The old saying that nothing is certain except death and taxes could be extended to video piracy, which continually defies all attempts to confine it to the margins of pay TV and streaming services. It does though come in waves, with periods of relative stability punctuated by spikes amid expressions of heightened concern both by services providers and their suppliers of counter piracy technology. Vendors though have to be careful to underline the progress they have made and argue that the latest increases in piracy are the result of factors beyond their immediate control, such as service developments, new technology and societal trends.

Around a decade ago the growth of streaming as a serious alternative to traditional linear TV delivery prompted a rash of such announcements, with a number of vendors purveying technologies designed to counter the new risk of illicit redistribution from sources that may themselves be legitimate. Indeed, at that time, content redistribution was on the way to taking over from traditional threats, which began in the pay TV age with smart card cloning and then control word sharing over the internet.

Consensus emerged that streaming piracy, especially of valuable live content such as premium sports whose value degrades rapidly after the event, required a combination of network forensics, fingerprinting, and forensic watermarking, to minimize business damage. Various pay TV operators adopted these technologies, in combination with legal measures and where possible going after pirates.

The truism that piracy has to be countered on multiple fronts held then and still does more than ever. These include legal, law enforcement, publicity of the risks of consuming pirated content, disruption of payment mechanisms, and those technical defenses against the distribution or redistribution itself.

Network forensics involves various measures to identify when infringements are occurring, an example being when an authorized streaming subscriber is continually transmitting content for large periods of time. Fingerprinting entails taking a snapshot of particular content such that it can then be searched for without too much computation as it is streamed. In this way fingerprinting contributes to network forensics.

Watermarking then involves insertion of data within the video payload, designed as far as possible both to be imperceptible to the viewer, and resistant to tampering or removal by pirates in order to circumvent the protection. Naturally pirates have attempted to disable watermarking protection with varying degrees of success, in turn countered by advances in technologies, so becoming a branch of the cybersecurity arms race.

Unlike fingerprinting, which is merely focused on identifying content, distinguishing between one TV series or movie from another, forensic watermarking tags individual streams with unique identifiers. These can then be used to trace the path taken, and originating points such as a subscriber’s PC, laptop, tablet or even smartphone.

PhDs and books have been devoted to forensic watermarking, which is a complex and continually evolving field, but boils down to some fundamentals. In the event of piracy being detected there are four stages – mark injection, pirate distribution, network forensics, and watermark extraction.

Firstly, the watermark is embedded into the payload data as some digital data sequence sufficiently light not to tarnish the image perceptibly. This mark identifies the stream instance itself and the ID of the originating device. It is possible to inject multiple marks identifying devices at different stages of the distribution chain.

Pirates may then intercept the video for redistribution, as well as manipulating it in attempts to disable the watermarks. In some cases, the pirate is an authorized subscriber but then redistributes the content illicitly in breach of the service conditions, for distribution in turn to its paying “customers”.

Thirdly, network forensics is applied, sometimes with the help of fingerprinting, to identify content that appears to be subject to illicit redistribution. Finally, the watermark is extracted from streams identified as being distributed illegally, in order to determine the source of the leak.

At that stage actions are taken which depend partly on the nature of the content, such as its value and lifespan. Such considerations also affect the choice of watermarking technology. For premium live sports it is imperative that illicit redistribution is identified quickly so that action such as blocking infringing streams is taken before business damage has been done. In effect that means within minutes, before consumers of pirated streams have been able to watch much of the event.

Watermarking breaks down into two principal categories on the basis of whether the marks are inserted at the headend, when it is sometimes called server based, and at the device level, when it is client based. In the latter case the watermark data could be delivered over the network after being generated at the headend, or created at the client. The mark will then be overlaid on the existing payload or blended with it, without taking account of the content itself, because there will usually not be enough computational power conveniently available on a client device to do the latter.

This then has the disadvantage that the mark, being overlaid without consideration for the content, is slightly more perceptible, as well as giving more scope for pirates to manipulate to prevent subsequent recovery of the mark. But client-side watermarking benefits from faster execution and mark recovery, which can be critical for combating live stream piracy.

By contrast, head end or server-side watermarking can be made content aware, which allows marks to be more imperceptible to the viewer. It can as a result be more studio-compliant, meaning it can meet requirements for artefact-free watermarking. Equally importantly, it can optimize robustness against attacks to disable the protection, by taking account of the underlying content.

While vendors tend to argue in favor of one or the other, depending on which ones they have developed or support, there has been some success in mitigating the deficiencies of both. Client-side watermarking can largely overcome the issue of artefacts, while extraction of marks from server-side systems has become faster.

However, traditional video watermarking from various long-standing players in the content security field has run into the buffers of Generative AI. At first sight it might appear machine learning based AI would be a boon for watermarking by optimizing further that balance between robustness, avoidance of noticeable artefacts and efficiency of extraction.

But the susceptibility of these watermarking schemes to AI based attacks was demonstrated by a study at the University of Maryland published in October 2023. Lead author on the study, Soheil Feizi, computer science professor at the university, bluntly summed up the current state of watermarking images generated by AI methods. “We don’t have any reliable watermarking at this point,” he says. “We broke all of them.”

The study concluded the situation was worst for so called “low perturbation” watermarks designed to be invisible even to the expert naked eye. The study showed AI made it possible to “wash out” watermarks through image manipulation and regeneration, as well as being capable of adding convincing marks to confer false legitimacy to computer generated images.

As might be expected this study was not the last word, and indeed the issue has attracted the attention of the big technology players such as Google, Meta, Microsoft and Amazon concerned also over the need to identify fake content more effectively.

This has led to various developments often initially designed just to mark and protect AI-generated content, but with potential for expansion to traditionally produced material as well. Some of them, including offerings from Google and Meta, are published open source and so could be adopted by established content security vendors, either for R&D or extension of their existing watermarking systems.

Google has developed a watermarking technique for Veo, its most advanced generative video model, derived from its existing still image watermarking tool called SynthID. The idea is that the watermarks are embedded directly into frames as part of the generative process such that they are integral to the system and very hard to disrupt.

However, such an approach does not lend itself to adaption for video generated by humans rather than machines. Meta has developed a different process called VideoSeal designed for general video rather than AI generated material. Indeed, Meta openly concedes that its approach would be vulnerable to compromise in the case of AI-created material because the marks would have to be added after the generation. As Meta pointed out, this makes its approach more flexible but also less secure, for example in the case of an open-source generative model whose mechanisms are publicly available, making it easier to compromise the watermarking - albeit not trivial.

It is not clear how this will all play out over the next few years, but what is certain is that AI and especially Gen AI has thrown a spanner into the works of forensic watermarking as a method of video protection. Yet with the absence of clear alternatives this has merely increased the overall R&D effort.

It is important though not to lose sight of the key point that all these technological battles or arms races between pirates or cybercriminals and service providers hinge around willingness of consumers to view pirated content. This in turn boils down to the question of what consumers define as pirated content or illicit activity and their tendency to engage, as has been shown in various surveys. There are also the ongoing tensions around password or credentials sharing.

We have seen how Netflix has been followed by other providers or on demand subscription services such as Disney+, Hulu, and Max, stepping in to stop password sharing with friends outside a designated home. There have been various concessions such as allowance for temporary viewing outside while travelling or on holiday, but the clear aim is to curtail a practice the providers believed was dampening revenue growth at best. These concessions are necessary though to avoid provoking honest customers into churning by imposing too much friction on their legitimate usage.

Yet many normally law-abiding consumers see content they know has been pirated as fair game for consumption. Consumers of pirated content fall into three broad groups, the unashamed who do not care, the unaware who view pirated content but do not know it is illegal, and the unwilling, who know they are using pirated content and may be open to paying for it but cannot access their chosen content legally. That content may be unavailable from their provider, or geoblocked in their area, for example.

Not much can be done about the unashamed, but the unaware can be tackled to some degree by informing them they are consuming pirated content and perhaps warning them gently of the consequences.

The unwilling group is the main target for conversion to legal consumption, which can involve bundling offers, specialized packages, and making desired content available where possible. There is scope for application of audience analytics and personalization, as well as synergy with associated services like epsorts and gaming.

In these areas then content protection intersects with personalization, navigation and recommendation, and here again there is scope for AI. Experience has shown that vilification of average consumers who consume a little pirated content, or give away credentials to friends, is counter-productive.

Affordability is another factor, and here the fragmentation of streaming services over the last few years has played into pirates’ hands and increased the temptation to share credentials. This shows up in surveys such as Deloitte’s TMT Predictions 2025 report, which found that the number of SVOD services consumers are willing to acquire to ensure they can access the range of content they want has peaked. In the US it rose to four after earlier SVOD fragmentation but has then stayed there since 2020. In Europe it rose from 1.3 in 2018 to 2.35 in 2023 but has since plateaued at that lower level.

There are tentative signs of reaggregation as SVOD services start to be bundled with pay TV, and other less related offerings such as mobile and even financial services. There is also the growth of FAST (Free Ad Supported TV), as well intermediate ad-subsidized services. All these play into the security equation with potential for making at least as great a contribution to revenues as technical measures.