Towards Performance Improvement of Authorship Attribution

An accurate authorship attribution model can play a vital role in security domain by detecting fraudulent texts and combating plagiarism, online piracy, and cyber attacks. In this paper, we work on improving the performance of authorship attribution. To this end, we focus on generating effective samples and features towards creating an authorship attribution model. We did our experiments using a convolutional neural network (CNN). Two key findings from our experiments are as follows: first, our results consistently show that fusing n-grams and stylometric features yields a better performance than independently using each type of features. Notably, with fused features, we achieved an accuracy of 97.03%, a precision of 97.58%, and a recall of 97.03%. Second key finding is—when a sliding window is used in generating training samples, it is possible to improve performance by increasing the amount of overlap between samples, which can be achieved by decreasing the step length of the window. Our study shows that there is a linear relationship between performance metrics and the percent of overlap between training samples. Across three different types of features (n-grams, stylometric, and fused), the worst performance in our experiments was obtained when there was no overlap in the training samples. Inversely, the best performance was achieved when there was a 95% or a 99% overlap in the sliding windows.