Multimodal data consistency anomaly detection based on sparse subspace clustering

The consistency anomaly of multimodal data during transmission is a challenge faced by social networks, intelligent monitoring, and other applications. These existing research methods of multimodal data consistency are based on supervised training and learning, which require a large amount of training data and high computing power. To overcome the problem, a multimodal consistency anomaly detection algorithm based on sparse subspace clustering is proposed in this paper. First, to solve the problem that the performance of existing multimodal anomaly detection algorithms is usually highly correlated with the amount of anomalous data and has a limited application, a sparse representation matrix is obtained by using the sparse subspace clustering algorithm combined with correlation entropy for each data modality. Then, a joint sparse representation matrix of multimodal data is obtained by combining all sparse representation matrices, and the highdimensional multimodal data are mapped into the lowdimensional joint space. Next, the above matrices are used to cluster the multimodal data into their respective categories using spectral clustering methods to detect the consistent anomalous data. It is noted that the proposed algorithm is based on the unsupervised learning mode, which can reduce the training set data and decrease the model complexity, and then introduces the correlation entropy into subspace clustering for multimodal data processing, which improves the robustness of the algorithm and can effectively detect the consistency anomalous data. Finally, compared with five typical clustering algorithms, experimental results show that the proposed algorithm can improve the detection accuracy of 0.53%~4.88% and the normalized mutual information performance of 3.67%~17.22% for different degrees of the inconsistent multimodal data.