Extreme Value Theory in High Dimensional Statistical Models and its Application in Anomaly Detection

When comparing anomaly detection methods in high-dimensional data environments, PCA and One Class SVM are often associated with high computational complexity and low computational efficiency. This article selects the CIFAR-10 dataset as the experimental object, and uses extreme value theory, principal component analysis, and One Class SVM methods for anomaly detection. Then, the detection research results of 20 classes are analyzed and compared. In order to objectively evaluate and compare the efficiency of these three anomaly detection methods, the recognition time (in milliseconds) was tested. Experimental studies have shown that anomaly detection based on extreme value theory outperforms PCA and One Class SVM methods in recognition time. The reason why this technology has better recognition time is because the extreme value theory anomaly detection method relies on the distribution characteristics of high-dimensional data, which has a relatively simple and effective calculation process, while PCA and One Class SVM anomaly detection methods use complex model training and feature transformation, which increases the calculation time.