Filip/matrix-spec
1
0
Fork 0
mirror of https://github.com/matrix-org/matrix-spec synced 2026-02-20 13:03:42 +01:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

Move Olm & Megolm to separate subpages

Browse source 
Signed-off-by: Johannes Marbach <n0-0ne+github@mailbox.org>
This commit is contained in:
Johannes Marbach 2025-10-29 15:00:50 +01:00
parent 8f3769af30
commit 7fa012b67f
3 changed files with 717 additions and 704 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

709
content/olm-megolm/_index.md
View file

@ -4,708 +4,9 @@ weight: 61
type: docs
---
## Olm: A Cryptographic Ratchet
The Olm and Megolm cryptographic ratchets are not formally part
of the Matrix specification. In lack of a better place, their
specification is reproduced here, regardless.
An implementation of the double cryptographic ratchet described by
https://whispersystems.org/docs/specifications/doubleratchet/.
### Notation
This document uses \(\parallel\) to represent string concatenation. When
\(\parallel\) appears on the right hand side of an \(=\) it means that
the inputs are concatenated. When \(\parallel\) appears on the left hand
side of an \(=\) it means that the output is split.
When this document uses \(\operatorname{ECDH}\left(K_A,K_B\right)\) it means
that each party computes a Diffie-Hellman agreement using their private key
and the remote party's public key.
So party \(A\) computes \(\operatorname{ECDH}\left(K_B^{public},K_A^{private}\right)\)
and party \(B\) computes \(\operatorname{ECDH}\left(K_A^{public},K_B^{private}\right)\).
Where this document uses \(\operatorname{HKDF}\left(salt,IKM,info,L\right)\) it
refers to the [HMAC-based key derivation function][] with a salt value of
\(salt\), input key material of \(IKM\), context string \(info\),
and output keying material length of \(L\) bytes.
### The Olm Algorithm
#### Initial setup
The setup takes four [Curve25519][] inputs: Identity keys for Alice and Bob,
\(I_A\) and \(I_B\), and one-time keys for Alice and Bob,
\(E_A\) and \(E_B\). A shared secret, \(S\), is generated using
[Triple Diffie-Hellman][]. The initial 256 bit root key, \(R_0\), and 256
bit chain key, \(C_{0,0}\), are derived from the shared secret using an
HMAC-based Key Derivation Function using [SHA-256][] as the hash function
([HKDF-SHA-256][]) with default salt and ``"OLM_ROOT"`` as the info.
\[
\begin{aligned}
S&=\operatorname{ECDH}\left(I_A,E_B\right)\;\parallel\;
\operatorname{ECDH}\left(E_A,I_B\right)\;\parallel\;
\operatorname{ECDH}\left(E_A,E_B\right)\\
R_0\;\parallel\;C_{0,0}&=
\operatorname{HKDF}\left(0,S,\text{``OLM\_ROOT"},64\right)
\end{aligned}
\]
#### Advancing the root key
Advancing a root key takes the previous root key, \(R_{i-1}\), and two
Curve25519 inputs: the previous ratchet key, \(T_{i-1}\), and the current
ratchet key \(T_i\). The even ratchet keys are generated by Alice.
The odd ratchet keys are generated by Bob. A shared secret is generated
using Diffie-Hellman on the ratchet keys. The next root key, \(R_i\), and
chain key, \(C_{i,0}\), are derived from the shared secret using
[HKDF-SHA-256][] using \(R_{i-1}\) as the salt and ``"OLM_RATCHET"`` as the
info.
\[
\begin{aligned}
R_i\;\parallel\;C_{i,0}&=
\operatorname{HKDF}\left(
R_{i-1},
\operatorname{ECDH}\left(T_{i-1},T_i\right),
\text{``OLM\_RATCHET"},
64
\right)
\end{aligned}
\]
#### Advancing the chain key
Advancing a chain key takes the previous chain key, \(C_{i,j-1}\). The next
chain key, \(C_{i,j}\), is the [HMAC-SHA-256][] of ``"\x02"`` using the
previous chain key as the key.
\[
\begin{aligned}
C_{i,j}&=\operatorname{HMAC}\left(C_{i,j-1},\text{``\char`\\x02"}\right)
\end{aligned}
\]
#### Creating a message key
Creating a message key takes the current chain key, \(C_{i,j}\). The
message key, \(M_{i,j}\), is the [HMAC-SHA-256][] of ``"\x01"`` using the
current chain key as the key. The message keys where \(i\) is even are used
by Alice to encrypt messages. The message keys where \(i\) is odd are used
by Bob to encrypt messages.
\[
\begin{aligned}
M_{i,j}&=\operatorname{HMAC}\left(C_{i,j},\text{``\char`\\x01"}\right)
\end{aligned}
\]
### The Olm Protocol
#### Creating an outbound session
Bob publishes the public parts of his identity key, \(I_B\), and some
single-use one-time keys \(E_B\).
Alice downloads Bob's identity key, \(I_B\), and a one-time key,
\(E_B\). She generates a new single-use key, \(E_A\), and computes a
root key, \(R_0\), and a chain key \(C_{0,0}\). She also generates a
new ratchet key \(T_0\).
#### Sending the first pre-key messages
Alice computes a message key, \(M_{0,j}\), and a new chain key,
\(C_{0,j+1}\), using the current chain key. She replaces the current chain
key with the new one.
Alice encrypts her plain-text with the message key, \(M_{0,j}\), using an
authenticated encryption scheme (see below) to get a cipher-text,
\(X_{0,j}\).
She then sends the following to Bob:
* The public part of her identity key, \(I_A\)
* The public part of her single-use key, \(E_A\)
* The public part of Bob's single-use key, \(E_B\)
* The current chain index, \(j\)
* The public part of her ratchet key, \(T_0\)
* The cipher-text, \(X_{0,j}\)
Alice will continue to send pre-key messages until she receives a message from
Bob.
#### Creating an inbound session from a pre-key message
Bob receives a pre-key message as above.
Bob looks up the private part of his single-use key, \(E_B\). He can now
compute the root key, \(R_0\), and the chain key, \(C_{0,0}\), from
\(I_A\), \(E_A\), \(I_B\), and \(E_B\).
Bob then advances the chain key \(j\) times, to compute the chain key used
by the message, \(C_{0,j}\). He now creates the
message key, \(M_{0,j}\), and attempts to decrypt the cipher-text,
\(X_{0,j}\). If the cipher-text's authentication is correct then Bob can
discard the private part of his single-use one-time key, \(E_B\).
Bob stores Alice's initial ratchet key, \(T_0\), until he wants to
send a message.
#### Sending normal messages
Once a message has been received from the other side, a session is considered
established, and a more compact form is used.
To send a message, the user checks if they have a sender chain key,
\(C_{i,j}\). Alice uses chain keys where \(i\) is even. Bob uses chain
keys where \(i\) is odd. If the chain key doesn't exist then a new ratchet
key \(T_i\) is generated and a new root key \(R_i\) and chain key
\(C_{i,0}\) are computed using \(R_{i-1}\), \(T_{i-1}\) and
\(T_i\).
A message key,
\(M_{i,j}\) is computed from the current chain key, \(C_{i,j}\), and
the chain key is replaced with the next chain key, \(C_{i,j+1}\). The
plain-text is encrypted with \(M_{i,j}\), using an authenticated encryption
scheme (see below) to get a cipher-text, \(X_{i,j}\).
The user then sends the following to the recipient:
* The current chain index, \(j\)
* The public part of the current ratchet key, \(T_i\)
* The cipher-text, \(X_{i,j}\)
#### Receiving messages
The user receives a message as above with the sender's current chain index, \(j\),
the sender's ratchet key, \(T_i\), and the cipher-text, \(X_{i,j}\).
The user checks if they have a receiver chain with the correct
\(i\) by comparing the ratchet key, \(T_i\). If the chain doesn't exist
then they compute a new root key, \(R_i\), and a new receiver chain, with
chain key \(C_{i,0}\), using \(R_{i-1}\), \(T_{i-1}\) and
\(T_i\).
If the \(j\) of the message is less than
the current chain index on the receiver then the message may only be decrypted
if the receiver has stored a copy of the message key \(M_{i,j}\). Otherwise
the receiver computes the chain key, \(C_{i,j}\). The receiver computes the
message key, \(M_{i,j}\), from the chain key and attempts to decrypt the
cipher-text, \(X_{i,j}\).
If the decryption succeeds the receiver updates the chain key for \(T_i\)
with \(C_{i,j+1}\) and stores the message keys that were skipped in the
process so that they can decode out of order messages. If the receiver created
a new receiver chain then they discard their current sender chain so that
they will create a new chain when they next send a message.
### The Olm Message Format
Olm uses two types of messages. The underlying transport protocol must provide
a means for recipients to distinguish between them.
#### Normal Messages
Olm messages start with a one byte version followed by a variable length
payload followed by a fixed length message authentication code.
```nohighlight
+--------------+------------------------------------+-----------+
| Version Byte | Payload Bytes | MAC Bytes |
+--------------+------------------------------------+-----------+
```
The version byte is ``"\x03"``.
The payload consists of key-value pairs where the keys are integers and the
values are integers and strings. The keys are encoded as a variable length
integer tag where the 3 lowest bits indicates the type of the value:
0 for integers, 2 for strings. If the value is an integer then the tag is
followed by the value encoded as a variable length integer. If the value is
a string then the tag is followed by the length of the string encoded as
a variable length integer followed by the string itself.
Olm uses a variable length encoding for integers. Each integer is encoded as a
sequence of bytes with the high bit set followed by a byte with the high bit
clear. The seven low bits of each byte store the bits of the integer. The least
significant bits are stored in the first byte.
**Name**|**Tag**|**Type**|**Meaning**
:-----:|:-----:|:-----:|:-----:
Ratchet-Key|0x0A|String|The public part of the ratchet key, Ti, of the message
Chain-Index|0x10|Integer|The chain index, j, of the message
Cipher-Text|0x22|String|The cipher-text, Xi, j, of the message
The length of the MAC is determined by the authenticated encryption algorithm
being used. (Olm version 1 uses [HMAC-SHA-256][], truncated to 8 bytes). The
MAC protects all of the bytes preceding the MAC.
#### Pre-Key Messages
Olm pre-key messages start with a one byte version followed by a variable
length payload.
```nohighlight
+--------------+------------------------------------+
| Version Byte | Payload Bytes |
+--------------+------------------------------------+
```
The version byte is ``"\x03"``.
The payload uses the same key-value format as for normal messages.
**Name**|**Tag**|**Type**|**Meaning**
:-----:|:-----:|:-----:|:-----:
One-Time-Key|0x0A|String|The public part of Bob's single-use key, Eb.
Base-Key|0x12|String|The public part of Alice's single-use key, Ea.
Identity-Key|0x1A|String|The public part of Alice's identity key, Ia.
Message|0x22|String|An embedded Olm message with its own version and MAC.
### Olm Authenticated Encryption
#### Version 1
Version 1 of Olm uses [AES-256][] in [CBC][] mode with [PKCS#7][] padding for
encryption and [HMAC-SHA-256][] (truncated to 64 bits) for authentication. The
256 bit AES key, 256 bit HMAC key, and 128 bit AES IV are derived from the
message key using [HKDF-SHA-256][] using the default salt and an info of
``"OLM_KEYS"``.
\[
\begin{aligned}
AES\_KEY_{i,j}\;\parallel\;HMAC\_KEY_{i,j}\;\parallel\;AES\_IV_{i,j}
&= \operatorname{HKDF}\left(0,M_{i,j},\text{``OLM\_KEYS"},80\right)
\end{aligned}
\]
The plain-text is encrypted with AES-256, using the key \(AES\_KEY_{i,j}\)
and the IV \(AES\_IV_{i,j}\) to give the cipher-text, \(X_{i,j}\).
Then the entire message (including the Version Byte and all Payload Bytes) are
passed through [HMAC-SHA-256][]. The first 8 bytes of the MAC are appended to the message.
### Message authentication concerns
To avoid unknown key-share attacks, the application must include identifying
data for the sending and receiving user in the plain-text of (at least) the
pre-key messages. Such data could be a user ID, a telephone number;
alternatively it could be the public part of a keypair which the relevant user
has proven ownership of.
#### Example attacks
1. Alice publishes her public [Curve25519][] identity key, \(I_A\). Eve
publishes the same identity key, claiming it as her own. Bob downloads
Eve's keys, and associates \(I_A\) with Eve. Alice sends a message to
Bob; Eve intercepts it before forwarding it to Bob. Bob believes the
message came from Eve rather than Alice.
This is prevented if Alice includes her user ID in the plain-text of the
pre-key message, so that Bob can see that the message was sent by Alice
originally.
2. Bob publishes his public [Curve25519][] identity key, \(I_B\). Eve
publishes the same identity key, claiming it as her own. Alice downloads
Eve's keys, and associates \(I_B\) with Eve. Alice sends a message to
Eve; Eve cannot decrypt it, but forwards it to Bob. Bob believes the
Alice sent the message to him, whereas Alice intended it to go to Eve.
This is prevented by Alice including the user ID of the intended recpient
(Eve) in the plain-text of the pre-key message. Bob can now tell that the
message was meant for Eve rather than him.
### IPR
The Olm specification (this document) is hereby placed in the public domain.
### Feedback
Can be sent to olm at matrix.org.
### Acknowledgements
The ratchet that Olm implements was designed by Trevor Perrin and Moxie
Marlinspike - details at https://whispersystems.org/docs/specifications/doubleratchet/. Olm is
an entirely new implementation written by the Matrix.org team.
[Curve25519]: http://cr.yp.to/ecdh.html
[Triple Diffie-Hellman]: https://whispersystems.org/blog/simplifying-otr-deniability/
[HMAC-based key derivation function]: https://tools.ietf.org/html/rfc5869
[HKDF-SHA-256]: https://tools.ietf.org/html/rfc5869
[HMAC-SHA-256]: https://tools.ietf.org/html/rfc2104
[SHA-256]: https://tools.ietf.org/html/rfc6234
[AES-256]: http://csrc.nist.gov/publications/fips/fips197/fips-197.pdf
[CBC]: http://csrc.nist.gov/publications/nistpubs/800-38a/sp800-38a.pdf
[PKCS#7]: https://tools.ietf.org/html/rfc2315
## Megolm group ratchet
An AES-based cryptographic ratchet intended for group communications.
### Background
The Megolm ratchet is intended for encrypted messaging applications where there
may be a large number of recipients of each message, thus precluding the use of
peer-to-peer encryption systems such as [Olm][].
It also allows a recipient to decrypt received messages multiple times. For
instance, in client/server applications, a copy of the ciphertext can be stored
on the (untrusted) server, while the client need only store the session keys.
### Overview
Each participant in a conversation uses their own outbound session for
encrypting messages. A session consists of a ratchet and an [Ed25519][] keypair.
Secrecy is provided by the ratchet, which can be wound forwards but not
backwards, and is used to derive a distinct message key for each message.
Authenticity is provided via Ed25519 signatures.
The value of the ratchet, and the public part of the Ed25519 key, are shared
with other participants in the conversation via secure peer-to-peer
channels. Provided that peer-to-peer channel provides authenticity of the
messages to the participants and deniability of the messages to third parties,
the Megolm session will inherit those properties.
### The Megolm ratchet algorithm
The Megolm ratchet \(R_i\) consists of four parts, \(R_{i,j}\) for
\(j \in {0,1,2,3}\). The length of each part depends on the hash function
in use (256 bits for this version of Megolm).
The ratchet is initialised with cryptographically-secure random data, and
advanced as follows:
\[
\begin{aligned}
R_{i,0} &=
\begin{cases}
H_0\left(R_{2^{24}(n-1),0}\right) &\text{if }\exists n | i = 2^{24}n\\
R_{i-1,0} &\text{otherwise}
\end{cases}\\
R_{i,1} &=
\begin{cases}
H_1\left(R_{2^{24}(n-1),0}\right) &\text{if }\exists n | i = 2^{24}n\\
H_1\left(R_{2^{16}(m-1),1}\right) &\text{if }\exists m | i = 2^{16}m\\
R_{i-1,1} &\text{otherwise}
\end{cases}\\
R_{i,2} &=
\begin{cases}
H_2\left(R_{2^{24}(n-1),0}\right) &\text{if }\exists n | i = 2^{24}n\\
H_2\left(R_{2^{16}(m-1),1}\right) &\text{if }\exists m | i = 2^{16}m\\
H_2\left(R_{2^8(p-1),2}\right) &\text{if }\exists p | i = 2^8p\\
R_{i-1,2} &\text{otherwise}
\end{cases}\\
R_{i,3} &=
\begin{cases}
H_3\left(R_{2^{24}(n-1),0}\right) &\text{if }\exists n | i = 2^{24}n\\
H_3\left(R_{2^{16}(m-1),1}\right) &\text{if }\exists m | i = 2^{16}m\\
H_3\left(R_{2^8(p-1),2}\right) &\text{if }\exists p | i = 2^8p\\
H_3\left(R_{i-1,3}\right) &\text{otherwise}
\end{cases}
\end{aligned}
\]
where \(H_0\), \(H_1\), \(H_2\), and \(H_3\) are different hash
functions. In summary: every \(2^8\) iterations, \(R_{i,3}\) is
reseeded from \(R_{i,2}\). Every \(2^{16}\) iterations, \(R_{i,2}\)
and \(R_{i,3}\) are reseeded from \(R_{i,1}\). Every \(2^{24}\)
iterations, \(R_{i,1}\), \(R_{i,2}\) and \(R_{i,3}\) are reseeded
from \(R_{i,0}\).
The complete ratchet value, \(R_{i}\), is hashed to generate the keys used
to encrypt each message. This scheme allows the ratchet to be advanced an
arbitrary amount forwards while needing at most 1020 hash computations. A
client can decrypt chat history onwards from the earliest value of the ratchet
it is aware of, but cannot decrypt history from before that point without
reversing the hash function.
This allows a participant to share its ability to decrypt chat history with
another from a point in the conversation onwards by giving a copy of the
ratchet at that point in the conversation.
### The Megolm protocol
#### Session setup
Each participant in a conversation generates their own Megolm session. A
session consists of three parts:
* a 32 bit counter, \(i\).
* an [Ed25519][] keypair, \(K\).
* a ratchet, \(R_i\), which consists of four 256-bit values,
\(R_{i,j}\) for \(j \in {0,1,2,3}\).
The counter \(i\) is initialised to \(0\). A new Ed25519 keypair is
generated for \(K\). The ratchet is simply initialised with 1024 bits of
cryptographically-secure random data.
A single participant may use multiple sessions over the lifetime of a
conversation. The public part of \(K\) is used as an identifier to
discriminate between sessions.
#### Sharing session data
To allow other participants in the conversation to decrypt messages, the
session data is formatted as described in [Session-sharing format](#session-sharing-format). It is then
shared with other participants in the conversation via a secure peer-to-peer
channel (such as that provided by [Olm][]).
When the session data is received from other participants, the recipient first
checks that the signature matches the public key. They then store their own
copy of the counter, ratchet, and public key.
#### Message encryption
This version of Megolm uses [AES-256][] in [CBC][] mode with [PKCS#7][] padding and
[HMAC-SHA-256][] (truncated to 64 bits). The 256 bit AES key, 256 bit HMAC key,
and 128 bit AES IV are derived from the megolm ratchet \(R_i\):
\[
\begin{aligned}
\mathit{AES\_KEY}_{i}\;\parallel\;\mathit{HMAC\_KEY}_{i}\;\parallel\;\mathit{AES\_IV}_{i}
&= \operatorname{HKDF}\left(0,\,R_{i},\text{"MEGOLM\_KEYS"},\,80\right) \\
\end{aligned}
\]
where \(\parallel\) represents string splitting, and
\(\operatorname{HKDF}\left(\mathit{salt},\,\mathit{IKM},\,\mathit{info},\,L\right)\)
refers to the [HMAC-based key
derivation function][] using using [SHA-256][] as the hash function
([HKDF-SHA-256][]) with a salt value of \(\mathit{salt}\), input key material of
\(\mathit{IKM}\), context string \(\mathit{info}\), and output keying material length of
\(L\) bytes.
The plain-text is encrypted with AES-256, using the key \(\mathit{AES\_KEY}_{i}\)
and the IV \(\mathit{AES\_IV}_{i}\) to give the cipher-text, \(X_{i}\).
The ratchet index \(i\), and the cipher-text \(X_{i}\), are then packed
into a message as described in [Message format](#message-format). Then the entire message
(including the version bytes and all payload bytes) are passed through
HMAC-SHA-256. The first 8 bytes of the MAC are appended to the message.
Finally, the authenticated message is signed using the Ed25519 keypair; the 64
byte signature is appended to the message.
The complete signed message, together with the public part of \(K\) (acting
as a session identifier), can then be sent over an insecure channel. The
message can then be authenticated and decrypted only by recipients who have
received the session data.
#### Advancing the ratchet
After each message is encrypted, the ratchet is advanced. This is done as
described in [The Megolm ratchet algorithm](#the-megolm-ratchet-algorithm), using the following definitions:
\[
\begin{aligned}
H_0(A) &\equiv \operatorname{HMAC}(A,\text{``\char`\\x00"}) \\
H_1(A) &\equiv \operatorname{HMAC}(A,\text{``\char`\\x01"}) \\
H_2(A) &\equiv \operatorname{HMAC}(A,\text{``\char`\\x02"}) \\
H_3(A) &\equiv \operatorname{HMAC}(A,\text{``\char`\\x03"}) \\
\end{aligned}
\]
where \(\operatorname{HMAC}(A, T)\) is the HMAC-SHA-256 of ``T``, using ``A`` as the
key.
For outbound sessions, the updated ratchet and counter are stored in the
session.
In order to maintain the ability to decrypt conversation history, inbound
sessions should store a copy of their earliest known ratchet value (unless they
explicitly want to drop the ability to decrypt that history - see [Partial
Forward Secrecy](#partial-forward-secrecy)). They may also choose to cache calculated ratchet values,
but the decision of which ratchet states to cache is left to the application.
### Data exchange formats
#### Session sharing format
This format is used for the initial sharing of a Megolm session with other
group participants who need to be able to read messages encrypted by this
session.
The session sharing format is as follows:
```nohighlight
+---+----+--------+--------+--------+--------+------+-----------+
| V | i | R(i,0) | R(i,1) | R(i,2) | R(i,3) | Kpub | Signature |
+---+----+--------+--------+--------+--------+------+-----------+
0 1 5 37 69 101 133 165 229 bytes
```
The version byte, ``V``, is ``"\x02"``.
This is followed by the ratchet index, \(i\), which is encoded as a
big-endian 32-bit integer; the ratchet values \(R_{i,j}\); and the public
part of the Ed25519 keypair \(K\).
The data is then signed using the Ed25519 keypair, and the 64-byte signature is
appended.
#### Session export format
Once the session is initially shared with the group participants, each
participant needs to retain a copy of the session if they want to maintain
their ability to decrypt messages encrypted with that session.
For forward-secrecy purposes, a participant may choose to store a ratcheted
version of the session. But since the ratchet index is covered by the
signature, this would invalidate the signature. So we define a similar format,
called the *session export format*, which is identical to the [session sharing
format](#session-sharing-format) except for dropping the signature.
The Megolm session export format is thus as follows:
```nohighlight
+---+----+--------+--------+--------+--------+------+
| V | i | R(i,0) | R(i,1) | R(i,2) | R(i,3) | Kpub |
+---+----+--------+--------+--------+--------+------+
0 1 5 37 69 101 133 165 bytes
```
The version byte, ``V``, is ``"\x01"``.
This is followed by the ratchet index, \(i\), which is encoded as a
big-endian 32-bit integer; the ratchet values \(R_{i,j}\); and the public
part of the Ed25519 keypair \(K\).
#### Message format
Megolm messages consist of a one byte version, followed by a variable length
payload, a fixed length message authentication code, and a fixed length
signature.
```nohighlight
+---+------------------------------------+-----------+------------------+
| V | Payload Bytes | MAC Bytes | Signature Bytes |
+---+------------------------------------+-----------+------------------+
0 1 N N+8 N+72 bytes
```
The version byte, ``V``, is ``"\x03"``.
The payload uses a format based on the [Protocol Buffers encoding][]. It
consists of the following key-value pairs:
**Name**|**Tag**|**Type**|**Meaning**
:-----:|:-----:|:-----:|:-----:
Message-Index|0x08|Integer|The index of the ratchet, i
Cipher-Text|0x12|String|The cipher-text, Xi, of the message
Within the payload, integers are encoded using a variable length encoding. Each
integer is encoded as a sequence of bytes with the high bit set followed by a
byte with the high bit clear. The seven low bits of each byte store the bits of
the integer. The least significant bits are stored in the first byte.
Strings are encoded as a variable-length integer followed by the string itself.
Each key-value pair is encoded as a variable-length integer giving the tag,
followed by a string or variable-length integer giving the value.
The payload is followed by the MAC. The length of the MAC is determined by the
authenticated encryption algorithm being used (8 bytes in this version of the
protocol). The MAC protects all of the bytes preceding the MAC.
The length of the signature is determined by the signing algorithm being used
(64 bytes in this version of the protocol). The signature covers all of the
bytes preceding the signature.
### Limitations
#### Message Replays
A message can be decrypted successfully multiple times. This means that an
attacker can re-send a copy of an old message, and the recipient will treat it
as a new message.
To mitigate this it is recommended that applications track the ratchet indices
they have received and that they reject messages with a ratchet index that
they have already decrypted.
#### Lack of Transcript Consistency
In a group conversation, there is no guarantee that all recipients have
received the same messages. For example, if Alice is in a conversation with Bob
and Charlie, she could send different messages to Bob and Charlie, or could
send some messages to Bob but not Charlie, or vice versa.
Solving this is, in general, a hard problem, particularly in a protocol which
does not guarantee in-order message delivery. For now it remains the subject of
future research.
#### Lack of Backward Secrecy
[Backward secrecy](https://intensecrypto.org/public/lec_08_hash_functions_part2.html#sec-forward-and-backward-secrecy)
(also called 'future secrecy' or 'post-compromise security') is the property
that if current private keys are compromised, an attacker cannot decrypt
future messages in a given session. In other words, when looking
**backwards** in time at a compromise which has already happened, **current**
messages are still secret.
By itself, Megolm does not possess this property: once the key to a Megolm
session is compromised, the attacker can decrypt any message that was
encrypted using a key derived from the compromised or subsequent ratchet
values.
In order to mitigate this, the application should ensure that Megolm sessions
are not used indefinitely. Instead it should periodically start a new session,
with new keys shared over a secure channel.
<!-- TODO: Can we recommend sensible lifetimes for Megolm sessions? Probably
depends how paranoid we're feeling, but some guidelines might be useful. -->
#### Partial Forward Secrecy
[Forward secrecy](https://intensecrypto.org/public/lec_08_hash_functions_part2.html#sec-forward-and-backward-secrecy)
(also called 'perfect forward secrecy') is the property that if the current
private keys are compromised, an attacker cannot decrypt *past* messages in
a given session. In other words, when looking **forwards** in time towards a
potential future compromise, **current** messages will be secret.
In Megolm, each recipient maintains a record of the ratchet value which allows
them to decrypt any messages sent in the session after the corresponding point
in the conversation. If this value is compromised, an attacker can similarly
decrypt past messages which were encrypted by a key derived from the
compromised or subsequent ratchet values. This gives 'partial' forward
secrecy.
To mitigate this issue, the application should offer the user the option to
discard historical conversations, by winding forward any stored ratchet values,
or discarding sessions altogether.
#### Dependency on secure channel for key exchange
The design of the Megolm ratchet relies on the availability of a secure
peer-to-peer channel for the exchange of session keys. Any vulnerabilities in
the underlying channel are likely to be amplified when applied to Megolm
session setup.
For example, if the peer-to-peer channel is vulnerable to an unknown key-share
attack, the entire Megolm session become similarly vulnerable. For example:
Alice starts a group chat with Eve, and shares the session keys with Eve. Eve
uses the unknown key-share attack to forward the session keys to Bob, who
believes Alice is starting the session with him. Eve then forwards messages
from the Megolm session to Bob, who again believes they are coming from
Alice. Provided the peer-to-peer channel is not vulnerable to this attack, Bob
will realise that the key-sharing message was forwarded by Eve, and can treat
the Megolm session as a forgery.
A second example: if the peer-to-peer channel is vulnerable to a replay
attack, this can be extended to entire Megolm sessions.
### License
The Megolm specification (this document) is licensed under the Apache License,
Version 2.0 http://www.apache.org/licenses/LICENSE-2.0.
[Ed25519]: http://ed25519.cr.yp.to/
[HMAC-based key derivation function]: https://tools.ietf.org/html/rfc5869
[HKDF-SHA-256]: https://tools.ietf.org/html/rfc5869
[HMAC-SHA-256]: https://tools.ietf.org/html/rfc2104
[SHA-256]: https://tools.ietf.org/html/rfc6234
[AES-256]: http://csrc.nist.gov/publications/fips/fips197/fips-197.pdf
[CBC]: http://csrc.nist.gov/publications/nistpubs/800-38a/sp800-38a.pdf
[PKCS#7]: https://tools.ietf.org/html/rfc2315
[Olm]: https://gitlab.matrix.org/matrix-org/olm/blob/master/docs/olm.md
[Protocol Buffers encoding]: https://developers.google.com/protocol-buffers/docs/encoding
- [Olm: A Cryptographic Ratchet](/olm-megolm/olm/)
- [Megolm group ratchet](/olm-megolm/megolm/)

378
content/olm-megolm/megolm.md Normal file
View file

@ -0,0 +1,378 @@
---
title: "Megolm group ratchet"
weight: 20
type: docs
---
An AES-based cryptographic ratchet intended for group communications.
## Background
The Megolm ratchet is intended for encrypted messaging applications where there
may be a large number of recipients of each message, thus precluding the use of
peer-to-peer encryption systems such as [Olm][].
It also allows a recipient to decrypt received messages multiple times. For
instance, in client/server applications, a copy of the ciphertext can be stored
on the (untrusted) server, while the client need only store the session keys.
## Overview
Each participant in a conversation uses their own outbound session for
encrypting messages. A session consists of a ratchet and an [Ed25519][] keypair.
Secrecy is provided by the ratchet, which can be wound forwards but not
backwards, and is used to derive a distinct message key for each message.
Authenticity is provided via Ed25519 signatures.
The value of the ratchet, and the public part of the Ed25519 key, are shared
with other participants in the conversation via secure peer-to-peer
channels. Provided that peer-to-peer channel provides authenticity of the
messages to the participants and deniability of the messages to third parties,
the Megolm session will inherit those properties.
## The Megolm ratchet algorithm
The Megolm ratchet \(R_i\) consists of four parts, \(R_{i,j}\) for
\(j \in {0,1,2,3}\). The length of each part depends on the hash function
in use (256 bits for this version of Megolm).
The ratchet is initialised with cryptographically-secure random data, and
advanced as follows:
\[
\begin{aligned}
R_{i,0} &=
\begin{cases}
H_0\left(R_{2^{24}(n-1),0}\right) &\text{if }\exists n | i = 2^{24}n\\
R_{i-1,0} &\text{otherwise}
\end{cases}\\
R_{i,1} &=
\begin{cases}
H_1\left(R_{2^{24}(n-1),0}\right) &\text{if }\exists n | i = 2^{24}n\\
H_1\left(R_{2^{16}(m-1),1}\right) &\text{if }\exists m | i = 2^{16}m\\
R_{i-1,1} &\text{otherwise}
\end{cases}\\
R_{i,2} &=
\begin{cases}
H_2\left(R_{2^{24}(n-1),0}\right) &\text{if }\exists n | i = 2^{24}n\\
H_2\left(R_{2^{16}(m-1),1}\right) &\text{if }\exists m | i = 2^{16}m\\
H_2\left(R_{2^8(p-1),2}\right) &\text{if }\exists p | i = 2^8p\\
R_{i-1,2} &\text{otherwise}
\end{cases}\\
R_{i,3} &=
\begin{cases}
H_3\left(R_{2^{24}(n-1),0}\right) &\text{if }\exists n | i = 2^{24}n\\
H_3\left(R_{2^{16}(m-1),1}\right) &\text{if }\exists m | i = 2^{16}m\\
H_3\left(R_{2^8(p-1),2}\right) &\text{if }\exists p | i = 2^8p\\
H_3\left(R_{i-1,3}\right) &\text{otherwise}
\end{cases}
\end{aligned}
\]
where \(H_0\), \(H_1\), \(H_2\), and \(H_3\) are different hash
functions. In summary: every \(2^8\) iterations, \(R_{i,3}\) is
reseeded from \(R_{i,2}\). Every \(2^{16}\) iterations, \(R_{i,2}\)
and \(R_{i,3}\) are reseeded from \(R_{i,1}\). Every \(2^{24}\)
iterations, \(R_{i,1}\), \(R_{i,2}\) and \(R_{i,3}\) are reseeded
from \(R_{i,0}\).
The complete ratchet value, \(R_{i}\), is hashed to generate the keys used
to encrypt each message. This scheme allows the ratchet to be advanced an
arbitrary amount forwards while needing at most 1020 hash computations. A
client can decrypt chat history onwards from the earliest value of the ratchet
it is aware of, but cannot decrypt history from before that point without
reversing the hash function.
This allows a participant to share its ability to decrypt chat history with
another from a point in the conversation onwards by giving a copy of the
ratchet at that point in the conversation.
## The Megolm protocol
### Session setup
Each participant in a conversation generates their own Megolm session. A
session consists of three parts:
* a 32 bit counter, \(i\).
* an [Ed25519][] keypair, \(K\).
* a ratchet, \(R_i\), which consists of four 256-bit values,
\(R_{i,j}\) for \(j \in {0,1,2,3}\).
The counter \(i\) is initialised to \(0\). A new Ed25519 keypair is
generated for \(K\). The ratchet is simply initialised with 1024 bits of
cryptographically-secure random data.
A single participant may use multiple sessions over the lifetime of a
conversation. The public part of \(K\) is used as an identifier to
discriminate between sessions.
### Sharing session data
To allow other participants in the conversation to decrypt messages, the
session data is formatted as described in [Session-sharing format](#session-sharing-format). It is then
shared with other participants in the conversation via a secure peer-to-peer
channel (such as that provided by [Olm][]).
When the session data is received from other participants, the recipient first
checks that the signature matches the public key. They then store their own
copy of the counter, ratchet, and public key.
### Message encryption
This version of Megolm uses [AES-256][] in [CBC][] mode with [PKCS#7][] padding and
[HMAC-SHA-256][] (truncated to 64 bits). The 256 bit AES key, 256 bit HMAC key,
and 128 bit AES IV are derived from the megolm ratchet \(R_i\):
\[
\begin{aligned}
\mathit{AES\_KEY}_{i}\;\parallel\;\mathit{HMAC\_KEY}_{i}\;\parallel\;\mathit{AES\_IV}_{i}
&= \operatorname{HKDF}\left(0,\,R_{i},\text{"MEGOLM\_KEYS"},\,80\right) \\
\end{aligned}
\]
where \(\parallel\) represents string splitting, and
\(\operatorname{HKDF}\left(\mathit{salt},\,\mathit{IKM},\,\mathit{info},\,L\right)\)
refers to the [HMAC-based key
derivation function][] using using [SHA-256][] as the hash function
([HKDF-SHA-256][]) with a salt value of \(\mathit{salt}\), input key material of
\(\mathit{IKM}\), context string \(\mathit{info}\), and output keying material length of
\(L\) bytes.
The plain-text is encrypted with AES-256, using the key \(\mathit{AES\_KEY}_{i}\)
and the IV \(\mathit{AES\_IV}_{i}\) to give the cipher-text, \(X_{i}\).
The ratchet index \(i\), and the cipher-text \(X_{i}\), are then packed
into a message as described in [Message format](#message-format). Then the entire message
(including the version bytes and all payload bytes) are passed through
HMAC-SHA-256. The first 8 bytes of the MAC are appended to the message.
Finally, the authenticated message is signed using the Ed25519 keypair; the 64
byte signature is appended to the message.
The complete signed message, together with the public part of \(K\) (acting
as a session identifier), can then be sent over an insecure channel. The
message can then be authenticated and decrypted only by recipients who have
received the session data.
### Advancing the ratchet
After each message is encrypted, the ratchet is advanced. This is done as
described in [The Megolm ratchet algorithm](#the-megolm-ratchet-algorithm), using the following definitions:
\[
\begin{aligned}
H_0(A) &\equiv \operatorname{HMAC}(A,\text{``\char`\\x00"}) \\
H_1(A) &\equiv \operatorname{HMAC}(A,\text{``\char`\\x01"}) \\
H_2(A) &\equiv \operatorname{HMAC}(A,\text{``\char`\\x02"}) \\
H_3(A) &\equiv \operatorname{HMAC}(A,\text{``\char`\\x03"}) \\
\end{aligned}
\]
where \(\operatorname{HMAC}(A, T)\) is the HMAC-SHA-256 of ``T``, using ``A`` as the
key.
For outbound sessions, the updated ratchet and counter are stored in the
session.
In order to maintain the ability to decrypt conversation history, inbound
sessions should store a copy of their earliest known ratchet value (unless they
explicitly want to drop the ability to decrypt that history - see [Partial
Forward Secrecy](#partial-forward-secrecy)). They may also choose to cache calculated ratchet values,
but the decision of which ratchet states to cache is left to the application.
## Data exchange formats
### Session sharing format
This format is used for the initial sharing of a Megolm session with other
group participants who need to be able to read messages encrypted by this
session.
The session sharing format is as follows:
```nohighlight
+---+----+--------+--------+--------+--------+------+-----------+
| V | i | R(i,0) | R(i,1) | R(i,2) | R(i,3) | Kpub | Signature |
+---+----+--------+--------+--------+--------+------+-----------+
0 1 5 37 69 101 133 165 229 bytes
```
The version byte, ``V``, is ``"\x02"``.
This is followed by the ratchet index, \(i\), which is encoded as a
big-endian 32-bit integer; the ratchet values \(R_{i,j}\); and the public
part of the Ed25519 keypair \(K\).
The data is then signed using the Ed25519 keypair, and the 64-byte signature is
appended.
### Session export format
Once the session is initially shared with the group participants, each
participant needs to retain a copy of the session if they want to maintain
their ability to decrypt messages encrypted with that session.
For forward-secrecy purposes, a participant may choose to store a ratcheted
version of the session. But since the ratchet index is covered by the
signature, this would invalidate the signature. So we define a similar format,
called the *session export format*, which is identical to the [session sharing
format](#session-sharing-format) except for dropping the signature.
The Megolm session export format is thus as follows:
```nohighlight
+---+----+--------+--------+--------+--------+------+
| V | i | R(i,0) | R(i,1) | R(i,2) | R(i,3) | Kpub |
+---+----+--------+--------+--------+--------+------+
0 1 5 37 69 101 133 165 bytes
```
The version byte, ``V``, is ``"\x01"``.
This is followed by the ratchet index, \(i\), which is encoded as a
big-endian 32-bit integer; the ratchet values \(R_{i,j}\); and the public
part of the Ed25519 keypair \(K\).
### Message format
Megolm messages consist of a one byte version, followed by a variable length
payload, a fixed length message authentication code, and a fixed length
signature.
```nohighlight
+---+------------------------------------+-----------+------------------+
| V | Payload Bytes | MAC Bytes | Signature Bytes |
+---+------------------------------------+-----------+------------------+
0 1 N N+8 N+72 bytes
```
The version byte, ``V``, is ``"\x03"``.
The payload uses a format based on the [Protocol Buffers encoding][]. It
consists of the following key-value pairs:
**Name**|**Tag**|**Type**|**Meaning**
:-----:|:-----:|:-----:|:-----:
Message-Index|0x08|Integer|The index of the ratchet, i
Cipher-Text|0x12|String|The cipher-text, Xi, of the message
Within the payload, integers are encoded using a variable length encoding. Each
integer is encoded as a sequence of bytes with the high bit set followed by a
byte with the high bit clear. The seven low bits of each byte store the bits of
the integer. The least significant bits are stored in the first byte.
Strings are encoded as a variable-length integer followed by the string itself.
Each key-value pair is encoded as a variable-length integer giving the tag,
followed by a string or variable-length integer giving the value.
The payload is followed by the MAC. The length of the MAC is determined by the
authenticated encryption algorithm being used (8 bytes in this version of the
protocol). The MAC protects all of the bytes preceding the MAC.
The length of the signature is determined by the signing algorithm being used
(64 bytes in this version of the protocol). The signature covers all of the
bytes preceding the signature.
## Limitations
### Message Replays
A message can be decrypted successfully multiple times. This means that an
attacker can re-send a copy of an old message, and the recipient will treat it
as a new message.
To mitigate this it is recommended that applications track the ratchet indices
they have received and that they reject messages with a ratchet index that
they have already decrypted.
### Lack of Transcript Consistency
In a group conversation, there is no guarantee that all recipients have
received the same messages. For example, if Alice is in a conversation with Bob
and Charlie, she could send different messages to Bob and Charlie, or could
send some messages to Bob but not Charlie, or vice versa.
Solving this is, in general, a hard problem, particularly in a protocol which
does not guarantee in-order message delivery. For now it remains the subject of
future research.
### Lack of Backward Secrecy
[Backward secrecy](https://intensecrypto.org/public/lec_08_hash_functions_part2.html#sec-forward-and-backward-secrecy)
(also called 'future secrecy' or 'post-compromise security') is the property
that if current private keys are compromised, an attacker cannot decrypt
future messages in a given session. In other words, when looking
**backwards** in time at a compromise which has already happened, **current**
messages are still secret.
By itself, Megolm does not possess this property: once the key to a Megolm
session is compromised, the attacker can decrypt any message that was
encrypted using a key derived from the compromised or subsequent ratchet
values.
In order to mitigate this, the application should ensure that Megolm sessions
are not used indefinitely. Instead it should periodically start a new session,
with new keys shared over a secure channel.
<!-- TODO: Can we recommend sensible lifetimes for Megolm sessions? Probably
depends how paranoid we're feeling, but some guidelines might be useful. -->
### Partial Forward Secrecy
[Forward secrecy](https://intensecrypto.org/public/lec_08_hash_functions_part2.html#sec-forward-and-backward-secrecy)
(also called 'perfect forward secrecy') is the property that if the current
private keys are compromised, an attacker cannot decrypt *past* messages in
a given session. In other words, when looking **forwards** in time towards a
potential future compromise, **current** messages will be secret.
In Megolm, each recipient maintains a record of the ratchet value which allows
them to decrypt any messages sent in the session after the corresponding point
in the conversation. If this value is compromised, an attacker can similarly
decrypt past messages which were encrypted by a key derived from the
compromised or subsequent ratchet values. This gives 'partial' forward
secrecy.
To mitigate this issue, the application should offer the user the option to
discard historical conversations, by winding forward any stored ratchet values,
or discarding sessions altogether.
### Dependency on secure channel for key exchange
The design of the Megolm ratchet relies on the availability of a secure
peer-to-peer channel for the exchange of session keys. Any vulnerabilities in
the underlying channel are likely to be amplified when applied to Megolm
session setup.
For example, if the peer-to-peer channel is vulnerable to an unknown key-share
attack, the entire Megolm session become similarly vulnerable. For example:
Alice starts a group chat with Eve, and shares the session keys with Eve. Eve
uses the unknown key-share attack to forward the session keys to Bob, who
believes Alice is starting the session with him. Eve then forwards messages
from the Megolm session to Bob, who again believes they are coming from
Alice. Provided the peer-to-peer channel is not vulnerable to this attack, Bob
will realise that the key-sharing message was forwarded by Eve, and can treat
the Megolm session as a forgery.
A second example: if the peer-to-peer channel is vulnerable to a replay
attack, this can be extended to entire Megolm sessions.
## License
The Megolm specification (this document) is licensed under the Apache License,
Version 2.0 http://www.apache.org/licenses/LICENSE-2.0.
[Ed25519]: http://ed25519.cr.yp.to/
[HMAC-based key derivation function]: https://tools.ietf.org/html/rfc5869
[HKDF-SHA-256]: https://tools.ietf.org/html/rfc5869
[HMAC-SHA-256]: https://tools.ietf.org/html/rfc2104
[SHA-256]: https://tools.ietf.org/html/rfc6234
[AES-256]: http://csrc.nist.gov/publications/fips/fips197/fips-197.pdf
[CBC]: http://csrc.nist.gov/publications/nistpubs/800-38a/sp800-38a.pdf
[PKCS#7]: https://tools.ietf.org/html/rfc2315
[Olm]: https://gitlab.matrix.org/matrix-org/olm/blob/master/docs/olm.md
[Protocol Buffers encoding]: https://developers.google.com/protocol-buffers/docs/encoding

334
content/olm-megolm/olm.md Normal file
View file

@ -0,0 +1,334 @@
---
title: "Olm: A Cryptographic Ratchet"
weight: 10
type: docs
---
An implementation of the double cryptographic ratchet described by
https://whispersystems.org/docs/specifications/doubleratchet/.
## Notation
This document uses \(\parallel\) to represent string concatenation. When
\(\parallel\) appears on the right hand side of an \(=\) it means that
the inputs are concatenated. When \(\parallel\) appears on the left hand
side of an \(=\) it means that the output is split.
When this document uses \(\operatorname{ECDH}\left(K_A,K_B\right)\) it means
that each party computes a Diffie-Hellman agreement using their private key
and the remote party's public key.
So party \(A\) computes \(\operatorname{ECDH}\left(K_B^{public},K_A^{private}\right)\)
and party \(B\) computes \(\operatorname{ECDH}\left(K_A^{public},K_B^{private}\right)\).
Where this document uses \(\operatorname{HKDF}\left(salt,IKM,info,L\right)\) it
refers to the [HMAC-based key derivation function][] with a salt value of
\(salt\), input key material of \(IKM\), context string \(info\),
and output keying material length of \(L\) bytes.
## The Olm Algorithm
### Initial setup
The setup takes four [Curve25519][] inputs: Identity keys for Alice and Bob,
\(I_A\) and \(I_B\), and one-time keys for Alice and Bob,
\(E_A\) and \(E_B\). A shared secret, \(S\), is generated using
[Triple Diffie-Hellman][]. The initial 256 bit root key, \(R_0\), and 256
bit chain key, \(C_{0,0}\), are derived from the shared secret using an
HMAC-based Key Derivation Function using [SHA-256][] as the hash function
([HKDF-SHA-256][]) with default salt and ``"OLM_ROOT"`` as the info.
\[
\begin{aligned}
S&=\operatorname{ECDH}\left(I_A,E_B\right)\;\parallel\;
\operatorname{ECDH}\left(E_A,I_B\right)\;\parallel\;
\operatorname{ECDH}\left(E_A,E_B\right)\\
R_0\;\parallel\;C_{0,0}&=
\operatorname{HKDF}\left(0,S,\text{``OLM\_ROOT"},64\right)
\end{aligned}
\]
### Advancing the root key
Advancing a root key takes the previous root key, \(R_{i-1}\), and two
Curve25519 inputs: the previous ratchet key, \(T_{i-1}\), and the current
ratchet key \(T_i\). The even ratchet keys are generated by Alice.
The odd ratchet keys are generated by Bob. A shared secret is generated
using Diffie-Hellman on the ratchet keys. The next root key, \(R_i\), and
chain key, \(C_{i,0}\), are derived from the shared secret using
[HKDF-SHA-256][] using \(R_{i-1}\) as the salt and ``"OLM_RATCHET"`` as the
info.
\[
\begin{aligned}
R_i\;\parallel\;C_{i,0}&=
\operatorname{HKDF}\left(
R_{i-1},
\operatorname{ECDH}\left(T_{i-1},T_i\right),
\text{``OLM\_RATCHET"},
64
\right)
\end{aligned}
\]
### Advancing the chain key
Advancing a chain key takes the previous chain key, \(C_{i,j-1}\). The next
chain key, \(C_{i,j}\), is the [HMAC-SHA-256][] of ``"\x02"`` using the
previous chain key as the key.
\[
\begin{aligned}
C_{i,j}&=\operatorname{HMAC}\left(C_{i,j-1},\text{``\char`\\x02"}\right)
\end{aligned}
\]
### Creating a message key
Creating a message key takes the current chain key, \(C_{i,j}\). The
message key, \(M_{i,j}\), is the [HMAC-SHA-256][] of ``"\x01"`` using the
current chain key as the key. The message keys where \(i\) is even are used
by Alice to encrypt messages. The message keys where \(i\) is odd are used
by Bob to encrypt messages.
\[
\begin{aligned}
M_{i,j}&=\operatorname{HMAC}\left(C_{i,j},\text{``\char`\\x01"}\right)
\end{aligned}
\]
## The Olm Protocol
### Creating an outbound session
Bob publishes the public parts of his identity key, \(I_B\), and some
single-use one-time keys \(E_B\).
Alice downloads Bob's identity key, \(I_B\), and a one-time key,
\(E_B\). She generates a new single-use key, \(E_A\), and computes a
root key, \(R_0\), and a chain key \(C_{0,0}\). She also generates a
new ratchet key \(T_0\).
### Sending the first pre-key messages
Alice computes a message key, \(M_{0,j}\), and a new chain key,
\(C_{0,j+1}\), using the current chain key. She replaces the current chain
key with the new one.
Alice encrypts her plain-text with the message key, \(M_{0,j}\), using an
authenticated encryption scheme (see below) to get a cipher-text,
\(X_{0,j}\).
She then sends the following to Bob:
* The public part of her identity key, \(I_A\)
* The public part of her single-use key, \(E_A\)
* The public part of Bob's single-use key, \(E_B\)
* The current chain index, \(j\)
* The public part of her ratchet key, \(T_0\)
* The cipher-text, \(X_{0,j}\)
Alice will continue to send pre-key messages until she receives a message from
Bob.
### Creating an inbound session from a pre-key message
Bob receives a pre-key message as above.
Bob looks up the private part of his single-use key, \(E_B\). He can now
compute the root key, \(R_0\), and the chain key, \(C_{0,0}\), from
\(I_A\), \(E_A\), \(I_B\), and \(E_B\).
Bob then advances the chain key \(j\) times, to compute the chain key used
by the message, \(C_{0,j}\). He now creates the
message key, \(M_{0,j}\), and attempts to decrypt the cipher-text,
\(X_{0,j}\). If the cipher-text's authentication is correct then Bob can
discard the private part of his single-use one-time key, \(E_B\).
Bob stores Alice's initial ratchet key, \(T_0\), until he wants to
send a message.
### Sending normal messages
Once a message has been received from the other side, a session is considered
established, and a more compact form is used.
To send a message, the user checks if they have a sender chain key,
\(C_{i,j}\). Alice uses chain keys where \(i\) is even. Bob uses chain
keys where \(i\) is odd. If the chain key doesn't exist then a new ratchet
key \(T_i\) is generated and a new root key \(R_i\) and chain key
\(C_{i,0}\) are computed using \(R_{i-1}\), \(T_{i-1}\) and
\(T_i\).
A message key,
\(M_{i,j}\) is computed from the current chain key, \(C_{i,j}\), and
the chain key is replaced with the next chain key, \(C_{i,j+1}\). The
plain-text is encrypted with \(M_{i,j}\), using an authenticated encryption
scheme (see below) to get a cipher-text, \(X_{i,j}\).
The user then sends the following to the recipient:
* The current chain index, \(j\)
* The public part of the current ratchet key, \(T_i\)
* The cipher-text, \(X_{i,j}\)
### Receiving messages
The user receives a message as above with the sender's current chain index, \(j\),
the sender's ratchet key, \(T_i\), and the cipher-text, \(X_{i,j}\).
The user checks if they have a receiver chain with the correct
\(i\) by comparing the ratchet key, \(T_i\). If the chain doesn't exist
then they compute a new root key, \(R_i\), and a new receiver chain, with
chain key \(C_{i,0}\), using \(R_{i-1}\), \(T_{i-1}\) and
\(T_i\).
If the \(j\) of the message is less than
the current chain index on the receiver then the message may only be decrypted
if the receiver has stored a copy of the message key \(M_{i,j}\). Otherwise
the receiver computes the chain key, \(C_{i,j}\). The receiver computes the
message key, \(M_{i,j}\), from the chain key and attempts to decrypt the
cipher-text, \(X_{i,j}\).
If the decryption succeeds the receiver updates the chain key for \(T_i\)
with \(C_{i,j+1}\) and stores the message keys that were skipped in the
process so that they can decode out of order messages. If the receiver created
a new receiver chain then they discard their current sender chain so that
they will create a new chain when they next send a message.
## The Olm Message Format
Olm uses two types of messages. The underlying transport protocol must provide
a means for recipients to distinguish between them.
### Normal Messages
Olm messages start with a one byte version followed by a variable length
payload followed by a fixed length message authentication code.
```nohighlight
+--------------+------------------------------------+-----------+
| Version Byte | Payload Bytes | MAC Bytes |
+--------------+------------------------------------+-----------+
```
The version byte is ``"\x03"``.
The payload consists of key-value pairs where the keys are integers and the
values are integers and strings. The keys are encoded as a variable length
integer tag where the 3 lowest bits indicates the type of the value:
0 for integers, 2 for strings. If the value is an integer then the tag is
followed by the value encoded as a variable length integer. If the value is
a string then the tag is followed by the length of the string encoded as
a variable length integer followed by the string itself.
Olm uses a variable length encoding for integers. Each integer is encoded as a
sequence of bytes with the high bit set followed by a byte with the high bit
clear. The seven low bits of each byte store the bits of the integer. The least
significant bits are stored in the first byte.
**Name**|**Tag**|**Type**|**Meaning**
:-----:|:-----:|:-----:|:-----:
Ratchet-Key|0x0A|String|The public part of the ratchet key, Ti, of the message
Chain-Index|0x10|Integer|The chain index, j, of the message
Cipher-Text|0x22|String|The cipher-text, Xi, j, of the message
The length of the MAC is determined by the authenticated encryption algorithm
being used. (Olm version 1 uses [HMAC-SHA-256][], truncated to 8 bytes). The
MAC protects all of the bytes preceding the MAC.
### Pre-Key Messages
Olm pre-key messages start with a one byte version followed by a variable
length payload.
```nohighlight
+--------------+------------------------------------+
| Version Byte | Payload Bytes |
+--------------+------------------------------------+
```
The version byte is ``"\x03"``.
The payload uses the same key-value format as for normal messages.
**Name**|**Tag**|**Type**|**Meaning**
:-----:|:-----:|:-----:|:-----:
One-Time-Key|0x0A|String|The public part of Bob's single-use key, Eb.
Base-Key|0x12|String|The public part of Alice's single-use key, Ea.
Identity-Key|0x1A|String|The public part of Alice's identity key, Ia.
Message|0x22|String|An embedded Olm message with its own version and MAC.
## Olm Authenticated Encryption
### Version 1
Version 1 of Olm uses [AES-256][] in [CBC][] mode with [PKCS#7][] padding for
encryption and [HMAC-SHA-256][] (truncated to 64 bits) for authentication. The
256 bit AES key, 256 bit HMAC key, and 128 bit AES IV are derived from the
message key using [HKDF-SHA-256][] using the default salt and an info of
``"OLM_KEYS"``.
\[
\begin{aligned}
AES\_KEY_{i,j}\;\parallel\;HMAC\_KEY_{i,j}\;\parallel\;AES\_IV_{i,j}
&= \operatorname{HKDF}\left(0,M_{i,j},\text{``OLM\_KEYS"},80\right)
\end{aligned}
\]
The plain-text is encrypted with AES-256, using the key \(AES\_KEY_{i,j}\)
and the IV \(AES\_IV_{i,j}\) to give the cipher-text, \(X_{i,j}\).
Then the entire message (including the Version Byte and all Payload Bytes) are
passed through [HMAC-SHA-256][]. The first 8 bytes of the MAC are appended to the message.
## Message authentication concerns
To avoid unknown key-share attacks, the application must include identifying
data for the sending and receiving user in the plain-text of (at least) the
pre-key messages. Such data could be a user ID, a telephone number;
alternatively it could be the public part of a keypair which the relevant user
has proven ownership of.
### Example attacks
1. Alice publishes her public [Curve25519][] identity key, \(I_A\). Eve
publishes the same identity key, claiming it as her own. Bob downloads
Eve's keys, and associates \(I_A\) with Eve. Alice sends a message to
Bob; Eve intercepts it before forwarding it to Bob. Bob believes the
message came from Eve rather than Alice.
This is prevented if Alice includes her user ID in the plain-text of the
pre-key message, so that Bob can see that the message was sent by Alice
originally.
2. Bob publishes his public [Curve25519][] identity key, \(I_B\). Eve
publishes the same identity key, claiming it as her own. Alice downloads
Eve's keys, and associates \(I_B\) with Eve. Alice sends a message to
Eve; Eve cannot decrypt it, but forwards it to Bob. Bob believes the
Alice sent the message to him, whereas Alice intended it to go to Eve.
This is prevented by Alice including the user ID of the intended recpient
(Eve) in the plain-text of the pre-key message. Bob can now tell that the
message was meant for Eve rather than him.
## IPR
The Olm specification (this document) is hereby placed in the public domain.
## Feedback
Can be sent to olm at matrix.org.
## Acknowledgements
The ratchet that Olm implements was designed by Trevor Perrin and Moxie
Marlinspike - details at https://whispersystems.org/docs/specifications/doubleratchet/. Olm is
an entirely new implementation written by the Matrix.org team.
[Curve25519]: http://cr.yp.to/ecdh.html
[Triple Diffie-Hellman]: https://whispersystems.org/blog/simplifying-otr-deniability/
[HMAC-based key derivation function]: https://tools.ietf.org/html/rfc5869
[HKDF-SHA-256]: https://tools.ietf.org/html/rfc5869
[HMAC-SHA-256]: https://tools.ietf.org/html/rfc2104
[SHA-256]: https://tools.ietf.org/html/rfc6234
[AES-256]: http://csrc.nist.gov/publications/fips/fips197/fips-197.pdf
[CBC]: http://csrc.nist.gov/publications/nistpubs/800-38a/sp800-38a.pdf
[PKCS#7]: https://tools.ietf.org/html/rfc2315