Peningkatan Akurasi Deteksi Intrusi Jaringan dengan Model Hybrid Convolutional Neural Network dan Long Short-Term Memory

Abstract

The evolving cyber threats demand more sophisticated and accurate intrusion detection systems (IDS). This research develops a hybrid CNN-LSTM model with comprehensive data preprocessing techniques to enhance network attack detection accuracy. The UNSW-NB15 dataset consisting of nine attack categories and 49 features was used as research data. The methodology begins with data preprocessing including data cleaning, categorical transformation using categorical codes, class balancing with upsampling, StandardScaler normalization, and 80:20 data splitting. The hybrid model architecture combines three CNN blocks for spatial feature extraction with two LSTM layers for modeling temporal dependencies. The model was compiled using Adam optimizer with 0.0005 learning rate and equipped with EarlyStopping, ReduceLROnPlateau, and ModelCheckpoint callbacks. Evaluation results show the CNN-LSTM model achieves 99% accuracy, precision, recall, and F1-score, significantly outperforming the standard CNN model which only reaches 96%. Learning curves demonstrate rapid convergence without overfitting indication. This research proves that the combination of CNN's spatial feature extraction capability and LSTM's temporal dependency modeling is highly effective for anomaly detection in complex sequential data such as network traffic.