@jeffchiang.bsky.social
No, you cannot hide the key rotation info from the inference server, because it is the server who performs the actual key rotation. The server can always keep track of the mapping between the word embeddings and the ids.
Looking at your paper, I bet no reviewer will allow it to be published.
Looking at your paper, I bet no reviewer will allow it to be published.
February 21, 2025 at 11:06 PM
No, you cannot hide the key rotation info from the inference server, because it is the server who performs the actual key rotation. The server can always keep track of the mapping between the word embeddings and the ids.
Looking at your paper, I bet no reviewer will allow it to be published.
Looking at your paper, I bet no reviewer will allow it to be published.
No, in their EE they have to use the same codebook for every inference, because they have to modify the word embeddings in a predefined way, so it is not possible to use different permutation for different inference. This EE method is completely bullshit
February 7, 2025 at 5:58 AM
No, in their EE they have to use the same codebook for every inference, because they have to modify the word embeddings in a predefined way, so it is not possible to use different permutation for different inference. This EE method is completely bullshit
No, this is NOT neat at all. The "inventor" of this method should receive some basic (even 30 minutes can work) education in cryptography. Just encrypt the tokenizer cannot protect anything, this can be very broken with a simple statistical analysis.
February 7, 2025 at 5:57 AM
No, this is NOT neat at all. The "inventor" of this method should receive some basic (even 30 minutes can work) education in cryptography. Just encrypt the tokenizer cannot protect anything, this can be very broken with a simple statistical analysis.