3 ms·

This is a false-positive rate of 2 in 2 trillion image pairs (1,431,168^2). Assuming the NCMEC database has more than 20,000 images, this represents a slightly

This is a false-positive rate of 2 in 2 trillion image pairs (1,431,168^2). Assuming the NCMEC database has more than 20,000 images, this represents a slightly higher rate than Apple had previously reported. But, assuming there are less than a million images in the dataset, it's probably in the right ballpark. If the author was comparing 2 trillion pictures of people, or children specifically, I think this false-positive rate would be different and arguably much higher. The reasons are obvious: humans are similar in dimensions to eachother and are much more likely to match in the same way the hatchet and nematode matched. I do not presume such a finding of photos is easy to come by but I wish the author put details on the sample set.

yeldarb 5y ago
The sample set is ImageNet, which is a well-known dataset in Computer Vision and is available for download here: https://www.kaggle.com/c/imagenet-object-localization-challenge https://www.kaggle.com/c/imagenet-object-localization-challe... I'd love to see this work extended; if you find additional collisions in the wild please submit a PR to the repo (please do not submit artificially generated adversarial images): https://github.com/roboflow-ai/neuralhash-collisions https://github.com/roboflow-ai/neuralhash-collisions For what it's worth, Apple claimed to find a _lower_ incidence of false-positives when it used pornographic images in its test[1] (which makes sense; images containing humans is probably more aligned with what the model was trained on than nematodes) [1] https://tidbits.com/2021/08/13/new-csam-detection-details-emerge-following-craig-federighi-interview/ https://tidbits.com/2021/08/13/new-csam-detection-details-em... > In Apple’s tests against 100 million non-CSAM images, it encountered 3 false positives when compared against NCMEC’s database. In a separate test of 500,000 adult pornography images matched against NCMEC’s database, it found no false positives.
- smithza 5y ago
  For what it's worth, Apple claimed to find a _lower_ incidence of false-positives when it used pornographic images in its test[1] (which makes sense; images containing humans is probably more aligned with what the model was trained on than nematodes) This is an important note. Is it the case that this algorithm is trained for humans or not? the 1/trillion false-positive rate might imply it is trained with a broader set. Thank you for those helpful tidbits.
- tgsovlerkhgsel 5y ago
  0 out of 0.5 million vs. 3 out of 100 million does not imply with any reasonable confidence that the incidence in porn is lower. You'd expect the same result even if the incidence in porn was 10x higher than in typical images (30 in 100 million = 0.15 in 0.5 million).
- fallingknife 5y ago
  500K is not a large enough dataset to determine that. The collision rate could plausibly be 1 in 500K (or even a bit higher) and have no collisions in the sample.
- contravariant 5y ago
  Well to know for sure if that's lower we'd need to know the size of the NCMEC database. They're the same if the NCMEC contains around 10 000 images. Though knowing that 1 in roughly 30 million images generates a false positive is the most important figure I suppose. Assuming 100 million iPhones with each 1000 pictures that would generate some 3000 phones with one or more false positives [1] and a roughly 5% chance that some phone has at least 2 false positives. [1]: https://www.wolframalpha.com/input/?i=%28100+million%29+e%5E-l+%28e%5El+-+1%29+where+l+%3D+%281000.0%2F30+million%29 https://www.wolframalpha.com/input/?i=%28100+million%29+e%5E... [2]: https://www.wolframalpha.com/input/?i=+%281+-+e%5E-l+%28e%5El+-+1+-+l%29%29+%5E+%28100+million%29+where+l+%3D+%281000.0%2F30+million%29 https://www.wolframalpha.com/input/?i=+%281+-+e%5E-l+%28e%5E...

yeldarb 5y ago

The sample set is ImageNet, which is a well-known dataset in Computer Vision and is available for download here: https://www.kaggle.com/c/imagenet-object-localization-challenge https://www.kaggle.com/c/imagenet-object-localization-challe... I'd love to see this work extended; if you find additional collisions in the wild please submit a PR to the repo (please do not submit artificially generated adversarial images): https://github.com/roboflow-ai/neuralhash-collisions https://github.com/roboflow-ai/neuralhash-collisions For what it's worth, Apple claimed to find a _lower_ incidence of false-positives when it used pornographic images in its test[1] (which makes sense; images containing humans is probably more aligned with what the model was trained on than nematodes) [1] https://tidbits.com/2021/08/13/new-csam-detection-details-emerge-following-craig-federighi-interview/ https://tidbits.com/2021/08/13/new-csam-detection-details-em... > In Apple’s tests against 100 million non-CSAM images, it encountered 3 false positives when compared against NCMEC’s database. In a separate test of 500,000 adult pornography images matched against NCMEC’s database, it found no false positives.

smithza 5y ago
For what it's worth, Apple claimed to find a _lower_ incidence of false-positives when it used pornographic images in its test[1] (which makes sense; images containing humans is probably more aligned with what the model was trained on than nematodes) This is an important note. Is it the case that this algorithm is trained for humans or not? the 1/trillion false-positive rate might imply it is trained with a broader set. Thank you for those helpful tidbits.
tgsovlerkhgsel 5y ago
0 out of 0.5 million vs. 3 out of 100 million does not imply with any reasonable confidence that the incidence in porn is lower. You'd expect the same result even if the incidence in porn was 10x higher than in typical images (30 in 100 million = 0.15 in 0.5 million).
fallingknife 5y ago
500K is not a large enough dataset to determine that. The collision rate could plausibly be 1 in 500K (or even a bit higher) and have no collisions in the sample.
contravariant 5y ago
Well to know for sure if that's lower we'd need to know the size of the NCMEC database. They're the same if the NCMEC contains around 10 000 images. Though knowing that 1 in roughly 30 million images generates a false positive is the most important figure I suppose. Assuming 100 million iPhones with each 1000 pictures that would generate some 3000 phones with one or more false positives [1] and a roughly 5% chance that some phone has at least 2 false positives. [1]: https://www.wolframalpha.com/input/?i=%28100+million%29+e%5E-l+%28e%5El+-+1%29+where+l+%3D+%281000.0%2F30+million%29 https://www.wolframalpha.com/input/?i=%28100+million%29+e%5E... [2]: https://www.wolframalpha.com/input/?i=+%281+-+e%5E-l+%28e%5El+-+1+-+l%29%29+%5E+%28100+million%29+where+l+%3D+%281000.0%2F30+million%29 https://www.wolframalpha.com/input/?i=+%281+-+e%5E-l+%28e%5E...