pixel-spreadsheet VS neural-hash-collider

It's not a problem that can be realistically solved with technology. The problem existed before electricity had been discovered[1][2] and continues to exist[3] in a significant way regardless of the presence of computers and the Internet.

I'd guess that the intent behind client-side scanning is to force child abusers away from mainstream technology towards niche technology[4][5][6], increasing the signal-to-noise ratio for finding and targeting child abusers. At present, there is nothing inherently suspicious about someone using WhatsApp, which means child abusers can hide amongst millions of other users. If CSAM is instead pushed onto Tor and a little-heard-of encrypted messaging app with servers located in eastern Europe, the theory is that it becomes easier to target just these niche users. If they're using Tor, they must have something to hide and therefore further investigation is warranted...?

A key problem with this approach is that, as has happened in other criminal circles, criminals find increasingly clever ways to remain 'boring' and not inviting of further scrutiny, or to otherwise remain hidden in the masses. For example, 9/11 hijackers communicated using boring methods, without encryption, using a coded language to make messages appear innocuous[7]. Even if general purpose computing was prohibited, would it be realistic to prevent users from using spreadsheet formulas to implement encryption and display of images[8][9][10], without impacting other users?. Another key problem is that innocent people will likely be targeted and harassed by association with niche technologies, which also wastes resources that could be used to limit child abuse via more effective means.

It would be possible to implement something that looks like North Korea's Kwangmyong[11] and largely solve the CSAM-sharing-via-computers problem, but at that point you've essentially banned computing altogether. This is actually a decision some countries have made[12] but good luck convincing the modern world that it'd be a good idea to eradicate the last 40 years of technological advancement whilst other countries continue to leap ahead technologically with the help of open and unrestricted access to general purpose computing.

[1] https://en.wikipedia.org/wiki/Pederasty

[2] https://en.wikipedia.org/wiki/Bacha_bazi

[3] https://europepmc.org/article/nbk/nbk554624

[4] https://en.wikipedia.org/wiki/ANOM

[5] https://en.wikipedia.org/wiki/EncroChat

[6] https://en.wikipedia.org/wiki/Tor_(network)

[7] https://hcgroups.wordpress.com/2010/02/13/july-2001-communic...

[8] https://poisonninja.github.io/2016/12/07/AES-in-Google-Sheet...

[9] https://www.iacis.org/iis/2017/2_iis_2017_103-112.pdf

[10] https://github.com/takacs/pixel-spreadsheet

[11] https://en.wikipedia.org/wiki/Kwangmyong_(network)

[12] https://en.wikipedia.org/wiki/2021_Burmese_protests#Internet...

That assumes the hash is base 26 which would be fun, but most hashing systems including NeuralHash are hexadecimal. NeuralHash also outputs 24 character hashes so there are 1624 possible NeuralHashes. However all of that ignores that NeuralHash is designed so that visually similar images will return the same hash on purpose. You can even provide two images of your choosing to this library and it will make them collide https://github.com/anishathalye/neural-hash-collider

This is incorrect.

The images in this link [1], are completely different. One is a cat, one is a dog. Same hash produced for both.

[1] https://github.com/anishathalye/neural-hash-collider

And here: https://github.com/anishathalye/neural-hash-collider

Also, Apple’s NeuralHash is rather prone to hash collisions, i.e. false positives, so you should absolutely expect at least some of your on-device pics to be viewed by unauthorized Apple/government personell.