I was waiting to get a haircut a while ago and took a look at the magazine rack. Normally, I don't bother because the material is out of my element (I went through a People magazine once and could only recognize a handful of the people in there). This time, I noticed a Scientific American (Oct, 2018). There was an interesting article titled as the subject line of the post https://www.scientificamerican.com/article/clicks-lies-and-videotape/ [preview only].

More recently, I read Lawmakers warn of 'deepfake' videos ahead of 2020 election.

I'm not interested in the political aspects, per se, but I am interested in learning if there are examples of "faked" videos that are beyond detection. I don't mean shaving a few frames or messing with timing and I don't mean Forrest Gump with JFK or Nat King Cole singing with his daughter. I mean legitimate and documented examples.

 

I was waiting to get a haircut a while ago and took a look at the magazine rack. Normally, I don't bother because the material is out of my element (I went through a People magazine once and could only recognize a handful of the people in there). This time, I noticed a Scientific American (Oct, 2018). There was an interesting article titled as the subject line of the post https://www.scientificamerican.com/article/clicks-lies-and-videotape/ [preview only].

More recently, I read Lawmakers warn of 'deepfake' videos ahead of 2020 election.

I'm not interested in the political aspects, per se, but I am interested in learning if there are examples of "faked" videos that are beyond detection. I don't mean shaving a few frames or messing with timing and I don't mean Forrest Gump with JFK or Nat King Cole singing with his daughter. I mean legitimate and documented examples.

If there aren't now, there will be at some point. What is a (digital) picture after all? It's a collection of numbers (color intensities for pixels). So if someone can say that they can determine that Picture B is a fake, it's because they can tell that the some of the numbers aren't what they should be. Well, once someone knows that, they can make those numbers be what they should be to pass that test. Every tool and technique that we devise to detect forgeries becomes the grounds for new tools and techniques to devise better forgeries.

 

If there aren't now, there will be at some point. What is a (digital) picture after all? It's a collection of numbers (color intensities for pixels). So if someone can say that they can determine that Picture B is a fake, it's because they can tell that the some of the numbers aren't what they should be. Well, once someone knows that, they can make those numbers be what they should be to pass that test. Every tool and technique that we devise to detect forgeries becomes the grounds for new tools and techniques to devise better forgeries.

We could use some sort of Cryptographic Image Authentication system that allows normal photographic transformations but not modifications to the image structures of some critical type with Proof-carrying data. Stenographic hashes and Digital Watermarking have been used to embed tracking and modification codes into sensitive images for ages. The main problem was key (Private key encryption) distribution to the image producers and users for authentication. Modern systems could use existing Public Key Infrastructure to provide the source Proof-carrying data keys needed to generate and authenticate digital images.

This won't stop bogus images and videos on the internet but it might provide a method for bogus source detection.

 

but it might provide a method for bogus source detection.

Nothing that a common Web user would know is available or care about, I guess.

 

We could use some sort of Cryptographic Image Authentication system that allows normal photographic transformations but not modifications to the image structures of some critical type with Proof-carrying data. Stenographic hashes and Digital Watermarking have been used to embed tracking and modification codes into sensitive images for ages. The main problem was key (Private key encryption) distribution to the image producers and users for authentication. Modern systems could use existing Public Key Infrastructure to provide the source Proof-carrying data keys needed to generate and authenticate digital images.

This won't stop bogus images and videos on the internet but it might provide a method for bogus source detection.

Some form of authentication makes sense as we use it elsewhere with, at least some, success. But, it seems that agreement on authenticating would/will be lacking because of the perception that we don't need authenticated video all the time.

https://hackernoon.com/detecting-fake-video-needs-to-start-with-video-authentication-224a988996ce

 

Nothing that a common Web user would know is available or care about, I guess.

For the common Web user it would be like HTTPS. Designed to verify signed 'official' media files have not been modified by a 'man in the middle'. Joe's cat videos won't need this but original source videos, embedded clips or even overlays from freelancers or news media productions would be historically traceable.

 

It is interesting to read about how "AI caused this problem and AI can fix this problem". Here is an example...how could this algorithm detect fakes with 95% accuracy? Well, only by "training" on samples of eye blink "errors" and developing the usual discriminatory outcome. We know how that goes - it's laughable.

A Neural Network Algorithm Analyzes Eye Blinking to Detect Fake Videos (with 95% accuracy!)

 

Some form of authentication makes sense as we use it elsewhere with, at least some, success. But, it seems that agreement on authenticating would/will be lacking because of the perception that we don't need authenticated video all the time.

https://hackernoon.com/detecting-fake-video-needs-to-start-with-video-authentication-224a988996ce

While good cryptography can result near absolute surety of the source, the authentication can only verify the original source encoder not the physical truth of the data used by the original source. There still must be trust involved.

For instance we could receive messages by any means including non-encrypted commercial radio broadcast bands. The key to trust was to have one-time pad type authentication biscuits and private codes that were personally delivered under seal by a three letter agency. If the source side provided the correct authentication sequence we were 100% sure it was from the holder of a biscuit but were unsure if the biscuit holder had a gun pointed at their head.

 

It is interesting to read about how "AI caused this problem and AI can fix this problem". Here is an example...how could this algorithm detect fakes with 95% accuracy? Well, only by "training" on samples of eye blink "errors" and developing the usual discriminatory outcome. We know how that goes - it's laughable.

A Neural Network Algorithm Analyzes Eye Blinking to Detect Fake Videos (with 95% accuracy!)

Yes, easily defeated like most AI systems that involve anything close to actual intelligence. AI or real 'I' won't solve this problem. It's like unscrambling a egg, after a while the 'tells' to the original configuration become smaller and smaller to the point you can't 'see' the transformation but you can track it as it's happening as a proof of an original whole egg.

Every nice eye movements and blinks with body changes.

 

If there aren't now, there will be at some point. What is a (digital) picture after all? It's a collection of numbers (color intensities for pixels). So if someone can say that they can determine that Picture B is a fake, it's because they can tell that the some of the numbers aren't what they should be. Well, once someone knows that, they can make those numbers be what they should be to pass that test. Every tool and technique that we devise to detect forgeries becomes the grounds for new tools and techniques to devise better forgeries.

Remember the good old days...when news traveled by messenger, whose skill was the ability to run long distances quickly. Kill the messenger, kill the story.

 

Remember the good old days...when news traveled by messenger, whose skill was the ability to run long distances quickly. Kill the messenger, kill the story.

Smoke/light/drum signal technology was far more reliable.

 

We could use some sort of Cryptographic Image Authentication system that allows normal photographic transformations but not modifications to the image structures of some critical type with Proof-carrying data. Stenographic hashes and Digital Watermarking have been used to embed tracking and modification codes into sensitive images for ages. The main problem was key (Private key encryption) distribution to the image producers and users for authentication. Modern systems could use existing Public Key Infrastructure to provide the source Proof-carrying data keys needed to generate and authenticate digital images.

This won't stop bogus images and videos on the internet but it might provide a method for bogus source detection.

I see this as a very different problem. Proving that an "official" photo has not been altered is trivially easy, the originator just digitally signs the hash of the image. If that concept can be extended to allow certain transformations, fine. But it still comes down to saying that Image 1 is no longer a faithful representation of Image 2 because I have information associated with Image 2 that I can verify.

But the problem here is that there's no Image 2 to begin with. Someone presents a picture of a two-headed alien playing chess with Bat Boy -- or of some politician cavorting with the wrong people -- and you only have the data in that picture to work from. There IS no source information.

It's like all of the high-tech anti-counterfeiting measures put into modern paper currency. I don't HAVE to figure out how to make a fake new $20 bill that can get by a bank teller, I only have to make a fake old $20 bill that I can get past the kid selling lemonade.

 

I see this as a very different problem. Proving that an "official" photo has not been altered is trivially easy, the originator just digitally signs the hash of the image. If that concept can be extended to allow certain transformations, fine. But it still comes down to saying that Image 1 is no longer a faithful representation of Image 2 because I have information associated with Image 2 that I can verify.

But the problem here is that there's no Image 2 to begin with. Someone presents a picture of a two-headed alien playing chess with Bat Boy -- or of some politician cavorting with the wrong people -- and you only have the data in that picture to work from. There IS no source information.

It's like all of the high-tech anti-counterfeiting measures put into modern paper currency. I don't HAVE to figure out how to make a fake new $20 bill that can get by a bank teller, I only have to make a fake old $20 bill that I can get past the kid selling lemonade.

The 'official' photo problem is not trivially easy because you don't need a original Image 2 with the correct image authentication system. That's one of that main points of Cryptographic Image Authentication.

 

The 'official' photo problem is not trivially easy because you don't need a original Image 2 with the correct image authentication system. That's one of that main points of Cryptographic Image Authentication.

Not really following you. How does not needing an original Image 2 with the correct image authentication system somehow make it nontrivial for the originator of an image to digitally sign that image so that no one else can alter it and claim that it is a faithful copy of the original?

But that's still beside the point. How can "the correct image authentication system" prove that the image I posted showing Politician A having a secret meeting with Mobster B is fake?

 

Not really following you. How does not needing an original Image 2 with the correct image authentication system somehow make it nontrivial for the originator of an image to digitally sign that image so that no one else can alter it and claim that it is a faithful copy of the original?

But that's still beside the point. How can "the correct image authentication system" prove that the image I posted showing Politician A having a secret meeting with Mobster B is fake?

Because the digital signature/authentication must remain intact after image transformations (resolution, encoding type for web presentation) that might alter pixels in a way that changes the signature match of the original hash and it should provide things like zero knowledge about images features (like a innocent child's face in a murder photo) that were removed by transformations like cropping and still provide authentication of the remaining image.

It can't prove an image was faked by a third party unless the image authentication system is designed into the capture sensor of the imaging device for the picture of Politician A and/or Mobster B but it is a way to trace the images to a reasonably trusted original source or not. Trust will always be part of the equation.

 

Because the digital signature/authentication must remain intact after image transformations (resolution, encoding type for web presentation) that might alter pixels in a way that changes the signature match of the original hash and it should provide things like zero knowledge about images features (like a innocent child's face in a murder photo) that were removed by transformations like cropping and still provide authentication of the remaining image.

It can't prove an image was faked by a third party unless the image authentication system is designed into the capture sensor of the imaging device for the picture of Politician A and/or Mobster B but it is a way to trace the images to a reasonably trusted original source or not. Trust will always be part of the equation.

As I suspected, we are talking about two different things -- actually three different things. Authenticating a copy of the original image is trivially easy with current technology. Authenticating signed images (for whatever definition of "signed" is applicable to the technology) that have been modified is what you are talking about, and the the third is detecting that an arbitrary image is not a true photograph is what the thread is about and what I was talking about. The guy making the fake will probably never knowingly use an image authentication system and the bulk of the people will never care that the image can't be authenticated to a trusted source any more than they care about that today.

What might be interesting to speculate about is whether official functions, such as legal proceedings, will ever care. Will we get to a point where untraceable images simply won't be accepted as evidence of any kind because they can so readily be faked without the ability to detect it? Will that ever be considered acceptable by the public? Does the accused go free because the video of them committing the crime was taken using an old camera? Does the defendant not get to enter into evidence the photo of them taken someplace a thousand miles away at the time of the crime?

 

Different aspects of the same general problem. I think we can design systems they don't need to care or think about much like HTTPS. You're right, detecting true photographs requires an infrastructure to produce and share images in a way the fake producer won't use by choice but even then a reasonable person will know the image is questionable when systems are in place to trace images. The reasonable person standard of evidence should still apply in legal cases with untraceable images when backed by other evidence of authenticity.

Even authenticating a copy of the original image is not trivially easy with current technology.
A Image Authentication System done badly.
https://blog.elcomsoft.com/2011/04/nikon-image-authentication-system-compromised/

What's actually easy is underestimating how hard it is to design secure systems.
https://www.schneier.com/essays/archives/1997/01/why_cryptography_is.html

The good news about cryptography is that we already have the algorithms and protocols we need to secure our systems. The bad news is that that was the easy part; implementing the protocols successfully requires considerable expertise. The areas of security that interact with people--key management, human/computer interface security, access control--often defy analysis. And the disciplines of public-key infrastructure, software security, computer security, network security, and tamper-resistant hardware design are very poorly understood.

Companies often get the easy part wrong, and implement insecure algorithms and protocols. But even so, practical cryptography is rarely broken through the mathematics; other parts of systems are much easier to break. The best protocol ever invented can fall to an easy attack if no one pays attention to the more complex and subtle implementation issues. Netscape's security fell to a bug in the random-number generator. Flaws can be anywhere: the threat model, the system design, the software or hardware implementation, the system management. Security is a chain, and a single weak link can break the entire system. Fatal bugs may be far removed from the security portion of the software; a design decision that has nothing to do with security can nonetheless create a security flaw.

 

I have been reading some articles about recent discoveries in analyzing half shadow reflections.

Analyzing multiple shadow flux. This detects both sources and objects....in reflections.

And they are getting good at it. So.....unless the 2 videos....are recorded from same angle and light source....they should be able to tell....and do it quickly.

 

I have been reading some articles about recent discoveries in analyzing half shadow reflections.

Analyzing multiple shadow flux. This detects both sources and objects....in reflections.

And they are getting good at it. So.....unless the 2 videos....are recorded from same angle and light source....they should be able to tell....and do it quickly.

And what prevents the bad guys from using the exact same technology to find out what is "wrong" with the image and then modifying it to make it "right"?

 

Well, I don't know. Perhaps speed and memory. They would not divulge all that much about it.

But evidently...most images are combinations of multiple light sources/reflections...even if designed not to.

These multiple sources cast multiple penumbra. These gradients can infer direction and shape of obstruction and source. And of course in action video, there would be a natural progression of these gradients. That there might prevent it.

Whether that could be reconfigured and still be viable and convincing......I don't know.