Research On Anomaly Detection Algorithm Based On Feature Adaptive Optimization

In the era of information society,data carries a lot of important information and has huge economic value.However,data also contains potential risks while serving human beings,which will bring huge security risks to production.As an effective protection method,anomaly detection can detect abnormal data that deviates from normal expected behavior,and provide important support for the normal operation of various systems.Therefore,researches on efficient anomaly detection algorithm have practical significance in many fields.In this thesis,the relevant theoretical knowledge of anomaly detection is introduced firstly.Then,under the condition that traditional anomaly detection methods are not effective for high-dimensional data,this thesis studies the anomaly detection algorithm combined with deep learning,an anomaly detection algorithm based on variational autoencoder with Gaussian mixture model and an anomaly detection algorithm based on sparse variational autoencoder is proposed.The main contents and achievements of this thesis are listed as follows:1.An anomaly detection model based on variational autoencoder with Gaussian mixture model is proposed in this thesis.First of all,the Gaussian mixture distribution is used as the prior distribution and approximate posterior distribution of the variational autoencoder in this method,which makes the hidden space more flexible and avoids the problem of posterior collapse caused by the too simple prior distribution which would influence the representation effect of features.Then a deep support vector data description network is built on the basis of the constructed variational autoencoder coding network,which could compresses the feature space and find the best hypersphere to separate normal data and abnormal data,then the Euclidean distance from the feature of data to the center of the hypersphere is calculated to measure the anomaly score of the data to detect anomalies.Finally,the algorithm is evaluated on the benchmark datasets MNIST and Fashion-MNIST,and achieves preferable effects.2.Considering that sparse coding can obtain more powerful features which could improve the effect when performing other tasks,an anomaly detection model based on sparse variational autoencoder is proposed.The discrete mixed model based on spike and slab distribution prior is used as the prior of the variational autoencoder to simulate the sparsity of the hidden space,and obtains the sparse representation of data features.Then a deep support vector data description network is built to realize anomaly detection.Finally,the algorithm is evaluated on the benchmark datasets MNIST and Fashion-MNIST.The experimental results show that this algorithm has better results than the algorithm proposed in Chapter 3 and other state-of-the-art algorithms.