“A destructive Turing test”
A Hollywood union has thrown its weight behind legislation in California’s Senate that would make pornographic deepfakes a crime.
Deepfakes are – increasingly prevalent and realistic – digital forgeries of videos or audio, created using machine-learning techniques.
Bill SB564, which has passed California’s Judiciary Committee, is being sponsored by The Screen Actors Guild‐American Federation of Television and Radio Artists (SAG-AFTRA).
It extends the definition of a consenting individual to include not just an actual act, or a performance, but a “realistic digitized performance in which the individual did not actually perform”.
Fears about the proliferation of deepfakes have led to an unusual alliance between the leading American union, which includes Hollywood royalty as members, and intelligence experts, as they seek to raise awareness of the threat.
Both sides sat down last week for discussions that brought in Adam Schiff, the powerful chairman of the House Permanent Select Committee on Intelligence, SAG-AFTRA President Gabrielle Carteris and actress and activisit Alyssa Milano.
As SAG-AFTRA member Heidi Johanningmeier put it: “Any one of us can have our image grabbed from us without consent. This is real. This is terrifying.”
Deepfakes: Researchers Building a Database of Presidential Candidates’ Mannerisms
The move comes as researchers are analysing the unique mannerisms and speaking styles of US presidential candidates, in the hope of identifying deepfakes and supporting digital forensics during the 2020 election.
Leading digital forensics expert Professor Hany Farid is developing “soft biometric models” of the current US presidential candidates, including Joe Biden, Elizabeth Warren, and Bernie Sanders. These models are built from analysis of videos of the candidates, mapping their unique talking styles and mannerisms.
As Deeptracelabs notes: “This can then be used to help identify inconsistencies with deepfakes of those candidates, such as poor speech to lip movement rendering.”
Deepfakes are created by a “generative adversarial network,” or GAN. Machine learning-powered GANs pit a “forger” and a “detective” algorithm against each other. In true Machine Learning fashion, over each iteration, the forger improves.
Capabilities are already convincing and easy to access (the most popular application is FakeApp, which leverages Google’s TensorFlow AI-framework), and as Parham Eftekhari, Executive Director, Institute for Critical Infrastructure Technology, puts it: “It’s a destructive Turing test in that it is a false construct tailored to mislead and deceive rather than emulate and iterate.”
“I am deeply concerned that deepfakes could be used to spread disinformation or interfere in our elections, and we have already seen these technologies used to harass, exploit and invade the privacy of private citizens, particularly women,” said Schiff during last week’s panel. “We have another election coming up and it’s more important than ever for the public to distinguish between what is real and what is fake. Our democracy depends on it.”
California is the first to move. It may take a high-profile incident for Federal action. Few doubt that during the upcoming presidential election, Deepfakes will proliferate.