한국통신학회 KICS

Popular Articles and Research Trend​

by Taewoon Kim, Pusan National University, South Korea

One effective way to identify the recent research directions and trends is to pay attention to the most popular research papers published recently. This article reviews such most accessed and downloaded papers in Journal of Communications and Networks (JCN) and Information & Communications Technology Express (ICT Express), both of which are published by The Korean Institute of Communications and Information Sciences (KICS).

<JCN>

 <ICT Express>

Machine learning (ML) and deep learning (DL) techniques are continuously advancing and their applications have extended to wireless communications and network security and several other fields. One notable research proposed by Kim et al. showed that deep learning can be an effective solution to the challenge of wireless channel estimation in wireless communication. It is well known that channel adaptive signaling can enhance the performance of wireless communication systems to a large extent. However, obtaining an accurate knowledge of channel state at the transmitter is challenging in practice. To address this challenge, especially for future wireless systems with high-frequency bands, Kim et al. presented a comprehensive study on (i) the choice of an adequate deep learning model, (ii) designing a neural network architecture, and (iii) data acquisition for training the model.

  This article begins with a summary of the conventional and DL-based approaches for wireless channel estimation. It then discusses commonly arising design issues in DL-based channel estimation, such as training data acquisition and design of the neural network architecture. Finally, in the DL-aided Channel Estimation and Exploration section, the authors explore diverse neural network models designed to solve the problems of channel estimation. For example, Fig. 1 shows the proposed LSTM-based parametric channel estimation model for mmWave/THz MIMO communications.

  Figure 1. (a) LSTM-based channel estimation (b) Evaluation results [Kim et al.]

  Furthermore, application of grid search and random forest machine learning model to the problem of intrusion detection in wireless sensor networks (WSN) was proposed by Subbiah et al. WSN are widely used networks dedicated to transmitting sensed readings in diverse applications, such as healthcare and manufacturing. Although many applications of WSN carry private and/or confidential information, they are vulnerable to security attacks because they operate in an unattended environment, communicate with each other in a broadcast manner, and are resource-constrained. Thus, to enhance the level of security of WSNs, an intrusion detection system (IDS) is required to protect legitimate user information from attackers.

  The use of large-scale WSNs over a long period of time results in a huge dataset that can be effectively handled by ML techniques. Many approaches use ML to enhance the performance of IDS; however, the importance of feature selection has been underestimated. Subbiah et al. proposed an effective feature selection algorithm with grid search and random forest to further enhance IDS for WSNs. The widely used KDDCUP dataset with 490 thousand samples and 42 features was used in their study, and the authors proposed two models that enhance the accuracy of IDS, referred to as Model 1 and Model 2. The proposed Model 1, called BFS-GSRF, uses the Boruta feature selection algorithm to select the features and then runs grid search random forest algorithm for detection classification. The detailed operation of BFS-GSRF is illustrated in Fig. 2. The Model 2, on the contrary, uses the Pearson’s correlation coefficient to select features. It then uses linear discriminant analysis, and classification and regression tree to detect attacks accurately.

Figure 2. BFS-GSRF algorithm [Subbiah et al.]

In this issue, we chose two papers from ICT Express that introduce the use of well-known DL models in computer vision applications. Jamtsho et al. proposed the use of YOLO v2 to detect the license plates of motorcycles if the driver is not wearing a helmet.

Figure 3. License plate detection [Jamtsho et al.]

  Fig. 3 shows the workflow of the proposed approach and its detection results. As shown in the flowchart, when a person is detected in a given image, the proposed approach first searches for a helmet. If the helmet is missing in the image, the next step is to find the license plate. Fig. 3(a) shows that license plate detection was not performed, given the presence of a helmet. Contrarily, Fig. 3(b) shows that the license plate detection step is carried out if the helmet is missing. However, the detection of license plate in Fig. 3(b) may be unnecessary if the helmet is found missing only because the helmet is outside the frame boundaries. To overcome such problems, the authors proposed using the centroid-tracking method that detects the center of a detected human object. If the center is above a predefined threshold, as indicated by the green horizontal line in Fig. 3(c), the proposed approach does not attempt to detect the license plate. This study reports that the overall accuracy of license plate detection was 98.52%.

 Figure 4. Soccer action categorization (SAC) scheme [Sen et al.]

​

  The main contributions of this work are twofold: First, it uses a pretrained VGG to extract features from the recorded soccer-playing image dataset. Subsequently, using the GRUs, the authors carry out transfer learning was performed to correctly categorize the player’s play. The extensive evaluation results reported in the study suggest that the proposed approach achieves an accuracy of 94%.