Idea for a new protocol based on NNTPchan #47

New Issue

2026-10-05T03:09:55+02:00

adam commented

2026-10-05 03:09:55 +02:00

I'm sorry to shit up your issue tracker @majestrate - I just don't know how else to reach you.

Recently I got an idea for a new protocol, but I want your take on it.

Basically, you would train a classification neural network to separate CP from non-CP. Perhaps you could aim for being able to classify images containing a child's face from not to make the problem easier. You could most likely use NNTPchan's daily stream of posts as a help.

Then once the AI was sufficiently strong, you could release a new protocol that has this, now open source, classifier built in.

It would be much harder to post CP now - brainlets probably couldn't figure it out. I know attacks against the classifier are possible, but definitely not easy for non-technical people.

This idea is maybe retarded. LMK fren

ty

I'm sorry to shit up your issue tracker @majestrate - I just don't know how else to reach you. Recently I got an idea for a new protocol, but I want your take on it. Basically, you would train a classification neural network to separate CP from non-CP. Perhaps you could aim for being able to classify images containing a child's face from not to make the problem easier. You could most likely use NNTPchan's daily stream of posts as a help. Then once the AI was sufficiently strong, you could release a new protocol that has this, now open source, classifier built in. It would be much harder to post CP now - brainlets probably couldn't figure it out. I know attacks against the classifier are possible, but definitely not easy for non-technical people. This idea is maybe retarded. LMK fren ty

adam closed this issue

2026-10-05 03:09:55 +02:00

adam commented

2026-10-05 03:09:56 +02:00

On Tue, Dec 04, 2018 at 01:16:57AM -0800, Fredrick Brennan wrote:

I'm sorry to shit up your issue tracker @majestrate - I just don't know how else to reach you.

Recently I got an idea for a new protocol, but I want your take on it.

Basically, you would train a classification neural network to separate CP from non-CP. Perhaps you could aim for being able to classify images containing a child's face from not to make the problem easier. You could most likely use NNTPchan's daily stream of posts as a help.
I have hooks for spamassasin in nntpchan, they just need tweaking to target attachments only and not the entire message.
spamassisn is a fuzzy logic spam filter that predates the AI buzzword hype but operates on the same basic principals, it has a Bayesian filter that can be trained to identify spam from a large dataset of spam and ham.

Then once the AI was sufficiently strong, you could release a new protocol that has this, now open source, classifier built in.
That's a lot of data you'll need to get it working, specifically you'd need an equally large number of ham as spam.

It would be much harder to post CP now - brainlets probably couldn't figure it out. I know attacks against the classifier are possible, but definitely not easy for non-technical people.

This idea is maybe retarded. LMK fren
a related solution that i worked on that is noteworthy is josh's preceptual hash thing

the implementation i came up with for it is basically the following:

compress the image down to a 32x32 1 byte deep greyscale bitmap
store all infringing entries in a N-tree
use Kademlia to "route" new images to the "closest" match in the N-tree using XOR to get a 1024 byte distance
find the closest entry in the N-tree with Kademlia
if the closest match in the N-tree has a that distance has a very few number of bits (under 512 bits set usually, but tweakable) set you put it into a moderation queue for review.

this worked to filter out simple watermarking and rescales but not rotations.

the implementation isn't done however.

ty

--
You are receiving this because you were mentioned.
Reply to this email directly or view it on GitHub:
https://github.com/majestrate/nntpchan/issues/164

On Tue, Dec 04, 2018 at 01:16:57AM -0800, Fredrick Brennan wrote: > I'm sorry to shit up your issue tracker @majestrate - I just don't know how else to reach you. > > Recently I got an idea for a new protocol, but I want your take on it. > > Basically, you would train a classification neural network to separate CP from non-CP. Perhaps you could aim for being able to classify images containing a child's face from not to make the problem easier. You could most likely use NNTPchan's daily stream of posts as a help. I have hooks for spamassasin in nntpchan, they just need tweaking to target attachments only and not the entire message. spamassisn is a fuzzy logic spam filter that predates the AI buzzword hype but operates on the same basic principals, it has a Bayesian filter that can be trained to identify spam from a large dataset of spam and ham. > > Then once the AI was sufficiently strong, you could release a new protocol that has this, now open source, classifier built in. That's a lot of data you'll need to get it working, specifically you'd need an equally large number of ham as spam. > > It would be much harder to post CP now - brainlets probably couldn't figure it out. I know attacks against the classifier are possible, but definitely not easy for non-technical people. > > This idea is maybe retarded. LMK fren a related solution that i worked on that is noteworthy is josh's preceptual hash thing the implementation i came up with for it is basically the following: * compress the image down to a 32x32 1 byte deep greyscale bitmap * store all infringing entries in a N-tree * use Kademlia to "route" new images to the "closest" match in the N-tree using XOR to get a 1024 byte distance * find the closest entry in the N-tree with Kademlia * if the closest match in the N-tree has a that distance has a very few number of bits (under 512 bits set usually, but tweakable) set you put it into a moderation queue for review. this worked to filter out simple watermarking and rescales but not rotations. the implementation isn't done however. > > ty > > -- > You are receiving this because you were mentioned. > Reply to this email directly or view it on GitHub: > https://github.com/majestrate/nntpchan/issues/164

adam commented

2026-10-05 03:09:57 +02:00

On Tue, Dec 04, 2018 at 01:16:57AM -0800, Fredrick Brennan wrote:

I'm sorry to shit up your issue tracker @majestrate - I just don't know how else to reach you.
also, feel free to email me (jeff at i2p dot rocks)

Recently I got an idea for a new protocol, but I want your take on it.

Basically, you would train a classification neural network to separate CP from non-CP. Perhaps you could aim for being able to classify images containing a child's face from not to make the problem easier. You could most likely use NNTPchan's daily stream of posts as a help.

Then once the AI was sufficiently strong, you could release a new protocol that has this, now open source, classifier built in.

It would be much harder to post CP now - brainlets probably couldn't figure it out. I know attacks against the classifier are possible, but definitely not easy for non-technical people.

This idea is maybe retarded. LMK fren

ty

--
You are receiving this because you were mentioned.
Reply to this email directly or view it on GitHub:
https://github.com/majestrate/nntpchan/issues/164

On Tue, Dec 04, 2018 at 01:16:57AM -0800, Fredrick Brennan wrote: > I'm sorry to shit up your issue tracker @majestrate - I just don't know how else to reach you. also, feel free to email me (jeff at i2p dot rocks) > > Recently I got an idea for a new protocol, but I want your take on it. > > Basically, you would train a classification neural network to separate CP from non-CP. Perhaps you could aim for being able to classify images containing a child's face from not to make the problem easier. You could most likely use NNTPchan's daily stream of posts as a help. > > Then once the AI was sufficiently strong, you could release a new protocol that has this, now open source, classifier built in. > > It would be much harder to post CP now - brainlets probably couldn't figure it out. I know attacks against the classifier are possible, but definitely not easy for non-technical people. > > This idea is maybe retarded. LMK fren > > ty > > -- > You are receiving this because you were mentioned. > Reply to this email directly or view it on GitHub: > https://github.com/majestrate/nntpchan/issues/164

adam commented

2026-10-05 03:09:58 +02:00

yeah OK thanks for the reply. I'm not sure exactly how this would work...I know great strides have been made in machine learning lately.

Also btw I gave @jaw-sh the idea to use perceptual hashes years ago. Not that this matters in the slightest, but I'm glad to see someone did it :^)

yeah OK thanks for the reply. I'm not sure exactly how this would work...I know great strides have been made in machine learning lately. Also btw I gave @jaw-sh the idea to use perceptual hashes _years_ ago. Not that this matters in the slightest, but I'm glad to see someone did it :^)

adam commented

2026-10-05 03:09:58 +02:00

Courtesy close since I got a reply

Courtesy close since I got a reply

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: adam/nntpchan#47