Search

Today, cloud environments are widely used as execution platforms for most applications. In these environments, virtualized applications often share computing resources. Although this increases hardware utilization, resources competition can cause performance degradation, and knowing which applications can run on the same host without causing too much interference is key to a better scheduling and performance. Therefore, it is important to predict the resource consumption profile of applications in their subsequent iterations. This work evaluates the use of machine learning techniques to predict the increase or decrease in computational resources consumption. The prediction models are evaluated through experiments using real and benchmark applications. Finally, we conclude that some models offer significantly better performance when compared to the current trend of resource usage. These models averaged up to 94% on the F1 metric for this task.

The study of Artificial Intelligence and, more specifically, Machine learning in games is of great interest to the gaming industry due to their wide application in several scenarios and their capabilities of simulating human behavior by non-playable characters. In online multiplayer scenarios, one of the greatest concerns is player disconnections and how to replace this player with a "good" replacement bot. In this work, we propose a machine learning based methodology to simulate the behavior of players that learns from their gaming history that generates more balance in the game. Our work uses Multi Layer Perceptron Neural Networks evaluated in a classic card game called Hearts, in order to emulate some previously defined behaviors. The results obtained in the experiments indicate that the proposed method has very good performance, since all generated models have managed to approach similar amount of victories when comparing to the behaviors that they were trained with. Through the evaluation of the results of 10,000 matches, with gameseeds of different matches used for training, the best result was for the model Shooter versus 3 Sheriffs getting 95.5% of the amount of wins compared to their particular bot. We also conclude that behavior learning is also clear on the difference of wins in all results depending on opponent skills, i.e., opponents who were difficult to win remain difficult to win in the simulated environment.

Detecting malicious code or categorizing it among families has become an increasingly difficult task. Malware1 exploits vulnerabilities and employ sophisticated techniques to avoid their detection and further classification, challenging cybersecurity teams, governments, enterprises, and the ordinary user, causing uncountable losses annually. Traditional machine learning algorithms have been used to attack the problem, although, these methods are heavily relying on domain expertise to be successful. Deep Learning methods requires less dependency on feature engineering, discovering the important features straightly from the raw data, recognizing patterns that humans usually can't. This work presents a deep learning approach for malware multi-class classification based on an unsupervised pre-trained classifier, using opcodes and its operands frequencies as raw data, ignoring knowledge that could be acquired from any known features from the malware families. The results confirmed that the approach is well succeeded and our best model achieved a MacroF1 of 93.14% a competitive result comparing to best-known classifier, since it uses less information about the malware.

Unmaned Aerial Systems (UAS) is a subject with great appeal, nowadays. Suitable applications for this kind of vehicles are countless. This work presents a study on scene matching (navigation with global positioning using image registration) in GPS denied environments. Usually, the problem is solved by fusing inertial data with other sources of displacement measurements like simultaneous localization and mapping or visual odometry. Although helpful, those methods establish a navigation system based on relative positioning, since they are limited to foreseeing displacements from previous states. Image registration can refresh the UAS position with georeferenced data, i.e., an absolute global position. This work reaches that goal through a detailed analysis of methods capable of extracting image information (e.g, edges or other invariant key points), enabling the registration process and consequently the estimation of position. Our results were quantified in terms of error estimation and time cost for computation. They have shown that good performance is achieved with a Convolutional Neural Network complemented with SURF approach.

This paper presents a real-time framework that combines depth data and infrared laser speckle pattern (ILSP) images, captured from a Kinect device, for static hand gesture recognition to interact with CAVE applications. At the startup of the system, background removal and hand position detection are performed using only the depth map. After that, tracking is started using the hand positions of the previous frames in order to seek for the hand centroid of the current one. The obtained point is used as a seed for a region growing algorithm to perform hand segmentation in the depth map. The result is a mask that will be used for hand segmentation in the ILSP frame sequence. Next, we apply motion restrictions for gesture spotting in order to mark each image as a `Gesture' or `Non-Gesture'. The ILSP counterparts of the frames labeled as "Gesture" are enhanced by using mask subtraction, contrast stretching, median filter, and histogram equalization. The result is used as the input for the feature extraction using a scale invariant feature transform algorithm (SIFT), bag-of-visual-words construction and classification through a multi-class support vector machine (SVM) classifier. Finally, we build a grammar based on the hand gesture classes to convert the classification results in control commands for the CAVE application. The performed tests and comparisons show that the implemented plugin is an efficient solution. We achieve state-of-the-art recognition accuracy as well as efficient object manipulation in a virtual scene visualized in the CAVE.

Statistics show that while large amounts of money are being invested in cybersecurity, the number of incidents continues to grow, with cyber attacks motivated by political and financial issues, many times funded by States as part of cyberwarfare. Although the general perception is that the occurrence of incidents is almost inevitable, the literature demonstrates that cybersecurity initiatives are often focused on prevention of incidents rather than its response, with many organizations often poorly prepared and ignoring key incident handling processes. Some initiatives were proposed in order to fill this gap, one of them being the VERIS framework, a "vocabulary for event recording and incident sharing." VERIS goal is to provide a basis for incident documentation, at the same time allowing the sharing of anonymized data to a community database, hence providing metrics for use within organizations or among external parties. As VERIS is a framework focused on information gathering and sharing, this work proposes the extension of the model from its original JSON representation to an OWL ontology, one of the main tools of the Semantic Web initiative, used for knowledge representation and strongly tied to the idea of information sharing. This work focus on the advantages of using such representation for incident handling.

Malware Analysis and Classification Systems use static and dynamic techniques, in conjunction with machine learning algorithms, to automate the task of identification and classification of malicious codes. Both techniques have weaknesses that allow the use of analysis evasion techniques, hampering the identification of malwares. In this work, we propose the unification of static and dynamic analysis, as a method of collecting data from malware that decreases the chance of success for such evasion techniques. From the data collected in the analysis phase, we use the C5.0 and Random Forest machine learning algorithms, implemented inside the FAMA framework, to perform the identification and classification of malwares into two classes and multiple categories. In our experiments, we showed that the accuracy of the unified analysis achieved an accuracy of 95.75% for the binary classification problem and an accuracy value of 93.02% for the multiple categorization problem. In all experiments, the unified analysis produced better results than those obtained by static and dynamic analyzes isolated.

A choice of the best computational solution for a particular task is increasingly reliant on experimentation. Even though experiments are often described through text, tables, and figures, their descriptions are often incomplete or confusing. Thus, researchers often have to perform lengthy web searches for reproducing and understanding the results. In order to minimize this gap, vocabularies and ontologies have been proposed for representing data mining and machine learning (ML) experiments. However, we still lack proper tools to export properly these metadata. To this end, we present an open-source library dubbed LOG4MEX which aims at supporting the scientific community to fulfill this gap.

Despite recent efforts to achieve a high level of interoperability of Machine Learning (ML) experiments, positively collaborating with the Reproducible Research context, we still run into problems created due to the existence of different ML platforms: each of those have a specific conceptualization or schema for representing data and metadata. This scenario leads to an extra coding-effort to achieve both the desired interoperability and a better provenance level as well as a more automatized environment for obtaining the generated results. Hence, when using ML libraries, it is a common task to re-design specific data models (schemata) and develop wrappers to manage the produced outputs. In this article, we discuss this gap focusing on the solution for the question: "What is the cleanest and lowest-impact solution, i.e., the minimal effort to achieve both higher interoperability and provenance metadata levels in the Integrated Development Environments (IDE) context and how to facilitate the inherent data querying task?". We introduce a novel and low-impact methodology specifically designed for code built in that context, combining Semantic Web concepts and reflection in order to minimize the gap for exporting ML metadata in a structured manner, allowing embedded code annotations that are, in run-time, converted in one of the state-of-the-art ML schemas for the Semantic Web: MEX Vocabulary.

We present Entropy Guided Transformation Learning models for three Portuguese Language Processing tasks: Part-of-Speech Tagging, Noun Phrase Chunking and Named Entity Recognition. For Part-of-Speech Tagging, we separately use the Mac-Morpho Corpus and the Tycho Brahe Corpus. For Noun Phrase Chunking, we use the SNR-CLIC Corpus. For Named Entity Recognition, we separately use three corpora: HAREM, MiniHAREM and LearnNEC06. For each one of the tasks, the ETL modeling phase is quick and simple. ETL only requires the training set and no handcrafted templates. ETL also simplifies the incorporation of new input features, such as capitalization information, which are sucessfully used in the ETL based systems. Using the ETL approach, we obtain state-of-the-art competitive performance in all six corpora-based tasks. These results indicate that ETL is a suitable approach for the construction of Portuguese corpus-based systems.

Transformation Based Learning (TBL) is a Machine Learning technique frequently used in some Natural Language Processing (NLP) tasks. TBL uses rule templates to identify error-correcting patterns. A critical requirement in TBL is the availability of a problem domain expert to build these rule templates. In this work, we propose an evolutionary approach based on Genetic Algorithms to automatically implement the template generation process. Additionally, we report our findings on five experiments with useful NLP tasks. We observe that our approach provides template sets with a mean loss of performance of 0.5% when compared to human built templates

The authentication of users on a Virtual Learning Environment (VLE) is, in general, punctual and intrusive, occurring when the user connects to the environment, by typing his password. Such approach allows, after the initial login, that unauthenticated users take the role of authenticated users and perform tasks in the environment, causing, among other things, distortions in the perception about the academic performance of students. The objective of this work is, thus, to propose a mechanism to execute periodic and non-intrusive authentications of users in VLEs. The proposed mechanism uses machine learning techniques to build recognition models based on the keystroke dynamics of users and it is also independent of the used VLE. A prototype of the proposed mechanism, integrated with Moodle, was implemented and applied to a postgraduate course with seventeen users. The recognition models generated by the prototype in the case study showed a performance above 92% of accuracy, which is a positive indication about the viability of the utilization of the proposed mechanism.

Human Computer Interaction (HCI) is a fundamental issue for virtual reality environments due to the need for natural approaches and comfortable devices. Such goals can be achieved using hand gestures to interact with the virtual reality engine. This paper presents a real-time system based on hand gesture recognition (HGR) for interaction with CAVE applications. The whole pipeline can be roughly divided into four steps: segmentation, feature extraction for bag-of-features construction, classification through multiclass support vector machine (SVM), generation of commands to control the application. We build a grammar based on the hand gesture classes to convert the classification results in control commands for an application running in a CAVE. The input is the depth stream data acquired from a Kinect device. The hand gesture recognition and command generation/execution approaches compose a client-server plug in that is part of a CAVE system implemented based on the Instant Reality architecture and the X3D standard. The results show that the implemented plug in is a promising solution. We achieve suitable recognition accuracy and efficient object manipulation in a virtual room representing a surgical environment visualized in the CAVE.

The malicious code analysis allows malware behavior characteristics to be identified, in other words how does it act in the operating system, what obfuscation techniques are used, which execution flows lead to the primary planned behavior, use of network operations, files downloading operations, user and system's information capture, access to records, among other activities, in order to learn how malware works, to create ways to identify new malicious softwares with similar behavior, and ways of defense. Manual scanning for signature generation becomes impractical, since it requires a lot of time compared to new malwares' dissemination and creation speed. Therefore, this paper proposes the use of sandbox techniques and machine learning techniques to automate software identification in this context. This paper, besides presenting a different and faster approach to malware detection, has achieved an accuracy rate of over 90% for the task of malware identifying.

The prediction of financial assets using either classification or regression models, is a challenge that has been growing in the recent years, despite the large number of publications of forecasting models for this task. Basically, the non-linear tendency of the series and the unexpected behavior of assets (compared to forecasts generated in studies of fundamental analysis or technical analysis) make this problem very hard to solve. In this work, we present for this task some modeling techniques using Support Vector Machines (SVM) and a comparative performance analysis against other basic machine learning approaches, such as Logistic Regression and Naive Bayes. We use an evaluation set based on company stocks of the BVM&F, the official stock market in Brazil, the third largest in the world. We show good prediction results, and we conclude that it is not possible to find a single model that generates good results for every asset. We also present how to evaluate such parameters for each model. The generated model can also provide additional information to other approaches, such as regression models.

Boosting is a machine learning technique that combines several weak classifiers to improve the overall accuracy. A well known algorithm based on boosting is AdaBoost. Boosting At Start (BAS) is a boosting framework that generalizes AdaBoost by allowing any initial weight distribution. BAS Committee is a scheme that uses feature clustering to determine the best weight assignments in the BAS framework. One of the drawbacks of BAS Committee is its final step which uses a simple Majority Voting approach over the chosen classifiers. Entropy Guided Transformation Learning (ETL) is a machine learning strategy that combines Decision Trees and Transformation Based Learning avoiding the explicit need of Template Design. Here, we present ETL Voting BAS Committee, a scheme that combines ETL and BAS Committee in order to determine the best combination for the classifiers of the ensemble. Besides that, since no extra assumption is made, ETL Voting is generic and can be used in any committee approach. Our empirical findings indicate that the BAS performance can be improved with a new combination of the classifiers determined by ETL Voting.

Named Entity Recognition (NER) is an important task in Natu- ral Language Processing. It provides key features that help on more ela- borated document management and information extraction tasks. In this paper, we propose seven machine learning approaches that use HMM, TBL and SVM to solve Portuguese NER. The performance of each model- ing approach is empirically evaluated. The SVM-based extractor shows a 88.11%F-score, which is our best observed value, slightly better than TBL. This is very competitive when compared to state-of-the-art extractors for similar Portuguese NER problems. Our HMM has reasonable precision and accuracy and does not require any additional expert knowledge. This is an advantage for our HMM over the other approaches. The experimen- tal results suggest that Machine Learning can be useful in Portuguese NER. They also indicate that HMM, TBL and SVM perform well in this natural language processing task.

Appositives are structures composed by semantically related noun phrases. In Natural Language Processing, the identification of appositives contributes to the building of semantic lexicons, noun phrase coreference resolution and information extraction from texts. In this paper, we present an appositive identifier for the Portuguese language. We describe experimental results obtained by applying two machine learning techniques: Transformation-based learning (TBL) and Hidden Markov Models (HMM). The results obtained with these two techniques are compared with that of a full syntactic parser, PALAVRAS. The TBL-based system outperformed the other methods. This suggests that a machine learning approach can be beneficial for appositive identification, and also that TBL performs well for this language task.

Semi-supervised learning is frequently used when we have a small labeled training set but a large set of unlabeled samples. In this paper, we combine Hidden Markov Models and Transformation Based Learning in a semi-supervised learning approach. Self-training and Co-training are the two semi-supervised techniques that we apply to our scheme in order to classify Portuguese noun phrases. Our main goal here is to show that we can achieve effective noun phrase extraction using fewer tagged examples by applying a semi-supervised technique. Our models show good improvement with a small labeled corpus and little with a large one.

Summary form only given. Prefix codes allow text to be decoded without ambiguity, since this code is a variable-length type where no codeword is a prefix of the other. The problem of improving the decoding speed has received special attention in the data compression community. A scheme that employs length-restricted codes to generate the codewords and table look-up is proposed. In order to reduce the bit manipulation, lexical expansion is introduced to decode more than one symbol in a single decoding step.