The advancement of internet applications over the last decade makes network security a priority, in which Intrusion Detection System (IDS) is responsible for detecting threat in network throughput. Nevertheless, current IDS encounter challenges because (i) the training dataset is seriously class-imbalance, (ii) conventional method requires a lot of hand-crafted feature selection. To address these problems, we propose a novel method that combines data augmentation method as well as convolutional autoencoder to build up an IDS. Our method uses SMOTE to balance the distribution of each attack in the training dataset, then uses a convolutional autoencoder to detect the intrusion, which is free of hand-crafted feature design. Experimentally, we show that our proposed method outperforms conventional method both in binary and multi-class classification in terms of accuracy. We also have an ablation study on how the ratio of minority class in dataset effect the f1-score of the model.