[strangecasts]

Good encryption schemes are designed so that ciphertexts are effectively indistinguishable from random data -- you should not be able to see any pattern in the encrypted text without knowledge of the key and the algorithm.

If your encryption scheme satisfies this, there are no patterns for the LLM to learn: if you only know the ciphertext but not the key, every continuation of the plaintext should be equally likely, so trying to learn the encryption scheme from examples is effectively trying to predict the next lottery numbers.

This is why FHE for ML schemes [1] don't try to make ML models work directly on encrypted data, but rather try to package ML models so they can run inside an FHE context.

[1] It's not for language models, but I like Microsoft's CryptoNets - https://www.microsoft.com/en-us/research/wp-content/uploads/... - as a more straightforward example of how FHE for ML looks in practice

[reliabilityguy]

I am confused: you can implement LLM learning with FHE. It’s a different problem than learning on encrypted data.

[strangecasts]

I didn't mean to suggest otherwise! That's why I also linked the CryptoNets paper - to show that you're transforming the inference to happen inside an FHE context, not trying to learn encrypted data

[reliabilityguy]

Yes, you can do Cryptonets. What I’m saying is that you don’t have to do cryptonets, you can simply use FHE to train the network in fully encrypted manner: both the network and the data are FHE-encrypted, so the training itself is an FHE application. It would be insanely slow and I doubt it can be done today even for “small” LLMs due to high overheads of FHE.

[derangedHorse]

> This is why FHE for ML schemes [1] don't try to make ML models work directly on encrypted data, but rather try to package ML models so they can run inside an FHE context.

I don't think @strangecasts was trying to say you couldn't. I believe their point was that you can't have a model learn to coherently respond to encrypted inputs with just traditional learning mechanisms (so without FHE). Doing so would require an implicit breaking of the encryption scheme by the model because it would need a semantic understanding of the plaintext to provide a cogent, correctly encrypted response.

[strangecasts]

That's a much better way of phrasing what I was trying to get at :)