Deep learning is a promising technique that has been used in many areas, including pattern recognition, natural language processing, etc. The deep learning reveals more potential than traditional machine learning methods. In this paper, sequential model is the key point, and new methods are proposed by the features of the model. The model can collect features from the network layer via tcpdump packets and application layer via system routines. Text-CNN and GRU methods are chosen because the can treat sequential data as a language model. The advantage compared with the traditional methods is that they can extract more features from the data and the experiments show that the deep learning methods have higher F1-score. We conclude that the sequential model-based intrusion detection system using deep learning method can contribute to the security of the IoT servers.
BibTeX
@inproceedings{zhong2021sequential,
  author = {Ming Zhong and Yajin Zhou and Gang Chen},
  title = {{Sequential Model Based Intrusion Detection System for IoT Servers Using Deep Learning Methods}},
  booktitle = {Sensors},
  year = {2021},
  doi = {10.3390/s21041113},
}
Figure