You people realize that if Apple convinced governments that this hash-based limited-database system is ENOUGH we could get MORE privacy out of this, since Apple would have an argument to not open up our data to ACTUAL scanning asked by the “think of the children” and “think of the terrorists” crowds?
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Apple Addresses CSAM Detection Concerns, Will Consider Expanding System on Per-Country Basis
- Thread starter MacRumors
- Start date
- Sort by reaction score
How do we know they haven't, I doubt anyone on this forum is privy to what the powers in charge in various countries ask Apple and what Apple gives in return tbh. Also I was talking really about iMessage communication here, as well as maybe notes and pages etc, how do we know what can and what cant be scanned for ultimately? We as I said are not privy to those conversations but looking around I'm pretty sure they have been asked, governments hate encryption and Apple just opened things up.
I saw a comment on here which made me smile, someone said "Its not a backdoor its a front door or maybe a side door" something like that, I cant be bothered to go back and look for it, but those are all doors, and they are not things you leave open for people to look though let alone walk though or take snapshots though, also any door needs to be secure which is why in this industry they are frowned on so much and to announce to the world and its cat that you've added one, well nothing is hack proof and the fears of government abuse are very real I would imagine. After all when you get major privacy advocates and organisations like the EFF and professors in this field that have way more knowledge than most of us here on this subject, unless you work in that field of course. When they say this is a bad move, I sit up and take notice.
Unfortunately this not the way it works. What Apple is doing is actually neural matching (they call it neural hashing because of hashed data is matched which can be done). They have used aprox. 200 000 images to teach the AI. All the images are scanned and if they fall within a threshold the report is being made. Effectively they are not looking for exact matches but similarities.
The main problem is “the black box nature” of the process. There is no independent vetting of the process. Actually we have no idea what data is used to train the AI. Also once this system is in place it can be trained to look for anything. This makes iOS devices the most effective CCTV network in the world especially if combined with NSO Group tech.
However you slice it, it’s matching against a limited known database and it’s said to make an error every 1176 years (I converted Apple’s 1 in a trillion claim in years in a previous post), it’s not “scanning” to try to figure out if an image constitutes a completely new offence on the spot.
That’s sniffing blindfolded, not looking at our pics.
The 7 months must be blamed on the Spanish police and court system. And for American policy and Spanish police not having agreed to a common date format.
The problem here is that Facebook didn't have any verifiable personal information about the user uploading the material. Only the IP address. You are much more prone to error when you don't have any personal information about the suspect.
Apple can in most cases provide mobile phone number, credit/debet card information and in many cases also the real name and address. This information can be cross-checked with cell provider and credit card provider.
Apple don't have the IP address so this hunt for unknown people can't really happen.
The biggest problem would be someone using a stolen identity. But this is something law enforcement agencies must handle. We can't stop reporting crimes because maybe the police does a bad job sometimes.
No matter the system, some innocent will suffer.
This system will work if these people all have the same iconic pictures on the phone in their Photo Library. But it will not work if you have a bunch of general pictures from a protest.
If the Chinese government wants to catch these people there are much more efficient technologies from Apple TODAY.
The Chinese government could force Apple today to do this much more efficient and Apple wouldn't have to change ANYTHING on the iPhones.
Professors are humans and have agendas too.
We’ve seen climate change denying professors and professors signing the Great Barrington Declaration last year.
Just saying.
The thing is people with much more knowledge than you and I consider this a really bad move for privacy, you seem blinkered to ulterior government uses for this door for some reason. Well go enjoy a path I hoped Apple would not go down, its not really a free world, but enjoy thinking this will help, because it wont. Its much harder to hack a letter given by hand or photos swapped at a meet I imagine, which is why the criminals of this world are not dumb enough to use these systems I imagine to purvey their disgusting images or terrorist ideologies.
Even Russia uses typewriters I'm sure our various government's do too, good old fashioned mechanical ones which by their nature are nicely air gapped because they can't go online. Dont get fooled this is for our own good to get the baddies and save the planet, its cracking privacy wide open bit by bit and the advocates and organisations who supported Apple on their stance towards privacy are now shouting that this is very wrong, I'm sure they know more than we do as its their job to. So if you stand back from Apples good old distortion field long enough you can see the very negative and dangerous implications for people fighting for free speech or LGTPQ+rights in countries run by the likes of Alexander Lukashenko, or Vladimir Putin to name just two. But what do security experts around the world, the EFF and even an academic at the US Johns Hopkins University know about this hey?
That's a pretty good point - the amount of independent scrutiny Apple should subject their actual implemention of the algorithms must match the gravity of their possible accusations against their users, if they want to regain trust. We have to be able (a) evaluate the source code and (b) subject devices to white-box attacks to test against possible misuses.
Sadly, based on experience Apple will probably never allow that leaving us all guessing and worrying.
Don't like the idea of clock-cycles and battery being used for purposes I didn't opt-in to. No matter how good the cause... And the fact that this tech can (and probably will) at some point be abused for other purposes should have made it a big no-go if apple is serious on privacy.
According to Apple's technical description:
"Perceptual hashing bases this number on features of the image instead of the precise values of pixels in the image. (...) ensures that perceptually and semantically similar images have close descriptors (...) The main purpose of the hash is to ensure that identical and visually similar images result in the same hash..."
Semantically similar - yes, Apple tries (to some undisclosed degree) to figure out what your pictures show. In the strictest sense it may not be trying to figure out if an image constitutes a completely new offence. But it is checking whether your picture is visually similar to one of ts set of 200,000 depictions of known offences, so the difference is academic. Better hope none of your legal pictures happens to have the same wallpaper in the background as CSA-pic #136,274.
They scan every app however there is an infinite way of obfuscating the use of private api's. The mouse always wins...
That’s supposed to catch edited (cropped, altered, etc.) versions of the known offending pics.
It won’t get your unrelated pic and in the rare chance it will get it, it won’t even make it to human review ‘cause you need multiple matches to raise a red flag…
Suppose you’re super unlucky and get multiple false positive matches (implausible), human review to the rescue..
The NeuralHash and MicrosoftDNA algorithms were designed to do the opposite of what you are describing. NeuralHash is optimised to not catch pictures of a similar "feel".
Both of these systems are optimised only to catch changes to a specific photo: cropping, changing colour, hue contrast, mirroring.
Here's an example:
Picture 1: Someone takes a picture of you where you live. Lets say you standing in front of a window.
You walk away for a few minutes.
Picture 2: Someone takes a picture of you standing in front of the same window and you have approx. the same pose as in the last picture.
Somehow picture 1 becomes part of the CSAM database. If NeuralHash is any good it should just flag picture 1 and not picture 2.
That's why its do difficult to misuse it. You can't just provide similar pictures of a protest and hope for the system to get people with "pictures of protests".
Still, you will get collisions and it's why Apple has a threshold on collisions before they are notified. With the threshold they have chosen they have estimated it to be 1 in 1 trillion accounts per year.
Yep an automated algorithm paired with human review is why YouTube’s copyright system works so well and no one has any complaints of false positives.
Have Snowden read the technical brief? I don't consider him an authority on this subject which are hashing algorithms.
The NeuralHash algorithms together with MicrosoftDNA which do the same time are AFAIKcontent neutral. They are optimised to identify specific photos and photos derived from those specific photos.
If you replaced the CSAM database with a database of general copyrighted images, the system would also be good at finding those.
Can you point me to where Snowden says these algorithms are optimised to catch child pornography and his arguments for his conclusions?
Yeah, you never acknowledged your misguided probability estimate I corrected so not sure what game are you playing here by citing random unrelated examples of fails of human review in totally unrelated contexts.
You seem more interested in making noise towards your view than looking at actual facts and numbers.
You are assuming that the “governments” are some uniform entities where everybody is always on the same page and no power struggle exists between different powerful groups who have very different interests. Who is exactly going to provide “a list of hashes” is a very big question.
This is a difficult question, IMO. As a grown-up, I personally do not need a mix between a nanny and a police officer looking over my shoulder every time I use my iPhone. To make it more ironic, I would pay to them to look over my shoulder with my iCloud subscription.
Most people might not be aware of what Apple are doing to begin with, even though this story has made it to some mainstream news by now:
Apple wants to check your phone for child abuse images – what could possibly go wrong? | Arwa Mahdawi
On the surface Apple’s new features sound both sensible and commendable – but they also open a Pandora’s box of privacy and surveillance issues
Apple criticised for system that detects child abuse
There are privacy concerns that the system will be expanded to scan phones for political speech.
www.bbc.com
For me, personally, abandoning iCloud would be quite difficult, as I store quite a bit of data there, including my iPhone & iPad backups. Of course I could go back to saving my stuff on the computer, external discs, etc., but this would feel like a big step back, technically speaking. For now I will see how this story develops.
I am hopeful that as the news gains traction, highlighting obvious paradoxes between being a “privacy advocate” and actively scanning our devices, Apple will have to either abandon it altogether or give us some solid guarantees that this new tech cannot be used the wrong way.
A small add-on. If Apple want to involve any of their “AI” into it, I am quite worried, seeing that they have rendered Siri almost unusable.
Last edited:
What actual facts did you present? Apple made a claim they refuse to back up with any validation.
Youtube's problem is much more difficult and is of a different nature. There are so many factors increasing the complexity and difficulty for YouTube. It seems to me to be almost an unsurmountable problem.
Apple's problem was solved by Microsoft (and Google) 10 years ago. It is a much simpler problem.
The review process is extremely simpler. Deciding if something is an extreme form of child pornography is much easier if some content violate copyright case law.
You don't need an attacker per se, you just need someone inside Apple. In the end, they're humans, and they can be coerced. As farfetched that may sound, the opening has been created in such a Blackbox system.
What you are describing is not perceptual hashing. You seem to describe some kind of machine learning.
Perceptual hashes for photos are designed to not change much when an image undergoes minor modifications such as compression, some cropping, doing colour correction, changing hue and brightness.
It's not good for finding pictures with similar "feel" which is what you are describing.
Anyway, Apple is using Hyperplane locality sensitivity hashing. It doesn't require any dataset to learn AFAIK.
In fact, Apple never got access to the CSAM-database so there would be some problem for Apple to train its algorithms.