7 ms·

> 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

• the8472 10y ago
  > I'm confused. I can configure a tracker to only communicate with trusted peers, and do so over a confidential channel. The tracker is assumed to not leak peer information to external parties. A DHT can do neither of these. But then you are running a private tracker for personal/closed group use and have a trust source. If you have a trust source you could also run a closed DHT. But the bittorrent DHT is public infrastructure and best compared to public trackers. > I'm assuming that peers are arbitrarily long-lived. Real-world distributed systems like BitTorrent and Bitcoin aspire to this. Physical machines are. Their identities (node IDs, IP addresses) and the content they participate in at any given time don't need to be. > If I can Sybil a DHT, then I can spin up arbitrarily many evil nodes. 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. > Are we assuming that no more than one third of the DHT's nodes are evil? Assuming is the wrong word. I think approaching BFT is simply part of what you do to harden a DHT against attackers. > Second, "nobody has actually needed that level of defense yet" does not mean that it is a sound decision for an application to use a DHT with the expectation that the problems will never occur. 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. > Neither (a) nor (b) address this; they simply increase the number of samples required. (b) is orthogonal to sampling vs. noise. > I'm not okay with using it as the sole mechanism or even the primary mechanism, since they're so easy to break when compared to other mechanisms. 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.
  • jude- 10y ago
    > But then you are running a private tracker for personal/closed group use and have a trust source. If you have a trust source you could also run a closed DHT. But the bittorrent DHT is public infrastructure and best compared to public trackers. 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-aad748f46d 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