29 releases (15 breaking)

0.17.0 Apr 2, 2024
0.14.1 Mar 4, 2024
0.14.0 Feb 5, 2024
0.12.0 Dec 4, 2023
0.0.2 Nov 30, 2021

#220 in Network programming

Download history 169/week @ 2023-12-20 127/week @ 2023-12-27 167/week @ 2024-01-03 429/week @ 2024-01-10 326/week @ 2024-01-17 124/week @ 2024-01-24 178/week @ 2024-01-31 207/week @ 2024-02-07 177/week @ 2024-02-14 232/week @ 2024-02-21 1429/week @ 2024-02-28 1155/week @ 2024-03-06 920/week @ 2024-03-13 1440/week @ 2024-03-20 781/week @ 2024-03-27 714/week @ 2024-04-03

3,971 downloads per month
Used in 18 crates (4 directly)


50K SLoC


Guard node selection for Tor network clients.


This crate is part of Arti, a project to implement Tor in Rust.

"Guard nodes" are mechanism that Tor clients uses to limit the impact of hostile relays. Approximately: each client chooses a small set of relays to use as its "guards". Later, when the client picks its paths through network, rather than choosing a different first hop randomly for every path, it chooses the best "guard" as the first hop.

This crate provides GuardMgr, an object that manages a set of guard nodes, and helps the tor-circmgr crate know when to use them.

Guard nodes are persistent across multiple process invocations.

More Arti users won't need to use this crate directly.


What's the point? By restricting their first hops to a small set, clients increase their odds against traffic-correlation attacks. Since we assume that an adversary who controls both ends of a circuit can correlate its traffic, choosing many circuits with random entry points will eventually cause a client to eventually pick an attacker-controlled circuit, with probability approaching 1 over time. If entry nodes are restricted to a small set, however, then the client has a chance of never picking an attacker-controlled circuit.

(The actual argument is a little more complicated here, and it relies on the assumption that, since the attacker knows statistics, exposing any of your traffic is nearly as bad as exposing all of your traffic.)


The real algorithm for selecting and using guards can get more complicated because of a variety of factors.

  • In reality, we can't just "pick a few guards at random" and use them forever: relays can appear and disappear, relays can go offline and come back online, and so on. What's more, keeping guards for too long can make targeted attacks against those guards more attractive.

  • Further, we may have particular restrictions on where we can connect. (For example, we might be restricted to ports 80 and 443, but only when we're on a commuter train's wifi network.)

  • We need to resist attacks from local networks that block all but a small set of guard relays, to force us to choose those.

  • We need to give good, reliable performance while using the guards that we prefer.

These needs complicate our API somewhat. Instead of simply asking the GuardMgr for a guard, the circuit-management code needs to be able to tell the GuardMgr that a given guard has failed (or succeeded), and that it needs a different guard in the future (or not).

Further, the GuardMgr code needs to be able to hand out provisional guards, in effect saying "You can try building a circuit with this guard, but please don't actually use that circuit unless I tell you it's safe."

For details on the exact algorithm, see guard-spec.txt (link below) and comments and internal documentation in this crate.


  • Our circuit blocking algorithm is simplified from the one that Tor uses. See comments in GuardSet::circ_usability_status for more information. See also proposal 337.


Guard nodes were first proposed (as "helper nodes") in "Defending Anonymous Communications Against Passive Logging Attacks" by Matthew Wright, Micah Adler, Brian N. Levine, and Clay Shields in the Proceedings of the 2003 IEEE Symposium on Security and Privacy. (See https://www.freehaven.net/anonbib/#wright03)

Tor's current guard selection algorithm is described in Tor's Guard Specification document.

Compile-time features

  • bridge-client: Build with support for bridges. (Bridges are relays that are not listed in the Tor network directory, which can be used for anti-censorship purposes.)

  • pt-client: Build with support for guards that can be contacted using pluggable transports. (A pluggable transport is an alternative mechanism for contacting a Tor relay, for censorship avoidance.)

  • full: Enable all features above.

License: MIT OR Apache-2.0


~598K SLoC