Scholarlink

Detecting zero-day attacks remains a critical challenge in cybersecurity because they exploit previously unknown vulnerabilities, rendering traditional security tools that rely on known signature patterns ineffective. In this study, we present an unsupervised deep learning (DL) detection system that uses an autoencoder to effectively capture the zero-day attacks. The main objective was to build a system with a high recall and low false-negative rates. We used the CIC-IDS2017 dataset for the evaluation. We compared the results of our model with those of the Isolation Forest to demonstrate the efficacy of our model. The results indicate that autoencoders are effective for detecting zero-day attacks. The proposed system exhibited an accuracy of 98.9%, low false negatives, and scalable potential for real-world deployment. This study contributes a robust and adaptable framework for next-generation zero-day detection systems.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Zero-Day Attack Detection System Using Autoencoders and Isolation Forest: An Unsupervised Machine Learning Approach

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Mujeeb Ur Rehman,
  • Margaret Zita,
  • Muhammad Abrar,
  • Muhammad Kazim,
  • Sohail Khalid

摘要

Detecting zero-day attacks remains a critical challenge in cybersecurity because they exploit previously unknown vulnerabilities, rendering traditional security tools that rely on known signature patterns ineffective. In this study, we present an unsupervised deep learning (DL) detection system that uses an autoencoder to effectively capture the zero-day attacks. The main objective was to build a system with a high recall and low false-negative rates. We used the CIC-IDS2017 dataset for the evaluation. We compared the results of our model with those of the Isolation Forest to demonstrate the efficacy of our model. The results indicate that autoencoders are effective for detecting zero-day attacks. The proposed system exhibited an accuracy of 98.9%, low false negatives, and scalable potential for real-world deployment. This study contributes a robust and adaptable framework for next-generation zero-day detection systems.