Deepfakes

With the Russian invasion of Ukraine, propaganda raises its head. Propaganda has always been a tool of war (and peace) but we're seeing escalation to new levels.

Recently, a video was shared on social media of the Ukrainian President - Volodymyr Zelendskyy - asking citizens to lay down their arms against the Russian invaders. Analysis by experts has concluded that it's a poorly done deepfake video. Zelendskyy's head didn't really fit his neck, and his head was disproportionate to his body.

So, what is a deepfake video? Deepfake, as Wikipedia describes it, is a portmanteau of “deep learning” and fake. Deep learning, in this instance, is referring to a technique in computer science that uses machine learning and artificial intelligence to create new content by analysing existing content.

The technology allows videos and other media to be created where the video appears to show somebody doing or saying something when they were not involved in the video at all.

There have been low quality instances of this for online porn, where celebrity faces are substituted for paid porn actors in an attempt to bolster the popularity of a video.

And there have been “fun” apps built which allow you to substitute a person's face into scenes from movies and TV shows, using just a single photograph. With more data, such as more photographs with open and closed eyes, and open and closed mouths, the video can become more realistic.

Ever since the creation of media there have been people who have experimented with faking. In Victorian times, camera double exposures were a common technique to generate a ghostly image of another person in a shot. But surely, all but the most gullible or invested viewer would spot that these were faked. (Although perhaps I'm over-estimating people - we regularly see images with photographic artefacts, such as lens flaring, being claimed as evidence of the paranormal!)

With the advent of film and video, creative editing techniques could be used to fool the viewer, to an extent.

But recently, the advance of AI and ubiquity of powerful computing technology has advanced to a point where it's becoming very difficult to tell whether something is fake or not.

As an example of this, I tried out a website which specialises in producing great-sounding voice overs for promotional videos. The site allows you to paste in your text, then have it rendered as a realistic sounding voice of your choosing - say Ava, a young adult female, or Ethan, a middle-aged male. I tried this out with the opening paragraphs of this item, and it's surprisingly realistic. Have a listen.

While that site allows you to choose from predefined voices, with a suitable budget, AI can analyse anybody's voice and then produce a facsimile of that voice saying words they never uttered.

As an aside, I experimented with speech synthesiser chips interfaced with 8-bit microcomputers in the mid 1980s. Back then, you could make a computer talk, with an extremely robotic sounding voice (that you often had to strain to understand), by figuring out what phonemes to send to the synthesiser chip to make the words. There's an example of a project using the SP0256-AL2 chip interfaced to a Zilog Z-80 processor in modern times. (Back in the day, the processor I cut my teeth on was a MOS Technology 6502). So, the speech synthesis of today just blows me away!

Anyway, the skeptical implications of deepfakes are huge. It used to be that seeing was believing - at least to a degree. The resources and technology required to fake somebody doing or saying something they never did were next to impossible. Now, it's certainly possible to produce something that is good enough to convince enough people that it's real.

The use of altered video in films has been around for a while. Forrest Gump had amusing use of obviously faked video content. More recently, the For All Mankind series on Apple TV+ included very realistic video of Ronald Reagan, and of The Johnny Carson Show. (The series portrays an alternate reality where the Russians beat the USA to the moon.) And, at the moment, I'm enjoying the Amazon Prime series The Man in the High Castle, which shows an alternate reality where Germany and Japan won World War II, but features films of alternate realities where the allies won. Characters in the series comment on the seeming impossibility that the films could have been faked!

The danger is that future elections will be fought and won using these deepfake techniques. It will be relatively easy to produce an incriminating video of a political opponent saying something distasteful. The converse of this is that nobody will trust anything they see in the media or online (and, are we getting to that point already?) so that, even if a video is real, there will be those that doubt its veracity. And while most people might recognise a fake video, enough might think it's real to sway opinion enough to affect the outcome of an election.

We should always be very wary about believing something when that thing plays into our own personal biases. The thing is that determining what is real, and what's generated, is becoming increasingly difficult.

Cycling back to Snopes - their recommendation is that it's possible to spot a deepfake video by looking at it, and also by considering the source where it's posted. I contend that simply looking at the video and figuring out whether it's real or not will become increasingly difficult over time. Checking where it is posted is a good indicator though - did it appear on the usual channels, or has it arisen from somewhere unknown or suspicious. Of course, hackers might also have the ability to post content through seemingly trustworthy channels.