A Language-Independent Approach to Classification of Textual File Fragments: Case Study of Persian, English, and Chinese Languages

With the advent of communications systems in recent decades, the transmission of electronic files on computer networks has dramatically increased. In this situation, identifying the type of files is important in many applications such as digital forensics and file carving. The state-of-the-art methods for identifying the file type of a file fragment are based on the content of the fragments. To the best of the authors' knowledge, there is no study addressing the effect of context language in identifying the file type of textual file fragments. In this paper, we have considered a machine learning approach for the classification among five types of common text file formats: PDF, DOC, DOCX, RTF, and TXT. Also, we have examined the effect of context language on the classification of the file fragments. Two scenarios are considered. In the first one, the language for both training and testing phases are the same, that the best results are achieved; the accuracies of the test for Persian, English, and Chinese languages are 85.6%, 76.4%, 86.1%, respectively. In the second scenario, the languages of training and testing sets are not the same, in which the training is done using one language and the evaluation is performed on the two other languages. In this case, the average accuracy values for Persian, English, and Chinese languages are 60.0%, 58.5%, and 71.4%, respectively. The evaluations of the second scenario show that the language-independent machine learning approach is robust in the identification of DOC, DOCX, and RTF formats.