Analysis of Autoencoders for Network Intrusion Detection.

Area of Science:

• Cybersecurity
• Artificial Intelligence
• Machine Learning

Background:

• Network attacks are evolving, necessitating advanced Network Intrusion Detection Systems (NIDS).
• Deep learning, particularly autoencoders, shows promise for detecting novel (zero-day) attacks and reducing manual labeling efforts in NIDS.
• Optimizing autoencoder models for NIDS is time-consuming, hindering practical deployment.

Purpose of the Study:

• To investigate the impact of autoencoder architecture, specifically latent size, on NIDS performance.
• To identify key factors influencing the effectiveness of autoencoder-based NIDS.
• To provide insights for optimizing autoencoder models for practical intrusion detection.

Main Methods:

• Utilized benchmark datasets: NSL-KDD, IoTID20, and N-BaIoT.
• Evaluated a simple autoencoder model with various combinations of model structures and latent sizes.
• Rigorously studied autoencoder performance across different configurations.

Main Results:

• The latent size of an autoencoder model demonstrated a significant impact on Intrusion Detection System (IDS) performance.
• Different combinations of model structures and latent sizes yield varying detection capabilities.
• The study identified latent size as a critical hyperparameter for autoencoder-based NIDS.

Conclusions:

• Autoencoder latent size is a crucial factor for achieving optimal performance in NIDS.
• Further research into hyperparameter optimization can accelerate the practical application of autoencoder NIDS.
• Findings guide the development of more effective and efficient deep learning-based NIDS for evolving cyber threats.