Hacker News Reader

Comment by sega_sai | Hacker News Reader

Is there some kind of private/public key signing technique for videos, when the camera creating a video signs it with some key providing a signature that can be verified for authenticity ? so you would be able to at least verify that the video was created at certain date on a certain camera ?

View on HN

You can sign the data, but it won't help.

Basically, the ability to fake any video means that other data is needed to prove that the person/events in the video was in fact the person/events that that video is claimed to depict.

Other proof, other measurements of the same events, etc.

Deep fakes mean people will know they can't trust their eyes, they need to think critically. That's been true for a while, but now they'll know it.

Which I presume is why Parker/Stone are doing this now, other than the obvious (make money). They're going to force people to recognize this is possible and how perfect the fake can be.

As a side note regarding the money: My understanding is they had formed a studio to make a deep faked movie some time ~2019? I think? But covid rolled around and the studio dissolved. They took advantage of the contacts and assets they had initially made to produce this youtube video. They think it might be the most expensive youtube video ever made.

Maybe if we start the habit of signing one's own appearances, this at some day will make unsigned video considered unreliable.

But I guess this is wishful thinking given that we don't even generally sign written comments yet.

There's the Trusted News Initiative [0] and Project Origin [1,2]. Too little, and not enough support imo.

[0] https://www.bbc.co.uk/mediacentre/latestnews/2020/trusted-ne...

[1] https://www.bbc.co.uk/blogs/aboutthebbc/entries/46f5eb33-b7b...

[2] https://www.originproject.info/

So I can cryptographically prove that at a particular date at a particular time with a particular camera that I authentically pointed my camera at a screen projecting faked content?

I think this only works if you're signing something currently much harder to fake, like a full light-field or something, and even then it won't stay hard to fake forever.

You could have the camera record the lidar also and crypto sign that too. Assuming it is all done in camera and the lidar is high resolution enough it should be possible to completely tamper proof a video recording at least at the “talking head” level.

I think Apple could possibly start providing this for iPhone videos if they decided to as the hardware is all there already both the lidar sensor as well as the Secure Enclave etc

I think this idea of adding additional layers is on the right track because it confounds the deepfake generation problem by adding a curse of dimensionality. You can't just fake 2D. You need to fake 3D which is much harder.

I'm curious what the equivalent of "3D" to "2D" will be for deepfaked voice would be.

There is already the color channel dimension in addition to the spatial dimensions and time dimension. Adding depth doesn't make things much harder, especially because there is so much correlation between depth and color and time.

I know they don't currently do this, but if GPS satellites cryptographically signed their timing signals, could that potentially be usable to establish that, in addition to the particular camera capturing the images at a particular time, that it also did so in a particular-ish place ?

It's possible to record and replay signals from gps satellites with different delays, so maybe that could be used to spoof the location, but if the camera also has an internal clock, perhaps it could detect if there was too much discrepancy there? But I don't know how much that could constrain the location. Also gps only has limited precision, especially indoors? Uh, maybe using cell towers instead of, or in addition to, gps would be better? Hm.

I mean presumably you could just have the subject or subjects of the video sign the content. A full unedited interview for example could be signed by the interviewer and interviewee.

Edit: This would of course be predicated on PKI being safe and easy for most people to use which decades has proven otherwise though.

I think this solves the wrong problem. Most of these deepfakes have obvious places where you can tell they are renders. I think it's more important to have tooling that can tell when a video has been altered.

I'm more concerned with the manipulation of grainy low-res video. Police body cameras are an incredibly tool for police departments to fight misinformation (well, only when people watch the whole thing). Is editing these types of videos as obvious has high resolution video, or is it more easy manipulable?

> I think it's more important to have tooling that can tell when a video has been altered.

It's probably going to be a cat and mouse game. Think about how far along green screen technology has come. You watch a movie in the 80s and 90s today and the green screening is super obvious to our 2020 trained eyes. Imagine how someone from the 80s and 90s would perceive the greenscreening from 2020?

I imagine there will always be a cutting edge of deepfakes that will travel halfway around the world before the truth comes out and won't spread to all the people that believed the fake.

I can't help but to think that as a society, we're entering the "schizophrenic" phase of popular reality.

Read these quora answers on what it's like to have schizophrenia and think "What would a society collectively afflicted in the same way behave?"

https://www.quora.com/What-is-it-like-to-have-schizophrenia-...

It's going to be weird to see society have to operate in a state of constant disbelief of so many things that were previously accepted as fact.

It's a good question, crytpo can be used for this sort of thing. It does preclude editing, unless you also establish a cryptographic trust chain where each successive editor also has a trusted key.

Probably never going to work with cameraphone journalism, but if we're talking about newsroom & press conference footage you could make it happen.

Why not? Can't there be "TLS for cameras"?

Expect from editing, which, as you say should be able to be solved somehow (encrypt every frame?), I think the biggest problem is the equivalent of the "analog hole".

How do you know if a signed recording is from the actual event, or from a camera being pointed at a screen in a pitch black room recording the playback of some malicious video trying to show said event?

Trust. If the root signatory is CNN, you decide whether you trust CNN not to do that.

Not really any different from other encryption. Your bank could be defrauding you on the back end despite the verified https session. But you trust they aren't.

Why do we need to sign the video then? Isn't it enough for me to ensure that I have a secure TLS connection to https://cnn.com? Maybe for external platforms? But then the challenge is just proving that https://www.youtube.com/user/CNN/ is controlled by CNN. Much simpler than involving the cameras.

The hope would be that if video clips were verifiable, they would be widely expected to be verified, such that when someone links to "old clip of Senator says outlandish thing", you habitually check to see if it's verified- no matter who or what is re-sharing it.

I've done some thinking about how to prove that a video was filmed at a particular time. I've written parts of the design but haven't written any code yet. The idea is to have a group of public Internet servers that create a hash chain mesh. Every second, each server produces a new block, containing the SHA3-512 hash (digest) of the previous block, hashes of many other servers' previous block, and hashes of user documents.

To show that a video was created after a particular time, a person can include a hash from any public server in the video. The easiest way is to have somebody hold a phone in view of the camera, running an app that shows a QR code of a recent hash, updating in realtime. They could also read the hash digits. And the phone app could play the hash as audio tones.

Afterward, anyone who has the video can use software to extract the hash, query public servers for the hash's position in the immutable historical hash mesh, and get the associated timestamp. The existence of the hash in the video proves that the video was created (or edited) after the timestamp.

To show that a video was created before a particular time, use software to send the file's hash to a public server for inclusion in the mesh. A good recording app can publish repeated hashes of the file as it is recorded and gets longer. When the user stops recording, the app adds an "end of recording" mark to the file and publishes a final hash.

Afterward, anyone with good enough software can determine the time interval between the hash embedded in the video and the published hash. This is the interval when the video was produced. A malicious actor would have that much time to edit the video. With a good enough network, this could be 2 seconds, insufficient time for a human to make any editing decisions.

Hashes apply only to the original file. Anyone distributing a file with hashes must provide the original file that was used to create the hashes. Low-res resampled files can't be verified. Shortened or spliced files can't be verified. Only the alleged original file from the camera can be verified.

To fake a video, malicious software would need to insert a hash QR code into the video, process the video so it looks like it was recorded from a camera, and publish the hashes, all with a few seconds of delay. Fakers could also use a physical camera to re-record a video on a monitor. Either technique should be detectable. Fakery software will get better and may produce undetectable fakes. I hope that people will invent new physical anti-fake techniques to thwart fakery software.

I'm thinking of calling the system "Livestamp". It would work for any type of file or recording, not just videos.

Such a system exists, it’s called OpenTimestamps

Thanks for letting me know. It seems to rely on bitcoin's block timestamps which are accurate within tens of minutes or hours.

https://petertodd.org/2016/opentimestamps-announcement#how-o...

I believe Canon has special cameras that the police uses. It’s been a while but they use some encoding scheme I think.

Canon's image signing scheme was half-assed and was comprehensively broken, they didn't even store their signing keys on a secure element:

https://www.digitalartsonline.co.uk/news/creative-lifestyle/...

>In Canon's second version of its ODD system, the HMAC code is 256 bits. The code is the same for all cameras of the same model. Knowing the HMAC code for one particular model allows the ODD to be forged for any camera within that model range, Sklyarov wrote

Ack...not so good then :/

The toughest issue here, IMO, is being able to sign and authenticate a document without proprietary DRM. Been thinking about this a lot.

DRM is indeed what it is; you need to verify a signature somewhere to ensure it hasn't been tampered or altered, which requires an internet connection to verify or download.

This seems absolutely inevitable.

More practically, PGP.

Reply on news.ycombinator.com