Various neural network technologies appear to be able to already generate convincing images, audio, and video of events that didn't happen, or happened differently to their portrayal in that video. Soon that generated multimedia will become very difficult to distinguish from authentic recordings. This changes the mental paradigm that possession of multimedia purporting to show an event means that event most likely occured -- seeing will no longer be believing.
Generation of fake multimedia already exists today, but its generation involves a lot of time and work, so authentic media occurs far more regularly than faked media. The work needed to generate this media is likely to diminish significantly in the next few years. Some form of certifying and then authenticating multimedia would be helpful, and a blockchain-based solution is probably suitable. This is increasingly a well-recognized problem: Y Combinator's RFS asks for startup ideas for safeguarding against fake video, and the problem is starting to get mainstream media coverage.
Beyond simply detecting digitally altered or created media, an underserved angle is mis-labelling of media: a video of South Asians burning an American flag labelled as a street protest in Karachi is a rather different piece of media than the identical video labelled as post 9-11 celebration in Brooklyn, even if no media alteration (other than cropping) has occured.
A photographer can already create a digital signature for their work using public/private keys. However, there's no central repository for these signatures, and no infrastructure for verifying them. Further, the general public is unlikely to recognize the names of individual photographers -- it would be useful to have a web of trust where provenance of media can be recorded, and where institutional certifiers of multimedia can also signal their trust in the multimedia. If a newspaper publishes an image, or a speech, they're (usually) signalling that they believe it to be authentic. This information may well be useful to a consumer of the media trying to find out more about it.
A Snopes-esque service allowing people to track the provenance and pedigree of the media they find online may be useful.
Below are some free-form thoughts by myself and some friends on this problem. It started work in the Summer of '17 as thoughts that might become a PhD proposal before I remembered I already had two jobs. Occasionally I may try and rewrite portions of this to make it less of a stream of consciousness.
Follows a brain dump that will hopefully coalesce into something more substantial with time...
-
Media will need a description of the media. The NYT publishing a photo of missiles that they believe has been doctored will want to authenticate not that the image is a true depiction of what happened, but that it's an image they've received from a state news source
-
Cryptographically signing media will work reasonably well for small media, say photos, but for larger items - an hour-long video of a speech - some kind of fingerprinting may be needed. This doesn't feel like it should form part of any specification, it feels like the system should be flexible to allow the certifier to choose the form of the certification. Should a certification of an image work for the same image with a slightly different colour cast? Perhaps, perhaps not ... (@AdamL points out that darkening or lightening skin colour could well be a particular example of subverting it if allowed)
-
Some new thoughts:
One key insight here is that the system isn't meant to verify the integrity of a given set of bytes -- we already have cryptographic digests for that. It's also not intended to verify any form of ownership for the media; cryptographic digests can be signed already.
Instead, it exists to facilitate retrieval of meta-data about multimedia. Below I cover legitimate reasons for the NYTimes to publish known doctored and faked images. The important functionality is not that the NYTimes can verify that you have an authentic copy of an image they published, it's that they can verify their description and understanding of both the image, and its provenance. It's that they can certify they believe the source of the image, and what it represents -- where it was taken, when it was taken, what it appears to show, and who the believed author is.
Similarly any fingerprinting of that image is not to show that you necessarily have a perfectly exact copy of it. It's to allow bidirectional mapping between the image and its meta-data, and ideally other copies that have been certified -- even if they're subtly different. If a newsorg has published a pool photo that has been altered to show someone's skintone subtly altered, the utility is in being able to easily find that many other newsorgs have published the same image unaltered, and to find the actual original.
In some ways then, you have a system that's a bit like Google Image search (and presumably they'll add video search at some point), only decentralized, and with stronger identification of trustworthy sources. Currently, given a suspicious image, I can upload it to Google Image search and see where it's been published, and get a good idea as well of if the copy I have has been altered. However, I'm relying on a centralized service that doesn't attempt to certify results at all; a publisher has no recourse if Google decide not to index their media, and I may struggle to determine if an image is authentic if it only shows up on sources I don't immediately recognize.
-
It would be useful for this to be decentralized
-
Much as papers can be cited, it'd be useful to have an alphanumeric ID to use as the certification handle for some media; when presented with multimedia, it'd be useful to ask for an ID that the certification process has created
-
It'd be useful to have a mechanism whereby derivative portions of media can be certified deterministically -- eg: if you have the certification id for an image, it should be easy for a non-trusted user to generate an id that still works for 30 specific seconds of that media
-
As @Tynan points out, I forgot to mention that this sounds a lot like a good application for a distributed blockchain of some type. Ideally you should be able to download the whole ledger, but hosted versions are also available as per bitcoin.
-
If you're signing something cryptographically, keys will get lost and stolen. Some way to revoke keys is needed, but also some kind of trust authority to say "this keypair really does belong to Library of Congress". To what degree could current TLS keypairs be re-used at least initially?
-
That a given image on the NYtimes website is hosted with an EV certificate already suggests a fair amount of trust in the image, albeit without any embedded meta-data or caption. This could be a good bootstrapping method, although the lack of captioning is already problematic. Consider this well-known doctored image (see this article). That it's served by the NYTimes webservers currently only says it's newsworthy, not that it's an authentic depiction of events.
-
@RobP points out that this is distinct from copyright; copyright assigns ownership and credit. If a newsorg sends you out to take a photo, you certify your photo AND the newsorg certifies the photo AND perhaps the newsorg's ownership trust signs it, and so on. News organizations become not just a general record of events, they become certifiers of the authenticity of an event, which gives them power.
-
Does it cost anything to add to the chain? How does that work? Who pays? What motivates them to pay? What's the return on investment?
-
Wikisource / Wikimedia Commons maintains (currently) pretty credible provenance and authenticity data for a lot of assets.
-
What about text and facts, as per perigrin on IRC? Texts of speeches or official announcements are absolutely fair game, I think. Unemployment figures released by a government body can absolutely be certified. But a fact in terms of "unemployment has gone down", much less so, because that's an interpretation. But again, a politician releasing a statement saying that -- it's unambiguous that they released that statement, so that's fair game for certifying. LeeJo points me at this article, which is interesting
Especially when it comes to video where by necessity it’s chopped/cropped/etc and thus it does legitimately change due to various visual & audio treatments, even something as simple as volume leveling.
- Keyframe signing for video?
- Chain of trust - Does the editing/raw-importing software from the individual photog/cameraguy sign it, and it’s then endorsed up the chain of trust in the organization as it moves around? e.g should organizations be saying “we” produced this, don’t worry about how, or is it John Smith, personally reputable photographer, says he took this photo, and his parent org also certifies it, but he’s putting his personal reputation on the line as well as the org. Thinking about situations where state security services may direct an organization to produce certain media and certify it as genuine
- Reputational insight - does each piece of work stand on its own, or is there a trust metric derived from how often assets are used/verified by others? (i.e a bastardization of pagerank)
- Cross-certification. If NYT takes a video of a building being bombed, and CNN also takes a video independently, do their cameramen certify each others distinct copies, sort of like witnesses who are testifying the content is correct.
@pjlsergeant says: this is interesting. I think the answer is probably no. There's a split here between data provenance, and data trust. The cameraman or their device certifying the image, and what they say it is is one thing, as are "votes of trust" by the commissioning (or purchasing) newsorg. But it should also be possible for an anonymous document to be certified as trusted, as per WikiLeaks, who don't produce content, they just (claim to) verify it. NGOs research and verify reports. I think a journalist who's sourced a piece of media may also want to sign it. One would hope that if you have two non-identical videos, one certified by CNN and one certified by FoxNews, of the same event, the meta-data that they certified with the multimedia (where it is, what it represents, when it occured) one would be able to then easily find the other piece of media, watch it from the other angle, while seeing that the two angles have been independently certified
I realize I’ve veered a bit here into authenticity vs provenance, but saying something is “definitely from CNN” by itself isn’t enough, there needs to be a way to say it’s “definitely from CNN, and this 4 second clip of trump punching a baby is certified digitally as true by the cameraman who shot it, the boom guy who got the audio, the runner on site, and the visual effects person who took the original and performed the standard CNN color adjustment/crop/stabilization treatments on it”
22:06 <•perigrin> One other issue that I see that hasn't been mentioned is the need to verify provenance but maintain anonymity
22:07 <•perigrin> "this picture is accurate but telling you I took it will cause me to go to jail|the gallows"
22:12 <•pete> I think that has its place but is not directly the problem this is trying to solve
22:12 <•pete> What if WikiLeaks signs it?
22:12 <•pete> You're not trying to show ownership
22:12 <•pete> You're asking several people to show trust
22:13 <•pete> Anonymous signing is basically pointless as an anonymous signer has no trust in the system
22:13 <•perigrin> that's rather my point
22:14 <•perigrin> some of the most important pictures are the ones that won't be signed ... the videos that Chelsea Manning released
22:14 <•perigrin> for example.
22:14 <•pete> But that's good because we shouldn't trust random images that show up
22:14 <•pete> HOWEVER
22:14 <•pete> WikiLeaks saying "We think this is authentic" is huge
22:15 <•perigrin> I'm not so sure it is anymore but how is that any different than Wikileaks just posting that image?
22:16 <•pete> That's discussed -- images being served by an organisation under an EV cert already have some trust
22:16 <•pete> But the problem being solved here is:
22:16 <•pete> Someone on Facebook says "this image is in the 200gb trove of WikiLeaks images"
22:17 <•pete> And you saying "What's its certification ID?"
22:17 <•pete> And oh look, also signed by the Guardian and a bunch of other news media who received that particular image
22:17 <•pete> A jpeg suddenly has an evidence chain
...
22:22 <•perigrin> I wonder though why stop at media other than they're discrete and can be easily checksummed? Why not just create a generic block chain for information ... a snopes or wikipedia with a blockchain backing it for auditing purposes?
22:31 <•pete> perigrin: I think fundamentally we've always considered word of mouth to be untrustworthy
22:32 <•pete> And now there's the internet, the written word is far more suspect than it was when something had to be physically printed
22:32 <•pete> But we've had several hundred years to get used to that now
22:32 <•pete> Where the technology to create convincing fake multimedia is new and developing very quickly
22:32 <•pete> So I think there are inherently different trust levels there