Abstract. Authorship verification is one of the most challenging tasks in stylebased text categorization. Given a set of documents, all by the same author, and another document of unknown authorship the question is whether or not the latter is also by that author. Recently, in the framework of the PAN-2013 evaluation lab, a competition in authorship verification was organized and the vast majority of submitted approaches, including the best performing models, followed the instance-based paradigm where each text sample by one author is treated separately. In this paper, we show that the profile-based paradigm (where all samples by one author are treated cumulatively) can be very effective surpassing the performance of PAN-2013 winners without using any information from external sources. The proposed approach is fully-trainable and we demonstrate an appropriate tuning of parameter settings for PAN-2013 corpora achieving accurate answers especially when the cost of false negatives is high.

No abstract

Cyberbullying is a new phenomenon resulting from the advance of new communication technologies including the Internet, cell phones and Personal Digital Assistants. It is a challenging bullying problem occurring in a new territory. Online bullying can be particularly damaging and upsetting because it's usually anonymous or hard to trace. In this paper, the proposed method is utilizing a dataset of real world conversations (i.e. pairs of questions and answers between cyber predator and the victim), in which each predator question is manually annotated in terms of severity using a numeric label. We approach the issue as a sequential data modelling approach, in which the predator's questions are formulated using a Singular Value Decomposition representation. The motivation of this procedure is to study the accuracy of predicting the level of cyberbullying attack using classification methods and also to examine potential patterns between the lingustic style of each predator. More specifically, unlike previous approaches that consider a fixed window of a cyber-predator's questions within a dialogue, we exploit the whole question set and model it as a signal, whose magnitude depends on the degree of bullying content. Using feature weighting and dimensionality reduction techniques, each signal is straightforwardly parsed by a neural network that forecasts the level of insult within a question given a window between two and three previous questions. Throughout the time series modeling experiments, an interesting discovery was made. By applying SVD on the time series data and taking into account the second dimension (since the first is usually modeling trivial dependencies between instances and attributes) we observed that its plot was very similar to the plot of the class attribute. By applying a Dynamic Time Warping algorithm, the similarity of the aforementioned signals was proved to exist, providing an immediate indicator for the severity of cyberbullying within a given dialogue.

The openness of the Android operating system and its immense penetration into the market makes it a hot target for malware writers. This work introduces Androtomist, a novel tool capable of symmetrically applying static and dynamic analysis of applications on the Android platform. Unlike similar hybrid solutions, Androtomist capitalizes on a wealth of features stemming from static analysis along with rigorous dynamic instrumentation to dissect applications and decide if they are benign or not. The focus is on anomaly detection using machine learning, but the system is able to autonomously conduct signature-based detection as well. Furthermore, Androtomist is publicly available as open source software and can be straightforwardly installed as a web application. The application itself is dual mode, that is, fully automated for the novice user and configurable for the expert one. As a proof-of-concept, we meticulously assess the detection accuracy of Androtomist against three different popular malware datasets and a handful of machine learning classifiers. We particularly concentrate on the classification performance achieved when the results of static analysis are combined with dynamic instrumentation vis-à-vis static analysis only. Our study also introduces an ensemble approach by averaging the output of all base classification models per malware instance separately, and provides a deeper insight on the most influencing features regarding the classification process. Depending on the employed dataset, for hybrid analysis, we report notably promising to excellent results in terms of the accuracy, F1, and AUC metrics.

No abstract