Scholarlink

<p>The escalating volume and sophistication of cyber threats make robust network Intrusion Detection Systems (NIDS) a critical line of defense for modern networks. Class imbalance in network traffic datasets severely hinders the performance of IDS, causing critical minority attack classes to be overlooked. This study explores the application of ResNet for multiclass intrusion detection using the CICIDS2017 dataset. We propose a framework integrating ResNet for robust feature learning and classification, coupled with SMOTE-ENN (Synthetic Minority Over-sampling Technique combined with Edited Nearest Neighbors) for effective data resampling. This hybrid approach strategically oversamples minority classes while cleaning noisy samples, ensuring robust representation of rare attacks. Our experiments demonstrate that this combined ResNet and SMOTE-ENN approach significantly enhances detection accuracy, particularly for minority attack classes like Web Attacks (XSS, SQL Injection), BotNet and Infiltration attacks, compared to baseline models and conventional sampling techniques. The results highlight the potential of ResNet’s skip connections and hierarchical feature learning combined with advanced hybrid resampling to build more reliable and effective IDS in imbalanced network environments.</p>

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

Enhancing intrusion detection with ResNet and SMOTE-ENN: a deep learning approach to class imbalance in CICIDS2017

  • 期刊论文

  • 发表日期：

    2025年11月6日

查看全文
  • Prabjot Singh,
  • Divya Nehra,
  • Veenu Mangat,
  • Krishan Kumar

摘要

The escalating volume and sophistication of cyber threats make robust network Intrusion Detection Systems (NIDS) a critical line of defense for modern networks. Class imbalance in network traffic datasets severely hinders the performance of IDS, causing critical minority attack classes to be overlooked. This study explores the application of ResNet for multiclass intrusion detection using the CICIDS2017 dataset. We propose a framework integrating ResNet for robust feature learning and classification, coupled with SMOTE-ENN (Synthetic Minority Over-sampling Technique combined with Edited Nearest Neighbors) for effective data resampling. This hybrid approach strategically oversamples minority classes while cleaning noisy samples, ensuring robust representation of rare attacks. Our experiments demonstrate that this combined ResNet and SMOTE-ENN approach significantly enhances detection accuracy, particularly for minority attack classes like Web Attacks (XSS, SQL Injection), BotNet and Infiltration attacks, compared to baseline models and conventional sampling techniques. The results highlight the potential of ResNet’s skip connections and hierarchical feature learning combined with advanced hybrid resampling to build more reliable and effective IDS in imbalanced network environments.