Social media provides a free space to users to post their information, opinions, feelings, etc. Also, it allows users to easily and simultaneously communicate with each other. As a result, threat detection in social media is critical for ensuring the user’s safety and preventing suspicious activities such as criminal behavior, hate speech, ethnic conflicts and terrorist plots. These suspicious activities have a negative impact on the community’s life and cause tension and social unrest among individuals in both inside and outside of cyberspace. Furthermore, with the recent popularity of social networking sites, the number of discussions containing threats is increasing, causing fear in various parties, whether at the individual or state level. Moreover, these social networking service providers do not have complete control over the content that users post. In this paper, we propose to design a threat detection model on Twitter using a semantic network. To achieve this aim, we designed a threat semantic network, named, ThrNet that will be integrated in our proposed threat detection model called, DetThr. We compared the performance of our model (DetThr) with a set of well-known Machine Learning algorithms. Results show that the DetThr model achieves an accuracy of 76% better than Machine Learning algorithms. It works well with an error rate of forecasting threatening tweet messages as non-threatening (false negatives) is about 29%, while the error rate of forecasting non-threatening tweet messages as threatening (false positives) is about 19%.

Online social networks, such as Facebook, Twitter, LinkedIn, etc., have grown exponentially in recent times with a large amount of information. These social networks have huge volumes of data especially in structured, textual, and unstructured forms which have often led to cyber-crimes like cyber terrorism, cyber bullying, etc., and extracting information from these data has now become a serious challenge in order to ensure the data safety. In this work, we propose a new, supervised approach for Information Extraction (IE) from Web resources based on remote dynamic editing, called EIDED. Our approach is part of the family of IE approaches based on masks extraction and is articulated around three algorithms: (i) a labeling algorithm, (ii) a learning and inference algorithm, and (iii) an extended edit distance algorithm. Our proposed approach is able to work even in the presence of anomalies in the tuples such as missing attributes, multivalued attributes, permutation of attributes, and in the structure of web pages. The experimental study, which we conducted, on a standard database of web pages, shows the performance of our EIDED approach compared to approaches based on the classic edit distance, and this with respect to the standard metrics recall coefficient, precision, and F1-measure.