On Improving the Accuracy and Performance of Content-Based File Type Identification

Types of files (text, executables, Jpeg images, etc.) can be identified through file extension, magic number, or other header information in the file. However, they are easy to be tampered or corrupted so cannot be trusted as secure ways to identify file types.In the presence of adversaries, analyzing the file content may be a more reliable way to identify file types, but existing approaches of file type analysis still need to be improved in terms of accuracy and speed. Most of them use byte-frequency distribution as a feature in building a representative model of a file type, and apply a distance metric to compare the model with byte-frequency distribution of the file in question. Mahalanobis distance is the most popular distance metric. In this paper, we propose 1) the cosine similarity as a better metric than Mahalanobis distance in terms of classification accuracy, smaller model size, and faster detection rate, and 2) a new type-identification scheme that applies recursive steps to identify types of files. We compare the cosine similarity to Mahalanobis distance using Wei-Hen Li et al.’s single and multi-centroid modeling techniques, which showed 4.8% and 13.10% improvement in classification accuracy (single and multi-centroid respectively). The cosine similarity showed reduction of the model size by about 90% and improvement in the detection speed by 11%. Our proposed type identification scheme showed 37.78% and 31.47% improvement over Wei-Hen Li’s single and multi-centroid modeling techniques respectively.