In March of this year, a flurry of emails and text messages from a half dozen reporters around the world disrupted my quiet Saturday morning. An Iranian news agency had aired a week-old video of what appeared to be a U.S.-made Tomahawk missile hitting a school. News reports claimed that the strike killed more than 175 people, most of them children. The U.S. president and military denied responsibility. In the age of generative AI and all things deepfake, everyone—including me—was skeptical of the video’s authenticity.
I get this type of request a lot. For the past three decades, I’ve been developing mathematical and computational techniques to analyze and detect manipulation in digital media. Over this time, I think it’s fair to say that I, like most people, have approached visual recordings with a “presumption of authenticity”: If a skeptic claims that a recording is fake, the burden of proof is on them.
In today’s world of generative AI, however, I think we need to shift this stance to a “presumption of fake.” That is, the burden is now on the creator to prove that their content is real. While this new stance may seem unfair to legitimate creators, I think it both reflects reality and points to a future where we will once again be able to trust that most online media is what it appears to be.
The Forensic Approach to Deepfake Detection
Even last March, I was already tilting toward a presumption of fake. My initial response to the reporters was something along the lines of, “I am suspicious of the video because the quality is low, it emerged weeks after the bombing, and it is only being distributed by an Iranian news agency with a less-than-stellar record of accurate reporting.”
Given the potential implications—if the video was indeed real, however, I resigned myself to spending my Saturday morning analyzing it.
I first reasoned that if the Tomahawk missile was digitally added to each frame of the video without taking into consideration the camera motion, we might see a physically implausible missile path. So I aligned the video frames using a stationary background landmark as a reference. In the updated video, the missile did, in fact, move in a nearly linear trajectory. (See the yellow dots in the composite shown below.)
I next asked if the size (in pixels) of the missile in the video was plausible. Under some reasonable assumptions, I estimated that the missile in the video corresponded to a 6-meter Tomahawk at a distance of approximately 100 meters. The distance seemed plausible from the scene and was also consistent with the delay of 0.3 seconds from visible impact to the sound of impact, given that sound travels at 343 meters per second.
Composite photo composed of frames from a video of the Tomahawk strike reveal’s the missile’s trajectory.Hany Farid; Source: Mehr News
Despite the mounting evidence, though, I struggled to refine my answer for the reporters.
My skepticism was fueled by a recent study I worked on with colleagues at the University of California, Berkeley, that showed how hard it is to distinguish real from fake videos. Observers are now so suspicious of videos in general that they have trouble convincing themselves that a video is real.
But after a few more hours and a few more analyses with the missile video, I finally decided that it was authentic, and I gave the reporters my updated view.
This analysis marked a turning point for me. I realized I was just like the participants in our study: The lack of artifacts was not enough to readily convince me a video is real. I had flipped from a presumption of real to a presumption of fake.
Signaling a Video’s Provenance
When the visual record is no longer reliable, what can we do?
The types of forensic analyses that I routinely perform do not currently operate at internet scale and speed. But provenance signals can. These signals include content credentials and watermarks that are inserted or embedded at the time of creation or recording.
Credentials that follow the C2PA standard, for example, attach a cryptographically signed manifest to a photo or video at the moment of capture; this metadata records the device, the time and location of the recording, and any subsequent edits. Because the manifest is signed, any tampering with either the content or its history can be detected later by machine. Watermarks, like Google’s SynthID, embed an imperceptible signal directly into the pixels or audio samples that identifies the content’s origin.
Apple recently took these ideas a step further. On some of its newest iPhones, the camera sensor can cryptographically sign the raw pixel data as it’s captured. That signed data is then developed into a reference image saved alongside the normal, editable photo. Anyone can then compare the two and see exactly what, if anything, was changed.
These provenance signals are often criticized as being only a partial solution or vulnerable to attack. This is true if we think of them as mechanisms for catching bad actors. But that’s not what they’re supposed to do. Their purpose is to let good actors prove that their audiovisual recording is what it claims to be.
Provenance signals thus need not be adversary-proof. Instead, these signals are part of a new reality in which consumers, journalists, and the courts must adapt to a presumption of fakeness. If enough good actors adopt these mechanisms, they can begin to drown out bad actors. In other words, we have to switch from a model of “catch the bad actors” to “empower the good actors.”
Real-Time Verification
This switch from trust to distrust of digital media has expanded to include routine video calls.
In September 2024, for example, Senator Ben Cardin (D-Md.), then chair of the Foreign Relations Committee, received a video meeting request from Dmytro Kuleba, who until a few weeks earlier had been Ukraine’s foreign minister. During the call, Cardin became suspicious when, out of character, Kuleba pressed him on whether the U.S. should allow Ukraine to fire long-range missiles into Russia. Cardin ended the call. Later, the State Department confirmed that Kuleba had never requested the meeting, and the Senate security office warned staff that foreign adversaries, possibly Russian, would likely try similar faked meetings again.
All it takes to conduct this type of impersonation is a face-swap or avatar deepfake, along with voice conversion, both of which substitute the speaker’s face and voice for deepfaked versions in real time. These technologies are readily available in commercial products and open-source libraries and run on nothing more powerful than a modern laptop.
This video illustrates the current state of face-swap deepfakes.
Farid’s Company GetReal Shows How Real Deepfake Videos Look Today
Firms that verify provenance have to move fast to detect real-time deepfakes. Hany Farid, GetReal
Unlike the Tomahawk video, which I could sit with for a morning analyzing frame by frame, a live video call offers no time for human analysis. On the other hand, real-time deepfakes pose new challenges for their creators: A fake video call has to be generated at around 24 frames per second to fool the eye into perceiving realistic motion. As a result, the deepfake-video-call software has to take shortcuts, such as simplifying faces and how they interact with their surroundings, and allowing some drift between audio and video or between video frames. The human eye doesn’t notice the shortcuts, but we can forensically expose them.
Provenance signals should eventually play a role here. Video-calling software could apply the same cryptographic signing to each frame of a live video stream that Apple now applies to still photographs. A face swap necessarily occurs between a speaker’s camera and the videoconferencing software, and so the software could verify in real time whether the frames it’s receiving were signed by a physical camera sensor or whether they were altered in transit. As before, the burden shifts from catching the impersonator to allowing real participants to prove they are who they claim to be.
For the first two decades of the digital revolution, forensic analysis—finding the telltale traces of manipulation—was enough to adjudicate most questions of authenticity. The AI revolution, however, has changed the scale and speed of manipulation beyond what forensics alone can handle. To believe a recording is real now, it must also carry a provenance signal from the point of recording. The two approaches work together, with provenance allowing good actors to be trusted, while forensic analyses root out the bad actors.
Source: IEEE Spectrum




