A large number of attacks are launched daily in the era of the internet and with a large number of users. Nowadays, effective detection of numerous attacks using the Intrusion Detection System (IDS) is an emerging research technique. Machine learning methodologies show effective results in intrusion detection system. We proposed a two-stage hybrid methodology for intrusion detection. In the first stage, the unsupervised Sparse autoencoder (SAE) with smoothed l1 regularization. We employ smoothed l1 regularization to enforce a sparsity of autoencoder. The smoothed l1 regularization is indeed able to learn sparse representations of features. In the second stage, the Deep Neural Network (DNN) was used to predict and classify attacks. The classifier classifies multi attack classification from the extracted features. Unsupervised SAE was optimized to train an efficient model. The experimental results demonstrate that proposed model better than the conventional models in terms of overall performance in detection rate and low false positive rate. The proposed model was assessed on the datasets KDDCup99, NSL-KDD and UNSW-NB15. The model attained the accuracy 99.98% , and detection rate 99.99% on UNSW-NB15 dataset.

互联网时代，每天都在发起大量的攻击，并且拥有大量的用户。如今，使用入侵检测系统（IDS）对众多攻击进行有效检测是一种新兴的研究技术。机器学习方法在入侵检测系统中显示出有效的结果。我们提出了一种用于入侵检测的两阶段混合方法。在第一阶段，具有平滑 l1 正则化的无监督稀疏自动编码器 (SAE)。我们采用平滑的 l1 正则化来强制执行自动编码器的稀疏性。平滑的 l1 正则化确实能够学习特征的稀疏表示。第二阶段，使用深度神经网络（DNN）对攻击进行预测和分类。分类器从提取的特征中对多攻击类别进行分类。无监督 SAE 被优化以训练一个有效的模型。实验结果表明，所提出的模型在检测率和低误报率方面的整体性能优于传统模型。所提出的模型在数据集 KDDCup99、NSL-KDD 和 UNSW-NB15 上进行了评估。该模型在 UNSW-NB15 数据集上达到了 99.98% 的准确率和 99.99% 的检测率。