Feature Extraction using LSTM Autoencoder in Network Intrusion Detection System

With the intent growth of web-based data, document classification has become an important task that can be used in many real-time applications to handle and organize text documents. In the traditional approaches, text documents are encoded using fixed length feature vector representation. Compared to the static, fixed length representation, a text document can be better represented using variable length feature vector representation, where text documents are allowed to have variable number of features. In this paper, SVM-based classification is used as it does not suffer much from the curse of dimensionality. User-defined kernel functions such as Jaccard coefficient kernel, n-gram kernel, and string subsequence kernels are used to find the kernel value between a pair of documents. To prove the performance of the proposed method, benchmark datasets like Reuters-21578 and Reuters-8, the most widely used datasets for text classification, are used for our experimental studies. Based on our experimental studies,
we claim that SVM using N-gram kernel gives better performance on Reuters-21578 dataset and SVM using string
subsequence kernel gives better performance on Reuters-8 dataset. We also observed that minor modifications to the
user-defined kernel improve the performance of the model.