Thread by @SarahJamieLewis

Since there was some "debate" over the impact of malicious nodes on decentralized consensus this week, I decided to update my consensus microworld to allow byzantine nodes (they always vote for "red")

Interestingly, global consensus falls apart at ~60% honest nodes.

Global consensus is readily compromised up to around 75% honest nodes (with un-coordinated malicious nodes - it seems an obvious result that coordinated malicious nodes could impact local consensus regardless of honest %)

There arises a trade-off in the algorithm between the stickyness of a vote v.s. the influence of other nodes.

Stickyness protects global consensus (so later clusters of red don't flip honest nodes, but also means local clusters are more susceptible to malicious influence)

Any decentralized consensus mechanism must, therefore, define an explicit control which caps the number / distribution of malicious nodes (or explicitly define non-malicious nodes but good luck doing that in a way that won't be labelled as centralization)

In any case, if your entire use case is fast transactions, then even mild local consensus failures are an issue for you (because "mild local consensus failure" is a nice way of saying "doublespend vulnerability")

There are some interesting debates to be had about the relative decentralization (/effective security) of certain sybil-resistance mechanisms.

But you definitely do need a robust sybil-resistance mechanism and that choice does impact your claims of decentralization.

PoW at least has a good argument for theoretical decentralized sybil-resistance (and also many arguments about its practical centralization).

I've yet to see even a theoretical argument of decentralization applied to non-PoW sybil resistance mechanisms (that work)

I feel I should also say at this point that I consider decentralization as a core property that gives cryptocurrency (practical/economic/moral) value.

Decentralization-theater, where cryptocurrencies add layers of complexity to hide the fact that they could have been implemented as a single server php web app and maintained the same security properties, bugs the hell out me.

Decentralization is ultimately about distributing power within a system (fieldnotes.resistant.tech/what-is-decent…). Consensus is ultimately about converging the system to a single decision. They are fundamentally at odds, and that's what makes some of the economics & trade-offs interesting.

(I've now pushed the updates for this model here: git.openprivacy.ca/sarah/microwor…)

Some further notes: both the honest nodes and the malicious nodes in these simulations move randomly throughout the space, meaning they engage with a new set of random neighbors way more often than any practical system could.

I suspect this is a double edged sword: while random walks minimizes the concentration of malicious nodes, they also allow honest nodes to wander into the influence of malicious nodes (allowing them to be corrupted and, as a consequence, contagious)