Science & TechnologyLOT 261122 · 5 min read
Deepfake media moves the authenticity burden to creators
A Tomahawk video test, C2PA credentials and Apple sensor signatures show why frame-by-frame forensics alone cannot settle authenticity as deepfakes move into live calls.
Key takeaways
- A March video alleged a U.S.-made Tomahawk hit a school; reports said more than 175 people died, mostly children.
- Farid estimated a 6-meter missile about 100 meters away; the 0.3-second sound delay matched a speed of 343 meters per second.
- Real-time deepfakes require roughly 24 frames per second, forcing creators to simplify faces and tolerate some audio-video drift.
- C2PA attaches a signed manifest recording device, time, location and edits; Apple signs raw pixels on some newest iPhones.
- A September 2024 Dmytro Kuleba impersonation attempt prompted a warning that foreign adversaries could try similar fake meetings.

What changed in March?
A March video distributed by Iran’s Mehr News appeared to show a U.S.-made Tomahawk missile hitting a school. News reports said the strike killed more than 175 people, most of them children; the U.S. president and military denied responsibility.
A half-dozen reporters around the world sent Hany Farid emails and text messages, disrupting his Saturday morning. After several hours of analysis, he concluded that the week-old footage was authentic.
Farid has spent three decades developing mathematical and computational techniques to detect manipulation in digital media. His old assumption gave skeptics the burden of proof; he now argues for a presumption of fake, with creators required to prove reality.
Farid acknowledges that placing the burden on creators can be unfair to legitimate producers. He argues that the stance reflects generative-AI reality and points toward a future in which most online media can again be trusted.
What made the missile footage credible?
Before testing the clip, Farid said: “I am suspicious of the video because the quality is low, it emerged weeks after the bombing.” Its distribution through an Iranian news agency with a less than stellar accuracy record also gave him pause.
He aligned frames against a stationary background landmark to test whether a digitally added missile would betray camera motion. The missile followed a nearly linear path, and an eight-frame composite did not expose an obvious insertion.
Under reasonable assumptions, Farid estimated that the missile appeared to be a 6-meter Tomahawk about 100 meters away. The scene supported that range. A 0.3-second delay between visible and audible impact matched sound traveling at 343 meters per second.
Evidence accumulated, yet Farid struggled to give reporters a firm answer. A recent study with University of California, Berkeley colleagues showed how poorly observers distinguished real video from fakes. Skepticism itself had made genuine recordings harder to accept.
After more hours and analyses, he told the reporters that the video was authentic. He said: “The lack of artifacts was not enough to readily convince me a video is real.” That reversal became his turning point.
Can provenance replace frame-by-frame forensics?
Forensic methods Farid routinely uses do not operate at internet scale and speed. Provenance signals can: credentials and watermarks are inserted or embedded at capture or creation.
C2PA credentials attach a cryptographically signed manifest to a photo or video when it is captured. The manifest records the device, time, location and subsequent edits. Because it is signed, machines can later detect tampering with either the content or its history.
Google’s SynthID takes a different route, embedding an imperceptible signal into pixels or audio samples to identify content origin. Apple has taken a more direct route on some of its newest iPhones.
Apple’s camera sensor cryptographically signs raw pixel data as it is captured. That signed data becomes a reference image saved beside the normal, editable photo, allowing anyone to compare the two and see exactly what changed.
Critics often call these mechanisms partial or vulnerable to attack. Farid accepts the criticism when provenance is treated as a way to catch bad actors. Its purpose, he argues, is to let good actors prove that an audiovisual recording says what it claims.
The systems need not be adversary-proof, Farid says, and if enough good actors adopt them, their evidence can drown out bad actors. Consumers, journalists and courts must adapt to the resulting presumption of fakeness. The practical shift resembles a tamper-evident shipping seal: empower credible sources with a handling record before suspicion hardens, rather than chase every bad actor afterward.
What does a faked meeting require?
In September 2024, Sen. Ben Cardin (D-Md.), then chair of the Senate Foreign Relations Committee, received a video meeting request from Dmytro Kuleba. Kuleba had served as Ukraine’s foreign minister until a few weeks earlier.
During the call, Cardin became suspicious when Kuleba, out of character, pressed whether the United States should allow Ukraine to fire long-range missiles into Russia. Cardin ended the meeting.
The State Department later confirmed that Kuleba had never requested the call. The Senate security office warned staff that foreign adversaries, possibly Russian, would probably try similar fake meetings again.
A face-swap or avatar deepfake, paired with voice conversion, is enough to conduct such an impersonation. Those tools are available in commercial products and open-source libraries, and they run on nothing more powerful than a modern laptop. A demonstration from Farid’s company, GetReal, shows the current state of face-swap video.
Can forensic shortcuts expose a live fake?
The Tomahawk clip allowed hours of frame-by-frame work. A live video call offers no comparable time for human analysis, although real-time deepfakes create exploitable weaknesses of their own.
A convincing fake call must generate roughly 24 frames per second to create realistic motion. That requirement forces software to simplify faces and their interaction with surroundings, and to tolerate some drift between audio and video or between consecutive frames. The eye misses these compromises; forensic analysis can expose them.
What must change at the camera boundary?
Farid proposes that video-conferencing software eventually apply the same frame-level cryptographic signing that Apple now applies to still photographs. Because a face swap occurs between the physical camera and conferencing software, the application could check in real time whether incoming frames carry a sensor signature or were altered in transit.
That would move proof back to the point of recording. For the first two decades of the digital revolution, forensic analysis handled most authenticity questions. AI has increased manipulation’s scale and speed beyond what forensics alone can manage.
The near-term test is adoption, not invincibility. If cameras, publishers and calling platforms equip enough good actors with signed evidence, provenance may restore confidence that forensic analysis cannot supply alone.
Farid expects the two methods to work together: provenance builds trust, while forensics root out manipulation. Until then, seeing is no longer believing. The paperwork now comes with the pixels.
spectrum.ieee.org (Source)
More from Nathan Brooks
Show full bio
Senior reporter covering people, launches and day-to-day practice in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
1,201 articles