Deep fakes are yet another prime example of how the writing is on the wall (i.e. it's clear that this technology can/will be used for nefarious purposes - esp. by our enemies) and yet mitigations - through regulation and enhanced standards - won't be pursued until long after the horses have left the barn. I don't know how/what form this should take, but an older analogy might be how colour photocopier manufacturers i…
1. Origin tracing. This is the microdot example: it's not meant to catch deepfakes, but link nefarious instances to their source. But the existing technical options here seem to be a mix of privacy-eroding (printers are dumb compared to phones/computers, and you'd have to prevent or ban sharing non-marked media) and ineffective (bulk color copying requires physical access to a device that's hard to make or modify; deepfakes can be constructed on a server in another country, or have their identifiers scrubbed after the fact).
2. Reactive, general verification. This is just an arms race between fakers and observers, like catching art forgers. Existing fakes have tells like a modified 'halo' around faces. Right now we only see manual checks, but major content hosts like Youtube could flag "suspected deepfake" like they do "music copyright strike". (Or hopefully better than that...) But it depends on staying ahead of fakers, and once the tells are too subtle to simply watch for it will only work when hosts or viewers choose to validate content.
3. Proactive, general verification. This corresponds to hard-to-implement, easy-to-verify security features like UV watermarks on money or prescriptions. But those rely on controlled supply, and decades of DRM failures tell us that digital fakes are much easier to make and safer to pass than physical ones. I don't expect this to expand beyond closed groups like news orgs giving out auto-watermarking cameras.
4. Specific authentication: not eradicating fakes but proving certain videos are legitimate. This is the most plausible, interesting category. We can't even prevent manipulative photo edits today, so we're unlikely to prevent manipulative deepfakes, but today we can prove specific aspects of specific images are legitimate. People will still believe fakes and lack proof of some real events, but this prevents a more fundamental transition to a "post-truth" era; we'll still have known-good records of key events.
We've had low-tech authentication since the dawn of photography - think of hostage photos taken with a daily newspaper to prove "this image is newer than this day". As photo editing emerged, steganography developed to catch out altered elements. Cryptographic signing took that further, allowing us to prove that an image is unaltered from a specific keyholder. We even have the reverse of the old newspaper photo; publishing a hash or encrypted file lets us date an image back to a specific time without having to actually release it.
Proving that a file is authentic to the world, not just an owner, is trickier. But we already have some steps: deepfakes take time, so any livestreamed video is not being edited that way after the fact. Authenticating something time-specific like a Presidential speech would only require combining that rapid turnaround with proof that the video wasn't prepared in advance; until on-the-wire editing becomes convincing a newspaper in the background would suffice. Quite likely we'll see more complex arrangements eventually, like trusted hosts that issue random values and demand their use in rapid responses.
None of this is going to stop people from believing fakes, but nothing ever has. What's more significant is whether we maintain the ability to create records which can be verified and trusted.