The increasing prevalence of blockchain-based cryptocurrencies as a payment instrument in the past decade and the rewards earned by the cryptominers has resulted in a new class of cyber attacks,cryptojacking, which involves unauthorized mining of cryptocurrencies on someone's system. Spotting cryptojacking is difficult in many cases, since the relevant software tries to disguise its presence to evade detection, by mimicking benign software such as compression applications by performing similar bitwise, cryptographic, and encryption operations. In this paper, we propose the processing of assembly code—a fundamental and platform-independent programming language—as a natural language using deep learning for profiling applications, which we callDeepCodeProfiler (DeCode Pro). Our proposed solution leverages the immutable step of any cyber attack: the deployment of instructions in system memory to carry out the attack. Through extensive experimentation with different neural network architectures in the profiling stage, we show that DeCode Pro is highly effective in the detection of evasive cryptojacking attacks and achieves low false positive and false negative rates. We also show that the model achieves high classification accuracy even with limited training data, which can considerably reduce the computing resources required for training and retraining the deep learning model.

The quality of features is one of the main factors that affect classification performance. Feature selection aims to remove irrelevant and redundant features from data in order to increase classification accuracy. However, identifying these features is not a trivial task due to a large search space. Evolutionary algorithms have been proven to be effective in many optimization problems, including feature selection. These algorithms require an initial population to start their search mechanism, and a poor initial population may cause getting stuck in local optima. Diversifying the initial population is known as an effective approach to overcome this issue; yet, it may not suffice as the search space grows exponentially with increasing feature sizes. In this study, we propose an enhanced initial population strategy to boost the performance of the feature selection task. In our proposed method, we ensure the diversity of the initial population by partitioning the candidate solutions according to their selected number of features. In addition, we adjust the chances of features being selected into a candidate solution regarding their information gain values, which enables wise selection of features among a vast search space. We conduct extensive experiments on many benchmark datasets retrieved from UCI Machine Learning Repository. Moreover, we apply our algorithm on a real-world, large-scale dataset, i.e., Stanford Sentiment Treebank. We observe significant improvements after the comparisons with three off-the-shelf initialization strategies.

Abstract The COVID‐19 pandemic has huge effects on the global community and an extreme burden on health systems. There are more than 185 million confirmed cases and 4 million deaths as of July 2021. Besides, the exponential rise in COVID‐19 cases requires a quick prediction of the patients' severity for better treatment. In this study, we propose a Multi‐threaded Genetic feature selection algorithm combined with Extreme Learning Machines (MG‐ELM) to predict the severity level of the COVID‐19 patients. We conduct a set of experiments on a recently published real‐world dataset. We reprocess the dataset via feature construction to improve the learning performance of the algorithm. Upon comprehensive experiments, we report the most impactful features and symptoms for predicting the patients' severity level. Moreover, we investigate the effects of multi‐threaded implementation with statistical analysis. In order to verify the efficiency of MG‐ELM, we compare our results with traditional and state‐of‐the‐art techniques. The proposed algorithm outperforms other algorithms in terms of prediction accuracy.

Word embeddings have become the de-facto tool for representing text in natural language processing (NLP) tasks, as they can capture semantic and syntactic relations, unlike their precedents such as Bag-of-Words. Although word embeddings have been employed in various studies in recent years and proven to be effective in many NLP tasks, they are still immature for sentiment analysis, as they suffer from insufficient sentiment information. General word embedding models pre-trained on large corpora with methods such as Word2Vec or GloVe achieve limited success in domain-specific NLP tasks. On the other hand, training domain-specific word embeddings from scratch requires a high amount of data and computation power. In this work, we target both shortcomings of pre-trained word embeddings to boost the performance of domain-specific sentiment analysis tasks. We propose a model that refines pre-trained word embeddings with context information and leverages the sentiment scores of sentences obtained from a lexicon-based method to further improve performance. Experiment results on two benchmark datasets show that the proposed method significantly increases the accuracy of sentiment classification.

With the advance in technology, the volume of available data grows massively. Therefore, feature selection has become an essential preprocessing step to extract valuable information. Feature selection is the task of reducing the number of features by removing redundant features from data while preserving the classification accuracy. It is a multiobjective problem as there are two objectives. In general, multiobjective selection algorithms with machine learning techniques are utilized to find the most promising feature subsets; however, classification performances of these machine learning techniques are analyzed separately. In this study, we propose a new multiobjective selection model that dynamically searches for the best ensemble of five classifiers to extract the best representative feature subsets. We present the experiment results on 12 well-known datasets. The results show that the proposed method performs significantly better than all the machine learning techniques when they are executed separately. Moreover, the proposed method outperforms two existing ensemble algorithms, namely AdaBoost and Gradient Boosting.