An off-grid social network

You're ignoring the fact that with a public DHT, the eavesdropper has the power to reroute requests through networks (s)he can already watch. With a public tracker, the eavesdropper needs vantage points in the tracker's network to gain the same insights.

If we're going to do an apples-to-apples comparison between a public tracker and a public DHT, then I'd argue that they are equivalent only if:

(1) the eavesdropper cannot add or remove nodes in the DHT; (2) the eavesdropper cannot influence other nodes' routing tables in a non-random way.

> This can be made costly. In the extreme case you could require a bitcoin-like proof of work system for node identities. But that would be wasteful... unless you're running some coin network anyway, then you can tie your ID generation to that. In lower-value targets IP prefixes tend to be costly enough to thwart attackers. If an attacker can muster the resources to beat that he would also have enough unique machines at his disposal to perform a DoS on more centralized things.

Funny you should mention this. At the company I work part-time for (blockstack.org), we thought of doing this very thing back when the system still used a DHT for storing routing information.

We had the additional advantage of having a content whitelist: each DHT key was the hash of its value, and each key was written to the blockchain. Blockstack ensured that each node calculated the same whitelist. This meant that inserting a key/value pair required a transaction, and the number of key/value pairs could grow no faster than the blockchain.

This was not enough to address data availability problems. First, the attacker would still have the power to push hash buckets onto attacker-controlled nodes (it would just be expensive). Second, the attacker could still join the DHT and censor individual routes by inserting itself as neighbors of the target key/value pair replicas.

The best solution we came up with was one whereby DHT node IDs would be derived from block headers (e.g. deterministic but unpredictable), and registering a new DHT node would require an expensive transaction with an ongoing proof-of-burn to keep it. In addition, our solution would have required that every K blocks, the DHT nodes would deterministically re-shuffled their hash buckets among themselves in order to throw off any encroaching routing attacks.

We ultimately did not do this, however, because having the set of whitelisted keys growing at a fixed rate afforded a much more reliable solution: have each node host a 100% replica of the routing information, and have nodes arrange themselves into a K-regular graph where each node selects neighbors via a random walk and replicates missing routing information in rarest-first order. We have published details on this here: https://blog.blockstack.org/blockstack-core-v0-14-0-release-....

> Assuming is the wrong word. I think approaching BFT is simply part of what you do to harden a DHT against attackers.

If you go for BFT, you have to assume that no more than f of 3f+1 nodes are faulty. Otherwise, the malicious nodes will always be able to prevent the honest nodes from reaching agreement.

> I haven't said that. I'm saying that simply because this kind of defense was not yet needed nobody tried to build it, as simple as that. Sophisticated security comes with implementation complexity, that's why we had HTTP for ages before HTTPS adoption was spurred by the snowden leaks.

Right. HTTP's lack of security wasn't considered a problem, until it was. Websites addressed this by rolling out HTTPS in droves. I'm saying that in the distributed systems space, DHTs are the new HTTP.

> What other mechanisms do you have in mind? Most that I am aware of don't offer the same O(log n) node-state and lookup complexity in a distributed manner.

How about an ensemble of bootstrapping mechanisms?

* give the node a set of initial hard-coded neighbors, and maintain those neighbors yourself.

* have the node connect to an IRC channel you maintain and ask an IRC bot for some initial neighbors.

* have the node request a signed file from one of a set of mirrors that contains a list of neighbors.

* run a DNS server that lists currently known-healthy neighbors.

* maintain a global public node directory and ship it with the node download.

I'd try all of these things before using a DHT.

EDIT: formatting

But in the context of bittorrent that is not necessary if we're still talking about information leakage. The tracker + pex gives you the same, and more, information than watching the DHT.

> we thought of doing this very thing back when the system still used a DHT for storing routing information.

The approaches you list seem quite reasonable when you have a PoW system at your disposal.

> have each node host a 100% replica of the routing information, and have nodes arrange themselves into a K-regular graph

This is usually considered too expensive in the context of non-coin/-blockchain p2p networks because you want nodes to be able to run on embedded and other resource-constrained devices. The O(log n) node state and bootstrap cost limits are quite important. Otherwise it would be akin to asking every mobile phone to keep up to date with the full BGP route set.

> assume that no more than f of 3f+1 nodes are faulty. Otherwise, the malicious nodes will always be able to prevent the honest nodes from reaching agreement.

Of course, but for some applications that is more than good enough. If your adversary can bring enough resources to bear to take over 1/3rd of your network he might as well DoS any target he wants. So you would be facing massive disruption anyway. I mean blockchains lose some of their security guarantees too once someone manages to dominate 1/2 of the mining capacity. Same order of magnitude. It's basically the design domain "secure, up to point X".

> I'm saying that in the distributed systems space, DHTs are the new HTTP.

I can agree with that, but I think the S can be tacked on once people feel the need.

> How about an ensemble of bootstrapping mechanisms?

The things you list don't really replace the purpose of a DHT. A dht is a key-value store for many keys and a routing algorithm to find them in a distributed environment. What you listed just gives you a bunch of nodes, but no data lookup capabilities. Essentially you're listing things that could be used to bootstrap into a DHT, not replacing the next layer services provided by a DHT.

Not in a long time. I haven't found a situation in which I couldn't use Postfix in quite a while. Although the occasional sendmail.cf flashback still hits me.