Detection and Visual Inspection of Highly Obfuscated Plagiarisms

In this paper, we present a framework for the detection of highly obfuscated plagiarisms. In contrast to existing plagiarism detection systems, which typically consist of the steps "source retrieval" and "text alignment", we introduce an intermediate step to detect interfragment relationships. This additional step fills the gap between the document level result of the source retrieval step and the fine granular n-gram input of the text alignment step. In the new intermediate step, fragments of about 50 to 60 words are examined and compared using an extended Jaccard measure to handle synonyms/ hypernyms. Additionally, the Jaccard measure is weighted to consider different relevance of words. The framework also includes an intuitive graphical visualization using heatmaps, which can be used for the easy visual inspection of possible plagiarisms without the final text alignment step. A number of experiments, performed with our implemented prototype, show the suitability of our approach.