This repository has now been merged into the cggmp-threshold-ecdsa repository and any improvements and updates will happen there. Consider this repository deprecated.
In this note we aim to re-purpose the Fouque-Stern Distributed Key Generation (DKG) to support a secure Distributed Key Refresh (DKR). As we claim, FS-DKR is well suited for rotation of threshold ECDSA keys.
The FS-DKG protocol is a one round DKG based on Publicly Verifiable Secret Sharing (PVSS) and the Paillier cryptosystem. There are two major security shortcomings to FS-DKG:
- It introduces a factoring assumptions (DCRA)
- it is insecure against rushing adversary
Rushing adversary is a common assumption in Multiparty Computation (MPC). In FS-DKG, an adversary waiting to receive messages from all other parties will be able to decide on the final public key. In the worst case it can lead to a rogue-key attack, giving full control of the secret key to the attacker. This is the main reason, in our opinion, why FS-DKG, altough with prominent features, was over-looked for the past 20 years. in this write-up we show how by adjusting FS-DKG to key rotation for threshold ecdsa the above shortcomings are avoided.
We use standard proactive security assumptions. The protocol will be run by
Each party calls RefreshMessage::distribute(key) on their LocalKey and broadcasts the RefreshMessage while saving their new DecryptionKey.
After recieving all the refresh messages each party calls RefreshMessage::collect(..) with a vector of all the refresh messages, a mutable reference to their own key, and their new DecryptionKey, This will validate all the refresh messages, and if all the proofs are correct it will update the local key to contain the new decryption keys of all the parties.
Example:
// All parties should run this
let mut party_i_key: LocalKey<_>;
let (party_i_refresh_message, party_i_new_decryption_key) = RefreshMessage::distribute(party_i_key);
broadcast(party_i_refresh_message);
let vec_refresh_messages = recv_from_broadcast();
RefreshMessage::collect(&vec_refresh_messages, &mut party_i_key, party_i_new_decryption_key, &[])?;Each party that wants to join first generates a JoinMessage via JoinMessage::distribute() and broadcasts it to the current parties.
The existing parties choose the index(who are they replacing) for the joining party.
Note that this part is delicate and needs to happen outside of the library because it requires some kind of mutual agreement, and you cannot trust the new party to communicate which party are they replacing.
After agreeing on the index each party modifies the join message to contain the index join_message.party_index = Some(index).
Each existing party calls RefreshMessage::replace(join_message, local_key) with the join message and its own local key, this returns a refresh message and a new decryption key, just like in a Key Refresh, and they all broadcast the RefreshMessage.
Each existing party recieves all the broadcasted refresh messages and calls RefreshMessage::collect(..) with a vector of all the refresh messages, a mutable reference to their own key, the new DecryptionKey, and a slice of all the join messages(JoinMessage)
This will validate both the refresh messages and the join messages and if all the proofs are correct it will update the local key both as a refresh(new decryption keys) and replace the existing parties with the new ones.
The new party calls join_message.collect(..) with the broadcasted RefreshMessage of the existing parties and all the join messages which returns a new LocalKey for the new party.
Example:
// PoV of the new party
let (join_message, new_party_decryption_key) = JoinMessage::distribute();
broadcast(join_message);
let new_party_index = recv_broadcast();
let vec_refresh_messages = recv_from_broadcast();
let new_party_local_key = join_message.collect(&vec_refresh_messages, new_party_decryption_key, &[])?;
// PoV of the other parties
let mut party_i_key: LocalKey<_>;
let mut join_message = recv_broadcast();
let new_index = decide_who_to_replace(&join_message);
broadcast(new_index);
assert!(recv_from_broadcast().iter().all(|index| index == new_index));
join_message.party_index = Some(new_index);
let (party_i_refresh_message, party_i_new_decryption_key) = RefreshMessage::replace(join_message, party_i_key)?;
broadcast(party_i_refresh_message);
let vec_refresh_messages = recv_from_broadcast();
RefreshMessage::collect(&vec_refresh_messages, &mut party_i_key, party_i_new_decryption_key, &[join_message])?;Here we give a short description of the FS-DKG protocol.
FS-DKG works in one round. This round includes a single broadcast message from each party
Verification proceeds as follows. Each party
- all broadcasted proofs of fairness
- all secret sharing schemes - computing the polynomial points "at the exponent"
The parties define the set
We will now highlight the adjustments required for FS-DKR. In a key refresh protocol the parties start with their inputs equal to the outputs of a DKG done in the past or the output of previous DKR. Meaning, as opposed to FS-DKG protocol in which the inputs are pseudorandom such that the attacker can bias the output, for example in a rushing adversary attack, FS-DKR avoids this potential attack on FS-DKG because of the added restriction over the inputs of the attacker. Concretely, in the case the parties must reshare their DKG/DKR output secret share, all other parties already know a public commitment to the attacker secret share and can check for it. Recall that FS-DKG is secure assuming Paillier is secure (what we called DCRA assumption). Moreover, we assumed a setup phase in which all parties generate paillier keys and share them. This fits well with threshold ECDSA: First, GG20 already requires us to assume Paillier security, therefore in this particular case, no new assumption is needed. The setup phase actually happens as part of GG20 DKG. We will use this to our advantage, running the FS-DKR using the GG20-DKG paillier keys. Obviously because we need to refresh the paillier keys as well we will also add a step to FS-DKR to generate new paillier keys and prove they were generated correctly. This is a standard proof, that can be made non-interactive. See the zk-paillier lib for an implementation.
Adding/Removing parties: There is a clear distinction between parties with secret shares (”Senders”) and new parties (”Receivers”). The FS-DKR protocol therefore supports adding and removing parties in a natural way: Define
Identifiable Abort: A nice property of FS-DKR is that if a party misbehaves all honest parties learn about it. This is due to the nature of PVSS used in the protocol. As GG20, our reference threshold ECDSA protocol, also have this property, it is important that identifiable abort can be guaranteed throughout the DKR as well.
For completeness, Below is the FS-DKR protocol, written as FS-DKG with changes in red for DKR. 
Our main requirement from FS-DKR is minimal round-count. In FS-DKR the parties can pre-process all the data they need to send. Our main bottleneck is
In this section we focus on alternative protocols for DKR. Three recent results come to mind. The first one, CGGMP20, is another threshold ECDSA protocol with a companion refresh protocol, see figure 6 in the paper. Their protocol has the most resemblance to FS-DKR, with few notable differences. First, while FS-DKR is publicly verifiable, CGGMP20-DKR current version suffers from a technichal issue with its Identifiable Abort (acknowledged by the authors). Second, the paillier keys used in the CGGMP20-DKR are the new ones, while in FS-DKR, we use the old ones, already known to all, which helps us save a round of communication. Finally, CGMMP20-DKR key refresh is done by adding shares of zero while in FS-DKR we re-share existing shares. Overall we treat the similarities between the protocols as a positive signal of validation for FS-DKR. A second protocol, by Gurkan et. al., uses gossip for aggregating transcripts from the parties. However, their DKG is generating group elements secret shares and we need field elements secret shares for our threshold ECDSA. The third relevant work is Jens Groth' Non interactive DKG and DKR. There, instead of paillier encryption, they use El-Gamal based encryption scheme that offers forward security. Their DKR makes the assumption that the El-Gamal decryption keys are long-term and not rotated. This assumption seems crucial for the Groth-DKG construction. In our context it means that we need to let the parties generate, store and use a new set of keypair,in addition to the Paillier keypair, and that this new keypair poses a security risk against the classical mobile adversary, which our model does not allow. As opposed to Groth-DKR, FS-DKR is reusing the existing paillier keypair and rotate it as well. In terms of efficiency - there is no complexity analysis given in the paper, however, from inspection we estimate the asymptotic complexity is comparable to FS-DKR (quadratic in the number of parties).
We thank Claudio Orlandi, Kobi Gurkan and Nikolaos Makriyannis for reviewing the note