-
Notifications
You must be signed in to change notification settings - Fork 41
Mithril Certificate Discussions
This document covers an open discussion on some important details required to build the certificates in general, and we give particular importance into how to create the certificate chain. More specifically, this document presents some ideas on the following topic:
- Role of the aggregator,
- Certificate granularity,
- Not collect enough signatures, and
- How messages to be signed are obtained,
Open questions:
- Define a rule to handle the case where we receive valid certificates for two different messages in a single "mithril epoch".
- Define a rule to handle mithril skipping a full Cardano epoch. This is unlikelly, but we should define what to do in that case.
- Is it an undesirable burden for users to produce themselves the message to be signed?
What would happen if some aggregator claims not enough signatures were received? This doesn’t really matter, as there will be a different aggregator that will collect sufficient signatures and aggregate them into a valid certificate. Similarly, different aggregators might have different views of the signatures submitted (one aggregator might receive 10 signatures, and a different one could receive 11), which would result in different certificates signing the same message. Again, this is not a problem, as long as they certify the same message (we should specify what to do if we receive more than one message signed in the same "mithril epoch").
As a result, aggregator should not be responsible for distributing the snapshot. There should be a separate entity that receives aggregated certificates, and then distributes the snapshot.
There is no requirements on the certificate granularity. However, it is desirable that there always is at least one certificate per new stake distribution. Given that in Cardano the stake distribution is updated at each epoch, we should aim towards guaranteeing one valid certificate per epoch. However, given that certificate generation is pseudorandom, there exists the possibility of skipping some certificates (we present below the probability of this happening). Therefore, having more than one certificate per epoch seems a good option, specially because when bootstrapping a new node, one only needs to verify one certificate per epoch.
Recall that when you are bootstrapping you only really care about the signature of the next stake distribution (and not about the state of the node). Therefore, if we have more than one certificate per epoch, when we are bootstrapping we don’t need to verify all these certificates. We only need to verify the certificates that certify the next stake distribution, and not those that maintain the distribution of the previous certificate. A node that is bootstrapping only really needs to verify the first certificate (or the last, really depends how we organise the message signed), of each different stake distribution, hence of each epoch. The other certificates are not required, as what a node is doing while bootstrapping is verifying the stake changes to be able to validate the next certificate. Therefore, even if we produce 10 certificates per epoch, when a node is bootstrapping it can verify one certificate per epoch, until the last one, where it verifies all existing certificates.
As mentioned above, the creation of certificates is pseudorandom, so we
might find ourselves in a situation where the submitted signatures are not
enough. As presented in the paper, the goal is to make the probability of this
happening as small as possible, while guaranteeing that an adversary will not
be able to generate invalid certificates. Table 2 of the paper is interesting
to discuss this (it shows required values of k, n so that an adversarial
quorum is formed with P ≤ 2 ^{−128}). The table shows different values, but if
we take for instance φ(.55), and assume adversarial stake of 40%, the
probability of reaching a quorum is 99.999% (assuming adversarial stake does not
participate). If we consider an adversarial stake of 33%, then with φ(.60) we
would reach a quorum with probability of 99.667%. We can also think of adversarial
stake as nodes which are not available.
Nonetheless, what to do when these events happen, and we don’t have a certificate.
The paper mentions the possibility of using a counter:
"if the probability of an honest quorum remains significant it can be boosted by
allowing retries (e.g by attaching a short counter to the message)".
However, there would also be the possibility of skipping a certificate, and
signing the next one. The probabilities of not reaching a quorum are so small
that it is unlikely that we will skip several blocks in a row. Skipping one (or
even a few) block is not problematic (as long as we have one per epoch, in the
opposite case we should reconsider what to do, as this becomes a bit uglier wrt
long range attacks).
The other mitigation approach is to have a sort of hybrid parametrization, where we would run a big (k,m) pair and a small one at the same time. If aggregators can find enough signatures in the small param set, they prefer that, if not they use the larger params. Might be clunkier for rewards though.
If we allow signers to also be aggregators, and therefore listen and accumulate signatures, then they will know how many are being received and whether the quorum is being reached or not. If we use the index solution proposed above, a signer could get a grasp of whether the quorum is going to be reached or not, and start trying with the next index. If we don’t follow the index solution, then signers don’t really care whether the quorum was reached or not, they simply proceed with the next signature.
The message to be signed is chosen by the signer itself. It is not received externally. If I’m a signer, I should have sufficient information to select what message I need to sign--this is, the stake distribution and aggregate key, and the state of the node at the block number for which we are producing a signature. Given that choosing the message is a deterministic procedure, all honest nodes will generate a signature for the same message. Following this, there is no room for adversaries to request signatures for adversarial messages, nor the need for signers to validate the message to be signed.
To guarantee that users have the same view of what is being signed, the snapshots are created past k blocks (security parameter), which already gives high guarantees we have reached consensus.