Stacked Autoencoder based Intrusion Detection System using One-Class Classification

The paper presents a study of deep learning based approach for Intrusion Detection System. Already existing models for classification were based on supervised learning methods which fails to classify instances of unknown attacks. Effective prediction of network packets as normal or attack, known and unknown to the model, is imperative requiring detection with minimal false alarm rate. Even for the attacks not known to the model, Stacked Autoencoder turns out to be one such deep learning architecture which identifies complex pattern leading to generation of the best latent representation of inputs. The proposed model was trained on single labeled instances from KDD Cup 99 dataset along with standardizing the inputs using batch normalization to minimize the problem of internal covariance shift and vanishing gradient to some extent. Experimental results obtained show that the proposed method outperforms all the other algorithms giving accuracy of 98.17% and false alarm rate of 0.38%.