mohammadehsan basiri

Today, the mobile phone is one of the smart devices that have become a necessity in everyday life and are used for various tasks such as shopping, banking, communicating with friends, family, etc. In recent years, the Android operating system has been able to gain more popularity than other mobile phone operating systems. The number of software related to this operating system is also expanding at a remarkable speed. Unfortunately, this issue is not hidden from the profit-seeking people, and the production of malware of this operating system has also grown in parallel with its development. Third-party Android app stores that have emerged in recent years have become a very strong source of malware distribution, as these stores have weak to non-existent measures to prevent malicious apps from being uploaded and distributed to users' devices. Therefore, one of the challenges that programmers are dealing with in this field is to find solutions to establish security in these types of devices, in such a way that it provides powerful security analysis capabilities while consuming few resources on the device itself. Software products such as Lookout, Norton, and Comodo Mobile Security mainly use signature-based methods to detect malware threats. However, malware attackers use techniques such as repackaging and obfuscation to circumvent signatures and defeat attempts to analyze their internal mechanisms. The ever-increasing sophistication of Android malware requires new defense techniques that can protect users against new threats while not using up all of a mobile device's processing and storage resources. Therefore, in the current research, a computational offloading method is presented in the cloud structure to identify Android malware. The solution proposed by this research first extracts the features of Android applications during installation and execution on the mobile phone, then sends these extracted features to the cloud servers. On the cloud server side, these features are analyzed and using machine learning algorithms, malware is distinguished from clean programs. The proposed approach is trained and tested using the Drebin dataset. The obtained results show that the proposed approach has achieved 96.44% accuracy for malware detection.

Today, despite the tremendous advances in medical science and technology, access to a specialist doctor is still considered a major challenge. This challenge is of great importance for diseases such as cancer. Skin cancer is the 13th most common cancer in men and the 15th most common cancer in women. While some skin problems are benign and harmless, some of them can be malignant masses, which will remain harmless if they are diagnosed in time. When consulting a specialist doctor may be time-consuming and expensive, an intelligent system can be a fast alternative or, at least, an efficient preliminary treatment solution. For skin cancer, such intelligent system may utilize the images of suspicious skin masses labeled according to their benign or malignant state by specialist physicians. These labeled images are useful for training intelligent systems which should diagnose the potential problems in unseen new images.In this research, a novel deep learning-based approach is proposed for the problem of classifying skin cancer images into two categories of benign and malignant images. In the proposed model, powerful deep learning models for image classification including VGG, ResNet, and Inception are used in two levels. Specifically, we formed two ensembles; VGG ensemble which consists of VGG-16 and VGG-19 models and ResNet ensemble which consists of ResNet152, ResNet50, and Inception models. CatBoost algorithm is used in each level to combine the models on that ensemble. Finally, at the next level, two ensembles were combined using the CatBoost algorithm. The proposed ensemble model tries to improve the accuracy and consistency of the results by aggregating the deep models at its two levels. In order to show the utility of the proposed model, a subset of ISIC public dataset for skin cancer images is used for training and evaluation of models. The performance of the proposed ensemble model is compared with several deep neural networks and previous similar researches. Specifically, we compared the results achieved by the proposed model with those obtained by existing similar deep models and those used as building blocks of the proposed model. The results show that the proposed model performs better in classifying skin cancer images. The performance of the proposed model, both in each of the classes and in general, has been better than all independent deep learning models. It has also been shown that using VGG ensemble along with this proposed model by combining its results with the help of CatBoost and forming a two-level ensemble has improved its independent performance in each class.

With the increasing desire of companies and organizations to employ interns in various situations, choosing the right person to participate in internships has become very important. Although the person who is selected for an internship must have relative knowledge and skills in the desired work fields; it should not be expert and experienced; because such people usually demand high wages. Community inquiry websites with many users can be used as one of the sources of intern knowledge. In previous research, statistical characteristics such as the number of answers, the number of specialized areas, the length of answers, and similar features have been proposed to identify potential interns; but the content of the user's answers has not been used to recognize the interns. This textual content is a rich resource for determining the breadth or depth of user knowledge and can be of great help in identifying potential trainees. In this research, a deep learning model called CNN-BiLSTM has been proposed to identify suitable people for internships based on the text of the answers they send to community inquiry websites. In addition, three machine learning models and four widely used deep learning models have also been used for comparison. Based on the obtained results, deep learning models have performed better in comparison with machine learning algorithms based on accuracy and F1 criteria. Also, among deep learning models, the proposed model has been able to show at least 7% higher accuracy and 2% higher F1 criterion than other models used to identify potential trainees.

With the spread of Covid-19 disease, quarantine, and social isolation, people are increasingly posting their opinions about the coronavirus on social networks such as Twitter. However, no study has yet been reported to analyze online opinions of individuals in order to understand their feelings about the Covid-19 epidemic in Iran. This study analyzes the emotions in the opinions of the Iranian people on the social network Twitter during the Corona crisis. For this purpose, a deep neural network model is presented. As there is no labeled dataset of Covid-19 tweets, the proposed model is first trained on the Stanford University Sentiment140 dataset, which contains 1.6 million tweets, and then used to classify the two classes of emotions contained in the collected corona-related tweets in Iran. The results show that the percentage of tweets with negative emotions is significantly higher than positive tweets. Also, the change in negative emotions of people in different months is proportional to the change in patient statistics.

Q&A forums are designed to help users in finding useful information and accessing high-quality content posted by other users in text forums. Automatically identifying high-quality replies posted in response to the initial posts not only provides users with appropriate content, but also saves their time. Existing methods for classifying user replies based on their quality, try to extract quality features from both the textual content and metadata of the replies. This feature engineering step is a time and labor-intensive task. The current study addresses this problem by proposing new model based on deep learning for detecting quality user replies using only raw textual content. Specifically, we propose a long short-term memory (LSTM) model that exploits the embeddings from language models (ELMo) for representing words as contextual numerical vectors. We compared the effectiveness of the proposed model with four traditional machine learning models on the TripAdvisor for New York City (NYC) and the Ubuntu Linux distribution online forums datasets. Experimental results indicated that the proposed model significantly outperformed the four traditional algorithms on both datasets. Moreover, the proposed model achieved about 16% higher accuracy compared to that obtained by the traditional algorithms trained on both textual and quality dimension features.

With the fast growth of social media, they have become the most important platform for posting multimodal content generated by users. Much of the data on social networks such as Instagram and Telegram is multimodal data. With the aim of analyzing such multimodal data in social networks, multimodal sentiment analysis has become one of the most significant subjects for researchers in the field of emotion recognition and data mining. Although multimodal sentiment analysis of social media data for English language has been addressed in several researches recently, few studies addressed the problem for the Persian language which is the official language of more than 120 million of people around the word. In this study, a multimodal deep learning model is proposed to address this problem. The proposed method utilizes a bi-directional long short-term memory (bi-LSTM) for processing text posts and a VGG16 convolutional network for analyzing images. A new dataset of Instagram and Telegram posts, MPerSocial, containing 1000 pairs of images and Persian comments is introduced in the current study and used for evaluating the proposed method. The results of experiments show that using the fusion of textual and image modalities improves sentiment polarity detection accuracy by 20% and 8% compared with the scenario in which image and text modalities in isolation. Also, the performance of the proposed model is better than three similar deep and four traditional machine learning models. All codes and dataset used in the current study are publicly available at GitHub.