No abstract

High-resolution range profile (HRRP) has obtained intensive attention in radar target recognition and convolutional neural networks (CNNs) are among predominant approaches to deal with HRRP recognition problems. However, most CNNs are designed by the rule-of-thumb and suffer from much more computational complexity. Aiming at enhancing the channels of one-dimensional CNN (1D-CNN) for extracting efficient structural information oftargets form HRRP and reducing the computation complexity, we propose a novel framework for HRRP-based target recognition based on 1D-CNN with channel attention and channel pruning. By introducing an aggregationperception-recalibration (APR) block for channel attention to the 1D-CNN backbone, channels in each 1D convolutional layer can adaptively learn to recalibrate the extracted features for enhancing the structural information captured from HRRP. To avoid rule-ofthumb design and reduce the computation complexity

In the field of intrusion detection, there is often a problem of data imbalance, and more and more unknown types of attacks make detection difficult. To resolve above issues, this paper proposes a network intrusion detection model called CWGAN-CSSAE, which combines improved conditional Wasserstein Generative Adversarial Network (CWGAN) and cost-sensitive stacked autoencoders (CSSAE). First of all, the CWGAN network that introduces gradient penalty and L2 regularization is used to generate specified minority attack samples to reduce the class imbalance of the training dataset. Secondly, the stacked autoencoder is used to intelligently extract the deep abstract features of the network data. Finally, a cost-sensitive loss function is constructed to give a large misclassification cost to a minority of attack samples. Thus, effective detection of network intrusion attacks can be realized. The experimental results based on KDDTest+, KDDTest-21, and UNSW-NB15 datasets show that the CWGAN-CSSAE network intrusion detection model improves the detection accuracy of minority attacks and unknown attacks. In addition, the method in this paper is compared with other existing intrusion detection methods, excellent results have been achieved in performance indicators such as accuracy and F1 score. The accuracy on the above datasets reached 90.34%, 80.78% and 93.27% respectively. The accuracy of U2R on the KDDTest+ and KDDTest-21 datasets both reached 42.50%. The accuracy of R2L on the KDDTest+ and KDDTest-21 datasets reached 54.39% and 52.51%, respectively. And the F1 score on the above datasets reached 91.01%, 87.18% and 93.99% respectively.

In order to prevent the overfitting and improve the generalization performance of Extreme Learning Machine (ELM), a new regularization method, Biased DropConnect, and a new regularized ELM using the Biased DropConnect and Biased Dropout (BD-ELM) are both proposed in this paper. Like the Biased Dropout to hidden nodes, the Biased DropConnect can utilize the difference of connection weights to keep more information of network after dropping. The regular Dropout and DropConnect set the connection weights and output of the hidden layer to 0 with a single fixed probability. But the Biased DropConnect and Biased Dropout divide the connection weights and hidden nodes into high and low groups by threshold, and set different groups to 0 with different probabilities. Connection weights with high value and hidden nodes with a high-activated value, which make more contribution to network performance, will be kept by a lower drop probability, while the weights and hidden nodes with a low value will be given a higher drop probability to keep the drop probability of the whole network to a fixed constant. Using Biased DropConnect and Biased Dropout regularization, in BD-ELM, the sparsity of parameters is enhanced and the structural complexity is reduced. Experiments on various benchmark datasets show that Biased DropConnect and Biased Dropout can effectively address the overfitting, and BD-ELM can provide higher classification accuracy than ELM, R-ELM, and Drop-ELM.

In order to exploit the potential intrinsic low-dimensional structure of the high-dimensional data from the manifold learning perspective, we propose a global graph embedding with globality-preserving property, which requires that samples should be mapped close to their low-dimensional class representation data distribution centers in the embedding space. Then we propose a novel local and global graph embedding auto-encoder(LGAE) to capture the geometric structure of data, its cost function have three terms, a reconstruction loss to reproduce the input data based on the learned representation, a local graph embedding regularization to enforce mapping the neighboring samples close together in the embedding space, a global embedding regularization to enforce mapping samples close to their low-dimensional class representation distribution centers. Thus in the learning process, our LGAE can map samples from same class close together in the embedding space, as well as reduce the scatter within-class and increase the margin between-class, it will also detect the local and global intrinsic geometric structure of data and discover the latent discriminant information in the embedding space. We build stacked LGAE for classification tasks and conduct comprehensive experiments on several benchmark datasets, the results confirm that our proposed framework can learn discriminative representation, speed up the network convergence process, and significantly improve the classification performance. INDEX TERMS Manifold learning, graph embedding, discriminative auto-encoder, deep learning.