Video provenance by motion vector analysis: A feasibility study

摘要

Motion vectors are used in digital video compression for the purpose of reducing file size, primarily by providing an estimate of a frame of interest as a motion-offset version of a reference frame or frames. Despite the name, a motion vector is not in and of itself a trace of object motion, but rather an index to a similar image block. The computational complexity of searching for a good motion vector means that many sub-optimal but computationally feasible approaches have been developed. It is conjectured that proprietary approaches to motion estimation lead to a provenance fingerprint. If the approach can be identified, then the identity of the source firmware or software can be inferred. This has further applications in edit and forgery detection. This paper describes the results of an early feasibility study in which the ability to distinguish between six motion vector algorithms is considered.