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's Proposed Phone-Scanning Child Safety Features 'Invasive, Ineffective, and Dangerous,' Say Cybersecurity Researchers in New Study
- Thread starter MacRumors
- Start date
- Sort by reaction score
The algorithms are all similar, the way the set of algorithms is applied and parameterized is different for Apple. Again, all discussed before in past threads, not going to repeat it. I have to repeat myself all day long for my computer science and math students, don't need to do it for MR members.
No, it doesn't recognise similar features in a picture.
The convoluted neural network has two outputs:
1) floating point descriptors of original/perturbed pairs and
2) floating point descriptors of original/distractor pairs.
1) is for recognising photos but 2) is for not recognising similar photos.
The hash is produced by feeding the hashing algorithm all the floating point descriptors.
It is 2) which makes the CSAM Detection System poorly suited to find pictures of similar nature.
Let's say you took picture of a police car, then moved 1 feet to either side and took another picture. If picture 1 was in the database but not picture 2, the CSAM Detection System would not recognise it even though it was almost the same picture. Again because of 2.
More to the point, Apple already DOESN'T refuse demands by Chinese, Arab and Russian authorities now (and probably a lot of other countries), and hides behind the 'we have to obey the laws of that country.'
“You’re going to find a way to do this or we’re going to go do it for you,” said Senator Lindsey Graham. “We’re not going to live in a world where a bunch of child abusers have a safe haven to practice their craft. Period. End of discussion.”
U.S. senators threaten Facebook, Apple with encryption regulation
U.S. senators grilled Apple Inc and Facebook Inc executives over their encryption practices on Tuesday and threatened to regulate the technology unless the companies make encrypted user data accessible to law enforcement.
Ran into this article which gives some great links and summarized it fairly well
Client-side content scanning is a disaster for democracy
Hopefully we're still listening to expertswww.theregister.com
#1
"CSS currently relies on one of two methods for image scanning: perceptual hashing, in which an algorithm generates a hash (a numeric identifier) that functions as a digital fingerprint and produces the same hash if the image is altered in a minor way;"
#2
"and machine learning, in which a machine learning model is trained to recognize target content, even images it hasn't seen before. Both methods, the paper points out, generate false positives, may rely on proprietary technologies that limit auditing, can be subverted, and can be evaded."
#1 is the CSAM Detection System and its ill equipped to be misused compared to #2.
Yet, almost everyone is concerned about #1 when they really should be concerned about #2.
Why would governments use method 1 when method 2 is already in all phones and can just be changed slightly to report it findings to the government?
That exception is a temporary measure which AFAIK is considered to be legally flawed and unable to hold in court even by the lawmakers who worked on it, who stated It was done quickly and under pressure and that there is need of a significantly better solution which addresses privacy considerations.
If they want to legislate something that holds in court they would likely need to add an explicit clause to the GDPR, or find a way to successfully argue that the scanning is necessary, which from a legal point of view is actually difficult.
You won't find it because Apple hasn't implemented the feature yet.
But you'll find Apple can scan iCloud content in the iCloud agreement.
So what kind of images could the government force into such a database which could be used to catch unwanted elements.
It would have to be images which are shared directly on a large number of devices to be useful.
The capability to do #1 foretells the ability to do #2. Remember that what is new on phones are the chips optimised to do AI computations.
I am pretty sure those of us who pointed out the problems of Apple's proposed CSAM surveillanceware and were accused of being ignorant have the right to say 'We told you so'. And our concerns the governments would take up systems like Apple's and use them for surveillance seems to be coming true.
We really do need to think about where technology is going, both with surveillance and robotics. Leaving important decisions about how technology is used to engineers won't do. We are at a crossroads now and the impact of our decisions we make now will have an overwhelming impact on the quality of life of our descendants. Time to think more carefully about whom we elect to office.
Link
It seems like a cop-out.
The generation of floating point descriptors for original/perturbed pairs by the convoluted neural network being fed into the hashing algorithm is exactly why NeuralHash is poor at recognising similar images.
You seem to think otherwise.
Yes, as far as I understand, but it's clear there is some political push to try to erode the privacy protections or introduce exceptions for some types of surveillance. The first attempt was flawed and likely ineffective, but I'm sure there will be other attempts.
That's why the more coverage and backlash these kind of initiatives get the better, even if the initiative currently might not impact your jurisdiction directly.
we know already, these researchers already know everybody in the world knows the score. They also know Apple postponed without stating anything further due to the already ‘more than enough’ backlash!
yet still none of them care to speak anything about how and why and lack of communication that Amazon, Microsoft, Facebook , and Google have done on this very same topic.
I find that like paid agents to level Apple as they turn their focus simply for excuse. Maybe researchers should become like sciences where they too should follow the scientific method and not exclude observation against existing parties.
The use of an image blaring "CHILD SAFETY" as the featured image for this article on MacRumors shows where the sympathies of the site's editorial staff lie. Not good, folks.
The algorithms are all similar, the way the set of algorithms is applied and parameterized is different for Apple. Again, all discussed before in past threads, not going to repeat it. I have to repeat myself all day long for my computer science and math students, don't need to do it for MR members.
I haven't seen any discussion about the distraction generation.
But even Microsofts PhotoDNA, just like Apple's CSAM Detection System, is poor at detecting similar images.
The purpose of the two systems is twofold:
1) Detect images which are exactly (or close derivatives) of the images in the CSAM databsae
2) Not detect images which are similar (and even dissimilar)
It's 2) which makes the system such a poor candidate for misuse when there already exist technologies on the iPhone who does detect similar images.
So then a pedophile could edit a few aspects of a CSAM photo and evade the system? If so, this will just lead to an arms race between image-editing pedophiles and those providing hashes, leading to increased storage requirements on the phone.
They’re talking about scanning the photos that are about to be uploaded to iCloud Photos (and apparently it’s disabled if you disable the iCloud Photo Library). Would you be happier if exactly the same photos were scanned, but the scanning was happening on Apple’s servers after the photos were uploaded?
The impression I get is that every service is already scanning, or will soon be required to scan, photos that end up on their servers, for CSAM, and Apple figured it improved user‘s privacy if that required scanning happened client-side before the photos were uploaded rather than server side after they’re uploaded.
Baking it into the OS makes it harder for a government to come along and ask for additions - it’d mean changing iOS rather than just changing a script that runs on the server. Seems like Google and others that do it all server-side would be easier for a government to arm-twist into scanning more, since they can just change one script that customers never see - and security researchers / privacy advocates can’t analyze - rather than having to add code to the OS that then needs to be updated on every user device.
We don't know much editing is needed to avoid detection. Probably changing into black and white, changing colours, hue, contrast, some minor cropping won't avoid detection. Maybe even mirroring is supported.
But yes, adding enough new elements to the image is a way to avoid detection.
It's not as big a problem as you might think when it comes to reducing the spread of CP through iCloud, but it clearly illustrates why it's such a stupid system for governments to use to catch "unwanted elements".
Last edited:
Sorry but I am not seeing your point. None of the afore mentioned companies do CSS.