The rapid evolution of network traffic and cyber-attack sophistication has necessitated robust Intrusion Detection Systems (IDS). Traditional machine learning methods often struggle to capture complex non-linear attack patterns. Deep learning, particularly Convolutional Neural Networks (CNNs), provides an effective alternative for automatic feature extraction and accurate classification. This paper presents the implementation of an IDS using the UNSW-NB15 dataset and explores various CNN architectures — including 1D-CNN, 2D-CNN, and hybrid CNN-LSTM models. A detailed experimental comparison is carried out in terms of accuracy, precision, recall, F1-score, false positive rate (FPR), and false negative rate (FNR). The results show that the hybrid CNN-LSTM model outperforms conventional CNN variants, achieving an accuracy of 98.6% for binary classification and 96.1% for multi-class detection. The study demonstrates the potential of CNN-based architectures to efficiently detect modern network intrusions.