Testbeds and Research Infrastructures for the Development of Networks and Communications

Feature selection is an important preprocessing technique used to determine the most important features that contributes to the classification of a dataset, typically performed on high dimension datasets. Various feature selection algorithms have been proposed for diabetes prediction. However, the effectiveness of these proposed algorithms have not been thoroughly evaluated statistically. In this paper, three types of feature selection methods (Sequential Forward Selection, Sequential Backward Selection and Recursive Feature Elimination) classified under the wrapper method are used in identifying the optimal subset of features needed for classification of the Pima Indians Diabetes dataset with an Artificial Neural Network (ANN) as the classifying algorithm. All three methods manage to identify the important features of the dataset (Plasma Glucose Concentration and BMI reading), indicating their effectiveness for feature selection, with Sequential Forward Selection obtaining the feature subset that most improves the ANN. However, there are little to no improvements in terms of classifier evaluation metrics (accuracy and precision) when trained using the optimal subsets from each method as compared to using the original dataset, showing the ineffectiveness of feature selection on the low-dimensional Pima Indians Diabetes dataset.

Monitoring and estimation of food intake is of great significance to health-related research, such as obesity management. Traditional dietary records are performed in manual way. These methods are of low efficiency and a waste of labor, which are highly dependent on human interaction. In recent years, some researches have made progress in the estimation of food intake by using the computer vision technology. However, the recognition results of these researches are usually for the whole food object in the image, and the accuracy is not high. In terms of this problem, we provide a method to the food smart recognition and automatic dietary assessment on the mobile device. First, the food image is processed by MASK R-CNN which is more efficient than traditional methods. And more accurate recognition, classification and segmentation results of the multiple food items are output. Second, the OpenCV is used to display the food category and the corresponding food information of unit volume on the recognition page. Finally, in order to facilitate daily use, TensorFlow Lite is used to process the model to transplant to the mobile device, which can help to monitor people’s dietary intake.

The micro-blog texts of the national grid provinces and cities will be analyzed as the main data, including the micro-blogs and corresponding comments, which will help us understand the events of power industry and people’s attitudes towards these events. In this work, the data set is composed of 420,000 micro-blog texts. Firstly, the professional vocabulary of electric power is extracted, and these vocabulary are manually labeled, thus proposing a new field dictionary closely related to the power industry. Secondly, using the new power domain dictionary to classify the 2018 electric micro-blogs, and we can find that classification accuracy increased from 88.7% to 95.2%. Finally, a classification model based on LSTM (Long Short-Term Memory) and RNN (Recurrent Neural Network) is used to deal with the comments under the micro-blog. The experimental result shows that the classification of the LSTM-RNN is more accurate. The rate was 83.1%, which was significantly better than the traditional LSTM and RNN text classification models of 78.4% and 73.1%.

The encrypting ransomware using public key cryptography is almost impossible to decrypt, so early detection and prevention is more important. Signature matching technology has low detection rate for unknown or polymorphic ransomware, and some intelligent algorithms have been proposed for solving this problem. Inspired by the Artificial Immune System (AIS), an improved double-layer negative selection algorithm (DL-NSA) was proposed which can reduce the number of holes in NSA and increase the detection rate. To obtain the behavior characteristics (e.g., files read or write, cryptography APIs call and network connection) of ransomware, a Cuckoo sandbox was built to simulate the malicious code running environment. After dynamic analysis, the behavior characteristics of ransomware were encoded to antigens. The improved double-layer negative selection algorithm has two sets of immune detectors. The first layer detectors set was generated by the original negative selection algorithm using r-contiguous bits matching. The second layer detectors set was directional generated holes’ detectors using r-chunk matching with variable matching threshold. Simulation result shows that comparing with NSA this algorithm can achieve high-rate space coverage for non-self, and can increase the detection rate of ransomware.

Simulating large-scale network experiments requires powerful physical resources. However, partitioning could be used to reduce the required power of the resources and to reduce the simulation time. Topology mapping is a partitioning technique that maps the simulated nodes to different physical nodes based on a set of conditions. In this paper, genetic algorithm-based mapping is proposed to solve the topology mapping problem. The obtained results prove a high reduction in simulation time, in addition to high utilization of the used resources (The number of used resources is minimum).

Traditional SDN has one controller, but more recent SDN approaches use multiple controllers on one network. However, the multiple controllers need to be synchronized with each other in order to guarantee a consistent network view, and complicated control management and additional control overhead are required. To overcome these limitations, Kandoo [5] has been proposed in which a root controller manages multiple unsynchronized local controllers. However, in this approach, loops can form between the local controllers because they manage different topologies. We propose a method for modeling a hierarchical design to detect loops in the topology and prevent them from occurring using UPPAAL model checker. In addition, the properties of multiple controllers are defined and verified based UPPAAL framework. In particular, we verify the following properties in a multiple controller: (1) elephant flows go through the root controller, (2) all flows go through the switch that is required to maintain security, and (3) they avoid unnecessary switches for energy efficiency.

This paper presented an energy management model for managing an active distribution network (ADN) consisting of multiple microgrids. The distribution system operator (DSO) of the ADN needs to coordinate the microgrids to achieve optimal energy management. This paper formulated the energy management of ADN with multiple microgrids as a mixed integer second-order cone programming (MISOCP), which considered network reconfiguration, on-load tap changer (OLTC) and static Var compensators (SVC). A case study on a modified IEEE 33-bus distribution network demonstrates the effectiveness of the proposed method.

A bivariate Fisher–Snedecor \( {\mathcal{F}} \) composite distribution with arbitrary fading parameters (not necessary identical) is presented in this paper. We derive novel theoretical formulations of the statistical characteristics for the correlated \( {\mathcal{F}} \) composite fading model, which include the joint probability density function, the joint cumulative distribution function, the joint moments and the power correlation coefficient. Capitalizing on the joint cumulative distribution function, the bit error rate for binary digital modulation systems and the outage probability of a correlated dual-branch selection diversity system, and the level crossing rate and the average fade duration of a sampled Fisher-Snedecor \( {\mathcal{F}} \) composited fading envelope are obtained, respectively. Finally, we employ numerical and simulation results to demonstrate the validity of the theoretical analysis under various correlated fading and shadowing scenarios.

Designing, coordinating and deploying repeatable experiments outside of a fully controlled environment pose a serious challenge when conducting network research. In particular, it can be difficult to correctly schedule experiments that collect bulk data using a shared resource. To address this problem, we introduce PLDetect, a simple testbed built on top of PlanetLab which simplifies configuring, scheduling, and deploying large scale Internet experiments for evaluating middlebox detection methods.

The most recent evolutionary steps in the development of mobile communication network architectures have introduced the concepts of virtualisation and slicing also into the Radio Access Network (RAN) part of the overall infrastructure. This trend has made RANs more flexible than ever before, facilitating resource sharing concepts which go far beyond the traditional infrastructure and RAN sharing schemes between commercial Mobile Network Operators (MNO). This paper introduces the EuWireless concept for a pan-European mobile network operator for research and presents its vision for RAN slicing and network resource sharing between the infrastructures of the EuWireless operator, commercial MNOs and research organisations around Europe. The EuWireless approach is to offer virtual large-scale testbeds, i.e., EuWireless experimentation slices, to European mobile network researchers by combining the experimental technologies from the local small-scale research testbeds with the commercial MNO resources such as licensed spectrum. The combined resources are configured and managed through the distributed EuWireless architecture based on inter-connected local installations, so-called Points of Presences (PoP).

In recent years, crowdsourced testing, which uses collective intelligence to solve complex software testing tasks has gained widespread attention in academia and industry. However, due to a large number of workers participating in crowdsourced testing tasks, the submitted test reports set is too large, making it difficult for developers to review test reports. Therefore, how to effectively process and integrate crowdsourced test reports is always a significant challenge in the crowdsourced testing process. This paper deals with the crowdsourced test reports processing, sorts out some achievements in this field in recent years, and classifies, summarizes, and compares existing research results from four directions: duplicated reports detection, test reports aggregation and classification, priority ranking, and reports summarization. Finally explored the possible research directions, opportunities and challenges of the crowdsourced test reports.

5G networks are expected to consist of multiple radio access technologies with a Software-defined networking (SDN) core, and so simulating these networks will require connecting multiple subnetworks with different technologies. Despite the availability of simulators for various technologies, there is currently no tool that can simulate a complete heterogeneous 5G network. In this work, we develop a novel SDN adapter to enable seamless inter-working between different simulation/emulation tools, such as NS-3, Mininet-WiFi, Omnet++, and OpenAirInterface5G. Using the adapter, we have built a large scale 5G simulator with multiple networking technologies by connecting existing simulators. We show that our adapter solution is easy-to-use, scalable, and can be used to connect arbitrary simulation tools. Using our solution, we show that Mininet-WiFi exhibits unreliable behaviour when connected to other networks. We compare our solution against other alternatives and show that our solution is superior both in terms of performance and cost. Finally, and for the first time, we simulate a large heterogeneous 5G network with all of the latest technologies using only a standard commodity personal computer.

Feature selection is an important step. It aims to filter some irrelevant features, improve the classifier speed and also reduce the interference during text classification process. Information gain (IG) feature selection algorithm is one of the most effective feature selection algorithms. But it is easy to filter out the characteristic words which have a low IG score but have a strong ability of text type identification. Because IG algorithm only considers the number of documents of feature items in each category. Aiming at this defect, we propose an improved information gain algorithm by introducing three parameters: intra-class word frequency, inter-class separation degree and intra-class dispersion degree. Then, the improved IG algorithm is used for feature selection, and important feature words with high IG value are selected according to the threshold value. Final, the important feature words in the text are expressed as two-dimensional word vectors and input into Convolutional Neural Network (CNN) to train and classify them. Therefore, a text classification model based on improved information gain and convolutional neural network is proposed and abbreviated as “I-CNN”. Through experiments, we achieve good experimental results in THUCNews Chinese text classification corpus. Experimental results prove that the improved IG algorithm is better than the traditional feature selection algorithm.

Emotions are physiological phenomena caused by complex cognitive activities. With the in-depth study of artificial intelligence and brain mechanism of emotion, affective computing has become a hot topic in computer science. In this paper, we used the existed emotional classification model based on electroencephalograph (EEG) to calculate the accuracy of emotion classification in 4 brain areas roughly sorted into frontal, parietal, occipital, and temporal lobes in terms of brain functional division, to infer the correlation between the emotion and 4 brain areas based on the accuracy rate of the emotion recognition. The result shows that the brain areas most related to emotions are located in the frontal and temporal lobes, which is consistent with the brain mechanism of emotional processing. This research work will provide a good guideline for selecting the most relevant electrodes with emotions to enhance the accuracy of emotion recognition based on EEG.

Based on the Wi-Fi widely separated in the world, Wi-Fi-based wireless activity recognition has attracted more and more research efforts. Now, device-based activity awareness is being used for commercial purpose as the most important solution. Such devices based on various acceleration sensors and direction sensor are very mature at present. With more and more profound understanding of wireless signals, commercial wireless routers are used to obtain signal information of the physical layer: channel state information (CSI) more granular than the RSSI signal information provides a theoretical basis for wireless signal perception. Through research on activity recognition techniques based on CSI of wireless signal and deep learning, the authors proposed a system for learning classification using deep learning, mainly including a data preprocessing stage, an activity detection stage, a learning stage and a classification stage. During the activity detection model stage, a correlation-based model was used to detect the time of the activity occurrence and the activity time interval, thus solving the problem that the waveform changes due to variable environment at stable time. During the activity recognition stage, the network was studied by innovative deep learning to conduct training for activity learning. By replacing the fingerprint way, which is used broadly today, with learning the CSI signal information of activities, we classified the activities through trained network.

Mysterious and uncertain deaths in the “Game of Thrones” novel-series have been stupefying to the vast pool of readers and hence interested researchers to come up with various models to predict the deaths. In this paper, we propose a Death-Prone Score model to predict if the candidate character is going to die or stay alive in the upcoming book in the series. We address the challenge of high-dimensional data and train our model on the most significant attributes by computing feature importance in the vector space. Further, we address the challenge of multiple interactions between characters and create a social network representing the weighted similarity between each character pair in the book. The proposed model takes similarity and proximity in a social network into account and generates a death-prone score for each character. To evaluate our model, we divide the characters data into training (characters died before year 300) and testing (characters died in the year 300 and characters alive till year 300). Our results show that the proposed Death-Prone Score model achieves an f-score of 86.2%.