Hijiri algorithms

Status	NOT STARTED
Stakeholders	Makoto Takemiya
Outcome
Due date
Owner	Ivan Rybin

Background

The hijiri reputation system is based on rounds. At each round, validating peers that are registered with the membership service perform the following tasks to establish trust (reliability) ratings for the peers:

data throughput test
version test
computational test
data consistency test

Which peers validate each other are based on the pairwise distance between hashes (e.g., sort(abs(hash && 0x0000ffff - publicKey && 0x0000ffff))). The hashes are computed based on the public keys of the peers that are concatenated with the round number and then SHA-3 hashed. Rounds occur whenever the Merkle root is less than some number (TODO:XXX). Results are shared in a separate Merkle tree, maintained independently of the transactions (so the systems can run in parallel).

Problem

Research hijiri algorithms and create literature review.

Solution

Decisions

As a decision I propose is trying to use both ideas from Incremental Eigentrust and metrics related to performance from GossipSub.

Alternatives

The most widely used general solution is Eigentrust++[1]. It relies upon the metrics of successful requests divided by successful and failed requests. Every peer propagates its scores to other peers and weighted transitive foreign scores are used to calculate trust to other peers.

In the stacking world of course stacking itself is used for reputation[2] although it can be used in other consensus approaches[3].

There is a class of blockchains which use simple local unshared trust:
- This kind of solution uses Substrate[4]. Every peer has a per peer trust counter which is not shared. If peer does something forbidden or not answering for a long time then its trust goes to minimum and the peer disconnects from the bad peer.
- GossipSub messaging solution[5] uses interesting formula for getting local trust. It uses several weighted parameters: time in mesh, first message deliveries, mesh message delivery rate, message delivery failures, invalid messages, application specific score, ip address collocation factor.

There are many variations of Eigentrust++:
- Incremental Eigentrust[6] uses snapshot comparison to reduce redundant calculations at each reputation cycle.
- HonestPeer[7] uses local trust to choose some honest peer (peer with maximum trust) and pulls data from it.
- AuthenticPeer++[8] uses Incremental Eigentrust relative to file storage.

Concerns

We should think about several issues. The main priority is to improve performance of the network as we have consensus algorithm to taking care of security. Performance itself is reducible to performance of the network and performance of the reputation system. We should minimize time spent in both sending messages and computing reputation, but generally sending messages takes more time then local calculation of reputation.

Assumptions

Assumptions is made that we need to address performance of the network and that local calculations take way less time then sending messages by network.

Page tree