Encryption and identity

What is end-to-end encryption?

End-to-end encryption is encryption in which a message is encrypted on the sender's device and can be decrypted only on the devices of its intended recipients. Servers, relays and networks in between carry the message but do not hold the keys to read it. It protects the content, not the fact that a message was sent, to whom or when.

TL;DR

End-to-end encryption keeps message content away from the servers, relays and networks that carry it, which TLS and server-side encryption at rest do not. It depends on knowing that a public key belongs to the recipient, and good schemes keep changing keys. Metadata, the devices themselves, delivery and unencrypted traffic stay exposed.

Learning objectives

After reading this article you will be able to:

  • Distinguish end-to-end encryption from TLS and from encryption at rest on a server
  • Describe how end-to-end encryption handles offline recipients and checks whose key is whose
  • Identify what end-to-end encryption leaves exposed, such as metadata and the devices themselves

Endpoints in end-to-end encryption

The ends in end-to-end encryption are the devices of the people communicating. The sender’s device encrypts a message before it leaves, and only the recipients’ devices hold the keys to decrypt it. Everything in between, such as the app’s servers, an internet relay, a Wi-Fi network or another phone passing the message along, handles only ciphertext.

Apple’s description of iMessage is a typical statement of the goal: message content and attachments are secured with end-to-end encryption so that no one but the sender and receiver can access them, and Apple cannot decrypt the data. The IETF’s Messaging Layer Security standard opens the same way, describing end-to-end security as making messages accessible only to the communicating endpoints, not to the servers that deliver them.

End-to-end encryption vs encryption in transit

Encryption in transit protects data while it moves between two machines, and when one of them is a server, it decrypts everything it receives. TLS 1.3, which secures web and app connections, lets a client and a server communicate in a way designed to prevent eavesdropping, tampering and forgery, but the server is one of the two ends of that connection. If a message goes from your phone to a chat server over TLS and on to your friend over another TLS connection, the server sees it in the clear in the middle.

Kind of encryptionWho can read the content
In transit (TLS)Each server the data reaches
At rest on a serverThe service, which holds the storage keys
End to endOnly the sender and the intended recipients

Data stored on a server is often encrypted too, but with keys the service holds. Apple’s Advanced Data Protection for iCloud shows the difference: when a user turns it on, service keys that were held in hardware security modules in Apple’s data centres are deleted there and kept only on the user’s trusted devices, which is what makes that data end-to-end encrypted.

End-to-end encryption does not replace TLS. Apps usually use both: TLS protects each connection, and end-to-end encryption protects the content across all of them.

End-to-end encryption key exchange

End-to-end encryption rests on public and private keys. Each device has a key pair and keeps the private half to itself. To send a message, the sender’s device needs the recipient’s public key. The two devices use their keys to agree on a shared secret, derive symmetric keys from it, and encrypt each message with authenticated encryption, which also detects tampering.

Two practical problems follow.

  • Recipients are often offline. Signal’s X3DH specification solves this with prekeys: the recipient publishes public keys to a server ahead of time, and a sender can use them to set up an encrypted session and send a first message while the recipient is away.
  • You have to know the key is really theirs. A service usually runs a directory that maps phone numbers or usernames to public keys; iMessage registers each device’s public keys with Apple’s identity service. A directory that lies could hand out its own key instead. X3DH notes that unless people compare key fingerprints, for example by scanning a QR code, they get no cryptographic guarantee of who they are talking to.

When there is no server at all, devices exchange and verify keys directly. How does end-to-end encryption work when there is no server? covers that case.

Group end-to-end encryption and key changes

For a group, every member needs the current keys, and the group changes as people join and leave, which makes it harder than encrypting for one recipient. MLS, published as RFC 9420, is the IETF’s standard for this: it establishes group keys asynchronously, with forward secrecy and post-compromise security, for groups ranging from two members to thousands. How does MLS work? explains the mechanics.

Good schemes also change keys over time, so that stealing today’s keys does not expose last month’s messages, and a device that was compromised can recover once it refreshes its keys.

End-to-end encryption limits

End-to-end encryption hides content. It does not protect metadata, the devices themselves, delivery, or anything the app does not encrypt.

  • Metadata. NIST’s definition notes that routing information remains visible. Whoever carries the message can usually see who sent it, to whom, when and how large it was. RFC 9420 lists the parts of MLS that are not confidential, including the group identifier and how often its keys change. See what metadata is, and whether encryption hides it.
  • The devices themselves. Messages are readable on the endpoints. Malware on a phone, an unlocked device or an unencrypted backup bypasses the encryption entirely.
  • Delivery. A server that cannot read messages can still delay or drop them.
  • Everything the app does not encrypt. Apps often mix encrypted and unencrypted traffic. In Offline Protocol’s mesh SDK, for example, MLS protects application messages and groups, while service advertisements are signed plaintext that observers can read but not alter undetected. Since version 0.29.0, service requests and responses are sealed with MLS once the two devices share a confirmed session; a first request to a provider never met is still signed plaintext.

Frequently asked questions

Is HTTPS end-to-end encryption?

Not in the messaging sense. HTTPS uses TLS to encrypt the connection between your device and a server, and the server decrypts what it receives. End-to-end encryption means the server never has the keys at all.

Can the service provider read end-to-end encrypted messages?

Not if the scheme is working as designed, because the provider never holds the decryption keys. It can still see metadata, refuse to deliver messages, or, if users never verify keys, try to substitute its own keys for someone else's.

Does end-to-end encryption hide who I am messaging?

Usually not. NIST's definition notes that routing information remains visible, and whoever carries the message can usually see who sent it, to whom, when and how large it was.

How can you send an end-to-end encrypted message to someone who is offline?

Signal's X3DH specification uses prekeys. The recipient publishes public keys to a server ahead of time, and a sender can use them to set up an encrypted session and send a first message while the recipient is away.

Why compare key fingerprints in an end-to-end encrypted app?

A directory that lies could hand out its own key instead of your contact's. X3DH notes that unless people compare key fingerprints, for example by scanning a QR code, they get no cryptographic guarantee of who they are talking to.

Does end-to-end encryption protect messages on a compromised phone?

No. Messages are readable on the endpoints, so malware on a phone, an unlocked device or an unencrypted backup bypasses the encryption entirely.

Sources

Build it with Offline Protocol

The security page describes how the Offline Protocol SDK uses MLS for application messages and groups, which control traffic is signed plaintext instead, and the residual risks recorded in its threat model.

Read the security model