TEACHING ARTIFICIAL INTELLIGENCE TO SPOT DEEPFAKES

The software for creating deepfakes – AI-generated videos of people seeming to say things they never actually did – is freely available online and relatively easy to learn to use. With an election not many weeks away, we could begin to see videos of Donald Trump admitting taking bribes from Vladimir Putin or Joe Biden confessing that he despises people of color, both of which never happened but would seem to the eye and ear to be absolutely genuine.
On a more personal level, the jealous coworker you bested for a promotion could create a false video of you confessing to embezzlement.
About 15,000 deepfakes had been sneaked onto the Internet from March to October in 2019, according to cybersecurity firm DeepTrace Labs. By now, the number is probably vastly higher.
Facebook sought to fight this new scourge by issuing its Deepfake Detection Challenge last October. The company hired more than 3,400 actors to make 100,000 short videos of them speaking, some of which were turned into deepfakes by substituting other faces into the speakers’ videos.
The challenge offered $1 million in prizes for the algorithm that could sort deepfakes from the real clips most accurately.
The company has chosen, but not announced, the winners, with the top prize going to an algorithm that flagged deepfakes accurately 65 percent of the time.
To up that percentage, Facebook is releasing the 100,000 videos and the winning software as open-source. It hopes that hackers and tinkerers will take up the challenge and find new and better ways of spotting deepfakes, perhaps using context or a video’s origins as well as its content, to render a judgment.
TRENDPOST: Deepfakes are yet another reason to not trust without question what you see or hear on social media. Hackers, fakers, paid liars, and passionate ideologues have claimed that space. Before believing or forwarding whatever appears on your social media feeds and pages, verify it first with a nonpartisan source such as FactCheck.org. The Reporters’ Lab at Duke University maintains a worldwide list of fact-checking organizations at https://reporterslab.org/fact-checking.