Pre-Investigation into Security Requirements for Future Brain Emulations
Very tentative steps towards promoting research efforts and technology standards
Dragonsphere Report
Aside from dragons, my big obsession is hell simulations. Specifically, PREVENTING hell simulations. To this end I am trying to get a general picture of computer information security, on the tentative assumption that the mathematical principles that apply to securing computational processes, inputs, outputs, and substrates will apply to securing whatever the processes, inputs, outputs, and substrates are that constitute conscious entities. That is, I am trying to figure out how to secure brain emulations. At present nobody even knows what a brain emulation might look like, so needless to say this is something of a challenge. However, it is best to start trying to resolve serious crises before they erupt, so trying to solve them before anyone knows what’s even involved seems like a good thing to do.
My own background is a modest IT and CS education that have equipped me with none of the tools necessary to understand these problems. I have only the scarcest clue what a neural network is, for instance. I am working mostly with secondhand descriptions of what different things are, so forgive me if I am off base or ignorant. One useful bit of information I recovered was that if the brain is based in computation, neural nets should be able to emulate it, since neural nets are computationally universal. Other bits of information follow.
I'm told that the thing I would need to make my primary security obsession possible (encrypted consciousness given that neural nets are capable of consciousness), is a form of differential privacy based on some unknown encryption method that is homomorphic under a desired set of relu "actions". I am also told this is a problem of algebraic or arithmetic geometry though, which is very hard and would require Top Mathematicians to work on it, who are all currently working on P=NP. I have no idea what most of this means. Also apparently it requires Class Field Theory.
Lots of statistical things can be computed fully homomorphically, including any statistical thing that runs on linear transformations, regressions, pca, etc. Very simple neural networks are probably also linear transformations
A lot of the popular things like sigmoid functions have been admitted into some reasonable-looking encryption methods, like encryption based on conjugacy apparently.
In terms of hardware security, the concept of a physically uncloneable function may help prevent copying as opposed to simply access.
Thanks to AstralCodexTen’s readerbase for this information, including “that's not very uwu of you” and “chuzz” from the Discord and “Max” from the comments of a hidden open thread, who writes the following:
"Okay! So, a few things.
Modern day artificial neural networks, as used in image recognition, don't try and approximate brains. A better way to think of them is as a series of matrix operations. Very simply, they receive inputs of the form of a 2D array (in the case of images), and this 2D array goes through a series of mathematical operations based on weights the network has "learned".
Homomorphic encryption is type of encryption that allows you to encrypt something, do math on it, and then decrypt it and have that math be preserved. So if I encrypt 4 and 5, then add the resulting ciphertext and decrypt, I'll get 9. So if I want to encrypt a neural network I can do something like encrypt the weights and input homomorphically, do all the math on them, and then decrypt the results.
The issue with this is that homomorphic encryption either limits the type of math operations you can do or takes a million years. Unfortunately, modern neural networks rely heavily on neural activation functions, one of which is called a 'relu', all of which are not easy to run on encrypted data.
If you want to understand more about this I'd just start with googling stuff about neural networks and homomorphic encryption. I know in the mid 2015s there was some buzz about running neural networks that were encrypted, but I don't think anyone in industry ever did anything with it.
Source: Wrote a prototype of an encrypted convolutional neural network for the DoD awhile back."
I have nothing substantive to say about any of this at the moment, except to say that I strongly believe this subject deserves investigation, and the concepts, fields, and so forth outlined above all seem to represent the best informed and most substantive elaborations on the very faint and limited ideas my own brain is capable of generating.
I guess I will start googling things now and keep you updated with future posts. Anyone willing to help contribute to my understanding in the comments would be more than welcome also.
Thus ends another Dragonsphere Report