An off-grid social network

FWIW, you can use pub.lua.cz:8008:@xYSW6eVu8gTS/nTSXZiH97dgKZ+wp7NkomR6WKK/PBI=.ed25519~iQ16RuvjKZqy/RhiXXmW9+6wuZNq+SBI8evG3PotxvI= if you have trouble connecting to the ones on github.

Feel free to add it to the wiki, I do plan to run it long term, but I am not a github user.

Replacing a tracker with a DHT trades having one server with all peer and chunk knowledge with N servers with partial peer and chunk knowledge. If the goal is to stop unwanted eavesdroppers, then the choice is between (1) trusting that a single server that knows everything will not divulge information, or (2) trusting that an unknown, dynamic number of servers that anyone can run (including the unwanted eavesdroppers) will not divulge partial information.

The paper I linked up the thread indicates that unwanted eavesdroppers can learn a lot about the peers with choice (2) by exploiting the ways DHTs operate. Heuristics can slow this down, but not stop it. With choice (1), it is possible to fully stop unwanted eavesdroppers if peers can trust the tracker and communicate with it confidentially. There is no such possibility with choice (2) if the eavesdropper can run DHT nodes.

> That's not what I said. Relying on a DHT for some parts, such as bootstrap and discovery is still... well... relying on it, for things it is good at.

> Well yes, but these limits arise naturally anyway since A stores data for B on C and you can't really incentivize C to manage anything more than small bits of data.

Thank you for clarifying. Would you agree that reliable bootstrapping and reliable stead-state behavior are two separate concerns in the application? I'm mainly concerned with the latter; I would never make an application's steady-state behavior dependent on a DHT's ability to keep data available. In addition, bootstrapping information like initial peers and network settings can be obtained through other channels (e.g. DNS servers, user-given configuration, multicasting), which further decreases the need to rely on DHTs.

> About DHTs in general, you can easily make reverse lookups difficult or impossible by hashing the keys (bittorrent doesn't because the inputs already are hashes), you can obfuscate lookups by making them somewhat off-target until they're close to the target and making data-lookups and maintenance lookups indistinguishable. You can further add plausible deniability by by replaying recently-seeing lookups when doing maintenance of nearby buckets.

I'm not quite sure what you're saying here, but it sounds like you're saying that a peer can obfuscate lookups by adding "noise" (e.g. doing additional, unnecessary lookups). If so, then my reply would be this only increases the number of samples an eavesdropper needs to make to unmask a peer. To truly stop an eavesdropper, a peer needs to ensure that queries are uniformly distributed in both space and time. This would significantly slow down the peer's queries and consume a lot of network bandwidth, but it would stop the eavesdropper. I don't know of any production system that does this.

"Host" is generally less ambiguous, referring to a specific thing given the context of the discussion. It's a pronoun for machines (kinda).

In practice trackers do divulge all the same information that can be gleaned from the DHT and so does PEX in a bittorrent swarm. Those are far more convenient to harvest.

> I'm not quite sure what you're saying here, but it sounds like you're saying that a peer can obfuscate lookups by adding "noise" (e.g. doing additional, unnecessary lookups).

That's only 2 of 4 measures I have listed. And I would mention encryption again as a 5th. The others: a) Opportunistically creating decoys by having others repeat lookups they have recently seen as part of their routing table maintenance b) storing data in the DHT in a way that requires some prior knowledge to be useful, which will ideally result in the only leaking information when the listener could obtain the information anyway if he has that prior knowledge.

There's a lot you can do to harden DHTs. I agree that naive implementations are trivial to attack, but to my knowledge it is possible to achieve byzantine fault tolerance in a DHT in principle, it's just that nobody has actually needed that level of defense yet, attacks in the wild tend to be fairly primitive and only succeed because some implementations are very sloppy about sanitizing things.

> To truly stop an eavesdropper, a peer needs to ensure that queries are uniformly distributed in both space and time.

Not quite. You only need to increase the number of samples needed beyond the number of samples a peer is likely to generate during some lifecycle, and that is not just done by adding more traffic.

> Would you agree that reliable bootstrapping and reliable stead-state behavior are two separate concerns in the application?

Certainly, but bootstrapping is a task that you do more frequently than you think. You don't just join a global overlay once, you also (re)join many sub-networks throughout each session or look for specific nodes. DHT is a bit like DNS. You only need it once a day for a domain (assuming long TTLs), and it's not exactly the most secure protocol and afterwards you do the heavy authentication lifting with TLS, but DNS is still important, even if it you're not spending lots of traffic on it.