Deep neural networks (DNNs) have widely used in various Internet-of-Things (IoT) applications. Pursuing superior performance is always a hot spot in the field of DNN modeling. Recently, feature reuse provides an effective means of achieving favorable nonlinear approximation performance in deep learning. Existing implementations utilizes a multilayer perception (MLP) to act as a functional unit for feature reuse. However, determining connection weight and bias of MLP is a rather intractable problem, since the conventional back-propagation learning approach encounters the limitations of slow convergence and local optimum. To address this issue, this paper develops a novel DNN considering a well-behaved alternative called reservoir computing, i.e., reservoir in network (ResInNet). In this structure, the built-in reservoir has two notable functions. First, it behaves as a bridge between any two restricted Boltzmann machines in the feature learning part of ResInNet, performing a feature abstraction once again. Such reservoir-based feature translation provides excellent starting points for the following nonlinear regression. Second, it serves as a nonlinear approximation, trained by a simple linear regression using the most representative (learned) features. Experimental results over various benchmark datasets show that ResInNet can achieve the superior nonlinear approximation performance in comparison to the baseline models, and produce the excellent dynamic characteristics and memory capacity. Meanwhile, the merits of our approach is further demonstrated in the network traffic prediction related to real-world IoT application.

Video services have hold a surprising proportion of the whole network traffic in wireless communication networks. Accurate prediction of video traffic can endow networks with intelligence in resource management, especially for the forthcoming beyond the fifth-generation (B5G) networks. However, the existing approaches fail to accurately predict video traffic with all types of frames, due to the natures of strong long-range dependence, self-similarity and burstiness. Obviously, it is unable to meet the QoS and QoE requirements of dynamic bandwidth allocation. In this paper, we propose the feasibility of advanced machine learning methodology applied in nonstationary video traffic prediction, i.e., smoothing-aided support vector machine (SSVM) model. The model utilizes classical smoothing methods to preprocess video traffic by relieving the drastic fluctuation of video stream. It can provide an effective association for the subsequent support vector regression, as the preprocessed data becomes more smooth and continuous than the original unprocessed one. Experimental results show that our proposed model significantly outperforms the state of the art model, i.e., logistic smooth transition autoregressive, in prediction performance. The superior nonlinear approximation capacity is further demonstrated by visualized statistical analysis.

Network traffic prediction is a great challenge due to complex statistical properties, generally covering the long-range correlations and self-similarity. To address this issue, this paper applies an integrated neural computing model to predict network traffic, namely enhanced echo-state restricted Boltzmann machine (eERBM). In structure, this model possesses the following functional components of feature learning, information compensation, input superposition and supervised nonlinear approximation. It is motivated by the introduction of information theory in modeling the hybrid architecture of echo state network and restricted Boltzmann machine. This is a first attempt that eERBM is applied in network traffic prediction tasks of different origin and characteristics, considering TCP/IP packet and variable-bit-rate video. By performing a theoretical analysis we show that eERBM achieves superior nonlinear approximation and robustness in comparison to baseline methods, and effectively preserves the self-similarity of network traffic traces.

Seawater quality prediction has a tremendous potential of enabling future smart ocean. However, this time-sensitive application puts forward a strict delay requirement, thus easily leading to overwhelmed networks. Edge computing is emerging as an effective means of solving network overload, due to its edge-based distributed processing. Therefore, we develop a hybrid multivariate prediction model for seawater quality assessment in an edge computing environment, considering the combination of principal component analysis (PCA) and relevance vector machine (RVM). The PCA method is employed for dimension reduction of ten seawater quality factors in advance. Six principal components are extracted from multiple features, used as input variables of the subsequent predictor. Finally, a RVM is developed to predict the future trends of dissolved oxygen and pH, measuring seawater quality. Experimental results on the real-world ocean sensor data show that our PCA-RVM based multivariate prediction model outperforms single RVM, SVM and its extended version in prediction accuracy and efficiency, meanwhile statistical testings confirm this finding.

Voice user interface (VUI) brings high efficiency and convenience for the applications of Internet of Things (IoT), meanwhile, it can also cause increasingly serious security issues. The word-level voice liveness detection is proved to be the promising solution to thwart spoofing attacks. However, the complex acoustic feature, diversified attacks, and different interaction distance can severely affects the improvement of detection accuracy. To alleviate this issue, we develop a novel pop noise-based word-level voice liveness detection framework. First, a new voice frame selection method is proposed for determining optimal frames, including short time Fourier transform, low-frequency average energy computation, and sequencing. Then, the acoustic features of the selected frames are calculated by the Gammatone frequency cepstral coefficient (GFCC). Finally, based on these features, a newly built joint voice detector, fusing the self-attentional residual network (ResNet), and light gradient boosting machine (LightGBM), can achieve accurate voice classification. On the popular voice spoofing attack data sets, experimental results show that our proposal significantly outperforming the baseline and the state-of-the-arts models, and it is gender dependent. Moreover, our proposal has good generalization ability for far-field replay voice attack, speech synthesis and voice conversion attacks, and partial fake voice attack. Finally, its effectiveness is verified by the ablation study.