4 ms·
Sorry, I should have specified WA means Wolfram Alpha. >barring huge breakthroughs in homomorphic encryption Actually this seems plausible to me without havin
by sebzim4500 4y ago
Sorry, I should have specified WA means Wolfram Alpha.
>barring huge breakthroughs in homomorphic encryption
Actually this seems plausible to me without having to use those general purpose algorithms that cause a 1000000000x slowdown.
For example, you can obscure the input to each matrix multiplication by shifting the input and outputs by a predetermined (user specific) constant. Maybe there are similar tricks for cross attention, although that might be a bit optimistic. Certainly you can do some amount of obfuscation by permuting the inputs, but I doubt that would achieve much.
- layer8 4y agoI’m pretty sure your query will effectively still be derivable on the server side, as it has to syntactically/semantically interact with the model. To truly disguise it, at least part of the model/weights would have to be transferred to the client side (so that the server doesn’t learn the adjustment), which would in turn leak the model.