融合互信息估计和对抗自编码器的异常检测

无监督深度学习网络的训练目标从信息论的角度可解释为最大化训练样本及其表示之间的互信息.对抗自编码器（AAE）通过生成对抗的方式学习训练样本集的分布，据此可以由AAE建立基于正常样本集的半监督异常检测模型，但是AAE无法显式最大化正常样本及其表示间的互信息.为此，提出了一种互信息估计网络和AAE相融合（IAAE）的异常检测方法，该方法首先以重构误差最小化为目标，训练编码器和解码器；其次，在对抗正则化阶段将正常样本低维表示的聚集后验分布约束为先验分布，并最大化正常样本与其表示之间的互信息；最后由全连接神经网络估计正常样本与其表示之间的互信息.由待测样本的重构误差及其表示在隐空间中的众数散度计算其异常得分值.公开数据集上的实验结果表明，与已有典型相关的深度异常检测模型相比，IAAE模型在F1取值上具有更好的表现.

An Anomaly Detection Method Combining Mutual Information Estimation with Adversarial Autoencoder

According to the information theory,the training objective of the unsupervised deep learning networks can be interpreted as maximizing the mutual information between the training samples and their representations. Adversarial autoencoder (AAE) learns the distribution of the training samples by the generative adversarial method. So the semi-supervised anomaly detection model based on the normal sample sets can be established using AAE. However,AAE cannot maximize the mutual information between the normal samples and their representations explicitly. A semi-supervised anomaly detection method based on mutual information estimation network and AAE (IAAE) is proposed. Firstly,the encoder and decoder of the AAE are trained to minimize the reconstruction error. Then,in the adversarial regularization stage,the aggregated posterior of the normal sample’s representations are matched to the arbitrary prior distribution,and the mutual information between normal samples and their representations is maximized. Finally,the mutual information between normal samples and their representations are estimated by fully connected neural network. The reconstruction error of the test sample and its mode divergence in the hidden space are used to calculate the abnormal score. The experimental results on public datasets show that the IAAE has better performance than the existing typical deep anomaly detection models in terms of F1 values.