3 ms·

I'm not big into "conspiracy" but you gotta wonder why this "NeuralHash" file on my computer exists if "NeuralHash" was supposedly delayed, according to Apple,

by evv 5y ago

I'm not big into "conspiracy" but you gotta wonder why this "NeuralHash" file on my computer exists if "NeuralHash" was supposedly delayed, according to Apple, on an undetermined timeline. If I did want to cook up conspiracy theory, it would be easy: Apple wants to distract from the fact that NeuralHash was broken by researchers. This project apparently is able to create CSAM collisions: https://github.com/AsuharietYgvar/AppleNeuralHash2ONNX https://github.com/AsuharietYgvar/AppleNeuralHash2ONNX If NeuralHash is really rolled out, and if this python project can really create collisions, the CSAM system could be DDoS'd by people on their own computers, jamming up Apple's internal censorship review system with false positives. Hence, Apple would be incentivized to sweep this under the rug by "delaying" rollout indeterminately.

detaro 5y ago
The work you link is based on the NeuralHash code shipping with iOS and macOS, and explicitly says so, so there is no reason to theorize if it might already have been in a shipped OS - if it hadn't, we wouldn't have research about its abilities. (What as far as is known - and I suspect it would have been found by people scrutinizing the updates - has not been rolled out is code that actually uses it to implement the reporting mechanisms)
- evv 5y ago
  Why would they distribute the ML weights file if it isn't being used for logging or reporting? Of course I do theorize that it is actually being used. There is so much network traffic between macOS services and the internet, I'm not sure how easy it would be for researchers to discover if CSAM reporting is enabled. Little snitch detects 292 macOS services on my M1, all talking to the internet. Network experts in Apple probably even have ways of bypassing little snitch.
  - detaro 5y ago
    The weights are part of the hash implementation, so not shipping them with the rest of the hashing code would make even less sense.
  - hitmyapi 5y ago
    It could potentially be "feature flagged". The client could ship a new version with all the resources/code they need for a specific feature, but a setting on the back-end determines if the feature is enabled or not. That way, Apple could easily toggle this feature on at any point by changing the flag's value on the back-end, without needing to ship an entirely new 15.x.x binary

by evv 5y ago

detaro 5y ago

The work you link is based on the NeuralHash code shipping with iOS and macOS, and explicitly says so, so there is no reason to theorize if it might already have been in a shipped OS - if it hadn't, we wouldn't have research about its abilities. (What as far as is known - and I suspect it would have been found by people scrutinizing the updates - has not been rolled out is code that actually uses it to implement the reporting mechanisms)

evv 5y ago
Why would they distribute the ML weights file if it isn't being used for logging or reporting? Of course I do theorize that it is actually being used. There is so much network traffic between macOS services and the internet, I'm not sure how easy it would be for researchers to discover if CSAM reporting is enabled. Little snitch detects 292 macOS services on my M1, all talking to the internet. Network experts in Apple probably even have ways of bypassing little snitch.
- detaro 5y ago
  The weights are part of the hash implementation, so not shipping them with the rest of the hashing code would make even less sense.
- hitmyapi 5y ago
  It could potentially be "feature flagged". The client could ship a new version with all the resources/code they need for a specific feature, but a setting on the back-end determines if the feature is enabled or not. That way, Apple could easily toggle this feature on at any point by changing the flag's value on the back-end, without needing to ship an entirely new 15.x.x binary