Signature based intrusion detection using latent semantic analysis

We address the problem of selecting and extracting key features by using singular value decomposition and latent semantic analysis. As a consequence, we are able to discover latent information which allows us to design signatures for forensics and in a dual approach for real-time intrusion detection systems. The validity of this method is shown by using several automated classification algorithms (Maxim, SYM, LGP). Using the original data set we classify 99.86% of the calls correctly. After feature extraction we classify 99.68% of the calls correctly, while with feature selection we classify 99.78% of the calls correctly, justifying the use of these techniques in forensics. The signatures obtained after feature selection and extraction using LSA allow us to class 95.69% of the calls correctly with features that can be computed in real time. We use Support Vector Decision Function and Linear Genetic Programming for feature selection on a real data set generated on a live performance network that consists of probe and denial of service attacks. We find that the results reinforce our feature selection method.