Research On Anomaly Detection Algorithm Based On Autoencoder

With the continuous development of artificial intelligence technology,anomaly detection technology has high application value,and it is widely used in intrusion detection,fault detection and security monitoring and other fields.However,in the era of big data,the data generated by information systems is not only large and complex,but also the unpredictable nature of anomalies,making anomaly detection a very challenging task.In recent years,deep learning has achieved great success in anomaly detection.Compared with traditional anomaly detection methods,deep learning can obtain complex patterns in large,high-dimensional data without complex feature engineering.The deep autoencoder is one of the most representative models in anomaly detection.It usually takes the reconstruction of normal samples as the training target,and takes the reconstruction error of samples as the standard to measure the degree of anomaly in the test stage.Recently,it is difficult for the autoencoder to capture the dependence between data.Some researchers solve this problem by saving training information through neural networks with memory ability.Among them,the memory-augmented autoencoder model(Mem AE)has attracted extensive attention.In order to improve the ability of the autoencoder to extract normal patterns,this paper presents memory-augmented adversarial autoencoder model(Mem AAE).Compared with the original model,the discriminator module is added to the model.Mem AAE treats the autoencoder and the memory module as a generator and matches the aggregate posterior distribution of the generator output with the prior distribution,which makes up for the deficiency of autoencoder,improves the reconstruction effect.Experimental results on several text datasets and MNIST show that the proposed model Mem AAE has better performance on the whole.In order to expand the application scope of anomaly detection algorithm based on autoencoder,an improved method for video anomaly detection based on dynamic prototype unit is proposed.Specifically,the autoencoder with u-net structure is regarded as a generator,and adds a discriminator that can distinguish the authenticity of input samples.The improved model is trained by the confrontation training framework of GAN,which improves the prediction ability of the autoencoder for future frames.In order to further improve the identification ability of the model for abnormal frames,in the test phase,the output value of the discriminator is added to the anomaly score based on the reconstruction error of the predicted frame.Through experiments on a public video anomaly detection dataset UCSD pedestrian,the results show that the improved method has higher accuracy than other methods.