exponent bit-length for hard DL (128-bit security)

gormatron3000

2/9/24, 4:18 PM

Following up on my previous post, I thought I might get a more concrete answer if I gave a more concrete question.

I require 128-bit security so I choose a 3072-bit RSA modulus ($\ell_n=3072$). Specifically I choose $n=pq=(2p'+1)(2q'+1)$ that is a product of safe primes $p$ and $q$.

Now, I want to choose $\ell_\Lambda$ such that finding DL with $\ell_\Lambda$-bit exponents is hard in $QR_n$, for an adversary who does not know the factorization of $n$, with 128 bits of security.

Note that the order of $QR_n=\phi(n)/4=p'q'$, so each element has large order.

Is choosing $\ell_\Lambda=256$ sufficient as suggested in the answer to my previous post?

1 + 3

quadratic-residuosity

domain-parameters

fgrieu

2/9/24, 5:42 PM

Is the factorization of $n$ secret? With $p$ and $q$ known, we can reduce a DLP modulo $n$ to two DLP modulo $p$ and $q$, and with each 1536-bit these would not be 128-bit secure (more like 100-bit, give or take a lot).

poncho

2/9/24, 6:32 PM

@fgrieu: as explained in his previous post, the factorization of $n$ is secret

gormatron3000

2/9/24, 7:00 PM

@fgrieu edited question to clarify $p$ and $q$ are unknown

Score:2

Crypto

fgrieu

2/9/24, 9:22 PM

I second that with the factorization of $n$ remaining secret, as assumed in the paper linked in that previous question (in particular by making the strong RSA assumption), the Discrete Logarithm Problem modulo $n$ is believed no easier that if $n$ was prime. And therefore, the best methods to solve that DLP have expected cost the lowest of:

a few times $2^{\ell_\Lambda/2}$ multiplications modulo $n$, for methods based on collision search like Pollard's rho/kangaroo and distributed variants.
about the cost of (G)NFS factorization of $n$; see this paper for about the state of the art, implying that the cost of factorization of $n$ and DLP for prime $n$ have roughly similar cost.

With $\ell_\Lambda=256$ and $\ell_n=3072$, (1.) is believed to be the least infeasible, and this parametrization is believed to give 128-bit security, disregarding hypothetical CRQC.

+ 0

Elon Musk

I sit in a Tesla and translated this thread with Ai:

EN: exponent bit-length for hard DL (128-bit security)

Post an answer

Most people don’t grasp that asking a lot of questions unlocks learning and improves interpersonal bonding. In Alison’s studies, for example, though people could accurately recall how many questions had been asked in their conversations, they didn’t intuit the link between questions and liking. Across four studies, in which participants were engaged in conversations themselves or read transcripts of others’ conversations, people tended not to realize that question asking would influence—or had influenced—the level of amity between the conversationalists.