> Someone who is "in" on encrypted content can observe the swarm anyway, thus gains very little from performing snooping on a DHT.
I can see that this thread is getting specific to Bittorrent, and away from DHTs in general. Regardless, I'm not sure if this is the case. Please correct me if I'm wrong:
* If I can watch requests on even a single copy of a single key/value pair in the DHT, I can learn some of the IP addresses asking for it (and when they ask for it).
* If I can watch requests on all copies of the key/value pair, then I can learn all the interested IP addresses and the times when they ask.
* If I can do this for the key/value pairs that make up a .torrent file, then I can (1) get the entire .torrent file and learn the list of file hashes, and (2) find out the IPs who are interested in the .torrent file.
* If I can then observe any of the key/value pairs for the .torrent file hashes, then I can learn which IPs are interested in and can serve the encrypted data (and the times at which they do so).
This does not strike me as "quite small," but that's semantics.
> There is no bumping in kademlia with unbounded node storage. And clients with limited storage can make bumping very hard for others with oldest-first and one-per-subnet policies, i.e. bumping the attackers instead of genuine keys.
Yes, the DHT nodes can employ heuristics to try to stop this, just like how BEP42 is a heuristic to thwart Sybils. But that's not the same as solving the problem. Applications that need to be reliable have to be aware of these limits, and anticipate them in their design.
> No, they should use DHT as a bootstrap mechanism of easy-to-replicate, difficult-to-disrupt small bits of information (e.g. peer contacts as in bittorrent) which then run their own content-specific gossip network for the critical content. In some contexts it can also make sense to make reverse lookups difficult, so attackers won't know what to disrupt unless they're already part of some group.
This kind of proves my point. You're recommending that applications not rely on DHTs, but instead use their own content-specific gossip network.
To be fair, I'm perfectly okay with using DHTs as one of a family of solutions for addressing one-off or non-critical storage problems (like bootstrapping). But the point I'm trying to make is that they're not good for much else, and developers need to be aware of these limits if they want to use a DHT for anything.
EDIT: formatting