Randomness is a crucial ingredient for Transport Layer Security (TLS) and related security protocols. Weak or predictable "cryptographically secure" pseudorandom number generators (CSPRNGs) can be abused or exploited for malicious purposes. An initial entropy source that seeds a CSPRNG might be weak or broken as well, which can also lead to critical and systemic security problems. This document describes a way for security protocol implementations to augment their CSPRNGs using long-term private keys. This improves randomness from broken or otherwise subverted CSPRNGs.
This document is a product of the Crypto Forum Research Group (CFRG) in the IRTF.
This document is not an Internet Standards Track specification; it is published for informational purposes.
This document is a product of the Internet Research Task Force (IRTF). The IRTF publishes the results of Internet-related research and development activities. These results might not be suitable for deployment. This RFC represents the consensus of the Crypto Forum Research Group of the Internet Research Task Force (IRTF). Documents approved for publication by the IRSG are not a candidate for any level of Internet Standard; see Section 2 of RFC 7841.
Information about the current status of this document, any errata, and how to provide feedback on it may be obtained at https://www.rfc-editor.org/info/rfc8937.
Copyright (c) 2020 IETF Trust and the persons identified as the document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document.
Secure and properly implemented random number generators, or "cryptographically secure" pseudorandom number generators (CSPRNGs), should produce output that is indistinguishable from a random string of the same length. CSPRNGs are critical building blocks for TLS and related transport security protocols. TLS in particular uses CSPRNGs to generate several values, such as ephemeral key shares and ClientHello and ServerHello random values. CSPRNG failures, such as the Debian bug described in [DebianBug], can lead to insecure TLS connections. CSPRNGs may also be intentionally weakened to cause harm [DualEC]. Initial entropy sources can also be weak or broken, and that would lead to insecurity of all CSPRNG instances seeded with them. In such cases where CSPRNGs are poorly implemented or insecure, an adversary, Adv, may be able to distinguish its output from a random string or predict its output and recover secret key material used to protect the connection.
This document proposes an improvement to randomness generation in security protocols inspired by the "NAXOS trick" [NAXOS]. Specifically, instead of using raw randomness where needed, e.g., in generating ephemeral key shares, a function of a party's long-term private key is mixed into the entropy pool. In the NAXOS key exchange protocol, raw random value x is replaced by H(x, sk), where sk is the sender's private key. Unfortunately, as private keys are often isolated in Hardware Security Modules (HSMs), direct access to compute H(x, sk) is impossible. Moreover, some HSM APIs may only offer the option to sign messages using a private key, yet offer no other operations involving that key. An alternate, yet functionally equivalent construction, is needed.
The approach described herein replaces the NAXOS hash with a keyed hash, or pseudorandom function (PRF), where the key is derived from a raw random value and a private key signature. Implementations SHOULD apply this technique a) when indirect access to a private key is available and CSPRNG randomness guarantees are dubious or b) to provide stronger guarantees about possible future issues with the randomness. Roughly, the security properties provided by the proposed construction are as follows:
This document represents the consensus of the Crypto Forum Research Group (CFRG).
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.
The output of a properly instantiated CSPRNG should be indistinguishable from a random string of the same length. However, as previously discussed, this is not always true. To mitigate this problem, we propose an approach for wrapping the CSPRNG output with a construction that mixes secret data into a value that may be lacking randomness.
Let G(n) be an algorithm that generates n random bytes, i.e., the output of a CSPRNG. Define an augmented CSPRNG G' as follows. Let Sig(sk, m) be a function that computes a signature of message m given private key sk. Let H be a cryptographic hash function that produces output of length M. Let Extract(salt, IKM) be a randomness extraction function, e.g., HKDF-Extract [RFC5869], which accepts a salt and input keying material (IKM) parameter and produces a pseudorandom key of L bytes suitable for cryptographic use. It must be a secure PRF (for salt as a key of length M) and preserve uniformness of IKM (for details, see [SecAnalysis]). L SHOULD be a fixed length. Let Expand(k, info, n) be a variable-length output PRF, e.g., HKDF-Expand [RFC5869], that takes as input a pseudorandom key k of L bytes, info string, and output length n, and produces output of n bytes. Finally, let tag1 be a fixed, context-dependent string, and let tag2 be a dynamically changing string (e.g., a counter) of L' bytes. We require that L >= n - L' for each value of tag2.
The construction works as follows. Instead of using G(n) when randomness is needed, use G'(n), where
G'(n) = Expand(Extract(H(Sig(sk, tag1)), G(L)), tag2, n)
Functionally, this expands n random bytes from a key derived from the CSPRNG output and signature over a fixed string (tag1). See Section 4 for details about how "tag1" and "tag2" should be generated and used per invocation of the randomness wrapper. Expand() generates a string that is computationally indistinguishable from a truly random string of n bytes. Thus, the security of this construction depends upon the secrecy of H(Sig(sk, tag1)) and G(L). If the signature is leaked, then security of G'(n) reduces to the scenario wherein randomness is expanded directly from G(L).
If a private key sk is stored and used inside an HSM, then the signature calculation is implemented inside it, while all other operations (including calculation of a hash function, Extract function, and Expand function) can be implemented either inside or outside the HSM.
Sig(sk, tag1) need only be computed once for the lifetime of the randomness wrapper and MUST NOT be used or exposed beyond its role in this computation. Additional recommendations for tag1 are given in the following section.
Sig MUST be a deterministic signature function, e.g., deterministic Elliptic Curve Digital Signature Algorithm (ECDSA) [RFC6979], or use an independent (and completely reliable) entropy source, e.g., if Sig is implemented in an HSM with its own internal trusted entropy source for signature generation.
Because Sig(sk, tag1) can be cached, the relative cost of using G'(n) instead of G(n) tends to be negligible with respect to cryptographic operations in protocols such as TLS (the relatively inexpensive computational cost of HKDF-Extract and HKDF-Expand dominates when comparing G' to G). A description of the performance experiments and their results can be found in [SecAnalysis].
Moreover, the values of G'(n) may be precomputed and pooled. This is possible since the construction depends solely upon the CSPRNG output and private key.
Both tags MUST be generated such that they never collide with another contender or owner of the private key. This can happen if, for example, one HSM with a private key is used from several servers or if virtual machines are cloned.
The RECOMMENDED tag construction procedure is as follows:
The PRF randomness wrapper can be applied to any protocol wherein a party has a long-term private key and also generates randomness. This is true of most TLS servers. Thus, to apply this construction to TLS, one simply replaces the "private" CSPRNG G(n), i.e., the CSPRNG that generates private values, such as key shares, with
G'(n) = HKDF-Expand(HKDF-Extract(H(Sig(sk, tag1)), G(L)), tag2, n)
Recall that the wrapper defined in Section 3 requires L >= n - L', where L is the Extract output length and n is the desired amount of randomness. Some applications may require n to exceed this bound. Wrapper implementations can support this use case by invoking G' multiple times and concatenating the results.
This document has no IANA actions.
A security analysis was performed in [SecAnalysis]. Generally speaking, the following security theorem has been proven: if Adv learns only one of the signature or the usual randomness generated on one particular instance, then, under the security assumptions on our primitives, the wrapper construction should output randomness that is indistinguishable from a random string.
The main reason one might expect the signature to be exposed is via a side-channel attack. It is therefore prudent when implementing this construction to take into consideration the extra long-term key operation if equipment is used in a hostile environment when such considerations are necessary. Hence, it is recommended to generate a key specifically for the purposes of the defined construction and not to use it another way.
The signature in the construction, as well as in the protocol itself, MUST NOT use randomness from entropy sources with dubious security guarantees. Thus, the signature scheme MUST either use a reliable entropy source (independent from the CSPRNG that is being improved with the proposed construction) or be deterministic; if the signatures are probabilistic and use weak entropy, our construction does not help, and the signatures are still vulnerable due to repeat randomness attacks. In such an attack, Adv might be able to recover the long-term key used in the signature.
Under these conditions, applying this construction should never yield worse security guarantees than not applying it, assuming that applying the PRF does not reduce entropy. We believe there is always merit in analyzing protocols specifically. However, this construction is generic so the analyses of many protocols will still hold even if this proposed construction is incorporated.
The proposed construction cannot provide any guarantees of security if the CSPRNG state is cloned due to the virtual machine snapshots or process forking (see [MAFS2017]). It is RECOMMENDED that tag1 incorporate all available information about the environment, such as process attributes, virtual machine user information, etc.
The construction proposed herein has similarities with that of [RFC6979]; both of them use private keys to seed a deterministic random number generator. Section 3.3 of [RFC6979] recommends deterministically instantiating an instance of the HMAC_DRBG pseudorandom number generator, described in [SP80090A] and Annex D of [X962], using the private key sk as the entropy_input parameter and H(m) as the nonce. The construction G'(n) provided herein is similar, with such difference that a key derived from G(n) and H(Sig(sk, tag1)) is used as the entropy input and tag2 is the nonce.
However, the semantics and the security properties obtained by using these two constructions are different. The proposed construction aims to improve CSPRNG usage such that certain trusted randomness would remain even if the CSPRNG is completely broken. Using a signature scheme that requires entropy sources according to [RFC6979] is intended for different purposes and does not assume possession of any entropy source -- even an unstable one. For example, if in a certain system all private key operations are performed within an HSM, then the differences will manifest as follows: the HMAC_DRBG construction described in [RFC6979] may be implemented inside the HSM for the sake of signature generation, while the proposed construction would assume calling the signature implemented in the HSM.
We thank Liliya Akhmetzyanova for her deep involvement in the security assessment in [SecAnalysis]. We thank John Mattsson, Martin Thomson, and Rich Salz for their careful readings and useful comments.