The Problem with Perceptual Hashes • Hacker News Board

Re: The Problem with Perceptual Hashes

#141

Post by giantrobot » Sat, Aug 07, 2021, 1:42 AM UTC

ris wrote: post #75

I agree with the article in general except part of the final conclusion > The simple fact that image data is reduced to a small number of bits leads to collisions and therefore false positives Our experience with regular hashes suggests this is not the underlying problem. SHA256 hashes have 256 bits and still there are no known collisions, even with people deliberately trying to find them. SHA-1 only has only 160 bit…

Perceptual hashes don't involve diffusion and confusion steps like cryptographic hashes. Perceptual hashes don't want decorrelation like cryptographic hashes. In fact they want similar but not identical images to end up with similar hash values.

Re: The Problem with Perceptual Hashes

#142

Post by dylan604 » Sat, Aug 07, 2021, 1:44 AM UTC

yellow_lead wrote: post #61

Regarding false positives re:Apple, the Ars Technica article claims > Apple offers technical details, claims 1-in-1 trillion chance of false positives. There are two ways to read this, but I'm assuming it means, for each scan, there is a 1-in-1 trillion chance of a false positive. Apple has over 1 billion devices. Assuming ten scans per device per day, you would reach one trillion scans in ~100 days. Okay, but not al…

knowing Apple, the initial scan of this will be done while the phone is on charge just like previous versions of scanning your library. However, according to Apple it is just the photos shared with iCloud. So since it's on a charger, it's minimal electron abuse.

Once you start adding new content from camera to iCloud, I'd assume the new ML chips of Apple Silicone will be calculating the phashes as part-and-parcel to everything else it does. So unless you're trying to "recreate" known CP, then new photos from camera really shouldn't need this hashing done to them. Only files not originated from the user's iDevice should qualify. If a CP creator is using an iDevice, then their new content won't match existing hashes, so what's that going to do?

So so many questions. It's similar yet different to mandatory metal detectors and other screening where 99.99% of people are innocent and "merely" inconvenienced vs the number of people any of that screening catches. Does the mere existence of that screening act as a deterent? That's like asking how many angels can stand on the head of a pin. It's a useless question. The answer can be whatever they want it to be.

Re: The Problem with Perceptual Hashes

#143

Post by fogof » Sat, Aug 07, 2021, 1:58 AM UTC

IfOnlyYouKnew wrote: post #20

Apple’s documents said they require multiple hits before anything happens, as the article notes. They can (and have) adjusted that number to any desired balance of false positive to negatives. How can they say it’s 1 in a trillion? You test the algorithm on a bunch of random negatives, see how many positives you get, and do one division and one multiplication. This isn’t rocket science. So, while there are many argum…

I was unhappy to find this comment so far down and even unhappier to see it downvoted. I'm not a fan of the decrease in privacy Apple is creating with this move but I think this forum has gotten its feelings for Apple caught up with its response to a completely valid criticism of an anti-Apple article.

To explain things even further, let's say that the perceptual algorithm makes a false positive 1% of the time. That is, 1 in every 100 completely normal pictures are incorrectly matched with some picture in the child pornography database. There's no reason to think (at least none springs to mind, happy to hear suggestions) that a false positive in one image will make it any more likely to see a false positive in another image. Thus, if you have a phone with 1000 pictures on it, and it takes 40 trigger a match, there's less than a 1 in a trillion probability that this would happen if the pictures are all normal.

Re: The Problem with Perceptual Hashes

#144

Post by Renaud » Sat, Aug 07, 2021, 1:59 AM UTC

Klonoar wrote: post #114

mrtksn wrote: post #72

The technical challenges aside, I’m very disturbed that my device will be reporting me to the authorities. That’s very different from authorities taking a sneak peek into my stuff. That’s like the theological concept of always being watched. It starts with child pornography but the technology is indifferent towards it, it can be anything. It’s always about the children because we all want to save the children. Soon t…

I would really like people to start answering this: what exactly do you think has changed? e.g, >That’s very different from authorities taking a sneak peek into my stuff. To be very blunt: - The opt out of this is to not use iCloud Photos. - If you _currently_ use iCloud Photos, your photos are _already_ hash compared. - Thus the existing opt out is to... not use iCloud Photos. The exact same outcome can happen regar…

What has changed is the inclusion of spyware technology on the device that can be weaponised to basically report on anything.

Today it's only geared toward iCloud and CSAM. How many lines of codes do you think it will take before it scans all your local pictures?

How hard do you think it will be for an authoritarian regime like China, that Apple bends over backwards to please, to start including other hashes that are not CSAM?

iCloud is opt-out. They can scan server-side like everyone does. Your device is your device, and it now contains, deeply embedded into it, the ability to perform actions that are not under your control and can silently report you directly to the authorities.

If you don't see a deep change there, I don't know what to say.

I live in a country that is getting more authoritarian by the day, where people are sent to prison (some for life) for criticizing the government, sometime just for chanting or printing a slogan.

This is the kind of crap that makes me extremely angry at Apple. Under the guise of something no-one can genuinely be against (think of the children!), they have now included a pretty generic snitch into your phone and made everyone accept it.

Re: The Problem with Perceptual Hashes

#145

Post by chipotle_coyote » Sat, Aug 07, 2021, 2:11 AM UTC

SavantIdiot wrote: post #66

This article covers three methods, all of which just look for alterations of a source image to find a fast match (in fact, that's the paper referenced). It is still a "squint to see if it is similar" test. I was under the impression there were more sophisticated methods that looked for types of images, not just altered known images. Am I misunderstanding?

Apple's proposed system compares against a database of known images. I can't think of a way to "look for types of images" other than trying to do it with machine learning, which strikes me as fraught with incredible fiasco potential. (The compare-to-a-known-database approach has its own issues, including the ones the article talks about, of course.)

Re: The Problem with Perceptual Hashes

#146

Post by cvwright » Sat, Aug 07, 2021, 2:19 AM UTC

starkd wrote: post #59

The method Apple is using looks more like a cryptographic hash. That's entirely different (and more secure) than a perceptual hash. From https://www.apple.com/child-safety/ "Before an image is stored in iCloud Photos, an on-device matching process is performed for that image against the known CSAM hashes. This matching process is powered by a cryptographic technology called private set intersection, which determines…

The crypto here is for the private set intersection, not the hash.

So your device has a list of perceptual (non-cryptographic) hashes of its images. Apple has a list of the hashes of known bad images.

The protocol lets them learn which of your hashes are in the “bad” set, without you learning any of the other “bad” hashes, and without Apple learning any of the hashes of your other photos.

Re: The Problem with Perceptual Hashes

#147

Post by Waterluvian » Sat, Aug 07, 2021, 2:26 AM UTC

I’m rather fascinated by the false matches. Those two images are very different and yet beautifully similar.

I want to see a lot more pairs like this!

Re: The Problem with Perceptual Hashes

#148

Post by thaumasiotes » Sat, Aug 07, 2021, 2:27 AM UTC

ezoe wrote: post #110

The problem of hash or NN based matching is, the authority can avoid explaining the mismatch. Suppose the authority want to false-arrest you. They prepare a hash that matches to an innocent image they knew the target has in his Apple product. They hand that hash to the Apple, claiming it's a hash from a child abuse image and demand privacy-invasive searching for the greater good. Then, Apple report you have a file th…

> like the dogs which conveniently bark at police's secret hand sign

This isn't necessary; the state of the art is for drug dogs to alert 100% of the time. They're graded on whether they ever miss drugs. It's easy to never miss.

Re: The Problem with Perceptual Hashes

#149

Post by klodolph » Sat, Aug 07, 2021, 2:29 AM UTC

mirker wrote: post #125

klodolph wrote: post #2

> Even at a Hamming Distance threshold of 0, that is, when both hashes are identical, I don’t see how Apple can avoid tons of collisions... You'd want to look at the particular perceptual hash implementation. There is no reason to expect, without knowing the hash function, that you would end up with tons of collisions at distance 0.

If images have cardinality N and hashes M and N > M, then yes, by pigeonhole principle you will have collisions regardless of hash function, f: N -> M. N is usually much bigger than M, since you have the combinatorial pixel explosion. Say images are 8 bit RGB 256x256, then you have 2^(8x256x256x3) bit combinations. If you have a 256-bit hash, then that’s only 2^256. So there is a factor of 2^(8x256x3) difference betw…

The number of possible different images doesn't matter, it's only the number of actually different images encountered in the world. This number cannot be anywhere near 2^256, that would be physically impossible.

Re: The Problem with Perceptual Hashes

#150

Post by wonnage » Sat, Aug 07, 2021, 2:32 AM UTC

Renaud wrote: post #144

Klonoar wrote: post #114

Earlier quoted context omitted.

I would really like people to start answering this: what exactly do you think has changed? e.g, >That’s very different from authorities taking a sneak peek into my stuff. To be very blunt: - The opt out of this is to not use iCloud Photos. - If you _currently_ use iCloud Photos, your photos are _already_ hash compared. - Thus the existing opt out is to... not use iCloud Photos. The exact same outcome can happen regar…

What has changed is the inclusion of spyware technology on the device that can be weaponised to basically report on anything. Today it's only geared toward iCloud and CSAM. How many lines of codes do you think it will take before it scans all your local pictures? How hard do you think it will be for an authoritarian regime like China, that Apple bends over backwards to please, to start including other hashes that are…

We gotta stop with the China bogeyman every time a privacy issue comes up. This is a feature designed by an American company for American government surveillance purposes. China is perfectly capable of doing the same surveillance or worse on its own citizens, with or without Apple. China has nothing to do with why American tech is progressively implementing more authoritarian features in a supposedly democratic country.