This paper is devoted to an actual concept of “Green IT” with the background of the “hardcore” resource consuming applications. The given work discusses the energy efficient constructions for computing centres and clusters, which provide optimisation of performance as well as PW, PUE and ERE factors. The following both approaches are examined: 1) workload consolidation and intelligent VM migration; 2) optimisation of the cooling process (guaranteeing PUE) and energy recycling (guaranteeing ERE and CO\(_2\) footprint reduction). The optimal ERE within the data centres and computing clusters provides advanced resource-consuming distributed applications like pharmacology simulations, Big Data analytics, Blockchain, crypto-currency mining.
Image stitching is one of the relevant operations useful in virtual reality as well as in remote sensing applications which is used for the generation of panoramic images. The quality of the obtained panoramic images may be decreased due to the presence of various factors, e.g., geometric distortions, blur, ghosting, or colour distortions. Since the distortion types specific for the stitched images are different than those that may be found in the databases used for general-purpose image quality assessment, the development of new quality metrics, as well as the testing image databases, are necessary. One of the recently proposed methods for the evaluation of the stitched images is based on the comparison of 36 features of the constituent and finally stitched images, including shape parameters of the Generalized Gaussian Distribution (GGD) and the eigenvalues of a bivariate distribution obtained from the Gaussian Mixture Model (GMM). Although this method, known as Stitched Image Quality Evaluation (SIQE), provides a high correlation with subjective scores for the ISIQA dataset developed by its authors, its verification has been originally made assuming the use of only 20% of images for testing and 80% for training. Nevertheless, a noticeably smaller correlation with subjective quality scores obtained for the whole dataset may be meaningfully increased, employing an additional image entropy analysis proposed in the paper. The obtained results are encouraging and the proposed extension of the SIQE metric increases significantly both the Pearson’s Linear Correlation Coefficient and Rank-Order Correlations with Mean Opinion Scores provided in the ISIQA database.
This paper analyzes the problem of remote text entry using a gyroscope. Its rotary movements are mapped to the screen coordinate system and employed for a virtual laser pointer behaviour. We analyze different placement for the unit and choose held in a hand smartphone with a buil-in gyroscope for further investigations. In the study, the analysis of different modes of text entry is examined, including simple point and click procedure on different layouts, and trajectory-based stroke or gesture typing.
In the paper, decision trees and Bayes classifiers are applied to detect malware traffic in computer networks and compared with the state-of-the-art convolutional neural network reference model. The experiments have been conducted using the known USTC-TFC2016 data set. The results obtained are close to 100% accuracy and are better than the original results produced by the reference method.
Dynamic programming of bioinformatics codes requires efficient programming on multi- and many-core platforms. In this paper, we propose an implementation of Nussinov’s RNA folding using the innovative Kokkos framework. Kokkos Core allows us to quickly code parallel programs, including bioinformatics tasks, for major HPC platforms. The Kokkos abstracts enable multi-threaded patterns like Posix Threads, OpenMP, or CUDA without significant code modifications. Our implementation consists of two steps. First, we parallelize Nussinov’s RNA folding code using well-known compilers like Traco and PLUTO. Second, we re-implement manually output to the Kokkos C++ modern and multi-platform code using templates and lambdas. We explain and practice the framework features for memory and threads management in terms of RNA folding programming. We carry out experiments to study the performance of Nussinov’s codes within Kokkos. We analyze speed-up as a ratio of serial run-time to the time taken by the parallel codes for various RNA sequences lengths. Related and future work are outlined.
Modeling through Boolean encoding is often used for the investigation of various kinds of algorithms or other computer systems. Such an approach can be used successfully to cryptanalysis of symmetric ciphers. In this case, cipher testing begins with encoding the cipher algorithm into a Boolean propositional formula. Then the randomly selected bits representing plaintext and cryptographic key are encoded as formulas too. Specially dedicated programs, called SAT solvers, which can solve an SAT problem, can then compute ciphertext values for such inputs. The final cryptanalysis is an attempt at computing a key value from plaintext and ciphertext. In this way, we perform cryptanalysis with plaintext and ciphertext. In this paper, we show how SAT techniques behave for the Salsa20 cipher which is one of a stream cipher widely used in many security systems of computer networks. In our work, we compared a number of selected SAT solvers. Some are relatively old but still very efficient and some are modern and popular.
Thresholding of document images is one of the most relevant operations that influence the final results of their further analysis. Although many image binarization methods have been proposed during recent several years, starting from global thresholding, through local and adaptive methods, to more sophisticated multi-stage algorithms and the use of deep convolutional neural networks, proper thresholding of degraded historical document images is still an open challenge. Due to the release of the recent challenging DIBCO 2019 dataset, containing two categories of images with the particularly demanding papyri dataset (track B), most of the earlier proposed methods turned out to be significantly less effective for them. Analysing the competition’s results of the DIBCO 2019, apart from the variety of submitted trendy methods based on the use of deep learning, the great importance of image preprocessing may be observed for such demanding images. Hence, in this paper, we present the applicability analysis of the previously proposed preprocessing methods based on the use of the GGD and GMM in combination with various image binarization algorithms for this purpose. The obtained results are comparable with the state-of-the-art methods, including the top algorithms submitted to the recent DIBCO series contest.
With the advancement of internet technologies, network traffic monitoring and cyber-attack detection are becoming more and more important for critical infrastructure. Unfortunately, there are still relatively few works in the literature that interpret the available benchmark data as data streams and take into account the dynamic characteristics of network packet tracking. The following work proposes an approach to generating data streams from the IoT-23 dataset and evaluates the resulting streams based on their suitability for use in network intrusion detection research.
The paper introduces (and investigates) a novel methodology for improving performances of detection of visual objects (keypoints are selected as objects for the feasibility study). Using any (i.e. either hand-crafted or deep-learned) detector in conjunction with the selected descriptor, we rank the extracted keypoints using stability of their descriptors. Stability is defined by the magnitude of partial derivatives of the descriptor over parameters determining deformations of the original images. In practice, discrete approximations of the derivatives are used. We have preliminarily verified that by using the keypoints top-ranked by the proposed criterion (i.e. those with least-sensitive descriptors) instead of keypoints top-ranked by the standard prominence criteria, performances of matching (we primarily focus on precision) can be substantially improved. The results have been obtained using a popular benchmark set of images.
The purpose of this study was to review and implementation of different forms of data extraction and their classification of classical music styles in a machine learning process. The most popular parameters describing audio signals were considered: spectrogram, melo-scale spectrogram and Mel-Frequency Cepstral Coefficients (MFCC). The influence of the machine learning algorithm on the obtained effectiveness of classical music style classification. In this paper, we analyze CNN (Convolutional Neural Network) and Recurrent Neural Network (RNN). All proposed algorithms were implemented in Python with various libraries to support the work with neural networks and audio signals such as TensorFlow, Keras, Librosa.
Many dynamic ensemble selection (DES) methods are known in the literature. A previously-developed by the authors, method consists in building a randomized classifier which is treated as a model of the base classifier. The model is equivalent to the base classifier in a certain probabilistic sense. Next, the probability of correct classification of randomized classifier is taken as the competence of the evaluated classifier. In this paper, a novel randomized model of base classifier is developed. In the proposed method, the random operation of the model results from a random selection of the learning set from the family of learning sets of a fixed size. The paper presents the mathematical foundations of this approach and shows how, for a practical application when learning and validation sets are given, one can determine the measure of competence and build a MC system with the DES scheme. The DES scheme with the proposed model of competence was experimentally evaluated on the collection of 67 benchmark datasets and compared in terms of eight quality criteria with two ensemble classifiers which use the previously-proposed concepts of randomized model. The proposed approach achieved the lowest ranks for almost all investigated quality criteria.
The human spine undergoes degenerative changes with age. One of the symptoms of aging is osteoporosis, which is manifested by loss of bone mass and reduced density of trabeculae. Previous studies have shown that some image features are correlated with age. In this research, image features’ values were obtained using textural analysis. CCTA images of the thoracic vertebra of patients of different ages were analyzed using qMaZda software. The correlation between features’ values and the age of the person was determined using Spearman and Pearson coefficients. The research was carried out on two groups: a group consisting of male and female patients, and a group consisting of only women. In the entire research group, the image attributes derived from GRLM and GLCM showed the highest correlation with age. In the female group, GLCM features were the most correlated with age; moreover, the values of the correlation’s strength were noticeably higher than the highest values obtained in the group of both sexes. The influence of sex and image parameters on the relationship between age and textural features is discussed.
This paper proposes a model of interaction of steganographic systems, which formalizes the description of the principles of construction of steganographic telecommunications. This model is based on the principle of decomposition of the OSI model and includes seven levels: application level, representation level, stegosession level, stegotransport level, stegonetwork level, stegochannel level and stegocontainer level. The main functions of all levels are determined. The article presents the order of interaction of steganographic systems at all levels. A method of redundant data encoding is proposed for the stegochannel level. It is based on an unconventional method of applying codes of the Residue Number System (RNS). The data represented by the RNS code is embedded in steganographic containers. The advantages of the proposed method of coding are increasing the stability of the steganographic channel to the destructive effects of interference and the attackers. An example of the functioning of a steganographic communication network is given. The process of encapsulation and decapsulation of steganographic data at seven levels of the interaction model of steganographic systems is presented. The order of addressing steganographic packets is formulated (in general outline.
There is a huge amount of news information on the Internet. This gives users the advantage of having access to a large amount of information. However, users are not always interested in all the news that is distributed. Therefore, when users read the news on the Internet, they need to select the information they are interested in. We have developed NEAR, a news application for smartphones, with the aim of solving the problem of information overload. However, in order for the app to learn the user’s interest trends, it needed to have the user sort through a lot of news, which was a burden for the user. Therefore, we propose a method that focuses on the user’s gaze information as a method for learning interests while reducing the burden of article selection by the user. By focusing on gaze, it is thought that it will be possible to analogize the degree of interest on a word-by-word basis rather than on an article-by-article basis, and it is expected that the number of article selection required for learning will be reduced.
In the paper, the real-world problem of predictive modeling is demonstrated. The objective is to associate automatically a clutter type to a telco network element (NE) using a limited input information on its local configuration and performance. For experimental data recorded in operator networks preprocessing actions have been applied which revealed the feature importance in the context of separation of network state space into several baskets (in reference to the clutter type systematics used in telecommunication). Then, selected machine learning models have been designed and applied to solve the problem of binary and multiclassification. Obtained results for the case of real network data measured in two city areas (North America) exhibited that it is possible to differentiate reliably (accuracy close to 90%) between rural and urban characterization of network elements, and even identify three classes of clutters, i.e. rural, urban (including suburban), and dense urban, with reliability significant for the telecommunication applications. Reported results can be further improved and used then for automation of network analytics, its optimization and management. The main contribution concerns the problem algorithmization and data processing in application to automation in telecommunication.
In a distributed system of nodes, leader election is the task of designating one of the nodes as the leader. In this work, we introduce ALEA, an anonymous leader election algorithm for distributed system with synchronous assumptions. The algorithm does not require a node to have knowledge of the identities of other nodes present in the network, ensuring privacy in the whole election process. Proof of other desirable properties such as uniqueness, agreement and termination are detailed as well. The overall number of exchanged messages is O(\(n^2\)), n being the number of nodes, and its actual performance is experimentally evaluated and discussed.
Whole heart segmentation significantly improves the diagnostic value of computed tomography images. Since manual segmentation is very time consuming, efforts have been made to automate this process, especially using deep learning methods such as convolutional neural networks (CNN). This paper considers how preprocessing of computed tomography with the Statistical Dominance Algorithm would affect the results of segmentation obtained with CNNs. Segmentation results were compared to original CTs and processed images. Compared to unprocessed data, improvements in segmentation accuracy were obtained after image processing. Using three-fold cross-validation, the average Dice similarity coefficient that was achieved was 0.811 for unprocessed images and 0.863 for processed images.
The paper presents the design and performance study of a fuzzy system for lip print identification. The problem applies to the general field of image analysis. It is a particular environment where uncertainty plays a very important role. The solution assumes a classical rule-based fuzzy system that automates the fuzzification of premises and facts based on appropriately selected ranges of learning data. Comparative analysis of several solutions has shown that the proposed approach obtains the best results.
The paper describes UAV application in automatic meter reading systems. The proposal increases the reliability of communications between SMs and TSC. The reading process must be improved because the communication capacity of the last mile RF networks has deteriorated significantly, as a result of the process of continuous replacement of RF smart meters to PLC units. The UAV is equipped with a communication node. The UAV communication node has a firmware dedicated for repeaters which allows it to operate as the mobile repeater. The additional moving node changes the topology of the wireless network with more connectivity introduced by the UAV. In this paper the algorithm of space track determination for the UAV is proposed. This algorithm is implemented on the UAV changing its trajectory according to the received sequences of packets types. Performance of the algorithm was tested in practice and compared to the efficiency of SMs reading process using fixed or random trajectories of the UAV. The application of the proposed algorithm makes the independence from the last mile RF network topology and changes in the propagation conditions.
The paper deals with the issue of hate speech and radicalization. Oftentimes, they are spread by means of social media. Twitter lets one express their views in a relatively anonymous way; however, it seems to be a simple, yet effective tool for disseminating offensive or radical contents, too. The paper proposes an effective solution which applies machine learning for detecting signs of radicalization and hate speech in Twitter posts. The authors decided to use the Polish language, which due to the level of its complexity is known to pose a challenge for automated sentiment analysis. The authors also needed to create their own dataset of posts containing hate speech, as prior to the experiment, there existed no such datasets in the language. In the paper, the underlying technologies are first presented, then the course of experiment is described and the final conclusions are given thereafter.
