censorship
sbr is designed against censors that block by server name, by TLS fingerprint, and by cutting connections some time after they start.
Which censors and surveillance programmes these are, what each one does, the published evidence for it, and what sbr still lacks are in two documents in the repository:
- threat model: Russia, Belarus, China, Egypt, the Gulf states, Israel, the United Kingdom and commercial data fusion, with a capability matrix and the known gaps ranked by impact.
- china: the Great Firewall's techniques one by one, what sbr and its reference server deployment do about each, and what has not been verified.
what sbr does
- server-name filtering
A handshake reset before the server answers is retried with the ClientHello split across TLS records and then TCP segments, cutting inside every label of the name. For QUIC, the name's halves travel in separate Initial packets.
- tls fingerprinting
TLS over TCP sends a ClientHello captured from Chrome, Safari or Firefox, including REALITY's, which carries the X25519MLKEM768 share current servers require.
- blocking after the handshake
Servers that complete a handshake and then stop carrying data are detected by a data probe and in live traffic, and failover moves to another server without waiting for the next probe interval.
- dns poisoning
DNS servers in a profile can be routed through the tunnel, and a query fails over to the next server when one stops answering.
- probe patterns
Latency probes to members that share a server are spaced apart and their period varies, so the client does not open several handshakes to one address at once on a fixed clock.
None of this has been measured from inside China; the china document says what has been tested and how.