Can Pseudorandom values be used as secure keys?

I want to derive $n$ keys from a master key.

I pick a master key $mk$ and use pseudo-random function (PRF) to generate $n$ pseudo-random values, $v_i=PRF(mk,i)$ where $1\leq i \leq n$. Next, I use key derivation function to generate $n$ keys: $k_i=KDF(v_i)$, where $1\leq i \leq n$. Note that key derivation must be deterministic.

Question 1: Is above scheme a secure way of generating $n$ keys?

Question 2: Can we consider values $v_i$ as secure cryptographic keys?

• why do you KDF(PRF()). All decent KDFs (=HKDF) will allow you to do this in one step. And the answers are "yes and yes"
– SEJPM
Nov 24 '15 at 19:28
• Actually KDF with only one parameter seems a bit strange. TLS just uses the PRF as KDF without explicitly defining it as one, which is cheating ever so slightly in my book. Nov 24 '15 at 22:16
• @Thank you for the answers. I also found below paper that in introduction says we can use outputs of PRF as a key: iacr.org/archive/crypto2010/62230625/62230625.pdf Nov 25 '15 at 9:36

Yes and yes, as mentioned in the comments.

It is worth noting that Bitcoin wallets use a scheme similar to this in BIP32, a method of creating n various EC keypairs from a single seed deterministically:

https://github.com/bitcoin/bips/blob/master/bip-0032.mediawiki

EDIT: This answer only holds true if 𝑃𝑅𝐹 is a CSPRNG. (This application is sort of why the definition for CSPRNG exists.)

• The technical term here for what PRF here is, is: PRF, short for pseudorandom function family. There's no need for the additional baggage that the term CSPRNG implies—backtracking resistance, rekeying, etc. Apr 17 '19 at 16:23

If your protocol were secure with independent uniform random choices for the $$v_i$$, then it can't be much less secure with $$v_i = \mathit{PRF}(\mathit{mk}, i)$$ for uniform random $$\mathit{mk}$$, as long as you never use $$\mathit{PRF}(\mathit{mk}, i)$$ for any other purpose—the only additional advantage an adversary can get in breaking the system with the PRF is the best possible advantage at distinguishing the PRF, which, for a good PRF, is negligible. This is the essential point of a PRF!

The term ‘KDF’ sometimes covers two steps, as in HKDF-Extract and HKDF-Expand:

1. extracting a short uniform string from a high-entropy but possibly nonuniform string (like a DH shared secret), sometimes with a salt to mitigate multi-target attacks, and
2. expanding a short uniform string by a PRF into many effectively independent keys with structured inputs (what you called $$i$$) to avoid collisions.

If you already have a short uniform string, you can safely skip the extraction step, and what you are left with is just a PRF! It doesn't hurt to do another extraction step afterward—hash the PRF output again as you described—but there's no need.