VEIL: A System for Certifying Video Provenance

Abstract

Traditionally, a consumer decided how much to trust a piece of data based on its source. As digital video cameras and editors become ubiquitous, an arbitrary video object is increasingly likely to be produced using a range of operations that combine clips from a multitude of sources. A consumer can determine the assurance level of the data by knowing its lineage. We describe a system to embed the provenance of the video into the data itself. As long as the video contains a predefined threshold of data (from the spatial and temporal domains), the entire lineage can be ascertained. We embed the metadata using subpixel linear interpolation between similar blocks in proximal frames. It can then be extracted in real time using a novel method for computing the embedded interpolation. We implemented the process in C and report the performance overhead it introduces for playing video files. We also characterize the tradeoff between the auxiliary channel's capacity (which limits the amount of provenance metadata that can be embedded) and the extent to which the video can be edited (in the spatial or temporal domains) while retaining complete lineage.