High-dimensional Anomaly Detection Based On Neural Networks Dimensionality Reduction And Support Vector Machine Classification

Anomaly detection is an important part of data mining.Under the playground of big data,the dimension of data has been growing in a “explosive” type,and the demand for anomalous detection of high-dimensional data has also increased dramatically.The excellent non-linear dimensionality reduction feature of deep neural networks(DNN)can reduce the dimension of data and solve the problem of “curse of dimensionality”.One-class Support Vector Machine(OCSVM)is an important method to solve the problem of abnormal data detection at present.Anomaly detection of high-dimensional data has become the focus and hot topic of scientific frontier research.1.Aiming at the difficulties of high-dimensional data anomaly detection,this paper proposes to use denosing Auto-Encoder(DAE)instead of ordinary Stacked Auto-Encoder(SAE)to detect high-dimensional abnormal data.DAE first randomly mixes the original data with a certain proportion of noise,and then reduces the dimension of data with noise by “encoding”,and then obtains the reconstructed value of the noise data by "decoding",and then minimizes the reconstruction between the original data without noise and the reconstructed value of data with noise.Experiments show that,for high-dimensional anomaly detection,the performance of DAE is better than SAE.Compared with traditional OCSVM algorithms,DAE solves the problem of anomaly detection in high dimension.2.In this paper,Deep Belief Networks(DBN)and OCSVM are combined to propose a hybrid model algorithm based on DBN and linear OCSVM for high-dimensional anomaly detection.The DBN,which is a stack of Restricted Boltzmann Machines at the lowest point of potential energy,has a good feature extraction function.After dimensionality reduction,the linear OCSVM is used for anomaly detection.The biggest feature of this algorithm is that by using the data after DBN dimension reduction,OCSVM has the same high accuracy whether it uses linear function or radial basis function as kernel function.While improving the recognition rate of high-dimensional anomaly detection,the use of linear kernel reduces the computational complexity of the algorithm.While improving the recognition rate of high-dimensional anomaly detection,the use of linear kernel reduces the computational complexity of the algorithm.At the same time,compared with AE algorithms,this algorithm is more efficient and the training and testing time are significantly decreased.In this paper,UCI datasets are used to detect the high-dimensional anomalies by using the auto-encoder and the combination algorithm of DNN and OCSVM,and compare them with the traditional anomaly detection algorithm.Experiments consider the following aspects:First,the selection of SVM kernel functions.We use linear function and radial basis function as the kernel functions of OCSVM respectively.Second,the hidden layers of DNN.Explore the impact of the experimental results when the number of hidden layers is different.Third,the efficiency of algorithms.That is,under the same precision,the algorithm with less training time and test time is the best.Experimental results show that the proposed algorithms improve the accuracy of anomaly detection while reducing the dimensionality and reduces the computational complexity of the algorithm.