The fight against deepfakes ramps up

From a faked video of Selena Gomez doling out kitchenware to phony Joe Biden robocalls, deepfake scams seem to be cropping up with alarming frequency.

And as we barrel toward what’s been dubbed the “generative elections,” companies and governments have been attempting to set standards and install guardrails to prevent such AI-aided fakery. There were some major steps toward those efforts in the past week:

• The Biden administration announced a new consortium dedicated to AI safety, with companies like OpenAI, Adobe, Google, and 200 others on board as partners. It’s yet another piece that builds on the sweeping executive order Biden signed last fall. One of the body’s stated goals is establishing guidelines for “watermarking synthetic content.”
• Meta said that it will begin attempting to identify and label all AI-generated content on Facebook, Instagram, and Threads. The company said it’s working with “industry partners on common technical standards” to do so.
• Google became the latest major company to sign on to an industry effort to label AI-created media called the Coalition for Content Provenance and Authenticity (C2PA), joining tech industry peers like Adobe and Microsoft. The group aims to bring together the Adobe-led Content Authenticity Initiative, which provides a “nutrition label” of sorts for content and news industry-focused Project Origin, according to its website.
• The FCC banned the use of AI-generated voices in robocalls.

The problem? It can be extremely hard to suss out whether a piece of media is AI-generated. Watermarking, one of the most-discussed fixes, can often be a cat-and-mouse game with bad actors that at least some technologists think could ultimately be a losing battle.

Karen Panetta, an electrical and computer engineering professor at the Tufts University School of Engineering, told Tech Brew that there should be a universal agreed-upon benchmark for watermarking, which doesn’t yet exist. She’s working on some of the standard-setting efforts in this area in tandem with the Institute of Electrical and Electronics Engineers (IEEE), a professional group where she’s a fellow, but said watermarking tech is still in its earliest stages.

“It is still in its infancy,” Panetta said. “With the push from regulation from the government to say, ‘You must have this,’ companies are now scrambling to address it. Right now, some of the state-of-the-art methods are trying to train a specific model to incorporate its own watermarking, which in itself sounds OK, but then the problem is, how many models are out there?...There’s no standard process in place that fits all scenarios…It’s not too difficult to get around—let’s put it that way.”

IT help: Panetta said that some of these efforts would do well to tap the expertise in existing fields, like cybersecurity.

“What folks are forgetting here is that there’s a whole field of cryptology, cybersecurity, in steganography, which is hiding information,” Panetta said. “An AI person might not be an expert in cybersecurity, so the thing is, now you’re going to see a marriage between those two fields.”

Now is the time: With a major presidential election looming and efforts outlined in Biden’s AI executive order beginning to take shape, the next several months could be key in hammering out standards around deepfake safeguards. When asked how soon she thinks it will take for standard-setting processes to start yielding results, Panetta said she expects it to happen in the coming year.

“Within the next year, we have to move—we have to,” Panetta said. “The bad actors are moving. You see it out there. And it’s bringing it to the attention of the public.”