Unsupervised Anomaly Detection Algorithm Based On One-Classification Support Vector Machine

With the development of modern information,computer,and wireless communication technologies,Wireless Sensor Network(WSN)is widely used in military,medical health,environmental monitoring,industrial inspection,and other fields.However,harsh environments,malicious hacker attacks,and energy limitations of sensor nodes will lead to unreliable sensor data,which further affects the quality of raw data and aggregated results.So anomaly detection of data collected by WSNs is needed to improve data quality and pave the way for further research.In this paper,unsupervised anomaly detection is performed on the data collected by wireless sensor networks based on one-classification support vector machines.The main work of this paper can be divided into two parts as follows.(1)In the anomaly detection algorithm of WSN,it is an open problem to dynamically adjust the hyperparameters of the algorithm in an unsupervised manner to detect anomalies quickly and effectively.In this paper,we divide the dataset into suspicious normal point dataset and suspicious outlier dataset based on KNN algorithm.The suspicious normal point dataset is used for OCSVM training and modeling,and for the suspicious outlier dataset,the D-S evidence theory is utilized to identify the normal data of them.The implementation results shows the DS-SVM algorithm can effectively separate the normal points and the anomalies,the mean AUC value of algorithm is 0.83 on the overall dataset,and 0.883 in the suspicious outlier dataset.(2)In the anomaly detection algorithm of WSN,it is an open problem how to dynamically adjust the hyperparameters of the algorithm in an unsupervised manner and detect the anomalies effectively.In this paper,based on the isolated forest algorithm,we use the isolated forest algorithm to calculate the anomaly scores of the input data and use the change point detection algorithm to identify and remove the suspicious anomalies in the data.For the setting of the hyperparameters of the OCSVM algorithm,this paper estimates the Gaussian kernel hyperparameters by selecting special data points by anomaly scores and minimizing the distance of this data in the high-dimensional space.After removing the suspicious anomalies,the data set is fed into OCSVM to model and predict the unlabeled data.In the experimental phase,we tested several benchmark datasets with different distributions and dimensions,and the experimental results show that the proposed technique statistically outperforms other tested unsupervised one-classification support vector machine anomaly detection algorithms on the test dataset.