You could almost certainly produce nearly photo realistic PhotoDNA inversions with a finetuned diffusion model now. Is it possible to create a perceptual hashing algorithm where this isn't possible?
You can certainly not produce inversions. The data that is left in the hash is not enough to produce anything vaguely photorealistic.
However, you can fill the gaps and generate photorealistic photos that fit to the extremely reduced information you get from the hash. You are generating believable (as defined by the training data) photos that fit the hash.
That’s a huge difference.
Statements like yours are extremely dangerous. Without proper understanding of what GenAI can and can not do, people start relying on things that are not there.
Imagine your photorealistic inversion AI putting a mole or a wrinkle in the face of somebody without any foundation in the actual hash. Just because it fits better to the trained data. Explain that to the judge, when the person with just the right facial features sits in front of them.
This modern narrative of people posting their opinions or assumptions somewhere being "dangerous" because someone could just believe it is much more dangerous because it can be applied to any opinion anywhere that was ever published.
No judge will ever rule on something based on a comment they read in the Internet.
Judges usually rely on experts in forensic science who, of course, are infallible and absolutely not influenced by what they read online during their day.
It is dangerous to push the narrative that GenAI can "put information back" where it was once removed. Especially dangerous, because most GenAI is built to put something there that is extremely believable. And while an innocent comment on HN might not play the biggest role, the linked project claims exactly what it can - by definition - not do ("a PhotoDNA hash can be used to produce thumbnail-quality reproductions of the original image") and it looks scientific, too.
Secretive, unaccountable, uncontrollably expanding, driven by shady “independent” NGO which is in fact completely in bed with certain branches of government. Russian DPI censorship system is really like that... Oh, I'm sorry, we're discussing something that is a decade older, and belongs to a “free world”, which is a completely different thing.
These things are simply selling their services to the highest bidder. It's a business model on a power connections market. They are made to be offered to, and controlled by, entities that enjoy having such tools. Sometimes they are also offered to smaller fish, like media corporations, to hurt competitors (pirates and foreign services). Also, social media corporations can proudly state that they themselves “censor nothing”, because it's outsourced.
What's novel about this? iirc, Apple withdrew its plan to hash photos client-side a couple years ago after an outcry. Dropbox has been hashing every file forever to save storage space. Store your shit with a cloud provider, expect it to get scanned, right?
Also, there are a million cute methods to make two different photos produce the same hash; that was actually what the outcry about Apple's version was about. The more the hash algorithm tried to produce the same hash for different variants of a photo, the more likely it was that someone could get their hands on a flagged hash and theoretically send you an innocuous looking photo that registered as CSAM. Pretty sure that's why Apple pulled it.
Technical details are actually irrelevant, people just like to show sophistication by discussing patterns on emperor's new clothes. It was, most likely, a second ad-hoc solution after first ad-hoc solution (most likely, having something to do with md5()) stopped working well enough.
What is important in this article is what isn't written there, and has to be deduced. How exactly “terrorist content” was included. What else has been discussed behind the closed doors. Who actually decides how the thing works.
> that was actually what the outcry about Apple's version was about. The more the hash algorithm tried to produce the same hash for different variants of a photo, the more likely it was that someone could get their hands on a flagged hash and theoretically send you an innocuous looking photo that registered as CSAM.
That was totally infeasible. There were two separate hashes, a public one and a private one, and there needed to be multiple false positives for the system to trigger. So not only would you need to generate collisions for two separate hashes simultaneously, including one for which there are no public details, you would need to do it for several images.
People made a lot of assumptions about how it would work without actually reading the papers Apple published on how it would work. So there’s this caricature of the system in people’s minds that is a lot simpler and easier to fool than the reality. That’s what Apple was forced to react to.
The users privacy is compromised at the point that the public hash had too many hits. An attacker that can implant one hit (e.g. by giving you an altered image that matched genuine illicit material or by using stable diffusion to generate fake illicit material and altering it to match the unaltered hash of an image you possess then submitting it to NCMEC) can obviously also implant multiple.
At that juncture the cryptographic keys are leaked to apple, and all further security depends on apple telling the truth about their process, not being compelled by administrative subpoena, and not ever being unwittingly compromised by hackers or intelligence operatives.
The extra steps of a second perceptual hash and human review are thus not all that relevant, and nor were they clearly enough defined for any analysis of their security properties. Particularly the second perceptual hash's security is apparently at least partially dependent on its obscurity, but you have no reason to believe that it won't be obtained by hackers, rogue employees, intelligence operatives, etc. (And if its obscurity isn't relevant, then why not publish it?).
Even if the hashes were flawless however, the system would be relatively straight forward to attack through less sophisticated means and would retain the overarching philosphical flaw:
Your computing device is your trusted agent-- you share with it material more confidential than your doctor, your lawyer, or your priest. You paid to purchase it. You pay to power it. Increasingly you cannot communicate with family, business partners, or carry out essential and mandatory interactions with you government without using it. Your computing device mediates almost every aspect of your life. As a trusted agent it has absolutely no business scanning your files against unaccountable secret databases, encrypted against your inspection, and undetectably phoning home matches like a KGB spy that you're forced to confide in and house. To do so is a gross betrayal, one that shouldn't just be a bad idea-- it ought to be unlawful.
Service providers scanning content is morally fraught itself, but in our unfortunate current legal standard you have little to no expectation of privacy in information to provide to a third party. And that against-your-own-interests scanning is done on computers owned and operated by the scanners, rather than you. And it's done using access to your information that they already have, so it's a realization of the consequences of existent poor privacy rather than an a new invasion.
The transition to your own devices scanning against you is a bridge to far, no matter how much technical obfuscation is layered onto it.
As someone who has developed privacy technology I found the entire presentation additionally offensive because apple misrepresented the PSI components as protecting the users privacy, when in reality the only purpose for their existence was concealing the list of hashes from the users and thus protecting it from review and criticism. It's one thing for a security scheme to provide insufficient protections, it's quite another to fraudulently present technology which is weaponized against the user as somehow being for them.
and as a minor point of clarification: The structure of the publish perceptual hash (and presumably the non-public one as well) were vulnerable to essentially arbitrary second preimage attacks, not just collisions.
This means that I can take an image and can usually adjust it to have an arbitrary hash, even a hash of an image I've never seen.
It's much more powerful than a collision attack (where the attacker must modify both images).
I doubt have any serious doubt that I would be unable to generate second preimages for two such hashes, but given the second hash was never published that just remains speculation. As AFAIK the first person to develop and demonstrate the second preimage attack against it, I'd like to think that my speculation on this is at least somewhat better than chance. :)
I would assert the only reason they pursued it in the first place is PR/optics, since the "optics" of not being able to proactively police what users do using E2EE services you provide is somewhat a problem. That said, I think the concept of having your own computer covertly report you to the authorities is a level too dystopian to accept even from Apple.
I agree the reason they pulled it was probably PR/optics. But given the problems with human reviews of apps on the app store, I wouldn't be confident that an underpaid employee somewhere wouldn't blindly agree with the algorithm.
Going from memory here but IIRC the deal was that on device they'd produce a hash using a known pHash, and if that was positive, they'd send the photo to check it against a second pHash that wasn't publicly-disclosed (to try to mitigate the problem of intentional collisions) and then if both of them were positive matches, they would have human reviewers in the loop.
Not "if then send" -- if that were the case you could detect that your images were matching and then use the system as an oracle to e.g. detect that the database might contain disfavored political content in addition to whatever they claim it contains.
The system was proposed as part of the upload process to your private encrypted cloud storage. So they already have the files. The system was designed so that if there were a sufficient number of hits against an encrypted database then a party possessing a particular private key would learn the private key for the encrypted files, as a passive effect of there being matches.
Like if I mutter random letters of my password while entering it, you'd eventually passively learn my password just by hanging around while I logged in each morning.
This way any matching is completely undetectable on the users end, at least until the private key possessing parties or the people they share the information with choose to take some action against the user.
It was a lot more advanced and abuse-resistant than people assumed. I really wish people had read how it worked instead of guessing it was something a lot simpler. There were two different perceptual hashes. If both matched, and the number of positive matches was high enough, a thumbnail would be able to be decrypted by Apple. Neither the device nor the server were able to independently check for a match, so the device wasn’t able to just scan all your files and flag the matches. It was tied into the iCloud upload process.
While this is understandable, the unfortunate issue was that Apple could be coerced into adding images certain authoritarian governments didn’t like to the list. Though imo it’s all moot if iCloud Photos aren’t end to end encrypted anyway.
The fact that it is CSAM makes it an even harder problem to solve. With e.g. copyright infringement, you could keep some kind of records why a particular file is in the system, potentially even letting trusted and vetted organizations audit the files, but doing that for CSAM would be highly illegal and defeat the purpose of the system.
“Coerced”? Check some recent news to see that for Apple rainbow-washing stops at the very moment they are held responsible for providing basic censorship circumvention tools.
I am amazed how people still cling to the hope that one day a corporation will do something nice for them without any hidden motive.
Yeah Hacker Factor's multi-post critiques are where I first saw it analyzed. For reference they run the popular fotoforensics.com image analysis site.
They also have scathing critique (eg [1]) about the Adobe-led C2PA digital provenance signing, having themselves been part of various groups that seek solutions to the provenance problem.
https://anishathalye.com/inverting-photodna/
You could almost certainly produce nearly photo realistic PhotoDNA inversions with a finetuned diffusion model now. Is it possible to create a perceptual hashing algorithm where this isn't possible?
You can certainly not produce inversions. The data that is left in the hash is not enough to produce anything vaguely photorealistic.
However, you can fill the gaps and generate photorealistic photos that fit to the extremely reduced information you get from the hash. You are generating believable (as defined by the training data) photos that fit the hash.
That’s a huge difference.
Statements like yours are extremely dangerous. Without proper understanding of what GenAI can and can not do, people start relying on things that are not there.
Imagine your photorealistic inversion AI putting a mole or a wrinkle in the face of somebody without any foundation in the actual hash. Just because it fits better to the trained data. Explain that to the judge, when the person with just the right facial features sits in front of them.
This modern narrative of people posting their opinions or assumptions somewhere being "dangerous" because someone could just believe it is much more dangerous because it can be applied to any opinion anywhere that was ever published.
No judge will ever rule on something based on a comment they read in the Internet.
Judges usually rely on experts in forensic science who, of course, are infallible and absolutely not influenced by what they read online during their day.
https://innocenceproject.org/misapplication-of-forensic-scie...
It is dangerous to push the narrative that GenAI can "put information back" where it was once removed. Especially dangerous, because most GenAI is built to put something there that is extremely believable. And while an innocent comment on HN might not play the biggest role, the linked project claims exactly what it can - by definition - not do ("a PhotoDNA hash can be used to produce thumbnail-quality reproductions of the original image") and it looks scientific, too.
Hold my beer: symmetric flip. Flip the photo horizontally and it’s essentially the “same” image without a hash collision
Secretive, unaccountable, uncontrollably expanding, driven by shady “independent” NGO which is in fact completely in bed with certain branches of government. Russian DPI censorship system is really like that... Oh, I'm sorry, we're discussing something that is a decade older, and belongs to a “free world”, which is a completely different thing.
These things are simply selling their services to the highest bidder. It's a business model on a power connections market. They are made to be offered to, and controlled by, entities that enjoy having such tools. Sometimes they are also offered to smaller fish, like media corporations, to hurt competitors (pirates and foreign services). Also, social media corporations can proudly state that they themselves “censor nothing”, because it's outsourced.
There's a great portrayal of people who run such services: https://www.newyorker.com/magazine/2019/11/04/a-cybersecurit...
What's novel about this? iirc, Apple withdrew its plan to hash photos client-side a couple years ago after an outcry. Dropbox has been hashing every file forever to save storage space. Store your shit with a cloud provider, expect it to get scanned, right?
Also, there are a million cute methods to make two different photos produce the same hash; that was actually what the outcry about Apple's version was about. The more the hash algorithm tried to produce the same hash for different variants of a photo, the more likely it was that someone could get their hands on a flagged hash and theoretically send you an innocuous looking photo that registered as CSAM. Pretty sure that's why Apple pulled it.
Technical details are actually irrelevant, people just like to show sophistication by discussing patterns on emperor's new clothes. It was, most likely, a second ad-hoc solution after first ad-hoc solution (most likely, having something to do with md5()) stopped working well enough.
What is important in this article is what isn't written there, and has to be deduced. How exactly “terrorist content” was included. What else has been discussed behind the closed doors. Who actually decides how the thing works.
> that was actually what the outcry about Apple's version was about. The more the hash algorithm tried to produce the same hash for different variants of a photo, the more likely it was that someone could get their hands on a flagged hash and theoretically send you an innocuous looking photo that registered as CSAM.
That was totally infeasible. There were two separate hashes, a public one and a private one, and there needed to be multiple false positives for the system to trigger. So not only would you need to generate collisions for two separate hashes simultaneously, including one for which there are no public details, you would need to do it for several images.
People made a lot of assumptions about how it would work without actually reading the papers Apple published on how it would work. So there’s this caricature of the system in people’s minds that is a lot simpler and easier to fool than the reality. That’s what Apple was forced to react to.
Plenty of commentary did address those points.
The users privacy is compromised at the point that the public hash had too many hits. An attacker that can implant one hit (e.g. by giving you an altered image that matched genuine illicit material or by using stable diffusion to generate fake illicit material and altering it to match the unaltered hash of an image you possess then submitting it to NCMEC) can obviously also implant multiple.
At that juncture the cryptographic keys are leaked to apple, and all further security depends on apple telling the truth about their process, not being compelled by administrative subpoena, and not ever being unwittingly compromised by hackers or intelligence operatives.
The extra steps of a second perceptual hash and human review are thus not all that relevant, and nor were they clearly enough defined for any analysis of their security properties. Particularly the second perceptual hash's security is apparently at least partially dependent on its obscurity, but you have no reason to believe that it won't be obtained by hackers, rogue employees, intelligence operatives, etc. (And if its obscurity isn't relevant, then why not publish it?).
Even if the hashes were flawless however, the system would be relatively straight forward to attack through less sophisticated means and would retain the overarching philosphical flaw:
Your computing device is your trusted agent-- you share with it material more confidential than your doctor, your lawyer, or your priest. You paid to purchase it. You pay to power it. Increasingly you cannot communicate with family, business partners, or carry out essential and mandatory interactions with you government without using it. Your computing device mediates almost every aspect of your life. As a trusted agent it has absolutely no business scanning your files against unaccountable secret databases, encrypted against your inspection, and undetectably phoning home matches like a KGB spy that you're forced to confide in and house. To do so is a gross betrayal, one that shouldn't just be a bad idea-- it ought to be unlawful.
Service providers scanning content is morally fraught itself, but in our unfortunate current legal standard you have little to no expectation of privacy in information to provide to a third party. And that against-your-own-interests scanning is done on computers owned and operated by the scanners, rather than you. And it's done using access to your information that they already have, so it's a realization of the consequences of existent poor privacy rather than an a new invasion.
The transition to your own devices scanning against you is a bridge to far, no matter how much technical obfuscation is layered onto it.
As someone who has developed privacy technology I found the entire presentation additionally offensive because apple misrepresented the PSI components as protecting the users privacy, when in reality the only purpose for their existence was concealing the list of hashes from the users and thus protecting it from review and criticism. It's one thing for a security scheme to provide insufficient protections, it's quite another to fraudulently present technology which is weaponized against the user as somehow being for them.
>That’s what Apple was forced to react to.
Perhaps apple could have made a better explanation available? Unless, there wasn't one?
> generate collisions
and as a minor point of clarification: The structure of the publish perceptual hash (and presumably the non-public one as well) were vulnerable to essentially arbitrary second preimage attacks, not just collisions.
This means that I can take an image and can usually adjust it to have an arbitrary hash, even a hash of an image I've never seen.
It's much more powerful than a collision attack (where the attacker must modify both images).
I doubt have any serious doubt that I would be unable to generate second preimages for two such hashes, but given the second hash was never published that just remains speculation. As AFAIK the first person to develop and demonstrate the second preimage attack against it, I'd like to think that my speculation on this is at least somewhat better than chance. :)
> What's novel about this?
Literally in the first paragraph you can see that it’s not novel.
> PhotoDNA was developed by Microsoft Research and Hany Farid, professor at Dartmouth College, beginning in 2009
Also this isn't novel.ycombinator.com
> theoretically send you an innocuous looking photo that registered as CSAM. Pretty sure that's why Apple pulled it.
Apple also had human reviewers in the mix. The only reason they pulled it was PR/optics.
I would assert the only reason they pursued it in the first place is PR/optics, since the "optics" of not being able to proactively police what users do using E2EE services you provide is somewhat a problem. That said, I think the concept of having your own computer covertly report you to the authorities is a level too dystopian to accept even from Apple.
I agree the reason they pulled it was probably PR/optics. But given the problems with human reviews of apps on the app store, I wouldn't be confident that an underpaid employee somewhere wouldn't blindly agree with the algorithm.
Wouldn't that require exfiltrating the original photo? I remember them swearing that wasn't part of the deal.
Going from memory here but IIRC the deal was that on device they'd produce a hash using a known pHash, and if that was positive, they'd send the photo to check it against a second pHash that wasn't publicly-disclosed (to try to mitigate the problem of intentional collisions) and then if both of them were positive matches, they would have human reviewers in the loop.
Not "if then send" -- if that were the case you could detect that your images were matching and then use the system as an oracle to e.g. detect that the database might contain disfavored political content in addition to whatever they claim it contains.
The system was proposed as part of the upload process to your private encrypted cloud storage. So they already have the files. The system was designed so that if there were a sufficient number of hits against an encrypted database then a party possessing a particular private key would learn the private key for the encrypted files, as a passive effect of there being matches.
Like if I mutter random letters of my password while entering it, you'd eventually passively learn my password just by hanging around while I logged in each morning.
This way any matching is completely undetectable on the users end, at least until the private key possessing parties or the people they share the information with choose to take some action against the user.
It was a lot more advanced and abuse-resistant than people assumed. I really wish people had read how it worked instead of guessing it was something a lot simpler. There were two different perceptual hashes. If both matched, and the number of positive matches was high enough, a thumbnail would be able to be decrypted by Apple. Neither the device nor the server were able to independently check for a match, so the device wasn’t able to just scan all your files and flag the matches. It was tied into the iCloud upload process.
While this is understandable, the unfortunate issue was that Apple could be coerced into adding images certain authoritarian governments didn’t like to the list. Though imo it’s all moot if iCloud Photos aren’t end to end encrypted anyway.
The fact that it is CSAM makes it an even harder problem to solve. With e.g. copyright infringement, you could keep some kind of records why a particular file is in the system, potentially even letting trusted and vetted organizations audit the files, but doing that for CSAM would be highly illegal and defeat the purpose of the system.
“Coerced”? Check some recent news to see that for Apple rainbow-washing stops at the very moment they are held responsible for providing basic censorship circumvention tools.
I am amazed how people still cling to the hope that one day a corporation will do something nice for them without any hidden motive.
Personally I think they had a lot of pressure from naive external and internal sources and had to take the PR hit on this one.
If Apple were to do something like that with any level of real intent, the first thing you’d find out about it was when it went live.
Is there any good article/paper that describes how it actually works or is implemented not just in high-level and hand-waving terms?
https://hackerfactor.com/blog/index.php?/archives/931-PhotoD... is linked from https://anishathalye.com/inverting-photodna/ and is likely a good starting point
Yeah Hacker Factor's multi-post critiques are where I first saw it analyzed. For reference they run the popular fotoforensics.com image analysis site.
They also have scathing critique (eg [1]) about the Adobe-led C2PA digital provenance signing, having themselves been part of various groups that seek solutions to the provenance problem.
[1] https://www.hackerfactor.com/blog/index.php?/archives/1013-C...
thanks!
There tends to be more information under the search term "perceptual hashing"
The secrecy of the inner tech is intentional.