deniz.in

Markets

Weather

Loading weather

· via Hacker News – Front Page (native)

One real missile video pushed a deepfake expert toward presuming all video is fake

IEEE Spectrum's Hany Farid describes spending hours verifying that a purported Tomahawk strike video was real, and argues creators must now prove their footage is authentic.

One real missile video pushed a deepfake expert toward presuming all video is fake

In a first-person account published by IEEE Spectrum, Hany Farid, a researcher who has spent three decades developing computational techniques for detecting manipulated media, describes the Saturday morning when reporters from around the world began contacting him about a video. An Iranian news agency had aired footage of what appeared to be a U.S.-made Tomahawk missile hitting a school; news reports claimed the strike killed more than 175 people, most of them children, while the U.S. president and military denied responsibility. In the age of generative AI, everyone, Farid included, suspected the footage might be fake.

A morning of frame-by-frame physics

Farid's first reaction, he writes, was skepticism: the video quality was low, it had surfaced well after the strike, and it was distributed only by an outlet with a poor record of accurate reporting. Given the stakes, he analyzed it anyway.

He reasoned that a missile digitally inserted into each frame without accounting for camera motion would trace a physically implausible path, so he aligned the frames against a stationary background landmark. The trajectory came out nearly linear. He then estimated the missile's apparent size and found it consistent with a 6-meter Tomahawk at roughly 100 meters, a distance that matched the 0.3-second gap between the visible impact and the sound of impact, given that sound travels at 343 meters per second. Hours of further analysis later, he judged the video authentic and told reporters so.

That morning, he says, marked a turning point. Research he had conducted with colleagues at the University of California, Berkeley found that observers are now so suspicious of video that they struggle to convince themselves footage is real. Farid realized the same applied to him: the absence of manipulation artifacts no longer counted as evidence of anything.

A presumption of fake, backed by provenance

The upshot, in Farid's view, is that the old default, a presumption of authenticity in which skeptics carried the burden of proof, must flip to a presumption of fake, with creators responsible for proving their content is real. He concedes this is unfair to legitimate creators but argues it reflects reality and points toward a future in which online media can be trusted again.

Forensic analysis like his does not run at internet scale or speed, he notes, but provenance signals can. He highlights C2PA content credentials, which attach a cryptographically signed manifest at the moment of capture recording the device, time, location and any subsequent edits, making tampering with the content or its history machine-detectable; watermarks such as Google's SynthID, which embed an imperceptible origin signal directly into pixels or audio samples; and Apple's newest iPhones, whose camera sensor can cryptographically sign raw pixel data as it is captured, saving a reference image alongside the normal, editable photo so anyone can compare the two.

These signals are often criticized as partial solutions vulnerable to attack, Farid acknowledges, but he argues that misses their purpose: not to catch bad actors, but to let good actors prove their recordings are what they claim to be. If enough legitimate creators adopt them, they could crowd out fakes, shifting the model from policing deceivers to equipping honest creators.

Deepfakes arrive in live video calls

The same distrust, Farid notes, has spread to routine video calls. He recounts the September 2024 episode in which Senator Ben Cardin of Maryland, then chair of the Foreign Relations Committee, accepted a meeting request that appeared to come from Dmytro Kuleba, who had stepped down as Ukraine's foreign minister a few weeks earlier. Cardin grew suspicious when the caller, out of character, pressed him on whether the U.S. should let Ukraine fire long-range missiles into Russia, and ended the call. The State Department confirmed Kuleba never requested the meeting, and the Senate security office warned staff that foreign adversaries, possibly Russian, would likely attempt similar faked meetings.

Such impersonations require only a face swap or avatar deepfake plus voice conversion, technology available in commercial products and open-source libraries that runs on a modern laptop, according to Farid. Unlike the missile video, a live call leaves no time for human analysis. But generating convincing motion at roughly 24 frames per second forces shortcuts: simplified faces, simplified interactions with surroundings, and drift between audio and video or between frames. Human eyes miss these; forensic tools can expose them. Farid also proposes that video-calling software could cryptographically sign each frame of a live stream, as Apple now does for still photos, letting the software verify in real time whether incoming frames were signed by a physical camera sensor or altered in transit, since a face swap necessarily happens between the camera and the app.

Why it matters

Farid's account documents a structural inversion in how visual evidence works. An authentic recording of an apparent mass-casualty strike took an expert hours to accept, and its credibility still rests largely on his analysis, while fabrications circulate freely and real events can be dismissed as fakes. Journalists, courts and ordinary viewers increasingly have to treat recordings as unverified until provenance says otherwise. That shifts responsibility onto device makers, platforms and calling software: C2PA credentials, watermarks and sensor-level signing are becoming the practical infrastructure of trust, and adoption by credible creators, rather than ever-better fake detection, is what could eventually restore the presumption that a video is real.

  • #deepfakes
  • #misinformation
  • #c2pa
  • #content-credentials
  • #media-forensics