Examining the performance for forensic detection of rare videos under time constraints

In many digital forensic investigations large amounts of material needs to be examined. Investigations involving video files are one instance where the amounts of material can be very large. To aid in examinations involving video, automated tools for video content classification can be employed. In this work we examine the performance of several different video classifiers in the context of forensic detection of a small number of relevant videos among a large number of irrelevant videos. The higher level task performance that is of interest is thus the ability to detect a relevant video in a limited amount of time. The performance on this higher level task is a combination of the classification performance, but also the run-time performance of the classifiers. A variety of video classification techniques are available in the literature. This work examines task performance for 6 video classification approaches from literature using Monte-Carlo simulations. The results illustrate the interdependence between run-time and classification performance, and show that high classification performance in terms of true positive and false positive rates not necessarily lead to high task performance.