Efficient Intrusion Detection Using CNNs on Imbalanced Data

Abstract

By detecting suspicious activity and cyberattacks in real time, Intrusion Detection Systems (IDS) protect modern networks. Highly unbalanced datasets, where regular traffic or frequent attack types outweigh uncommon but dangerous incursions, make IDS model development difficult. Traditional deep learning algorithms frequently favor majority classes, making minority attack categories extremely difficult to identify. To overcome this restriction, this paper offers an upgraded deep learning system that combines CNNs with a class-weighting technique to balance datasets. The model learns additional discriminative features and avoids majority sample dominance by weighting underrepresented attack types during training. Class weighting optimizes more reliably than oversampling, which can increase noise or redundancy. The model improves minority attack recall and F1 scores while reaching 98.2% accuracy. This improvement shows that the model can detect numerous intrusions and uncommon, complex assaults that traditional IDS systems miss. The results show that CNN architectures with targeted class balancing are realistic and successful for real-world intrusion detection applications, where network security requires dependability across all threat categories.