Scholarlink

<p>Rapidly evolving network traffic and sophisticated attack vectors reduce the detection accuracy of traditional Intrusion Detection Systems (IDS). We propose a CNN-based IDS enhanced with two complementary gradient-based explainability methods. Integrated Gradients and Grad-CAM—to improve interpretability and minority-attack detection. Although IG and Grad-CAM are primarily interpretability tools, the attribution feedback was also used to adjust class-balancing and feature refinement, improving minority-class recall. Using CIC-IDS2017, the model achieves 99.21% accuracy, 98.87 F1-score, 99.10 recall, and 0.93% false-positive rate, outperforming classical ML and baseline CNN models. The recorded 0.93% false-positive rate (macro-average across all classes) demonstrates that the model maintains low false alarms not only for majority traffic but also for minority attack categories. Unlike previous works that apply XAI only for visualization, we incorporate gradient attribution feedback into dataset balancing and feature refinement, producing a 12.6% improvement in minority-class recall. The system demonstrates 2.5 ms/flow inference latency, showing feasibility for real-time network environments. These results indicate that dual-gradient explainability offers both operational transparency and statistically significant improvements in intrusion detection performance.</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

Next-gen cyber defense: CNN-based intrusion detection with advanced gradient techniques

  • 期刊论文

  • 发表日期：

    2026年1月5日

查看全文
  • Aditya Mandloi,
  • Sravanthi R,
  • Rakesh S,
  • Vijayalakshmi Senniappan,
  • Anandan P,
  • Sabitha C H,
  • Athiraja A

摘要

Rapidly evolving network traffic and sophisticated attack vectors reduce the detection accuracy of traditional Intrusion Detection Systems (IDS). We propose a CNN-based IDS enhanced with two complementary gradient-based explainability methods. Integrated Gradients and Grad-CAM—to improve interpretability and minority-attack detection. Although IG and Grad-CAM are primarily interpretability tools, the attribution feedback was also used to adjust class-balancing and feature refinement, improving minority-class recall. Using CIC-IDS2017, the model achieves 99.21% accuracy, 98.87 F1-score, 99.10 recall, and 0.93% false-positive rate, outperforming classical ML and baseline CNN models. The recorded 0.93% false-positive rate (macro-average across all classes) demonstrates that the model maintains low false alarms not only for majority traffic but also for minority attack categories. Unlike previous works that apply XAI only for visualization, we incorporate gradient attribution feedback into dataset balancing and feature refinement, producing a 12.6% improvement in minority-class recall. The system demonstrates 2.5 ms/flow inference latency, showing feasibility for real-time network environments. These results indicate that dual-gradient explainability offers both operational transparency and statistically significant improvements in intrusion detection performance.