结合分层深度网络与双向五元组损失的跨模态异常检测

Cross-Modal Anomaly Detection via Hierarchical Deep Networks and Bi-Quintuple Loss

大数据环境下的跨模态异常检测是一个非常有价值且极具挑战性的工作.针对目前已有跨模态异常检测框架对数据异常值类型检测不全面以及数据利用率较低的问题,提出了一个结合分层深度网络与相似度双向五元组损失的跨模态异常检测方法.首先,提出的框架引入一个单视图异常检测网络层,通过模态内近邻样本相似度来检测数据样本中是否存在属性异常与部分属性类别异常点;接着,提出基于相似度双向五元组损失的双分支深度网络用于检测数据中的类别异常与剩余部分的属性类别异常,该损失一方面能够使不同属性数据正交化,另一方面使得相同属性数据之间线性相关,从而有效地加大了不同属性数据之间的特征差异性,以及增加了相同属性之间的特征相关性;同时,提出的双分支网络通过模态间双向约束和模态内的邻域约束,极大提高了数据利用率和模型的泛化能力.实验结果表明,所提出的框架可以全面检测出不同模态中所有的异常类型样本点,并且表现优于现有的可应用于跨模态异常检测的方法,优势明显.

Cross-modal anomaly detection in big data environment is a very valuable and challenging work.Existing cross-modal anomaly detection framework often suffers from the incomplete data anomaly detection and low data utilization problems.To alleviate these concerns,an efficient cross-modal anomaly detection framework is proposed via hierarchical deep networks and similarity based bi-quintuple loss.First,the proposed framework introduces a single-view anomaly detection network to detect the attribute anomaly and part of class-attribute anomaly in data samples.Then,the similarity bi-quintuple loss,integrated with double-branch deep networks,is efficiently developed to detect the class anomaly and the remaining part class-attribute anomaly in data samples.Meanwhile,this loss regularizes the different attribute data with orthogonal property,and ensures the linear correlation between the same attribute data,enlarges the feature difference between different attribute data and increases the feature correlation between the same attribute data.Meanwhile,the bidirectional constraint and neighborhood constraint can significantly improve the data utilization and the generalization ability of the model.Extensive experimental results show that the proposed framework is able to detect possible abnormal sample points in different modalities,and outperforms the state-of-the-art corresponding methods,with obvious advantages.