Ask HN: Any solution to photoshop and deepfakes?

It seems like a great opportunity to finally do something useful with blockchain.

I predict: No.

It's been possible for a long time for big-budget studios to make convincing fakes. The new thing is that it's being commoditized so anyone can do it. Studios had to worry about being sued for libel or appropriation of likenesses, but individuals with an axe to grind won't be deterred.

We lived for a short time in a world where amateur recordings were more convincing than professional ones. Before the 90s, fake-news photos were more likely to be amateur since professionals had some sense of ethics and, while they enjoyed making fictional photos, generally labeled then as such.

So we're going back to the world where only the professional reputation of the photographer is a guarantee of authenticity. It may be better overall.

I wrote a longish post explaining some of the methods we could use to differentiate between real from fake content — but this led me to think of a more interesting scenario.

As society integrates more and more with cyberspace, we will have new metadata to verify events.

This is even happening now. Is it more likely that you won’t get hired at a job because of a photo of you doing something bad, or from a social media post? Most likely the latter.

Imagine a world wherein all our interactions are digitized— we would likely have no choice other than radical transparency, as all of our actions would be as traceable as our current day social media activity.

This increased amount of data associated with our actions allows for more sophisticated verification techniques, and ultimately would render video obsolete as a form of evidence.

I'm sure there are image forensics that could be done, ex. thresholding the image and looking for odd patterns.

I don't believe much can be done with amateur videos, but official videos of celebrities could be cryptographically signed with a key only known by that celebrity/their PR manager. That of course could lead to celebrities not signing unfavorable videos. Also the average info consumer wouldn't bother downloading and confirming the signature.

An interesting solution I read, I think here on HN, was to make a blockchain-based system where the hashes of videos were placed, thus proving that a given video was produced before a certain date.

A "deepfake" generated on an nvidia gtx 1080 gaming laptop in hours is not fooling any expert level human. Or even a median one. And would undoubtedly contain compression artefacts detectable via image analysis.

Training a custom deep generative model on massive prior data that can then synthesize image output would require significant cloud GPU resources. And be beyond the level of commitment of all but the most determined of counterfeiters.

However, if this indeed becomes a problem. Propaganda from bad state actors, for example. The real solution is exhaustive "reverse" video and image search. Given the output, determine which patches were sampled by the generator. A perfectly reversible process ;)

No, because it's not a problem. It's just a thing that we have to acknowledge: it's possible and someone will do something nasty with it, as with everything and every new invention.

Also: you are already unable to "establish veracity of the images and movies that I am receiving from the Internet" so why bother?

Learn which sources to trust

I have been thinking about this problem for awhile.

The short answer to your question is yes, it is possible. We have a prototype, called Amber, that establishes authenticity for audio and video recordings by fingerprinting and tracking provenance. Even if the video is cut or combined with other content, we can tell you whether it is the original (notwithstanding the edit in length). It is lightweight and needs the recorder to participate.

For most people and most video consumption purposes, it is not pertinent. But for certain companies and certain areas of the government, establishing fact or minimizing trust in visual/aural evidence is critical, especially when you have stakeholders who are independent or even antagonistic.

We are early stages and seeking to test it out with potential customers and their use case applications. The splash page is here: https://ambervideo.co and we hope to release the demo web / mobile apps soon.