Text by C. Scott Wyatt
eepfakes, euphemistically called “synthetic media” by researchers, use artificial intelligence techniques to manipulate audio, video and still images so effectively that only another AI application can detect the deception.
As reported in the Washington Post and Forbes magazine, media production experts struggle to identify fakes created by combining “deep learning” algorithms with the best multimedia editing applications.
The Soviet Union edited people out of photographs so effectively that only decades later were some edits exposed. David King’s book “The Commissar Vanishes: The Falsification of Photo-graphs and Art in Stalin’s Russia” explains the painstaking efforts of Soviet photo editors to enhance some people while removing others from history.
The darkroom magicians would spend weeks and sometimes months making new negatives from manipulated prints.
Creative photo manipulation dates to the earliest days of chemical-based photography. When I learned black-and-white photography in high school, we learned both in-camera and darkroom manipulation. Exposing a film negative more than once required a tripod and careful planning. I ruined a lot of film with timed exposures, double exposures and other experiments gone wrong.
Scott Mutter, one of my favorite photographers, used traditional techniques to create amazing photo montages. Some of Mutter’s iconic works were collected in the book “Surrational Images,” published in 1992.
I have a mint-condition copy, alongside the works of other photographers.
Former Soviet photo-montage experts praised Mutter’s skill.
Mutter would place people and objects in impossible arrangements. He’d add waterfalls to escalators and swimming swans to marble floors. My favorite photo, Library, shows a library hidden beneath a busy street.
Adding people to images was one of the earliest forms of trick photography.
Georges Méliès brought double exposure and other photography tricks to motion pictures, giving us the first special effects in movies. Méliès would even carefully cut negatives to create new composite frames. You have likely seen his 1902 film “A Trip to the Moon” (“Le Voyage dans la Lune”).
By the end of the 1990s, Adobe’s Photoshop had become a verb. We use the term “photoshopped” for any manipulated image. However, those edits in the 1990s were still time-consuming tasks performed by humans.
Digital photography editing was merely the first step. Audio and video naturally followed, with non-linear editing offering new tools that replaced skilled manual labor.
In the 1980s and ’90s, I learned to edit audiotapes using razor blades and tape. Removing or rearranging words was a time-consuming task. Thankfully, listeners to AM radio didn’t hear the slight clicks and pops at the editing gaps.
Having learned the old, physical techniques for manipulating photo-graphs, 8 mm film and audiotapes, I appreciate digital editing tools.
I have a master of fine arts degree in film and media. Only a small handful of people I’ve met could edit media so well that only other skilled editors could detect the work. I cannot do it, despite several years of experience.
Adobe Photoshop, Audition and Premiere Pro now include basic AI plug-ins that offer to fix problems in photos, audio recordings and videos. You can add additional plug-ins from other publishers to extend the editing power of these media applications.
Skylum promotes Luminar AI as the best AI-powered photo editing solution. Replace the sky in a photo with a quick set of mouse clicks. Remove a person. Skylum claims its AI technologies can process high-resolution images in 12 seconds. Skin will look perfect. Eyes will sparkle. The sun will shine, even if you snapped the image on a cloudy day.
Although Skylum and Adobe promote their products as creative tools, the potential for nefarious should worry us.
Today’s propogandists, many located in Russia, China and North Korea, rely on software instead of darkroom techniques. Deepfakes are among their tools.
Skylum and Adobe argue that edits are marked in the metadata stored within media files, but those data can be edited, too.
Machine learning, integral to artificial intelligence, looks for patterns and then models possible mathematical equations for those patterns. Biometric software can recognize facial patterns unique to each person. Similar pattern learning can re-create faces and add them to any still picture or video.
If we have enough digital samples of a person, including the individual’s voice, AI can create a detailed deepfake: a synthetic digital rendering. The deepfake looks, sounds and even moves like the person it represents.
It’s one thing to digitally adjust an actor’s age and quite another to create an entirely fake video of a political leader.
There’s a deepfake video of Donald Trump in a clip from the series Breaking Bad. There are deepfakes of several world leaders giving speeches they never delivered. While entertaining on YouTube, it’s a short step toward causing a geopolitical controversy.