https://twitter.com/kanzure/status/1090777875253407745

This is joint work with George Danezis who couldn't be here today and also cofounded Chainspace. Among all the possible challenges we may have on a blockchain, this talk will be about how to build strong sybil resistance. We would consider a subset of the problem, how will we bootstrap a federated agreement system?

I am sure everyone knows about sybil attacks. Imagine you have a classic setting where you have an attacker creating fake identities. This is a sybil attack and eventually these nodes take over the system.

Traditionally you cap the ability of the adversary to create multiple identities. Traditional defenses include proof-of-work, which leverages scarce resources to put a cost on the sybil attack and creating each identity. We could force an adversary to burn or lock some of their money. But this means that it's not free for other users to join the system as well. In this case, the adversary would then need to be rich.

These attacks can become very real because the transactions can have large financial consequences. So an attacker can be very financially motivated, and there can be attacks where the adversary can financially come out on top by doing the attacks.

The key idea is still to leverage scarce resources, and require adversaries to burn or lock money. But we have a new concept: we also leverage trust, which is a new kind of scarce resource. We leverage trust by penalizing poor judgement between nodes. This is the core of how sybilquorum works and this is what we're going to talk about today.

This is generally how it works. There's two components. The first is traditional proof-of-stake. Then there's social network analysis. We bring the two together. Instead of locking some stake on particular social links, ... we ... and then we use social network analysis to do statistical analysis of relationships.

The steps are to attribute weights to people you trust. The goal is to have a stake-weighted trust relationship graph. This is the node's view of the network may look like. He will lock money on the links with other nodes. It does it in a particular way. Once you lock your money to a particular link, any vertex of the link can withdraw the money and take it out and add it to their accounts. This is the key idea of how we leverage trust and punish poor judgement. We imagine some nodes might put some money on a non-trustworthy node, which might take the money from the link and then disappear.

So bulk dishonesty protects against strategic dishonest. Say we have an adversary that creates a lot of sybils and puts a lot of money on the links. But if you're an honest node, you're going to be super careful about putting your money on a link with someone you don't know. Now we know that nodes will be much more careful and therefore their links to these malicious regions would be more rare.

After that, you run social network analysis. This relies on the fast mixing assumption. The idea is to say let's imagine the network looks like this. It's an idealistic attack where we had a lot of very well connected sybil region in grey, and then you have the normal network with all the honest nodes and potentially some are dishonest but they are not performing a sybil attack in theory. So the fast mixing assumption says mainly two things- the first is that if you're a node that wants to join the network, you will integrate into the network quite fast. It would be a long integration of the sybils connected to each other, to the real network. So we assume that people will trust one of the sybils or something, and this is an important assumption we use here. Slow 2 would work as following- each node performs a local judgement. It's the nodes view of the network, and it inputs this into a black box algorithm. This black box algorithm is the social network analysis and I'll come to that in a moment. The output of that would be a map between the nodes and weights. It's the probability of a node being a sybil. The black box in our case is Sybilinfer, SybilGuard, SybilLimit and other examples exist.

In the interest of time, I won't explain how SybilInfer works. But the general idea is that if you start with an honest node in the green here, and you take a short random walk in the network, when you end up with a higher chance of ending up with non-sybil nodes. If you start in the sybil region, you will take a random walk and in high probability you will stay in the sybil region. You can distinguish these regions statistically. If you choose your distributions right, you would have ... I'm happy to chat with you about that after the talk if you're interested in that.

Then you determine the quorum slices.

This is still work in progress. We want to evaluate the number of sybil nodes, number of links or stake between sybils, and number of links or stake between nodes and sybils, and also what fraction of naive nodes are in the system? What happens if you increase or decrease the stake that the adversary puts at the links?

Sybilquorum is a sybil resistant mechanism that leverages money by forcing nodes to lock it or to burn it, as is tradition in proof-of-stake. Instead of locking it on the node itself, you enter the system and lock it into links. The second thing it leverages is trust by penalizing poor judgement. If you connect to a fraudster, he can take the money from the link and disappear. We use proof-of-stake, weighted graphs, and social network analysis to determine sybil regions probabilistically.