Authors: Stefan Axelsson (Norwegian University of Science and Technology)
DFRWS USA 2010
We have applied the generalized and universal distance measure NCDdNormalised Compression Distancedto the problem of determining the type of file fragments. To enable a later comparison of the results, the algorithm was applied to fragments of a publicly available corpus of files. The NCD algorithm in conjunction with the k-nearest-neighbor (k ranging from one to ten) as the classification algorithm was applied to a random selection of circa 3000 512-byte file fragments from 28 different file types. This procedure was then repeated ten times. While the overall accuracy of the n-valued classification only improved the prior probability from approximately 3.5% to circa 32e36%, the classifier reached accuracies of circa 70% for the most successful file types. A prototype of a file fragment classifier was then developed and evaluated on a new set of data (from the same corpus). Some circa 3000 fragments were selected at random and the experiment repeated five times. This prototype classifier remained successful at classifying individual file types with accuracies ranging from only slightly lower than 70% for the best class, down to similar accuracies as in the prior experiment.