Abstract
File sorting is essential in a wide range of data management areas, including forensic sciences. In some cases, forensics specialists face situations when file extensions are absent or intentionally altered for some purposes. Nevertheless, such files might contain valuable information for the investigation at hand. This paper reports on an attempt to develop a deep learning system to sort files by categories using bit representation of file contents, unlike traditional file extension, header or metadata-based solutions. Training, validation and test datasets included 100 files of various document, media and audio formats. Overall, the system demonstrates accuracy rate that exceeds 95%, with highest inaccuracy when trying to classify files with similar contents (for instance, different video file types, etc.). Nonetheless, the solution serves a solid proof of the concept and a foundation for future research in this direction.