Internet-Draft | Merkle Tree Certificates for TLS | September 2024 |
Benjamin, et al. | Expires 9 March 2025 | [Page] |
This document describes Merkle Tree certificates, a new certificate type for use with TLS. A relying party that regularly fetches information from a transparency service can use this certificate type as a size optimization over more conventional mechanisms with post-quantum signatures. Merkle Tree certificates integrate the roles of X.509 and Certificate Transparency, achieving comparable security properties with a smaller message size, at the cost of more limited applicability.¶
This note is to be removed before publishing as an RFC.¶
The latest revision of this draft can be found at https://davidben.github.io/merkle-tree-certs/draft-davidben-tls-merkle-tree-certs.html. Status information for this document may be found at https://datatracker.ietf.org/doc/draft-davidben-tls-merkle-tree-certs/.¶
Discussion of this document takes place on the Transport Layer Security Working Group mailing list (mailto:tls@ietf.org), which is archived at https://mailarchive.ietf.org/arch/browse/tls/. Subscribe at https://www.ietf.org/mailman/listinfo/tls/.¶
Source for this draft and an issue tracker can be found at https://github.com/davidben/merkle-tree-certs.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 9 March 2025.¶
Copyright (c) 2024 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. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
Authors' Note: This is an early draft of a proposal with many parts. While we have tried to make it as concrete as possible, we anticipate that most details will change as the proposal evolves.¶
A typical TLS [RFC8446] handshake uses many signatures to authenticate the server public key. In a certificate chain with an end-entity certificate, an intermediate certificate, and an implicit trust anchor, there are two X.509 signatures [RFC5280]. Intermediate certificates additionally send an extra public key. If the handshake uses Certificate Transparency (CT) [RFC6962], each Signed Certificate Timestamp (SCT) also carries a signature. CT policies often require two or more SCTs per certificate [APPLE-CT] [CHROME-CT]. If the handshake staples an OCSP response [RFC6066] for revocation, that adds an additional signature.¶
Current signature schemes can use as few as 32 bytes per key and 64 bytes per signature [RFC8032], but post-quantum replacements are much larger. For example, Dilithium3 [Dilithium] uses 1,952 bytes per public key and 3,293 bytes per signature. A TLS Certificate message with, say, four Dilithum3 signatures (two X.509 signatures and two SCTs) and one intermediate CA's Dilithium3 public key would total 15,124 bytes of authentication overhead. Falcon-512 and Falcon-1024 [Falcon] would, respectively, total 3,561 and 6,913 bytes.¶
This document introduces Merkle Tree Certificates, an optimization that authenticates a subscriber key using under 1,000 bytes. See Section 5.6. To achieve this, it reduces its scope from general authentication:¶
Certificates are short-lived. The subscriber is expected to use an automated issuance protocol, such as ACME [RFC8555].¶
Certificates are only usable with relying parties that have contacted a transparency service sufficiently recently. See Section 7.¶
Certificates are issued after a significant processing delay of, in the recommended parameters (Section 5.1), about an hour. Subscribers that need a certificate issued quickly are expected to use a different mechanism.¶
To support the reduced scope, this document also describes a certificate negotiation mechanism. Subscribers send these more efficient certificates when available, and otherwise fall back to other mechanisms.¶
Merkle Tree Certificates are not intended to replace existing Public Key Infrastructure (PKI) mechanisms but, in applications where a significant portion of authentications meet the above requirements, complement them as an optional optimization. In particular, it is expected that, even within applications that implement it, this mechanism will not be usable for all TLS connections.¶
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.¶
This document additionally uses the TLS presentation language defined in Section 3 of [RFC8446].¶
All time computations in this document are represented by POSIX timestamps, defined in this document to be integers containing a number of seconds since the Epoch, defined in Section 4.16 of [POSIX]. That is, the number of seconds after 1970-01-01 00:00:00 UTC, excluding leap seconds. A UTC time is converted to a POSIX timestamp as described in [POSIX].¶
Durations of time are integers, representing a number of seconds not including leap seconds. They can be added to POSIX timestamps to produce other POSIX timestamps.¶
The current time is a POSIX timestamp determined by converting the current UTC time to seconds since the Epoch. One POSIX timestamp is said to be before (respectively, after) another POSIX timestamp if it is less than (respectively, greater than) the other value.¶
There are five roles involved in a Merkle Tree certificate deployment:¶
The party that authenticates itself in the protocol. In TLS, this is the side sending the Certificate and CertificateVerify message.¶
The service that issues Merkle Tree certificates to the subscriber, and publishes logs of all certificates.¶
The party authenticating the subscriber. In TLS, this is the side receiving the Certificate and CertificateVerify message.¶
The service that mirrors the issued certificates for others to monitor. It additionally summarizes the CA's activity for relying parties, in order for certificates to be accepted. This is conceptually a single service, but may be multiple services, run by multiple entities in concert. See Section 7. For example, if the relying party is a web browser, the browser vendor might run the transparency service, or it may trust a collection of third-party mirrors.¶
Parties who monitor the list of valid certificates, published by the transparency service, for unauthorized certificates.¶
Additionally, there are several terms used throughout this document to describe this proposal. This section provides an overview. They will be further defined and discussed in detail throughout the document.¶
A protocol-specific statement that the CA is certifying. For example, in TLS, the assertion is that a TLS signing key can speak on behalf of some DNS name or other identity.¶
A partially-hashed Assertion to save space. For example, in TLS, an abridged assertion replaces the subject public key by a hash.¶
A structure, generated by the CA, that proves to the relying party that the CA has certified some assertion. A certificate consists of the assertion itself accompanied by an associated proof string.¶
A collection of assertions certified at the same time. CAs in this proposal only issue certificates in batches at a fixed frequency.¶
A hash computed over all the assertions in a batch, by building a Merkle Tree. The Merkle Tree construction and this hash are described in more detail in Section 5.5.1.¶
A structure which proves that some assertion is contained in some tree head. See Section 5.5.3.¶
A range of consecutive batch tree heads. A relying party maintains a copy of the CA's latest validity window. At any time, it will accept only assertions contained in tree heads contained in the current validity window.¶
The process of issuing and using a certificate is as follows:¶
The subscriber requests a certificate from the CA. Section 9 describes ACME [RFC8555] extensions for this.¶
The CA collects certificate requests into a batch (see Section 5.1) and builds the Merkle Tree and computes the tree head (see Section 5.5.1). It then signs the validity window ending at this tree head (see Section 5.5.2) and publishes (see Section 8) the result.¶
The CA constructs a certificate using the inclusion proof. It sends this certificate to the subscriber. See Section 5.5.3.¶
The transparency service downloads the abridged assertions, recreates the Merkle Tree, and validates the window signature. It mirrors them for monitors to observe. See Section 7.¶
The relying party fetches the latest validity window from the transparency service. This validity window will contain the new tree head.¶
In an application protocol such as TLS, the relying party communicates its currently saved validity window to the subscriber.¶
If the relying party’s validity window contains the subscriber’s certificate, the subscriber negotiates this protocol and sends the Merkle Tree certificate. See Section 10.3 for details. If there is no match, the subscriber proceeds as if this protocol were not in use (e.g., by sending a traditional X.509 certificate chain).¶
Figure 1 below shows this process.¶
The remainder of this document discusses this process in detail, followed by concrete instantions of it in TLS [RFC8446] and ACME [RFC8555].¶
[[TODO: The protocol described in this document is broadly independent of the assertion format. We describe, below, one possible structure, but welcome feedback on how best to structure the encoding. The main aims are simplicity and to improve on handling cross-protocol attacks per Section 13.2.]]¶
TLS certificates associate some application-specific identifier with a TLS signing key. When TLS is used to authenticate HTTPS [RFC9110] servers, these identifiers specify DNS names or HTTP origins. Other protocols may require other kinds of assertions.¶
To represent this, this document defines an Assertion structure:¶
enum { tls(0), (2^16-1) } SubjectType; enum { dns(0), dns_wildcard(1), ipv4(2), ipv6(3), (2^16-1) } ClaimType; struct { ClaimType claim_type; opaque claim_info<0..2^16-1>; } Claim; struct { SubjectType subject_type; opaque subject_info<0..2^16-1>; Claim claims<0..2^16-1>; } Assertion;¶
An Assertion is roughly analogous to an X.509 TBSCertificate (Section 4.1.2 of [RFC5280]). It describes a series of claims about some subject. The subject_info
field is interpreted according to the subject_type
value. For TLS, the subject_type
is tls
, and the subject_info
is a TLSSubjectInfo structure. TLSSubjectInfo is defined in full in Section 10.1 below, but as an illustrative example, it is reproduced below:¶
struct { SignatureScheme signature; opaque public_key<1..2^16-1>; } TLSSubjectInfo;¶
This structure represents the public half of a TLS signing key. The semantics are thus that each claim in claims
applies to the TLS client or server. This is analogous to X.509's SubjectPublicKeyInfo structure (Section 4.1.2.7 of [RFC5280]) but additionally incorporates the protocol. Protocols consuming an Assertion MUST check the subject_type
is a supported value before processing subject_info
. If unrecognized, the structure MUST be rejected.¶
Other protocols aiming to integrate with this structure allocate a SubjectType codepoint and describe how it is interpreted.¶
Likewise, a Claim structure describes some claim about the subject. The claim_info
field is interpreted according to the claim_type
. Each Claim structure in an Assertion's claims
field MUST have a unique claim_type
and all values MUST be sorted in order of increasing claim_type
. Structures violating this constraint MUST be rejected.¶
When a relying party interprets an Assertion certified by the CA, it MUST ignore any Claim values with unrecognized claim_type
. When a CA interprets an Assertion in a certification request from a subscriber, it MUST reject any Claim values with unrecognized claim_type
.¶
This document defines claim types for DNS names and IP addresses, but others can be defined.¶
[[TODO: For now, the claims below just transcribe the X.509 GeneralName structure. Should these be origins instead? For HTTPS, it's a pity to not capture the scheme and port. We do mandate ALPN in Section 10.2, so cross-protocol attacks are mitigated, but it's unfortunate that subscribers cannot properly separate their HTTPS vs FTPS keys, or their port 443 vs port 444 keys. One option here is to have HTTPS claims instead, and then other protocols can have FTPS claims, etc. #35 ]]¶
The dns
and dns_wildcard
claims indicate that the subject is authoritative for a set of DNS names. They use the DNSNameList structure, defined below:¶
opaque DNSName<1..255>; struct { DNSName dns_names<1..2^16-1>; } DNSNameList;¶
DNSName values use the "preferred name syntax" as specified by Section 3.5 of [RFC1034] and as modified by Section 2.1 of [RFC1123]. Alphabetic characters MUST additionally be represented in lowercase. IDNA names [RFC5890] are represented as A-labels. For example, possible values include example.com
or xn--iv8h.example
. Values EXAMPLE.COM
and <U+1F50F>.example
would not be permitted.¶
Names in a dns
claim represent the exact DNS name specified. Names in a dns_wildcard
claim represent wildcard DNS names and are processed as if prepended with the string "*.
" and then following the steps in Section 6.3 of [I-D.draft-ietf-uta-rfc6125bis].¶
The ipv4
and ipv6
claims indicate the subject is authoritative for a set of IPv4 and IPv6 addresses, respectively. They use the IPv4AddressList and IPv6AddressList structures, respectively, defined below. IPv4Address and IPv6Address are interpreted in network byte order.¶
uint8 IPv4Address[4]; uint8 IPv6Address[16]; struct { IPv4Address addresses<4..2^16-1>; } IPv4AddressList; struct { IPv6Address addresses<16..2^16-1>; } IPv6AddressList;¶
This section describes the structure of Merkle Tree certificates and defines the process of how a Merkle Tree certification authority issues certificates for a subscriber.¶
A Merkle Tree certification authority is defined by the following values:¶
hash
:A cryptographic hash function. In this document, the hash function is always SHA-256 [SHS], but others may be defined.¶
issuer_id
:A trust anchor identifier (see Section 3 of [I-D.beck-tls-trust-anchor-ids]) that identifies the CA. See Section 5.2 for details.¶
public_key
:The public half of a signing keypair. The corresponding private key, private_key
, is known only to the CA.¶
start_time
:The issuance time of the first batch of certificates, represented as a POSIX timestamp (see Section 2.1).¶
batch_duration
:A number of seconds which determines how frequently the CA issues certificates. See details below.¶
lifetime
:A number of seconds which determines the lifetime of certificates issued by this CA. MUST be a multiple of batch_duration
.¶
validity_window_size
:An integer describing the maximum number of unexpired batches which may exist at a time. This value is determined from lifetime
and batch_duration
by lifetime / batch_duration
.¶
These values are public and known by the relying party and the CA. They may not be changed for the lifetime of the CA. To change these parameters, the entity operating a CA may deploy a second CA and either operate both during a transition, or stop issuing from the previous CA.¶
[[TODO: The signing key case is interesting. A CA could actually maintain a single stream of Merkle Trees, but then sign everything with multiple keys to support rotation. The CA -> Subscriber -> RP flow does not depend on the signature, only the CA -> Transparency Service -> RP flow. The document is not currently arranged to capture this, but it probably should be. We probably need to decouple the signing half and the Merkle Tree half slightly. #36 ]]¶
Certificates are issued in batches. Batches are numbered consecutively, starting from zero. All certificates in a batch have the same issuance time, determined by start_time + batch_duration * batch_number
. This is known as the batch's issuance time. That is, batch 0 has an issuance time of start_time
, and issuance times increment by batch_duration
. A CA can issue no more frequently than batch_duration
. batch_duration
determines how long it takes for the CA to return a certificate to the subscriber.¶
All certificates in a batch have the same expiration time, computed as lifetime
past the issuance time. After this time, the certificates in a batch are no longer valid. Merkle Tree certificates uses a short-lived certificates model, such that certificate expiration replaces an external revocation signal like CRLs [RFC5280] or OCSP [RFC6960]. lifetime
SHOULD be set accordingly. For instance, a deployment with a corresponding maximum OCSP [RFC6960] response lifetime of 14 days SHOULD use a value no higher than 14 days. See Section 13.3 for details.¶
CAs are RECOMMENDED to use a batch_duration
of one hour, and a lifetime
of 14 days. This results in a validity_window_size
of 336, for a total of 10,752 bytes in SHA-256 hashes.¶
To prevent cross-protocol attacks, the key used in a Merkle Tree CA MUST be unique to that Merkle Tree CA. It MUST NOT be used in another Merkle Tree CA, or for another protocol, such as X.509 certificates.¶
A Merkle Tree CA's issuer_id
is a trust anchor identifier, defined in Section 3 of [I-D.beck-tls-trust-anchor-ids]. However, unlike an X.509 CA, the entire OID arc rooted at the identifier is associated with the CA. OIDs under this arc are used to identify batches below.¶
An individual batch from a Merkle Tree CA also has an associated trust anchor identifier. It is determined by appending the batch number to the CA's issuer_id
.¶
For example, a Merkle Tree CA may have an issuer_id
of 32473.1
, in the ASCII representation. The batch with batch number 42 would then have a trust anchor identifier of 32473.1.42
.¶
Each batch is in one of three states:¶
The current time is before the batch's issuance time¶
The current time is not before the batch's issuance time, but the batch has not yet been issued¶
Certificates have been issued for this batch¶
The CA also maintains a latest batch number, which is the number of the last batch in the "issued" state. As an invariant, all batches before this value MUST also be in the "issued" state.¶
For each batch in the "issued" state, the CA maintains the following batch state:¶
The list of abridged assertions certified in this batch.¶
The tree head, a hash computed over this list, described in Section 5.5.1.¶
A validity window signature computed as described in Section 5.5.2.¶
The CA exposes all of this information in an HTTP [RFC9110] interface described in Section 8.¶
The CA additionally maintains an issuance queue, not exposed via the HTTP interface.¶
When a subscriber requests a certificate for some assertion, the CA first validates it per its issuance policy. For example, it may perform ACME identifier validation challenges (Section 8 of [RFC8555]). Once validation is complete and the CA is willing to certify the assertion, the CA appends it to the issuance queue.¶
The CA runs a regularly-scheduled issuance job which converts this queue into certificates. This job runs the following procedure:¶
If no batches are in the "ready" state, do nothing and abort this procedure. Schedule a new job to run sometime after the earliest "pending" batch's issuance time.¶
For each batch in the "ready" state other than the latest one, run the procedure in Section 5.5 with an empty assertion list, in order of increasing batch number. Batches cannot be skipped.¶
Empty the issuance queue into an ordered list of assertions. Run the procedure in Section 5.5 using this list and the remaining batch in the "ready" state. This batch's issuance time will be at or shortly before the current time.¶
This section describes how to certify a given list of assertions at a given batch number. The batch MUST be in the "ready" state, and all preceding batches MUST be in the "issued" state.¶
First, the CA then builds a Merkle Tree from the list as follows:¶
Let n
be the number of input assertions. If n > 0
, the CA builds a binary tree with l levels numbered 0
to l-1
, where l
is the smallest positive integer such that n <= 2^(l-1)
. Each node in the tree contains a hash value. Hashes in the tree are built from the following functions:¶
HashEmpty(level, index) = hash(HashEmptyInput) HashNode(left, right, level, index) = hash(HashNodeInput) HashAssertion(assertion, index) = hash(HashAssertionInput)¶
HashEmpyInput
, HashNodeInput
and HashAssertionInput
are computed by encoding the structures defined below:¶
struct { uint8 distinguisher = 0; opaque issuer_id<1..32>; uint32 batch_number; uint64 index; uint8 level; } HashEmptyInput; struct { uint8 distinguisher = 1; opaque issuer_id<1..32>; uint32 batch_number; uint64 index; uint8 level; opaque left[hash.length]; opaque right[hash.length]; } HashNodeInput; struct { SubjectType subject_type; opaque subject_info_hash[hash.length]; Claim claims<0..2^16-1>; } AbridgedAssertion; struct { uint8 distinguisher = 2; opaque issuer_id<1..32>; uint32 batch_number; uint64 index; AbridgedAssertion abridged_assertion; } HashAssertionInput;¶
issuer_id
and batch_number
are set to the CA's issuer_id
and the current batch number. HashAssertionInput.abridged_assertion.subject_info_hash
is set to hash(assertion.subject_info)
from the function input assertion
, and the remaining fields of HashAssertionInput.abridged_assertion
are taken unmodified from assertion
. The remaining fields, such as index
, are set to inputs of the function.¶
Tree levels are computed iteratively as follows:¶
Initialize level 0 with n elements. For j
between 0
and n-1
, inclusive,
set element j
to the output of HashAssertion(assertion[j], j)
.¶
For i
between 1
and l-1
, inclusive, compute level i
from level i-1
as
follows:¶
If level i-1
has an odd number of elements j
, append HashEmpty(i-1, j)
to the level.¶
Initialize level i
with half as many elements as level i-1
. For all j
,
set element j
to the output of HashNode(left, right, i, j)
where left
is
element 2*j
of level i-1
and right
is element 2*j+1
of level i-1
.
left
and right
are the left and right children of element j
.¶
At the end of this process, level l-1
will have exactly one root element. This element is called the tree head. Figure 2 shows an example tree for three assertions. The tree head in this example is t20.¶
If n
is zero, the CA does not build a tree and the tree head is HashEmpty(0, 0)
.¶
If n
is one, the tree contains a single level, level 0, and has a tree head of HashAssertion(assertion, 0)
.¶
Batches are grouped into consecutive ranges of validity_window_size
batches, called validity windows. As validity_window_size
is computed to cover the full certificate lifetime, a validity window that ends at the latest batch number covers all certificates that may still be valid from a CA.¶
Validity Windows are serialized into the following structure:¶
opaque TreeHead[hash.length]; struct { uint32 batch_number; TreeHead tree_heads[validity_window_size*hash.length]; } ValidityWindow;¶
batch_number
is the batch number of the highest batch in the validity window.¶
tree_heads
value contains the last validity_window_size
tree heads. (Recall the TLS presentation language brackets the total length of a vector in bytes; not the number of elements.) tree_heads
starts from batch_number
, in decreasing batch number order. That is, tree_heads[0]
is the tree head for batch batch_number
, tree_heads[1]
is the tree head for batch_number - 1
, and so on. If batch_number < validity_window_size - 1
, any tree heads for placeholder negative batch numbers are filled with HashEmpty(0, 0)
, computed with batch_number
set to 0.¶
After the CA builds the Merkle Tree for a batch, it constructs the ValidityWindow structure whose batch_number
is the number of the batch being issued. It then computes a signature over the following structure:¶
struct { uint8 label[32] = "Merkle Tree Crts ValidityWindow\0"; opaque issuer_id<1..32>; ValidityWindow window; } LabeledValidityWindow;¶
The label
field is an ASCII string. The final byte of the string, "\0", is a zero byte, or ASCII NULL character. The issuer_id
field is the CA's issuer_id
. Other parties can verify the signature by constructing the same input and verifying with the CA's public_key
.¶
The CA saves this signature as the batch's validity window signature. It then updates the latest batch to point to batch_number
. A CA which generates such a signature is considered to have certified every assertion contained in every value in the tree_heads
list, with expiry determined by batch_number
, the position of the tree head in the list, and the CA's input parameters as described in Section 5.1.¶
A CA MUST NOT generate signatures over inputs that are parseable as LabeledValidityWindow, except via the above process. If a LabeledValidityWindow structure that was not produced in this way has a valid signature by CA's public_key
, this indicates misuse of the private key by the CA, even if the preimages to the tree_heads
values, or intermediate nodes, or subject_info_hash
values are not known.¶
[[TODO: BikeshedCertificate is a placeholder name until someone comes up with a better one. #15 ]]¶
[[TODO: A subscriber has no way to know when a certificate expires. We need to define a mandatory expiration certificate property, or do #83, which, depending on how it's done could avoid that.]]¶
For each assertion in the tree, the CA constructs a BikeshedCertificate structure containing the assertion and a proof. A proof is a message that allows the relying party to accept the associated assertion, provided it trusts the CA and recognizes the tree head. The structures are defined below:¶
/* See Section 4.1 of draft-beck-tls-trust-anchor-ids */ opaque TrustAnchorIdentifier<1..2^8-1>; struct { TrustAnchorIdentifier trust_anchor; opaque proof_data<0..2^16-1>; } Proof; struct { Assertion assertion; Proof proof; } BikeshedCertificate;¶
A proof's trust_anchor
field is a trust anchor identifier (see Section 3 of [I-D.beck-tls-trust-anchor-ids] and Section 4.1 of [I-D.beck-tls-trust-anchor-ids]), which determines the proof's type and issuer. It is analogous to an X.509 trust anchor's subject name. When the issuer is a Merkle Tree CA, the trust_anchor
is the containing batch's trust anchor identifier, as described in Section 5.2.¶
The proof_data
is a byte string, opaque to the subscriber, in some format agreed upon by the proof issuer and relying party. If the issuer is a Merkle Tree CA, as defined in this document, the proof_data
contains a MerkleTreeProofSHA256, described below. Future mechanisms using the BikeshedCertificate may define other formats.¶
opaque HashValueSHA256[32]; struct { uint64 index; HashValueSHA256 path<0..2^16-1>; } MerkleTreeProofSHA256;¶
After building the tree, the CA constructs a MerkleTreeProofSHA256 for each assertion as follows. For each index i
in the batch's assertion list:¶
Set index
to i
. This will be a value between 0
and n-1
, inclusive.¶
Set path
to an array of l-1
hashes. Set element j
of this array to element
k
of level j
, where k
is (i >> j) ^ 1
. >>
denotes a bitwise
right-shift, and ^
denotes a bitwise exclusive OR (XOR) operation. This
element is the sibling of an ancestor of assertion i
in the tree. Note the
tree head is never included.¶
For example, the path
value for the third assertion in a batch of three assertions would contain the marked nodes in Figure 3, from bottom to top.¶
If the batch only contained one assertion, path
will be empty and index
will be zero.¶
For each assertion, the CA assembles a BikeshedCertificate structure and sends it to the subscriber.¶
This certificate can be presented to supporting relying parties as described in Section 6. It is valid until the batch expires.¶
Merkle Tree proofs scale logarithmically in the batch size. Section 11.2 recommends subscribers renew halfway through the previous certificate's lifetime. Batch sizes will thus, on average, be subscriber_count * 2 / validity_window_size
, where subscriber_count
is a CA's active subscriber count. The recommended parameters in Section 5.1 give an average of subscriber_count / 168
.¶
Some organizations have published statistics which can estimate batch sizes for the Web PKI. On March 7th, 2023, [LetsEncrypt] reported around 330,000,000 active subscribers for a single CA. [MerkleTown] reported around 3,800,000,000 unexpired certificates in Certificate Transparency logs, and an issuance rate of around 257,000 per hour. Note the numbers from [MerkleTown] represent, respectively, all Web PKI CAs combined and issuance rates for longer-lived certificates and may not be representative of a Merkle Tree certificate deployment.¶
These three estimates correspond to batch sizes of, respectively, around 2,000,000, around 20,000,000, and 257,000. The corresponding path
lengths will be 20, 24, and 17, given proof sizes of, respectively, 640 bytes, 768 bytes, and 544 bytes.¶
For larger batch sizes, 32 hashes, or 1024 bytes, is sufficient for batch sizes up to 2^33 (8,589,934,592) certificates.¶
This section describes how subscribers present and relying parties verify Merkle Tree certificates.¶
For each Merkle Tree CA it trusts, a relying party maintains a copy of the most recent validity window from the CA. This structure determines which certificates the relying party will accept. It is regularly updated from the transparency service, as described in Section 7.¶
Each batch in the relying party's validity window is a trust anchor for purposes of certificate verification (see Section 6.2) and certificate negotiation (see Section 10.3).¶
This section describes the verification process for a BikeshedCertificate. It describes error conditions with TLS alerts, defined in Section 6.2 of [RFC8446]. Non-TLS applications SHOULD map these error conditions to the corresponding application-specific errors. When multiple error conditions apply, the application MAY return any applicable error.¶
When a subscriber presents a BikeshedCertificate, the relying party runs the following procedure:¶
Determines if trust_anchor
corresponds to a supported trust anchor, and the type of that trust anchor. If trust_anchor
is unrecognized, the relying party rejects the certificate with an unknown_ca
error.¶
Run the verification subroutine corresponding to that trust anchor, defined below.¶
Optionally, perform any additional application-specific checks on the assertion and issuer. For example, an HTTPS client might constrain an issuer to a particular DNS subtree.¶
If all the preceding checks succeed, the certificate is valid and the application can proceed with using the assertion.¶
Step 2 in the above procedure runs a trust-anchor-specific verification subroutine. This subroutine is determined by the type of trust anchor. Each mechanism using the BikeshedCertificate format MUST define a verification subroutine.¶
For a Merkle Tree trust anchor, the trust anchor will identify a batch in the relying party's validity window. (See Section 5.2 and Section 6.1.) The batch's verification subroutine is defined below:¶
Compute the batch's expiration time, as described in Section 5.1. If this value is before the current time, abort this procedure with a certificate_expired
error.¶
Set hash
to the output of HashAssertion(assertion, index)
. Set remaining
to the certificate's index
value.¶
For each element v
at zero-based index i
of the certificate's path
field, in order:¶
If remaining
is non-zero, abort this procedure with an error.¶
If hash
is not equal to the batch's tree head in the relying party's saved validity window (see Section 6.1), abort this procedure with a bad_certificate
error.¶
If all the preceding checks succeed, the Merkle Tree certificate verification subroutine completes successfully.¶
This section describes the role of the transparency service. The transparency service ensures all certificates accepted by the relying party are consistently and publicly logged. It performs three functions:¶
Mirror all abridged assertions certified by the CA and present them to monitors¶
Validate all tree heads and validity windows produced by the CA¶
Provide the latest valid validity window to relying parties¶
In doing so, the transparency service MUST satisfy the following requirements:¶
The mirrored CA state is append-only. That is, the hashes, signatures, and assertions for a given batch number MUST NOT change.¶
All tree hashes sent to relying parties MUST be reflected in the mirrored CA state.¶
The transparency service publishes the mirrored CA state using the same interface as Section 8. The protocol between the relying party and transparency service is out of scope of this document. The relying party MAY use the interface defined here, or an existing application-specific authenticated channel.¶
As discussed in Section 13.1, relying parties MUST ensure that any validity windows obtained were asserted by the CA. This SHOULD be done by having the transparency service forward the CA's signature, with the relying party verifying it. However, if the transparency service already maintains a trusted, authenticated channel to the relying parties (e.g. a software or root store update channel), relying parties MAY rely on the transparency service to validate the signature on their behalf, rather than sending it over this channel.¶
Although described as a single service for clarity, the transparency service may be implemented as a combination of services run by multiple entities, depending on security goals. For example deployments, this section first describes a single trusted service, then it describes other possible models where trust is divided between entities.¶
Some relying parties regularly contact a trusted update service, either for software updates or to update individual components, such as the services described in [CHROMIUM] and [FIREFOX]. Where these services are already trusted for the components such as the trust anchor list or certificate validation software, a single trusted transparency service may be a suitable model.¶
The transparency service maintains a mirror of the CA's latest batch number, and batch state. Roughly once every batch_duration
, it polls the CA's HTTP interface (see Section 8) and runs the following steps:¶
Fetch the CA's latest batch number. If this fetch fails, abort this procedure with an error.¶
Let new_latest_batch
be the result and old_latest_batch
be the currently mirrored value. If new_latest_batch
equals old_latest_batch
, finish this procedure without reporting an error.¶
If new_latest_batch
is less than old_latest_batch
, abort this procedure with an error.¶
If the issuance time for batch new_latest_batch
is after the current time (see Section 5.1), abort this procedure with an error.¶
For all i
such that old_latest_batch < i <= new_latest_batch
:¶
Fetch the signature, tree head, and abridged assertion list for batch i
. If this fetch fails, abort this procedure with an error.¶
Compute the tree head for the assertion list, as described in Section 5.5.1. If this value does not match the fetched tree head, abort this procedure with an error.¶
Compute the ValidityWindow structure and verify the signature, as described in Section 5.5.2. Set tree_heads[0]
to the tree head fetched above. Set the other values in tree_heads
to the previously mirrored values. If signature verification fails, abort this procedure with an error.¶
Set the mirrored latest batch number to i
and save the fetched batch state.¶
[[TODO: If the mirror gets far behind, if the CA just stops publishing for a while, it may suddenly have to catch up on many batches. Should we allow the mirror to catch up to the latest validity window and skip the intervening batches? The intervening batches are guaranteed to have been expired #37 ]]¶
If the relying party has a trusted update service, but the update service does not have the resources to mirror the full batch state, the transparency service can be composed of this update service and several, less trusted mirrors. In this model, the mirrors are not trusted to serve authoritative trust anchor information to relying parties, but the update service trusts at least half of them to faithfully and consistently mirror the batch state.¶
Each mirror follows the procedure in Section 7.1 to maintain and publish a mirror of the CA's batch state.¶
The update server maintains the latest validity window validated to appear in all mirrors. It updates this by polling the mirrors and running the following steps:¶
For each mirror, fetch the latest batch number.¶
Let new_latest_batch
be the highest batch number that is bounded by the value fetched from at least half of the mirrors. Let old_latest_batch
be the batch number of the currently stored validity window.¶
If new_latest_batch
equals old_latest_batch
, finish this procedure without reporting an error.¶
If new_latest_batch
is less than old_latest_batch
, abort this procedure with an error.¶
If the issuance time for batch new_latest_batch
is after the current time (see Section 5.1), abort this procedure with an error.¶
Fetch the validity window with new_latest_batch
from each mirror that returned an equal or higher latest batch number. If any fetches fail, or if the results do not match across all mirrors, abort this procedure with an error.¶
Verify the validity window signature, as described in Section 5.5.2. If the signature is invalid, abort this procedure with an error.¶
If the old and new validity windows contain overlapping batch numbers, verify that the tree hashes match. If not, abort this procedure with an error.¶
Update the saved validity window with the new value.¶
Compared to Section 7.1, this model reduces trust in the mirror services, but can delay certificate usability if some of the mirrors consume CA updates too slowly. This can be tuned by adjusting the threshold in step 2.¶
In a transparency service using this model, each mirror independently publishes the batch state via Section 8.¶
Relying parties without a trusted update service can fetch from mirrors directly. Rather than relying on the update service to fetch the validity window state, the relying party runs the procedure described in Section 7.2, and uses the saved validity window to verify certificates.¶
Monitors in this document are analogous to monitors in [RFC6962]. Monitors watch an implementation of the HTTP APIs in Section 8 to verify correct behavior and watch for certificates of interest. This is typically the transparency service. A monitor needs to, at least, inspect every new batch. It may also maintain a copy of the batch state.¶
It does so by following the procedure in Section 7.1, fetching from the service being monitored. If the procedure fails for a reason other than the service availability, this should be viewed as misbehavior on the part of the service. If the procedure fails due to service availability and the service remains unavailable for an extended period, this should also be viewed as misbehavior. If the monitor is not maintaining a copy of the batch state, it skips saving the abridged assertions.¶
[RFC6962] additionally defines the role of auditor, which validates that Signed Certificate Timestamps (SCTs) and Signed Tree Heads (STHs) in Certificate Transparency are correct. There is no analog to SCTs in this document. The signed validity window structure (Section 5.5.2) is analogous to an STH, but consistency is checked simply by ensuring overlapping tree heads match, so this document does not define this as an explicit role. If two inconsistent signed validity windows are ever observed from a Merkle Tree CA, this should be viewed as misbehavior on the part of the CA.¶
[[TODO: This section hasn't been written yet. For now, this is just an informal sketch. The real text will need to define request/response formats more precisely, with MIME types, etc. #12 ]]¶
CAs and transparency services publish state over an HTTP [RFC9110] interface described below.¶
CAs and any components of the transparency service that maintain validity window information implement the following interfaces:¶
GET {prefix}/latest
returns the latest batch number.¶
GET {prefix}/validity-window/latest
returns the ValidityWindow structure and signature (see Section 5.5.2) for the latest batch number.¶
GET {prefix}/validity-window/{number}
returns the ValidityWindow structure and signature (see Section 5.5.2) for batch number
, if it is in the "issued" state, and a 404 error otherwise.¶
GET {prefix}/batch/{number}/info
returns the validity window signature and tree head for batch number
, if batch number
is in the "issued" state, and a 404 error otherwise.¶
CAs and any components of the transparency service that mirror the full abridged assertion list additionally implement the following interface:¶
GET {prefix}/batch/{number}/assertions
returns the abridged assertion list for batch number
, if number
is in the issued state, and a 404 error otherwise.¶
If the interface is implemented by a distributed service, with multiple servers, updates may propagate to servers at different times, which will cause temporary inconsistency. This inconsistency can impede this system's transparency goals (Section 13.4).¶
Services implementing this interface SHOULD wait until batch state is fully propagated to all servers before updating the latest batch number. That is, if any server returns a latest batch number of N in either of the first two HTTP endpoints, batch numbers N and below SHOULD be available under the last three batch-number-specific HTTP endpoints in all servers. If this property does not hold at any time, it is considered a service unavailability.¶
Individual servers in a service MAY return different latest batch numbers. Individual servers MAY also differ on whether a batch number has a response available or return a 404 error. Provided the above consistency property holds, these two inconsistencies do not constitute service unavailability.¶
Section 11.4 discusses service availability requirements.¶
[[TODO: Once a batch has expired, do we allow a CA to stop publishing it? The transparency service can already log it for as long, or as little, as it wishes. We effectively have CT log temporal sharding built into the system. #2 ]]¶
[[TODO: If we have the validity window endpoint, do we still need to separate "info" and "assertions"? #12]]¶
[[TODO: This section hasn't been written yet. Instead, what follows is an informal discussion. #13 ]]¶
Section 6 of [I-D.draft-beck-tls-trust-anchor-ids] defines the bulk of what's needed. The missing parts are:¶
Some way to specify that the client supports BikeshedCertificate. At minimum a separate MIME type, but it likely needs to be known at order creation.¶
Some way to accommodate MTC's long issuance time. ACME has the "processing" state, and the Retry-After header can tell the subscriber when to query again. But the fallback certificate will issue much faster, so they cannot be issued together in the same ACME order, as [I-D.draft-beck-tls-trust-anchor-ids] currently does.¶
Use [I-D.draft-ietf-acme-ari] to move the renewal logic in Section 11.2 from the subscriber to the ACME server.¶
We should also define a certificate request format, though it is broadly just reusing the Assertion structure. If the CA wishes to check possession of the private key, it'll need to come with a signature or do some online operation (e.g. if it's a KEM key). This is inherently protocol-specific, because the mechanism needs to coexist with the target protocol. (Signed CSRs implicitly assume the target protocol's signature payloads cannot overlap with that of a CSR.)¶
This section describes the SubjectType for use with TLS [RFC8446]. The SubjectType value is tls
, and the subject_info
field contains a TLSSubjectInfo structure, defined below:¶
enum { tls(0), (2^16-1) } SubjectType; struct { SignatureScheme signature; opaque public_key<1..2^16-1>; /* TODO: Should there be an extension list? #38 */ } TLSSubjectInfo;¶
A TLSSubjectInfo describes a TLS signing key. The signature
field is a SignatureScheme Section 4.2.3 of [RFC8446] value describing the key type and signature algorithm it uses for CertificateVerify.¶
The public_key
field contains the subscriber's public key. The encoding is determined by the signature
field as follows:¶
The public key is an RSAPublicKey structure [RFC8017] encoded in DER [X.690]. BER encodings which are not DER MUST be rejected.¶
The public key is a UncompressedPointRepresentation structure defined in Section 4.2.8.2 of [RFC8446], using the curve specified by the SignatureScheme.¶
The public key is the byte string encoding defined in [RFC8032]¶
This document does not define the public key format for other algorithms. In order for a SignatureScheme to be usable with TLSSubjectInfo, this format must be defined in a corresponding document.¶
[[TODO: If other schemes get defined before this document is done, add them here. After that, it's on the other schemes to do it. #39 ]]¶
[[TODO: Bikeshed is a placeholder name until someone comes up with a better one. #15]]¶
This section defines the Bikeshed
TLS certificate type, which may be negotiated with the client_certificate_type
, server_certificate_type
[RFC7250], or cert_type
[RFC6091] extensions. It can only be negotiated with TLS 1.3 or later. Servers MUST NOT negotiate it in TLS 1.2 or below. If the client receives a ServerHello that negotiates it in TLS 1.2 or below, it MUST abort the connection with an illegal_parameter
alert.¶
[[TODO: None of these three extensions is quite right for client certificates because the negotiation isn't symmetric. See discussion in Appendix B. We may need to define another one. #18]]¶
When negotiated, the Certificate message MUST contain a single CertificateEntry structure. CertificateEntry is updated as follows:¶
enum { Bikeshed(TBD), (255) } CertificateType; struct { select (certificate_type) { /* Certificate type defined in this document */ case Bikeshed: BikeshedCertificate certificate; /* From RFC 7250 */ case RawPublicKey: opaque ASN1_subjectPublicKeyInfo<1..2^24-1>; case X509: opaque cert_data<1..2^24-1>; /* Additional certificate types based on the "TLS Certificate Types" registry */ }; Extension extensions<0..2^16-1>; } CertificateEntry;¶
The subject_type
field in the certificate MUST be of type tls
(Section 10.1). The CertificateVerify message is computed and processed as in [RFC8446], with the following modifications:¶
The signature is computed and verified with the key described in the TLSSubjectInfo. The relying party uses the key decoded from the public_key
field, and the subscriber uses the corresponding private key.¶
The SignatureScheme in the CertificateVerify MUST match the signature
field in the TLSSubjectInfo.¶
The second modification differs from [RFC8446]. Where [RFC8446] allowed an id-rsaEncryption key to sign both rsa_pss_rsae_sha256
and rsa_pss_rsae_sha384
, TLSSubjectInfo keys are specific to a single algorithm. Future documents MAY relax this restriction for a new SignatureScheme, provided it was designed to be used concurrently with the value in TLSSubjectInfo. In particular, the underlying signature algorithm MUST match, and there MUST be appropriate domain separation between the two modes. For example, [I-D.draft-ietf-tls-batch-signing] defines new SignatureSchemes, but the same keypair can be safely used with one of the new values and the corresponding base SignatureScheme.¶
If this certificate type is used for either the client or server certificate, the ALPN [RFC7301] extension MUST be negotiated. If no application protocol is selected, endpoints MUST close the connection with a no_application_protocol
alert.¶
[[TODO: Suppose we wanted to introduce a second SubjectType for TLS, either to add new fields or capture a new kind of key. That would need to be negotiated. We could use another extension, but defining a new certificate type seems most natural. That suggests this certificate type isn't about negotiating BikeshedCertificate in general, but specifically SubjectType.tls and TLSSubjectInfo. So perhaps the certificate type should be TLSSubjectInfo or BikeshedTLS. #7 ]]¶
Merkle Tree certificates will only be accepted in up-to-date relying parties, and require a negotiation mechanism to use. Merkle Tree certificate implementations SHOULD use the trust_anchors
extension [I-D.draft-beck-tls-trust-anchor-ids] as described below:¶
For each Merkle Tree CA trusted by the relying party, the batches in the validity window each determine a trust anchor, as described in Section 6.1. The trust anchor's identifier is the batch identifier, as described in Section 5.2. Future mechanisms using the BikeshedCertificate format (see Section 5.5.3) MUST similarly define how relying parties determine trust anchor identifiers.¶
As even a single validity window results in validity_window_size
trust anchors, sending all trust anchors in the trust_anchors
extension would be prohitively large in most cases. Instead, relying parties SHOULD use the retry mechanism described in Section 4.3 of [I-D.draft-beck-tls-trust-anchor-ids] and the DNS hint described in Section 5 of [I-D.draft-beck-tls-trust-anchor-ids].¶
[[TODO: We could reduce the reliance on DNS by adding https://github.com/davidben/tls-trust-expressions/issues/62, either in this draft or the main trust anchor IDs draft.]]¶
The subscriber's list of candidate certification paths (see Section 3.3 of [I-D.beck-tls-trust-anchor-ids]) is extended to carry both X.509 and BikeshedCertificate credentials. The two types of credentials MAY appear in any relative preference order, based on the subscriber's policies. Like an X.509 credential, a BikeshedCertificate credential also has a CertificatePropertyList (see Section 3.1 of [I-D.beck-tls-trust-anchor-ids]).¶
For each of the subscriber's BikeshedCertificate credentials, the corresponding trust anchor identifier is the trust_anchor
field in the BikeshedCertificate structure. This differs from X.509 credentials, which require an out-of-band value in the CertificatePropertyList. It is an error for a BikeshedCertificate credential's CertificatePropertyList to contain the trust_anchor_identifier
property.¶
The subscriber then selects certificates as described in Section 4.2 of [I-D.draft-beck-tls-trust-anchor-ids]. In doing so, it SHOULD incorporate trust anchor negotiation and certificate type negotiation (see Section 10.2) into the selection criteria for BikeshedCertificate-based credentials.¶
[[TODO: Certificate type negotiation doesn't work right for client certificates. See Appendix B]]¶
Subscribers using Merkle Tree certificates SHOULD additionally provision certificates from another PKI mechanism, such as X.509. This ensures the service remains available to relying parties that have not recently fetched validity window updates, or lack connectivity to the transparency service.¶
If the pipeline of updates from the CA to the transparency service to relying parties is interrupted, certificate issuance may halt, or newly issued certificates may no longer be usable. When this happens, the optimization in this document may fail, but fallback mechanisms ensure services remain available.¶
When a subscriber requests a certificate, the CA cannot fulfill the request until the next batch is ready. Once published, the certificate will not be accepted by relying parties until the batch state is mirrored by their respective transparency services, then pushed to relying parties.¶
To account for this, subscribers SHOULD request a new Merkle Tree certificate significantly before the previous Merkle Tree certificate expires. Renewing halfway into the previous certificate's lifetime is RECOMMENDED. Subscribers additionally SHOULD retain both the new and old certificates in the certificate set until the old certificate expires. As the new tree hash is delivered to relying parties, certificate negotiation will transition relying parties to the new certificate, while retaining the old certificate for clients that are not yet updated.¶
Merkle Tree certificates' issuance delays make them unsuitable when rapidly deploying a new service and reacting to key compromise.¶
When a new service is provisioned with a brand new Merkle Tree certificate, relying parties will not yet have received a validity window containing this certificate from the transparency service and can therefore not validate this certificate until receiving them. The subscriber SHOULD, in parallel, also provision a certificate using another PKI mechanism (e.g. X.509). Certificate negotiation will then switch over to serving the Merkle Tree certificate as relying parties are updated.¶
If the service is performing a routine key rotation, and not in response to a known compromise, the subscriber MAY use the process described in Section 11.2, allowing certificate negotiation to also switch the private key used. This slightly increases the lifetime of the old key but maintains the size optimization continuously.¶
If the service is rotating keys in response to a key compromise, this option is not available. Instead, the service SHOULD immediately discard the old key and request a more immediate issuance mechanism. As in the initial deployment case, it SHOULD request a Merkle Tree certificate in parallel, which will restore the size optimization over time.¶
CAs and transparency services serve an HTTP interface defined in Section 8. This service may be temporarily unavailable, either from service outage or if the service does not meet the consistency condition mid-update. Exact availability requirements for these services are out of scope for this document, but this section provides some general guidance.¶
If the CA's interface becomes unavailable, the transparency service will be unavailable to update. This will prevent relying parties from accepting new certificates, so subscribers will need to use fallback mechanisms per Section 11.1. This does not compromise transparency goals per Section 13.4.2. However, a CA which is persistently unavailable may not offer sufficient benefit to be used by subscribers or trusted by relying parties.¶
However, if the transparency service's interface becomes unavailable, monitors will be unable to check for unauthorized certificates. This does compromise transparency goals. Mirrors of the batch state partially mitigate this, but service unavailability may prevent mirrors from replicating a batch that relying parties accept.¶
The Privacy Considerations described in Section 9 of [I-D.beck-tls-trust-anchor-ids] apply to its use with Merkle Tree Certificates.¶
In particular, relying parties that share a transparency service will fetch the same stream of updates. However, updates may reach different users at different times, resulting in some variation across users. This variation may contribute to a fingerprinting attack [RFC6973]. If the Merkle Tree CA trust anchors are sent unconditionally in trust_anchors
, this variation will be passively observable. If they are sent conditionally, e.g. with the DNS-mechanism, the trust anchor list will require active probing.¶
A key security requirement of any PKI scheme is that relying parties only accept assertions that were certified by a trusted certification authority. This is achieved by the following two properties:¶
The relying party MUST NOT accept any validity window that was not authenticated as coming from the CA.¶
For any tree head computed from a list of assertions as in Section 5.5.1, it is computationally infeasible to construct an assertion not this list, and some inclusion proof, such that the procedure in Section 6.2 succeeds.¶
Section 7 discusses achieving the first property.¶
The second property is achieved by using a collision-resistant hash in the Merkle Tree construction. The HashEmpty
, HashNode
, and HashAssertion
functions use distinct initial bytes when calling the hash function, to achieve domain separation.¶
Using the same key material in different, incompatible ways risks cross-protocol attacks when the two uses overlap. To avoid this, Section 5.1 forbids the reuse of Merkle Tree CA private keys in another protocol. A CA MUST NOT generate signatures with its private key, except as defined in Section 5.5.2, or an extension of this protocol. Any valid signature of a CA's public_key
that does not meet these requirements indicates misuse of the private key by the CA.¶
To reduce the risk of attacks if this guidance is not followed, the LabeledValidityWindow structure defined in Section 5.5.2 includes a label string, and the CA's issuer_id
. Extensions of this protocol MAY be defined which reuse the keys, but any that do MUST use a different label string and analyze the security of the two uses concurrently.¶
Likewise, key material included in an assertion (Section 4) MUST NOT be used in another protocol, unless that protocol was designed to be used concurrently with the original purpose. The Assertion structure is designed to facilitate this. Where X.509 uses an optional key usage extension (see Section 4.2.1.3 of [RFC5280]) and extended key usage extension (see Section 4.2.1.12 of [RFC5280]) to specify key usage, an Assertion is always defined first by a SubjectType value. Subjects cannot be constructed without first specifying the type, and subjects of different types cannot be accidentally interpreted as each other.¶
The TLSSubjectInfo structure additionally protects against cross-protocol attacks in two further ways:¶
A TLSSubjectInfo specifies the key type not with a SubjectPublicKeyInfo Section 4.1.2.7 of [RFC5280] object identifier, but with a SignatureScheme structure. Where [RFC8446] allows an id-rsaEncryption key to sign both rsa_pss_rsae_sha256
and rsa_pss_rsae_sha384
, this protocol specifies the full signature algorithm parameters.¶
To mitigate cross-protocol attacks at the application protocol [ALPACA], this document requires connections using it to negotiate the ALPN [RFC7301] extension.¶
Merkle Tree certificates avoid sending an additional signature for OCSP responses by using a short-lived certificates model. Per Section 5.1, Merkle Tree CA's certificate lifetime MUST be set such that certificate expiration replaces revocation. Existing revocation mechanisms like CRLs and OCSP are themselves short-lived, signed messages, so a low enough certificate lifetime provides equivalent revocation capability.¶
Relying parties with additional sources of revocation such as [CRLite] or [CRLSets] SHOULD provide a mechanism to express revoked assertions in such systems, in order to opportunistically revoke assertions in up-to-date relying parties sooner. It is expected that, in most deployments, relying parties can fetch this revocation data and Merkle Tree CA validity windows from the same service.¶
[[TODO: Is it worth defining an API for Merkle Tree CAs to publish a revocation list? That would allow automatically populating CRLite and CRLSets. Maybe that's a separate document. #41]]¶
The transparency service does not prevent unauthorized certificates, but it aims to provide comparable security properties to Certificate Transparency [RFC6962]. If a subscriber presents an acceptable Merkle Tree certificate to a relying party, the relying party should have assurance it was published in some form that monitors and, in particular, the subject of the certificate will be able to notice.¶
Although CAs in this document publish structures similar to a Certificate Transparency log, they do not need to function correctly to provide transparency.¶
A CA could violate the append-only property of its batch state, and present differing views to different parties. Unlike a misbehaving Certificate Transparency log, this would not compromise transparency. Whichever view is presented to the transparency service at the time of updates determines the canonical batch state for both relying parties and monitors. Certificates that are consistent with only the other view will be rejected by relying parties. If the transparency service observes both views, the procedures in Section 7 will prevent the new, conflicting view from overwriting the originally saved view. Instead, the update process will fail and further certificates will not be accepted.¶
A CA could also sign a validity window containing an unauthorized certificate and feign an outage when asked to serve the corresponding assertions. However, if the assertion list was never mirrored by the transparency service, the tree head will never be pushed to relying parties, so the relying party will reject the certificate. If the assertion list was mirrored, the unauthorized certificate continues to be available to monitors.¶
As a consequence, monitors MUST use the transparency service's view of the batch state when monitoring for unauthorized certificates. If the transparency service is a collection of mirrors, as in Section 7.2 or Section 7.3, monitors MUST monitor each mirror. Monitors MAY optionally monitor the CA directly, but this alone is not sufficient to avoid missing certificates.¶
This document divides CA and transparency service responsibilities differently from how [RFC6962] divides CA and Certificate Transparency log. The previous section describes the implications of a failure to meet the log-like responsibilities of a CA, provided the transparency service is operating correctly.¶
For the remainder of log-like responsibilities, the relying party trusts its choice of transparency service deployment to ensure the validity windows it uses are consistent with what monitors observe. Otherwise, a malicious transparency service and CA could collude to cause a relying party to accept an unauthorized certificate not visible to monitors. Where a single trusted service is not available, the Section 7 discusses possible deployment structures where the transparency service is a collection of mirrors, all or most of whom must collude instead.¶
Merkle Tree certificates are intended to be used as an optimization over other PKI mechanisms. More generally, certificate negotiation allows relying parties to support many kinds of certificates, to meet different goals. This document discusses the security properties of Merkle Tree certificates, but the overall system's security properties depend on all of a relying party's trust anchors.¶
In particular, in relying parties that require a publicly auditable PKI, the supported fallback mechanisms must also provide a transparency property, either with Certificate Transparency [RFC6962] or another mechanism.¶
IANA is requested to create the following entry in the TLS Certificate Types registry [RFC8447]. The "Reference" column should be set to this document.¶
Value | Name | Recommended |
---|---|---|
TBD |
Bikeshed
|
TBD |
[[ TODO: Define registries for the enums introduced in this document. #42]]¶
The following is an Assertion claiming example.com
for
a TLS subject with Ed25519 public key
c5d2080fa9a489a226b58166dad00be8120931a769c9c6f1f8eefafc38af9065
.¶
00000024 08070020 c5d2080f a9a489a2 26b58166 dad00be8 120931a7 69c9c6f1 f8eefafc 38af9065 00130000 000f000d 000b6578 616d706c 652e636f 6d¶
The corresponding AbridgedAssertion:¶
00000022 0807d8e2 c44fc82e 175e5698 b1c25324 6c9a996f c37bad29 fd59b6aa 838b0a93 0b000013 0000000f 000d000b 6578616d 706c652e 636f6d¶
Next, we have an Assertion claiming *.example.com
, 192.0.2.37
,
192.0.12.0
, 198.51.100.60
and 203.0.113.0
for
a TLS subject with RSASSA-PSS public key with modulus 264851…51544459
and exponent 65537.¶
00000112 0804010e 3082010a 02820101 00d1cd9c d613c050 929e6418 14b4957c 40f30d07 0927f653 bde7054c 06d53a89 36228b70 72fad4db a186c379 7e00300b a5b6de8e 7ab3fed4 cb5a537e 7674916a 130a0435 664428a9 7f1983b7 e028b9ab f24700de 1d6478c9 ae361176 daa64c2f 89b42ec0 270add68 85323401 35d22724 c7bd8f65 075b25b8 96a89ab8 2a2b2194 49b029b8 97e130dc dc96fce1 37351f2b 7a28f1d0 7b710afb 2c796211 d9ba1feb 43d30810 63f19afd b7ba2ab0 e19fd008 e719491d d10ed235 5d4790f0 3039e3a3 31aa2644 2d656716 ebe710f2 4260599a 2d082db1 eccfaa8f f51cfb8e 3dfca0eb e1af59c2 f007b35e 02b0582f 50090018 b78a6b06 c0188ab3 514d60d6 6243e017 8b020301 00010029 0001000f 000d000b 6578616d 706c652e 636f6d00 02001200 10c00002 25c0000c 00c63364 3ccb0071 00¶
The corresponding AbridgedAssertion:¶
00000022 08049a04 087a4d52 033a0a20 04333359 ccf29703 25684c5f a96f1ca1 35cb2ab1 f2670029 0001000f 000d000b 6578616d 706c652e 636f6d00 02001200 10c00002 25c0000c 00c63364 3ccb0071 00¶
[[TODO: We may need a new TLS certificate types extension, either in this document or a separate one. For now, this section just informally describes the problem. #18 ]]¶
The server certificate type is negotiated as follows:¶
The client sends server_certificate_type
in ClientHello with accepted certificate types.¶
The server selects a certificate type to use, It sends it in server_certificate_type
in EncryptedExtensions.¶
The server sends a certificate of the server-selected type in Certificate.¶
This model allows the server to select its certificate type based on not just server_certificate_type
, but also other ClientHello extensions like certificate_authorities
or trust_anchors
([I-D.beck-tls-trust-anchor-ids]). In particular, if there is no match in trust_anchors
, it can fallback to X.509, rather than staying within the realm of BikeshedCertificate.¶
However, the client certificate type is negotiated differently:¶
The client sends client_certificate_type
in ClientHello with certificates it can send¶
The server selects a certificate type to request. It sends it in client_certificate_type
in EncryptedExtensions.¶
The server requests a client certificate in CertificateRequest¶
The client sends a certificate of the server-selected type in Certificate.¶
Here, the client (subscriber) does not select the certificate type. The server (relying party) does. Moreover, this selection is made before the client can see the server's certificate_authorities
or trust_anchors
value, in CertificateRequest. There is no opportunity for the client to fallback to X.509.¶
The cert_types
extension behaves similarly, but additionally forces the client and server types to match. These extensions were defined when TLS 1.2 was current, but TLS 1.3 aligns the client and server certificate negotiation. Most certificate negotiation extensions, such as certificate_authorities
or compress_certificate
[RFC8879] can be offered in either direction, in ClientHello or CertificateRequest. They are then symmetrically accepted in the Certificate message.¶
A more corresponding TLS 1.3 negotiation would be to defer the client certificate type negotiation to CertificateRequest, with the server offering the supported certificate types. The client can then make its selection, taking other CertificateRequest extensions into account, and indicate its selection in the Certificate message.¶
Two possible design sketches:¶
We can have the subscriber indicate the certificate type in an extension of the first CertificateEntry. One challenge is the extensions come after the certificate, so the relying party must seek to the extensions
field independent of the certificate type. Thus all certificate types must be updated to use a consistent opaque cert_data<0..2^24>
syntax, with any type-specific structures embedded inside.¶
RawPublicKey and X509 already meet this requirement. OpenPGP and Bikeshed need an extra length prefix.¶
Alternatively, we can negotiate an extension that changes the syntax to Certificate to:¶
struct { CertificateType certificate_type; opaque certificate_request_context<0..2^8-1>; CertificateEntry certificate_list<0..2^24-1>; } Certificate;¶
The negotiation can be:¶
Client sends its accepted certificate types in ClientHello. Offering this new extension also signatures it is willing to accept the new message format. Unlike the existing extensions, an X.509-only client still sends the extension with just X509 in the list.¶
Server, if it implements the new syntax, acknowledges the syntax change with an empty extension in EncryptedExtensions. (It doesn't indicate its selection yet.)¶
If both of the above happen, Certificate's syntax has changed. Server indicates its selection with the certificate_type
field¶
Server can also send this extension in CertificateRequest to offer non-X.509 certificate types¶
Client likewise indicates its selection with the certificate_type
field.¶
This is a bit cleaner to parse, but the negotiation is more complex.¶
This document stands on the shoulders of giants and builds upon decades of work in TLS authentication and X.509. The authors would like to thank all those who have contributed over the history of these protocols.¶
The authors additionally thank Bob Beck, Ryan Dickson, Nick Harper, Dennis Jackson, Ryan Sleevi, and Emily Stark for many valuable discussions and insights which led to this document. We wish to thank Mia Celeste in particular, whose implementation of an earlier draft revealed several pitfalls.¶
RFC Editor's Note: Please remove this section prior to publication of a final version of this document.¶
Simpify hashing by removing the internal padding to align with block size. #72¶
Avoid the temptation of floating points. #66¶
Require lifetime
to be a multiple of batch_duration
. #65¶
Rename window to validity window. #21¶
Split Assertion into Assertion and AbridgedAssertion. The latter is used in the Merkle Tree and HTTP interface. It replaces subject_info
by a hash, to save space by not serving large post-quantum public keys. The original Assertion is used everywhere else, including BikeshedCertificate. #6¶
Add proper context to every node in the Merkle tree. #32¶
Clarify we use a single CertificateEntry
. #11¶
Clarify we use POSIX time. #1¶
Miscellaneous changes.¶
Replace the negotiation mechanism with TLS Trust Anchor Identifiers.¶