散列记忆网络增强的自编码器异常检测方法

Autoencoder Anomaly Detection Method Enhanced by Hash Memory Network

深度自编码器是异常检测的重要工具,通过异常样本由于分布的差异,无法在编码器中进行重构这一假设实现对异常的检测.而实际应用中,由于深度自编码器的泛化性较强,异常输入后也能实现较好重构,导致漏检情况发生.本文在改进注意力机制基础上,构建了一个散列记忆网络增强的自编码器异常检测方法,较好解决了这一问题.首先,模型将输入编码为编码信息,根据编码信息获取子查询向量,然后通过子查询向量获取子注意力权重及对应子索引,再将子权重交叉求和获得散列权重及索引并从记忆网络单元检索出解码信息,最后利用解码信息进行重构输出.重构的输出总是与正常数据相似,使得异常输入与重构输出之间的重构误差将被放大,从而让异常更容易被识别.仿真实验表明,本文提出方法在图像、视频监控、通用异常检测任务中,均取得了较好的检测效果.

Deep autoencoder is an important tool for anomaly detection,which detects anomalies by assuming that they cannot be reconstructed in the encoder due to their distributional differences.However,in practical applications,the generalization ability of deep autoencoders often leads to missed detections because even anomalous inputs can be reconstructed well.In this paper,an improved attention mechanism-based autoencoder with a hash memory network is proposed to address this issue.Firstly,the model encodes the input into encoded information,obtains sub-query vectors based on the encoded information,and then uses the sub-query vectors to obtain sub-attention weights and corresponding sub-indices.The sub-weights are cross-summed to obtain hash weights and indices,and the decoding information is retrieved from the memory network unit.Finally,the decoding information is used for reconstruction output,which is always similar to normal data,and the reconstruction error between the anomaly input and the reconstruction output is amplified,making anomalies easier to identify.Simulation experiments show that the proposed method achieves good detection results in image,video surveillance,and general anomaly detection tasks.