Intrusion Detection Methods Based On Manifold Learning And SVM With Parameter Optimization

With the development of information and network technology, the network infrastructure and a number of important web hosts grow particularly rapid under the drive of the political,economic, military and any other interests. Traditional network protection measures such as network firewalls can’t defend the new types of cyber-attacks in the environment in which attacking methods which are refurbished constantly. The appearance of the intrusion detection technology makes the network safe technology to develop towards the direction of initiative and passive combination, which provides multi-level protection to the network security.There are many shortcomings in the traditional intrusion detection, such as false positives, false negatives and low efficiency, which can’t meet the demand of network security. The data set of the intrusion detection always has the characteristic of high-dimension, small samples and inseparability at present and common classification algorithms are of low efficiency in intrusion detection field. Support Vector Machine(SVM)is a classification learning method based on the small samples learning which avoids the local optima and overcomes the dimensionality problems. The SVM exhibits many advantages in the solution of high-dimensional problems and small samples, so it is feasible that SVM is used for instruction detection no matter in theory or reality.Firstly, the paper describes the background and the meaning of the intrusion detection research and describes the technology of the intrusion detection, the theory of the SVM, and the algorithm of the statistical manifold in detail. Secondly, it describes two algorithms. One is the application of the LE-CV-SVM in intrusion detection, the other is improved MDS-GA-SVM applied in intrusion detection. In the first model, a classifier is adopted to estimate whether the action is an attack, MLE used to estimate the intrinsic dimensions, and the LE used as a preprocessor of SVM to reduce the dimension of feature vectors and shorten training time. In order to improve the performance of SVM, CV is used to optimize the parameters of RBF kernel function in SVM. By comparison with other detection algorithms,the experimental results show that the proposed model’s training time is shorter, accuracyrate is higher and false positive rate is lower than the other compared algorithms.In the second model, MDS is used to deduce the dimensions of the data set, the improved SVM used to classify the data set, and the GA used to optimize the parameters of the SVM at the same time. In this algorithm, the Kernel function is improved and the parameters of the SVM are optimized. The result of the experiments verify the effectiveness of the method mentioned above, which improves the detection efficiency.