DEEP RESIDUAL LEARNING FOR HIGH-ACCURACY NETWORK INTRUSION DETECTION USING CICIDS2017 AND CICIDS2018

Abstract

The increasing complexity of modern cyber-attacks has rendered traditional
intrusion detection systems inadequate for protecting contemporary network
infrastructures. Signature-based and shallow machine learning techniques
struggle to generalize to unseen attack patterns and high-dimensional traffic
features. Recent advances in deep learning have demonstrated promising results
for intrusion detection; however, increasing network depth often leads to
vanishing gradient and performance degradation problems. This study presents
a deep residual learning-based network intrusion detection model that leverages
the strengths of residual neural networks to enhance detection accuracy and
training stability. Using a merged benchmark dataset derived from CICIDS2017
and CICIDS2018, the approach learns discriminative traffic representations
capable of accurately classifying benign and malicious activities. Experimental
results demonstrate that residual learning significantly improves classification
performance and robustness, making it suitable for large-scale and complex
network environments.