Implementing End-to-End Encryption in Matrix clients

This guide is intended for authors of Matrix clients who wish to add support for end-to-end encryption. It is highly recommended that readers be familiar with the Matrix protocol and the use of access tokens before proceeding.

The libolm library

End-to-end encryption in Matrix is based on the Olm and Megolm cryptographic ratchets. The recommended starting point for any client authors is with the libolm library, which contains implementations of all of the cryptographic primitives required. The library itself is written in C/C++, but is architected in a way which makes it easy to write wrappers for higher-level languages.

Devices

We have a particular meaning for “device”. As a user, I might have several devices (a desktop client, some web browsers, an Android device, an iPhone, etc). When I first use a client, it should register itself as a new device. If I log out and log in again as a different user, the client must register as a new device. Critically, the client must create a new set of keys (see below) for each “device”.

The longevity of devices will depend on the client. In the web client, we create a new device every single time you log in. In a mobile client, it might be acceptable to reuse the device if a login session expires, provided the user is the same. Never share keys between different users.

Devices are identified by their device_id (which is unique within the scope of a given user). By default, the /login and /register endpoints will auto-generate a device_id and return it in the response; a client is also free to generate its own device_id or, as above, reuse a device, in which case the client should pass the device_id in the request body.

The lifetime of devices and access_tokens are closely related. In the simple case where a new device is created each time you log in, there is a one-to-one mapping between a device_id and an access_token. If a client reuses a device_id when logging in, there will be several access_tokens associated with a given device_id - but still, we would expect only one of these to be active at once (though we do not currently enforce that in Synapse).

Keys used in End-to-End encryption

There are a number of keys involved in encrypted communication: a summary of them follows.

Ed25519 fingerprint key pair

Ed25519 is a public-key cryptographic system for signing messages. In Matrix, each device has an Ed25519 key pair which serves to identify that device. The private part of the key pair should never leave the device, but the public part is published to the Matrix network.

Curve25519 identity key pair

Curve25519 is a public-key cryptographic system which can be used to establish a shared secret. In Matrix, each device has a long-lived Curve25519 identity key which is used to establish Olm sessions with that device. Again, the private key should never leave the device, but the public part is signed with the Ed25519 fingerprint key and published to the network.

Theoretically we should rotate the Curve25519 identity key from time to time, but we haven't implemented this yet.

Curve25519 one-time keys

As well as the identity key, each device creates a number of Curve25519 key pairs which are also used to establish Olm sessions, but can only be used once. Once again, the private part remains on the device.

At startup, Alice creates a number of one-time key pairs, and publishes them to her homeserver. If Bob wants to establish an Olm session with Alice, he needs to claim one of Alice’s one-time keys, and creates a new one of his own. Those two keys, along with Alice’s and Bob’s identity keys, are used in establishing an Olm session between Alice and Bob.

Megolm encryption keys

The Megolm key is used to encrypt group messages (in fact it is used to derive an AES-256 key, and an HMAC-SHA-256 key). It is initialised with random data. Each time a message is sent, a hash calculation is done on the Megolm key to derive the key for the next message. It is therefore possible to share the current state of the Megolm key with a user, allowing them to decrypt future messages but not past messages.

Ed25519 Megolm signing key pair

When a sender creates a Megolm session, he also creates another Ed25519 signing key pair. This is used to sign messages sent via that Megolm session, to authenticate the sender. Once again, the private part of the key remains on the device. The public part is shared with other devices in the room alongside the encryption key.

Creating and registering device keys

This process only happens once, when a device first starts.

It must create the Ed25519 fingerprint key pair and the Curve25519 identity key pair. This is done by calling olm_create_account in libolm. The (base64-encoded) keys are retrieved by calling olm_account_identity_keys. The account should be stored for future use.

It should then publish these keys to the homeserver, which is done by using the device_keys property of the /keys/upload endpoint.

In order to sign the device_keys payload as described in Signing JSON, clients should call olm_account_sign.

Creating and registering one-time keys

The client should keep track of how many one-time keys the homeserver has stored for it, and, if necessary, generate and upload some more.

This can be achieved by inspecting the device_one_time_keys_count property of a /sync/ response.

The maximum number of active keys supported by libolm is returned by olm_account_max_number_of_one_time_keys. The client should try to maintain about half this number on the homeserver.

To generate new one-time keys:

  • Call olm_account_generate_one_time_keys to generate new keys.

  • Call olm_account_one_time_keys to retrieve the unpublished keys. This returns a JSON-formatted object with the single property curve25519, which is itself an object mapping key id to base64-encoded Curve25519 key. For example:

    {
      "curve25519": {
        "AAAAAA": "wo76WcYtb0Vk/pBOdmduiGJ0wIEjW4IBMbbQn7aSnTo",
        "AAAAAB": "LRvjo46L1X2vx69sS9QNFD29HWulxrmW11Up5AfAjgU"
      }
    }
    
  • Each key should be signed in the same way as the previous identity keys payload, and uploaded using the one_time_keys property of the /keys/upload endpoint.

  • Call olm_account_mark_keys_as_published to tell the olm library not to return the same keys from a future call to olm_account_one_time_keys.

Configuring a room to use encryption

To enable encryption in a room, a client should send a state event of type m.room.encryption, and content { "algorithm": "m.megolm.v1.aes-sha2" }.

Handling an m.room.encryption state event

When a client receives an m.room.encryption event as above, it should set a flag to indicate that messages sent in the room should be encrypted.

This flag should not be cleared if a later m.room.encryption event changes the configuration. This is to avoid a situation where a MITM can simply ask participants to disable encryption. In short: once encryption is enabled in a room, it can never be disabled.

The event should contain an algorithm property which defines which encryption algorithm should be used for encryption. Currently only m.megolm.v1-aes-sha2 is permitted here.

The event may also include other settings for how messages sent in the room should be encrypted (for example, rotation_period_ms to define how often the session should be replaced). See the spec for more details.

Handling an m.room.encrypted event

Encrypted events have a type of m.room.encrypted. They have a content property algorithm which gives the encryption algorithm in use, as well as other properties specific to the algorithm [1].

[1]Note that a redacted event will have an empty content, and hence the content will have no algorithm property. Thus a client should check whether an event is redacted before checking for the algorithm property.

The encrypted payload is a JSON object with the properties type (giving the decrypted event type), and content (giving the decrypted content). Depending on the algorithm in use, the payload may contain additional keys.

There are currently two defined algorithms:

m.olm.v1.curve25519-aes-sha2

The spec gives details on this algorithm and an example payload .

The sender_key property of the event content gives the Curve25519 identity key of the sender. Clients should maintain a list of known Olm sessions for each device they speak to; it is recommended to index them by Curve25519 identity key.

Olm messages are encrypted separately for each recipient device. ciphertext is an object mapping from the Curve25519 identity key for the recipient device. The receiving client should, of course, look for its own identity key in this object. (If it isn't listed, the message wasn't sent for it, and the client can't decrypt it; it should show an error instead, or similar).

This should result in an object with the properties type and body. Messages of type '0' are 'prekey' messages which are used to establish a new Olm session between two devices; type '1' are normal messages which are used once a message has been received on the session.

When a message (of either type) is received, a client should first attempt to decrypt it with each of the known sessions for that sender. There are two steps to this:

  • If (and only if) type==0, the client should call olm_matches_inbound_session with the session and body. This returns a flag indicating whether the message was encrypted using that session.
  • The client calls olm_decrypt, with the session, type, and body. If this is successful, it returns the plaintext of the event.

If the client was unable to decrypt the message using any known sessions (or if there are no known sessions yet), and the message had type 0, and olm_matches_inbound_session wasn't true for any existing sessions, then the client can try establishing a new session. This is done as follows:

  • Call olm_create_inbound_session_from using the olm account, and the sender_key and body of the message.
  • If the session was established successfully:
    • Call olm_remove_one_time_keys to ensure that the same one-time-key cannot be reused.
    • Call olm_decrypt with the new session.
    • Store the session for future use.

At the end of this, the client will hopefully have successfully decrypted the payload.

As well as the type and content properties, the plaintext payload should contain a number of other properties. Each of these should be checked as follows [2].

sender
The user ID of the sender. The client should check that this matches the sender in the event.
recipient
The user ID of the recipient. The client should check that this matches the local user ID.
keys
an object with a property ed25519. The client should check that the value of this property matches the sender's fingerprint key when marking the event as verified.
recipient_keys
an object with a property ed25519. The client should check that the value of this property matches its own fingerprint key.
[2]These tests prevent an attacker publishing someone else's curve25519 keys as their own and subsequently claiming to have sent messages which they didn't.

m.megolm.v1.aes-sha2

The spec gives details on this algorithm and an example payload .

Encrypted events using this algorithm should have sender_key, session_id and ciphertext content properties. If the room_id, sender_key and session_id correspond to a known Megolm session (see below), the ciphertext can be decrypted by passing the ciphertext into olm_group_decrypt.

In order to avoid replay attacks a client should remember the megolm message_index returned by olm_group_decrypt of each event they decrypt for each session. If the client decrypts an event with the same message_index as one that it has already received using that session then it should treat the message as invalid. However, care must be taken when an event is decrypted multiple times that it is not flagged as a replay attack. For example, this may happen when the client decrypts an event, the event gets purged from the client's cache, and then the client backfills and re-decrypts the event. One way to handle this case is to ensure that the record of message_indexes is appropriately purged when the client's cache of events is purged. Another way is to remember the event's event_id and origin_server_ts along with its message_index. When the client decrypts an event with a message_index matching that of a previously-decrypted event, it can then compare the event_id and origin_server_ts that it remembered for that message_index, and if those fields match, then the message should be decrypted as normal.

The client should check that the sender's fingerprint key matches the keys.ed25519 property of the event which established the Megolm session when marking the event as verified.

Handling an m.room_key event

These events contain key data to allow decryption of other messages. They are sent to specific devices, so they appear in the to_device section of the response to GET /_matrix/client/r0/sync. They will also be encrypted, so will need decrypting as above before they can be seen. (These events are generated by other clients - see starting a megolm session).

The room_id, together with the sender_key of the m.room_key event before it was decrypted, and the session_id, uniquely identify a Megolm session. If they do not represent a known session, the client should start a new inbound Megolm session by calling olm_init_inbound_group_session with the session_key.

The client should remember the value of the keys property of the payload of the encrypted m.room_key event and store it with the inbound session. This is used as above when marking the event as verified.

Downloading the device list for users in the room

Before an encrypted message can be sent, it is necessary to retrieve the list of devices for each user in the room. This can be done proactively, or deferred until the first message is sent. The information is also required to allow users to verify or block devices.

The client should use the /keys/query endpoint, passing the IDs of the members of the room in the device_keys property of the request.

The client must first check the signatures on the DeviceKeys objects returned by /keys/query. To do this, it should remove the signatures and unsigned properties, format the remainder as Canonical JSON, and pass the result into olm_ed25519_verify, using the Ed25519 key for the key parameter, and the corresponding signature for the signature parameter. If the signature check fails, no further processing should be done on the device.

The client must also check that the user_id and device_id fields in the object match those in the top-level map [3].

The client should check if the user_id/device_id correspond to a device it had seen previously. If it did, the client must check that the Ed25519 key hasn't changed. Again, if it has changed, no further processing should be done on the device.

Otherwise the client stores the information about this device.

[3]This prevents a malicious or compromised homeserver replacing the keys for the device with those of another.

Sending an encrypted message event

When sending a message in a room configured to use encryption, a client first checks to see if it has an active outbound Megolm session. If not, it first creates one as per below. If an outbound session exists, it should check if it is time to rotate it, and create a new one if so.

The client then builds an encryption payload as follows:

{
  "type": "<event type>",
  "content": "<event content>",
  "room_id": "<id of destination room>"
}

and calls olm_group_encrypt to encrypt the payload. This is then packaged into event content as follows:

{
  "algorithm": "m.megolm.v1.aes-sha2",
  "sender_key": "<our curve25519 device key>",
  "ciphertext": "<encrypted payload>",
  "session_id": "<outbound group session id>",
  "device_id": "<our device ID>"
}

Finally, the encrypted event is sent to the room with POST /_matrix/client/r0/rooms/<room_id>/send/m.room.encrypted/<txn_id>.

Starting a Megolm session

When a message is first sent in an encrypted room, the client should start a new outbound Megolm session. This should not be done proactively, to avoid proliferation of unnecessary Megolm sessions.

To create the session, the client should call olm_init_outbound_group_session, and store the details of the outbound session for future use.

The client should then call olm_outbound_group_session_id to get the unique ID of the new session, and olm_outbound_group_session_key to retrieve the current ratchet key and index. It should store these details as an inbound session, just as it would when receiving them via an m.room_key event.

The client must then share the keys for this session with each device in the room. It must therefore download the device list if it hasn't already done so. Then it should build a unique m.room_key event, and send it encrypted using Olm to each device in the room which has not been blocked, .

Once all of the key-sharing event contents have been assembled, the events should be sent to the corresponding devices via PUT /_matrix/client/r0/sendToDevice/m.room.encrypted/<txnId>.

Rotating Megolm sessions

Megolm sessions may not be reused indefinitely. The parameters which define how often a session should be rotated are defined in the m.room.encryption state event of a room.

Once either the message limit or time limit have been reached, the client should start a new session before sending any more messages.

Encrypting an event with Olm

Olm is not used for encrypting room events, as it requires a separate copy of the ciphertext for each device, and because the receiving device can only decrypt received messages once. However, it is used for encrypting key-sharing events for Megolm.

When encrypting an event using Olm, the client should:

  • Build an encryption payload as illustrated in the spec.
  • Check if it has an existing Olm session; if it does not, start a new one. If it has several (as may happen due to races when establishing sessions), it should use the one with the first session_id when sorted by their ASCII codepoints (ie, 'A' would be before 'Z', which would be before 'a').
  • Encrypt the payload by calling olm_encrypt.
  • Package the payload into an Olm m.room.encrypted event.

Starting an Olm session

To start a new Olm session with another device, a client must first claim one of the other device's one-time keys. To do this, it should initiate a request to /keys/claim.

The client should check the signatures on the signed key objects in the response. As with checking the signatures on the device keys, it should remove the signatures and (if present) unsigned properties, format the remainder as Canonical JSON, and pass the result into olm_ed25519_verify, using the Ed25519 device key for the key parameter.

Provided the key object passes verification, the client should then pass the key, along with the Curve25519 Identity key for the remote device, into olm_create_outbound_session.

Handling membership changes

The client should monitor rooms which are configured to use encryption for membership changes.

When a member leaves a room, the client should invalidate any active outbound Megolm session, to ensure that a new session is used next time the user sends a message.

When a new member joins a room, the client should first download the device list for the new member, if it doesn't already have it.

After giving the user an opportunity to block any suspicious devices, the client should share the keys for the outbound Megolm session with all the new member's devices. This is done in the same way as creating a new session, except that there is no need to start a new Megolm session: due to the design of the Megolm ratchet, the new user will only be able to decrypt messages starting from the current state. The recommended method is to maintain a list of members who are waiting for the session keys, and share them when the user next sends a message.

Handling new devices

When a user logs in on a new device, it is necessary to make sure that other devices in any rooms with encryption enabled are aware of the new device, so that they can share their outbound sessions with it as they would with a new member.

The device tracking process which should be implemented is documented in the spec.

Blocking / Verifying devices

It should be possible for a user to mark each device belonging to another user as 'Blocked' or 'Verified', through a process detailed in the spec.

When a user chooses to block a device, this means that no further encrypted messages should be shared with that device. In short, it should be excluded when sharing room keys when starting a new Megolm session. Any active outbound Megolm sessions whose keys have been shared with the device should also be invalidated so that no further messages are sent over them.

Marking events as 'verified'

Once a device has been verified, it is possible to verify that events have been sent from a particular device. See the section on Handling an m.room.encrypted event for notes on how to do this for each algorithm. Events sent from a verified device can be decorated in the UI to show that they have been sent from a verified device.

Encrypted attachments

Homeservers must not be able to read files shared in encrypted rooms. Clients should implement a strategy described in the spec.

Currently, the files are encrypted using AES-CTR, which is not included in libolm. Clients have to rely on a third party library.

Key sharing

When an event cannot be decrypted due to missing keys, a client may want to request them from other clients which may have them. Similarly, a client may want to reply to a key request with the associated key if it can assert that the requesting device is allowed to see the messages encrypted with this key.

Those capabilities are achieved using m.room_key_request and m.forwarded_room_key events.

The session_key property of a m.forwarded_room_key event differs from the one of a m.room_key event, as it does not include the Ed25519 signature of the original sender. It should be obtained from olm_export_inbound_group_session at the desired message index, and the session can be restored with olm_import_inbound_group_session.

The forwarded_room_key property starts out empty, but each time a key is forwarded to another device, the previous sender in the chain is added to the end of the list. Consider the following example:

  • A -> B : m.room_key
  • B -> C : m.forwarded_room_key
  • C -> D : m.forwarded_room_key

In the message B -> C forwarded_room_key is empty, but in the message C -> D it contains B's Curve25519 key. In order for D to believe that the session came from A, D must trust the direct sender C and every entry in this chain.

In order to securely implement key sharing, clients must not reply to every key request they receive. The recommended strategy is to share the keys automatically only to verified devices of the same user. Requests coming from unverified devices should prompt a dialog, allowing the user to verify the device, share the keys without verifying, or not to share them (and ignore future requests). A client should also check whether requests coming from devices of other users are legitimate. This can be done by keeping track of the users a session was shared with, and at which message index.

Key requests can be sent to all of the current user's devices, as well as the original sender of the session, and other devices present in the room. When the client receives the requested key, it should send a m.room_key_request event to all the devices it requested the key from, setting the action property to "cancel_request" and request_id to the ID of the initial request.

How helpful was this page? Click to give a rating