By Steven Hawley
Watermarking is one of the main technologies used to identify whether audiovisual content is being distributed and consumed through legitimate channels – and if the content is not where it’s supposed to be, to help identify the source and the origin of the ‘leak.’
During IBC, I learned of two other interesting applications.
Watermarks to detect deepfakes
In these times of disinformation and media trickery, the doctoring of videos of public figures has become common and has become a tool of politics. Hence, another watermarking use case has emerged: to verify that the content is what it is supposed to be, has come from a legitimate source; and is not a deepfake itself.
I heard more than one story about legitimate content providers applying watermarks at the point of production or at release into distribution exactly for this purpose. If someone suspects the veracity of the video, the watermark can be extracted by forensic analysis to confirm it. Or, if a video clip or image is found in a deepfake video, its rights-holder can analyze it and potentially find who used it without having rights.
Technical barriers to producing deepfakes are falling: Samsung scientists in Moscow have developed a way to create convincing deepfake videos with just one still image frame.
On the detection side, researchers at NYU have been working on this use-case, and published a paper in February. Microsoft, the Partnership on AI, Facebook and others are running a contest called the Deepfake Detection Challenge to motivate further research.
Watermarks that carry value
Two industry friends referred me to a South African company called Custos, which uses watermarking in combination with blockchain to associate video content with monetary value and gather piracy intelligence directly from the file sharing communities.
For example, movies sold as downloads can contain a smart contract ‘security deposit’ which proves that the movie is still in the intended recipient’s control. If the movie finds its way onto a pirate site, Custos technology enables any downloader to extract the deposit as a bounty. This immediately identifies the source of the leak, while keeping the bounty hunter completely anonymous.
Another application is to embed license information inside stock photos, and offer a purchase credit as an incentive to buy more images legitimately.
Traditional watermarking applications
Watermarks are intended to be undetectable by the human visual system, but if content that is suspected of being distributed by a pirate is subjected to analysis, the watermark is read (“extracted”) and associated with its legitimate use, and with an end user or consumer video account. If the instance of the video represents an infringing use, the video provider can be signaled to take action – such as revoking the keys to terminate the streaming session, or forcing the user to log in again, or more.
The original classic watermarking application is with reviewer copies of movies, in case the reviewer leaks the copy into pirate distribution. Watermarks can also be used to identify a theatre screening or a production facility where the content may have ‘escaped’ into pirate distribution. Now it’s used to complement CAS and DRM to help identify content pre- or post-legitimate distribution when it’s “outside” of the reach of traditional pay TV and streaming content protection.
Two general approaches
For video streaming, there are two general approaches. One is client-side: a watermarking process or a set of libraries resides within the video player or device, used to composite a unique identifier (such as a device ID or session information) into the content. It’s considered to have little impact on a video provider’s delivery infrastructure.
The second is the A/B approach. With A/B, two copies of the video are generated, using a server-side process to apply one watermark to the “A” copy and another to the “B” copy and split into segments that are each a few seconds long. When a video streaming session is requested, “A” and “B” segments are then assembled in unique sequences “on the fly” for each session. Two copies means two sets of encoders, storage and origination resources – which makes A/B very resource intensive.
The technical requirements for watermarking are strict. Speed of extraction and recognition is of utmost importance – to detect piracy of a brief sporting event such as a prize fight or a race before it’s over. A watermark must also be sufficiently “robust” to resist removal. Approaches that pirates may use to remove a watermark include geometric and axial distortion, analog-digital-analog conversion, collusion (“averaging”), tampering with chroma and luma, and a variety of other attacks.
Have you heard any unusual use-cases for watermarking? Let me hear from you!