SAX-LSTM Approach for Network Intrusion Detection

This repository contains the source code for the paper: "Lightweight SAX-LSTM approach for IoT Network Intrusion Detection"

Our method combines Symbolic Aggregate approXimation (SAX) and Long Short-Term Memory (LSTM) networks to detect network intrusions efficiently in resource-constrained environments using the CICIoT2023 dataset.

Overview

We present a lightweight, explainable, and modular approach to intrusion detection by:

• Converting raw packet-level time-series data into symbolic sequences using SAX
• Training an LSTM model on the symbolic representations
• Achieving competitive classification performance with low inference time (0.46–2.44 ms/record)

Experimental Setup

• Dataset: CICIoT2023
• Runtime: Kaggle IPython notebooks (no GPU/TPU required)
• Frameworks: TensorFlow, pyts, sklearn, pandas

Results

Binary Classification

The SAX-LSTM model achieved a binary classification accuracy of 96.79%, with inference times ranging from 0.46 ms to 2.44 ms per record.

Multi-class Classification

In the multiclass setting with 18 classes, it reached an overall accuracy of 83.47% after employing random oversampling to handle the class imbalance problem.

Observations

• Our model matches or exceeds the performance of many recent approaches while requiring no GPU acceleration or heavy preprocessing.
• Inference times are consistently below 2.5 ms, making the approach deployable in latency-sensitive edge environments.
• Performance drop in multiclass is attributed to class imbalance, which was mitigated with random oversampling after SAX transformation.
• The symbolic abstraction using SAX improves model interpretability and lowers computational load.

Key Contribution

• Symbolic abstraction of time-series payload data for enhanced explainability
• Inference-efficient design enabling real-time detection with only payload length, timestamp, and protocol type
• Modular implementation with parameterized design for easy tuning

Credits

Code Credits to the team behind this work:

1. Purujit Srinivasan
2. Gertrude Nabasirye
3. Gainikamal Batayeva
4. E. Fatih Yetkin
5. Tuğçe Ballı