Table of Contents #
Introduction #
As providers of streamed content open next-gen streaming gateways to real-time, higher quality and more immersive user engagement, they need to make sure they’re deploying streaming platforms best suited to countering the intensifying AI deepfake scourge.
AI-driven fakery at unprecedented levels of verisimilitude in just about every use case category involving internet transmission of static images and A/V content has become a big money maker on the Dark Web. Purveyors of deception are generating payback by:
- executing fraudulent financial transactions using AI to create fakes that bypass facial and voice authentication systems,
- inserting faux online meeting presences to eavesdrop on enterprises and government agencies,
- blackmailing people with highly realistic distortions of reality, including puerile depictions of children and adults,
- manipulating political outcomes through fake ads and newscasts,
- undermining companies’ standing with investors and the public by putting false words in the mouths of executives,
- and much else.
There’s also growing use of deepfake technology by people engaged in self-promotional deception targeted to enhancing personal stature in art, e-commerce, scientific research and other career pursuits. On the auditory side, AI is now generating fully produced music productions that are topping pop music charts. And as reported by at least one trade publication, there’s even a growing acceptance of public deception among some TV broadcasters who see no harm in using AI to lip-sync on-air personalities for delivery of live commentary in different languages over sister outlets.
High-performing deepfake tools are abundantly available, often at prices and usage simplicity that put them in reach of just about anyone. Users can generate videos directly from text using replications of real people sampled from photos, create YouTube how-to videos from documents, replace people and distort scenes in existing videos.
Some deepfake platforms go so far as to eliminate the need for dedicated computing hardware by performing most of the processing in the cloud. But even the top-performing deepfake tools offered at higher prices reaching hundreds of dollars monthly present no barriers to tech-savvy users who are positioned to generate high ROIs on those outlays.
In light of these trends, combatting AI fakery has moved to the strategic front burner wherever the reliability of content being what it’s claimed to be matters, which is to say, just about everywhere. The multi-billion-dollar question is, what can be done to stem the tide?
With reality benders and defenders generating tit-for-tat AI-fueled tech advances at breakneck speed, expert opinion is neutral on the question of whether the anti-fake battle can be won with technology alone, especially in laissez-faire regulatory environments that prevail in the U.S. and elsewhere. But it’s clear that the best chances for success will lie with streamers who can rely on next-gen streaming platforms that optimize use of the tools at hand, whether they’re designed to detect fakes or to provide validation of provenance legitimacy.
In the discussion that follows we explain how Red 5’s support for next-gen streaming provides the framework providers need to maximize effectiveness of whatever approaches they take to combatting AI fakes. As shall be seen, whether our Experience Delivery Network (XDN) Architecture is used in deployments involving streaming via the new MOQ Transport standard, WebRTC or legacy Hypertext Transfer Protocol (HTTP) based platforms, it provides the edge-based intelligence essential to timely execution of anti-fake solutions, facilitated by our open-source API approach to integration with those solutions and our unique high-speed live-stream video frame extraction process.
But before getting into those details, it’s worth taking a look at where things stand in the use of fakes and anti-fake technologies. Research reports from numerous sources lump static image and video fakes together in painting a disturbing picture of disruptions to societal norms on multiple fronts against a backdrop of lagging industry commitment to remedial action.
Generally speaking, anti-fake tools apply in all circumstances but face bigger challenges to effectiveness when hackers target live A/V streams. This is where the capabilities of XDN-based streaming platforms are especially significant.
The Deepfake and Deterrence Landscape #
The Ascending Assault Vector
Given the exponential pace of the fake surge, researchers acknowledge there’s no way to track the volume of deepfake instances transpiring across the globe other than through projections based on samplings and surveys. But all such endeavors point to skyrocketing rates of deception and fraud worldwide.
For example, Reuters has cited research from deepfake detection supplier Deep Media that projected the number of deepfake files shared over the internet would jump from 500,000 in 2023 to eight million in 2025. “This sheer scale combined with greater sophistication and convincingness means finding ways to quickly detect and mitigate this ever-growing threat is an increasingly urgent priority,” Reuters reported. Separately, a 2025 press release from the Reuters Institute for the Study of Journalism reported results from a global consumer survey that found 74% of adults worldwide said they were encountering misinformation in news reports at least weekly. News-related deepfake tracking by detection supplier Resemble that focused on news outlets with an aggregate reach totaling 292.2 billion potential views in the first half of 2026 reported there were 821 deepfake attacks on the 1,760 news reports that were tracked during that timeframe.
Looking at the deepfake impact from the enterprise perspective, research reported by Gartner in fall 2025 found that 62% of recently surveyed organizations had experienced at least one deepfake attack in the previous 12 months. Gartner projects that by 2027 hackers leveraging AI agents will halve the time they need to execute account takeovers with more automated steps aimed at defeating the deepfake kill chain, including use of “deepfake voices to make social engineering more convincing” and new ways to compromise authentication channels.
Recent research conducted by IRONSCALES, another detection supplier, found the deepfake incidence rate to be even higher with 85% of surveyed IT and cybersecurity professionals reporting at least one deepfake attack over the previous year and 40% reporting attack rates at three or above. Over half of the attack victims said they’d lost money from the incidents, with 61% reporting losses of $100,000 or more over that timeframe, 19% registering losses of $500,000 or more, and 5% putting the totals at $1 million or more.
Deepfake Deterrence Moves to the Global Front Burner #
Alarm over the deepfake threat has fueled significant increases in cybersecurity budgets and a corresponding surge in suppliers devoted to supplying effective deterrents. Gartner predicts that by next year, 50% of enterprises will be investing in disinformation security products and strategies, up from less than 5% in 2025.
According to a U.K. government report, the number of entities addressing this demand with new deepfake detection solutions as of 2025 had jumped from just a handful in 2017 to 59 worldwide, led by the U.S. with 23 providers and the U.K. with 7. Investment funding averaging $34 million per startup as reported by the U.K. attests to the scale of the impact AI deepfake attacks are having on the global marketplace.
That’s an especially remarkable rate of capitalization in light of the hurdles to profitability faced by all these players. According to the U.K. report, projected ROIs are low for the those who survive owing to concerns over high technical costs, resource constraints, detection reliability, vulnerability to manipulation and variability in metrics and testing datasets.
Arguably, in contrast to deepfake detection, the approach with the broadest range of participation from streamers and suppliers alike involves validating the identity of the original sources, i.e., the provenance, of streamed and posted content as formulated through the Content Authenticity Initiative (CAI) undertaken by the 6,000-member Coalition for Content Provenance and Authenticity (C2PA).
C2PA Provenance Validation
C2PA establishes a common approach to registering and tracking content provenance identity based on data or assertions listing who created the asset, when it was created, the devices and software used, whether AI was involved, what edits such as cropping or color adjustment were applied in production, and other information, including new categories of assertions C2PA groups might want to add to the specifications over time. All of this is compiled in a manifest file that’s cryptographically tied to the content wherever it goes.
The system uses a C2PA tool to compute a hash in the form of a compact invisible forensic fingerprint or watermark and signs the manifest using a private key backed by a digital certificate that ties the content to the originating device and software. That signing is what makes the credentials trustworthy, insofar as any time someone alters the pixels or the recorded provenance, the match to the stored hatch is broken. The specifications also allow for legitimate downstream editing of the content with new manifests created by C2PA tools chained to the original through the signing process.
Pluses & Minuses in C2PA Support
As noted in a blog posted at the beginning of 2026 by senior CAI director Andy Parsons, C2PA moved into widescale deployment last year and has been rapidly gaining support ever since with the latest version of the specifications now supporting provenance verification with live-streamed video in real time. “Entering year six, we are no longer just defining principles and goals,” Parsons wrote. “We are seeing interoperable provenance take shape in the real world.”
A big boost to adoption was provided in late 2025 when Adobe, which, along with Intel, The New York Times, Twitter, Arm, BBC, Truepic and Microsoft, founded C2PA in 2021, introduced what it calls Content Authenticity for Enterprise to operationalize provenance verification in the production process. This provides an automated means by which brands using Adobe’s production tools can integrate provenance into custom workflows, digital asset management platforms and publishing systems, eliminating the need to apply C2PA on a per-project basis. (Adobe arch competitor Avid is also a member contributing to C2PA development but, as far as we know, has yet to take a similar step toward streamlining use of the technology.)
Adding to the momentum, as reported by C2PA applications supplier Lumethic, a large number of leading suppliers of video and photo cameras, smartphones and camcorders have automated registration of C2PA credentials at the point of capture. And providers of streaming platforms such as Amsterdam-based Unified Streaming are introducing support for C2PA by attaching certified source credentials in the pre-distribution packaging phase.
There are now two organizations devoted to ensuring consistency in the use of C2PA specifications:
- The Creator Assertions Working Group (CAWG) serves to standardize C2PA specifications for use by professional content creators, develop new modes of identity assertion, and help creators test their conformance to the standards.
- JPEG Trust is an initiative within the ISO’s JPEG ecosystem that defines how provenance information is annotated, extracted, evaluated and secured over the lifetime of JPEG media in conformity with ISO’s Content Credential standard, which is built on the C2PA specifications.
But with all this growing support on the credential registration side, the absence of equally broad support for detecting C2PA provenance validation on the receive end remains a serious drawback to effective protection against deepfake deception. It’s a classic chicken-and-egg situation where, until there’s more pervasive use of C2PA credentials in content creation, spending on technology that looks for needles in the haystack may not make much sense to many market players.
Nonetheless, recent progress in that direction offers some hope for a brighter outlook over time. For example, Google has added C2PA credential detection into its “About this Image” feature, which executes with a right click on any image opened in Chrome, and there are C2PA credential detection apps available in app stores that can be activated to work in various browsers.
But, given the sparse use of C2PA on the creator side, the credential detection apps more often than not generate a message in the C2PA space that says C2PA is not used with the chosen image. In cases where C2PA has been used by the content generator, clicking on the confirmation message will open the C2PA manifest to reveal all the identifying data.
It doesn’t appear that similar capabilities have been activated in browsers for C2PA credential detection in videos. But users do have the ability to search for C2PA credentials by clicking on a CAI link that takes them to a site where they can upload or paste the URL of any image, video or document. It takes about 30 seconds for the CAI tool to check for and display any C2PA credentials.
The Breakthrough in Live Stream Provenance Validation #
As for C2PA’s support for provenance validation in live video streams, the process incorporated in the recently issued Version 2.3 specifications is much simpler than preceding techniques used with stored files. But it requires compatible device player software, which is just emerging as early adopters like Bitmovin’ and Red5 incorporate C2PA plug-ins with their players.
The first demonstration of real-time C2PA provenance verification with live-streamed content occurred at the 2026 NAB Show in Las Vegas, where content protection provider EZDRM teamed with engineering, testing and consulting firm Qualabs to implement the solution the two helped C2PA develop for recently released Version 2.3 specifications. As described by Qualabs, cryptographically signed metadata formatted as a universally unique identifier (uuid) box is injected in-band into each multi-frame segment by a signing proxy during the packaging process.
Within 500ms of receiving a segment that’s missing the metadata, a client player equipped with a C2PA plug-in module known as a validator delivers an alert warning that the validation chain has been broken by interlopers.
Along with activating the live-stream C2PA demonstration using C2PA Rust SDKs (C2PA-RS) to work with CMAF in HLS and MPEG-DASH streams and tuning the C2PA JavaScript repository (C2PA-js) to ensure player-side verifications synchronize with playback, EZDRM and Qualabs collaborated with Ateme to bring the verification process into play with streaming formatted to the emerging MOQ standard. This entailed adapting processing in MOQ Relays to sign content in real time with the arrival of each stream segment.
As with the other streaming modes, EZDRM’s validator is integrated directly with the MOQ player to read the embedded metadata, replay the cryptographic hash chain, and verify signatures, thereby triggering tamper alerts without interrupting playback. All the capabilities demonstrated by EZDRM at NAB are embodied in the EZDRM Video Signature Service, the first Level 2 conformance generator to be verified by C2PA, which is offered as an option with EZDRM’s Universal Complete multi-DRM service.
The accomplishment bodes well for broader use of C2PA as other suppliers learn the ropes and adjust capabilities to those demonstrated by the EZDRM team, which is working on making the real-time live validation process available for use with VOD content. This would overcome the drawback in previous C2PA playback validations, which, according to documentation from AWS, can take anywhere from several seconds to minutes when C2PA labels are visible or hours to days when there’s no way to know whether C2PA is in use without searching for invisible forensic marks.
Vulnerabilities to Relying on C2PA
But vulnerabilities intrinsic to the design of C2PA as described in C2PA documentation and other reports, including a recent critique issued by a team of researchers from the University of Maryland, the National Security Agency and elsewhere. leave open the question of what more needs to be done to protect against AI deepfakes as malefactors get better at what they do. Some of the more likely points of attack against C2PA validation involve stripping unprotected metadata manifests and re-posting the content without provenance information or shifting the manifests to alternative content files while using purloined C2PA signing keys to assign the deepfakes the provenance.
Drilling down to specific instances, one new report names a remote code execution vulnerability tied to Adobe’s C2PA-based Content Authenticity Initiative. And there are other instances where attackers manage to compromise the C2PA validator or even the underlying validation logic. More fundamentally, C2PA doesn’t show whether content was manipulated before it was recorded, in which case a deepfake signed at the moment of creation is deemed authentic.
A number of entities have developed blockchain-based modes of media authentication that create immutable records of content provenance that aren’t subject to capture through distribution. If projects like Numbers Protocol and Starling Lab, which are described in this analysis, get significant traction they could be combined with C2PA standards to create a more trustworthy chain of provenance validation, but that’s a big “if,” which, in any case, would take quite a while to make a difference.
Uncertainties in the Detection Battle against Deepfakes #
Meanwhile, there’s a rapidly expanding supply of solutions that take the alternative route to battling deepfakes through ever-more sophisticated modes of detection. But here, too, as some of the developers themselves acknowledge, there are no slam dunks in the fast-paced race between deepfake detectors and perpetrators seeking to seize the technology edge. And, as in the case of C2PA only more so, scale of adoption is a big issue as detection modes battle for market traction.
As described in this overview by the AI Security & Safety Directory, an anonymously maintained site hosting a voluminous repository of information about AI companies and organizations, detection solutions from the likes of Imagera, Usefulai and Resemble AI identify deepfakes by looking for clues often undetectable by human observation. They detect things like inconsistency in lighting and shadow, unnatural skin texture, irregular eye reflections, and spectral patterns that differ from camera-captured imagines.
The Hacker Pushback
But, so far, no matter how sophisticated the solutions might be, the generators of deepfakes have been able to adapt their models to overcome newfound vulnerabilities. As the security consulting firm Adaptive Security put it in a recent blog, “Every improvement in deepfake detection triggers a corresponding countermeasure in generation systems.”
Noting that deepfake generators’ AI systems training against detector feedback loops are able to identify and suppress “the very artifacts detectors are taught to identify,” Adaptive Security says, “The result is a permanent cat-and-mouse dynamic with no stable equilibrium. When detectors learn to flag unnatural blinking patterns, generators retrain with eye-movement regularization; and when detectors target heartbeat signals in facial blood flow, generators begin modeling physiological signals.”
Misleading Performance Metrics & Other Issues
It doesn’t help that providers of deepfake detection solutions appear to be putting out misleading performance metrics. A recent analysis conducted by the U.K.’s Department for Science, Innovation and Technology highlighted research showing that accuracy rates reported under lab conditions typically drop 10-20% when solutions were tested under real-world conditions where variables like demographic diversity and changing lighting conditions complicate detection processes.
Analysis by the cyber security company Brightside recently put an even harsher light on the lab-vs.-real-world-results disparity. “Commercial deepfake detection tools face a harsh truth when they leave the controlled environment of research labs and enter the messy reality of business operations,” Brightside said. “That impressive 96% accuracy? It drops to somewhere between 50% and 65% in actual use. Suddenly, you’re barely doing better than a coin flip.”
There are other issues confronting detection providers, such as the fact that a lot of the nuances they’re tuned to look for as proofs of legitimacy are stripped by encoders in compression processes that do away with content elements that are considered unimportant to delivering good viewing experiences. And, given all the ways deepfakes are used to distort reality, detection systems which typically aren’t designed to cover all those bases are bound to miss sometimes.
While some of the more sophisticated multi-mode detectors now entering the market aim to address this problem, it’s too soon to get a read on their impact.
Preparing for What Comes Next #
Looking at the provenance validation and deepfake detection solutions at hand and the challenges attending widescale adoption, no one can say if and when the tide will be turned against the deepfake torrent. But it’s clear that streamed content providers should do everything they can to improve their chances of success by choosing streaming transport architectures and media layer platforms that can facilitate execution of deepfake deterrence.
The motivation in this direction is sure to intensify as regulatory bodies come to grips with the deepfake challenge, as has already happened with the EU AI Act and China’s new AI labeling rules and is queued up with bipartisan backing of proposed but stalled U.S. legislation.
However the rulemaking goes, it’s likely the global battle against deepfakes will take a multi-pronged approach in the immediate future by tapping into various mixes of provenance validation and detection platforms. It’s even possible a scenario suggested by Adaptive Security eventually prevails where provenance identification is so pervasive “we may move to a model where media is untrusted by default and authentication is required to establish trust.” the previously quoted Secure Analytics blog suggests.
But we’re a long way from getting there. “This paradigm shift would fundamentally change how media is consumed, shared, and used in decision-making,” the blog writer adds, noting this “would require widespread adoption of content authentication standards that are still in early stages.”
Consequently, for now, when it comes to choosing streaming platforms with deepfake mitigation in mind, the best recourse is to go with those that are optimized to cover all bases. This is where what Red5 has to offer comes into play.
Leveraging XDN Edge Intelligence
One aspect to our support for deepfake deterrence that’s enabled by our Experience Delivery Network (XDN) Architecture centers on the ways in which deep-edge positioning of intelligent XDN nodes expedites processing for both C2PA provenance and fake detection applications.
Intelligence orchestrated by the XDN Stream Manager enables output, no matter how many streams may be coming in from upstream Origin and Relay Nodes, to be tuned to requirements associated with each use case, whether the goal is delivery of a live stream to end users or transmission of a video segment for parsing in deepfake deterrence.
This XDN edge intelligence supports Red5’s approaches to ABR profile streaming, dynamic ad and other content insertions, watermark extractions, multiviewing and much else. In all cases we have engineered our XDN software to enable egress latency that doesn’t exceed 10ms with video delivery from deep edge locations to end points occurring at sub-50ms latencies.
At the same time, Origin Nodes can be co-located with Edge Nodes to accommodate ingestion of massive volumes of streams at minimum latency in interactive scenarios serving all end points. Whatever the use case might be, it doesn’t matter whether just a few, thousands or even millions of users are engaged or where they are.
Support for Real-Time C2PA Validation
In the case of Red5’s support for C2PA validation, the emergence of the previously discussed player-centered real-time C2PA solution is a big help. Here it’s important to note that we are partnering with EZDRM to bring these capabilities into play wherever XDN Architecture is deployed.
Look for more information soon about how this can be done with MOQ, WebRTC and conventional HTTP streaming on the managed Red5 Cloud service or with Red5 Pro DIY implementations of XDN infrastructures. At the same time, in instances where that solution isn’t supported, especially in the case of stored content viewing, Edge Node support for minimizing roundtrip latencies associated with cloud-based C2PA validation of content segments remains essential.
Making Deepfake Detection Effective in Live Streaming
As for deterrence involving deepfake detection, extraction of video for analysis is another application that calls for minimal latencies. While there are solutions designed to leverage device computing power in the detection process, the computing power required to execute the best approaches requires use of cloud resources. Or, as the Alliance for Forensic Integrity & Provenance puts it, “Multi-model ensemble approaches that achieve the highest accuracy in research settings are often too computationally expensive for real-time deployment.”
Latency imposed by common approaches to video frame extraction can be a major impediment to cloud-based deepfake detections involving live streams. As explained by Adaptive Security, “For viable live-call detection, such as flagging a deepfake participant in a Zoom or Teams meeting, the total detection pipeline must complete within roughly 300 milliseconds end-to-end. Beyond this threshold, the alert arrives after the conversation has moved on, rendering it operationally useless.”
One aspect to meeting this challenge is the ultra-low latency achieved with XDN Edge Node to cloud transmissions. At the same time, as we report in this blog, the Red5 video frame extraction service overcomes the latency barrier imposed by even the fastest frame extraction techniques, which typically take a few seconds and even at the lowest 500ms-1-second levels are impediments to effective deepfake detection with live streaming. In contrast, Red5’s extraction service can be used to extract frames for deepfake detection in milliseconds from any live stream, including high-latency HTTP-based streams as well as from MOQ and WebRTC streams delivered at real-time speeds over XDN infrastructures.
When HTTP Live Streaming (HLS), MPEG-Dynamic Adaptive Streaming over HTTP (DASH) or another conventional streaming mode is involved, the post-production playout stream is delivered as usual over CDNs to end users while, at the same time, the feed is ingested by the XDN Architecture for real-time frame extraction at whatever intervals users set on their Red5 extraction dashboards. In this approach, there’s no end-to-end distribution involving egress from a cloud XDN infrastructure to end users.
Instead, as dictated by the user, the extracted frames are pushed into S3 or another cloud-based object storage platform, where the user’s chosen mode of deepfake detection can be applied instantly to execute the relevant tasks. Depending on the speed of the detection process, those tasks can often be completed before the HTTP-streamed primary content reaches end users, especially in cases where the streaming entity employs low-latency backend connections to the storage repositories.
Alternatively, in instances where streamers have activated real-time streaming end to end over Red5 Cloud XDN infrastructures, there’s no need to push the extracted frames into storage. Instead, distributors can instantly apply any detection solution that’s been integrated with the Red5 Cloud service to execute the process within the real-time streaming parameters, typically at 250ms or lower end to end, that are supported by XDN Architecture.
Expediting Integration of Deepfake Detection Tools
Adding to the advantages Red5 brings to deepfake detection, our open-source API approach to integrating AI solutions into customers’ streaming applications facilitates their ability to leverage new detection solutions as they come to market. Such integrations build on the large ecosystem of AI Large Language and Vision Language Models (LLMs and VLMs) that are already available to Red5 customers, as described in this blog.
Conclusion #
The global battle against AI deepfakes is taking on ever greater urgency as the business and cultural toll exacted by deception intensifies. But anyone looking to deter the aggressors must deal with the fact that nothing yet has taken hold as an undefeatable deterrent.
That suggests the best course of action is to be positioned to take advantage of the best solutions at hand with the flexibility to adapt as better ones emerge. Contact us to learn more about how Red5 customers can be assured they are operating in a streaming environment that will maximize their deterrence impact at each stage of the anti-fake evolution ahead.
The Red5 Team brings together software, DevOps, and quality assurance engineers, project managers, support experts, sales managers, and marketers with deep experience in live video, audio, and data streaming. Since 2005, the team has built solutions used by startups, global enterprises, and developers worldwide to power interactive real-time experiences. Beyond core streaming technology, the Red5 Team shares insights on industry trends, best practices, and product updates to help organizations innovate and scale with confidence.