Abstract:

It puts forward a new challenge with intrusion detection, the continuous collection of traffic data by the network leads to the massive data problems. Therefore, a deep belief nets model oriented to the intrusion detection system (DBN-IDS) is proposed. First, an unsupervised, greedy algorithm is employed to train each restricted Boltzmann machine (RBM) at a time by a bottom-up approach, which makes large amounts of nonlinear high-dimensional unlabeled input data can be sampled as optimal low-dimensional feature representations. Second, using the labeled data at the top layer, the supervised back propagation （BP）algorithm is employed in classifying the learned low-dimensional representations and fine-tuning the RBM network simultaneously. The parameters and the performance of the model are deeply analyzed on NSL-KDD dataset. Experimental results demonstrate that the DBN-IDS outperforms the support vector machine (SVM) and neural network （NN）, and which is a feasible approach in intrusion classification for the high-dimensional, nonlinear and large-scale data.