anonymize Tub IDs when using Tahoe-LAFS with anonymity net like Tor or I2p

[dawuud]

I believe it was Leif Ryge who first pointed this out... We really need to randomize the client ID when using an anonymity network like Tor or I2p.

This ticket is definitely related to #517 ​https://tahoe-lafs.org/trac/tahoe-lafs/ticket/517

This ticket also seems to be a sub-ticket of #1010 ​https://tahoe-lafs.org/trac/tahoe-lafs/ticket/1010 in so far as we can turn this client ID randomization on and off with a configuration option.

[zooko]

Brian pointed out in Nuts & Bolts today that #510 would solve this.

[zooko]

Replying to zooko:

Brian pointed out in Nuts & Bolts today that #510 would solve this.

Or maybe it was Daira who pointed that out.

[daira]

Replying to zooko:

Replying to zooko:

Brian pointed out in Nuts & Bolts today that #510 would solve this.

Or maybe it was Daira who pointed that out.

It was. (More precisely, that an HTTP-based protocol wouldn't have this specific problem with client tub IDs, although that isn't the only problem involved in client anonymisation.)

[warner]

Current status: client connections to storage servers use ephemeral Tubs (thanks to #2759), so storage servers won't see tub-id correlations between subsequent boots of a single client.

The IntroducerClient, though, uses a static tub (with key stored in NODEDIR/private/node.pem) for all connections. So the Introducer can correlate connections across client reboots.

Do folks think we can close this ticket as is, or should we use an ephemeral Tub for the introducer client too?

(note that when Accounting happens, clients will get a persistent Ed25519 public key, and they'll sign their storage-server messages with it. But I think we can just declare that when anonymous=true, these keys are disabled, or regenerated at each boot, and you don't get to participate in any accounting-related tasks. Maybe we can add a pseudonymous=true flag which will allow persistent client pubkeys, but still enforce the other safety restrictions. In that world, servers could tell which clients were which, but that "identity" is still unlinked to the client's real IP address)

[warner]

In today's devchat, we decided to document the TubID linkages that clients currently reveal (introducer client + storage server, and multiple connections within a single client lifetime for everything else), and then close this ticket. We'll add a new ticket (with milestone=undecided) for either addressing the remaining linkages, or WONTFIXing them.

[warner]

#2828 added for addressing/accepting the remaining linkages.

[Brian Warner <warner@…>]

In 80acd56/trunk:

docs: describe known linkability

closes ticket:2384