Evidence

Evidence is an important component of CometBFT’s security model. Whilst the core consensus protocol provides correctness gaurantees for state machine replication that can tolerate less than 1/3 failures, the evidence system looks to detect and gossip byzantine faults whose combined power is greater than or equal to 1/3. It is worth noting that the evidence system is designed purely to detect possible attacks, gossip them, commit them on chain and inform the application running on top of CometBFT. Evidence in itself does not punish “bad actors”, this is left to the discretion of the application. A common form of punishment is slashing where the validators that were caught violating the protocol have all or a portion of their voting power removed. Evidence, given the assumption that 1/3+ of the network is still byzantine, is susceptible to censorship and should therefore be considered added security on a “best effort” basis.

This document walks through the various forms of evidence, how they are detected, gossiped, verified and committed.

NOTE: Evidence here is internal to CometBFT and should not be confused with application evidence

Detection

Equivocation

Equivocation is the most fundamental of byzantine faults. Simply put, to prevent replication of state across all nodes, a validator tries to convince some subset of nodes to commit one block whilst convincing another subset to commit a different block. This is achieved by double voting (hence DuplicateVoteEvidence). A successful duplicate vote attack requires greater than 1/3 voting power and a (temporary) network partition between the aforementioned subsets. This is because in consensus, votes are gossiped around. When a node observes two conflicting votes from the same peer, it will use the two votes of evidence and begin gossiping this evidence to other nodes. Verification is addressed further down.

type DuplicateVoteEvidence struct {
    VoteA Vote
    VoteB Vote

    // and abci specific fields
}

Light Client Attacks

Light clients also comply with the 1/3+ security model, however, by using a different, more lightweight verification method they are subject to a different kind of 1/3+ attack whereby the byzantine validators could sign an alternative light block that the light client will think is valid. Detection, explained in greater detail here, involves comparison with multiple other nodes in the hope that at least one is “honest”. An “honest” node will return a challenging light block for the light client to validate. If this challenging light block also meets the validation criteria then the light client sends the “forged” light block to the node. Verification is addressed further down.

type LightClientAttackEvidence struct {
    ConflictingBlock LightBlock
    CommonHeight int64

      // and abci specific fields
}

Verification

If a node receives evidence, it will first try to verify it, then persist it. Evidence of byzantine behavior should only be committed once (uniqueness) and should be committed within a certain period from the point that it occurred (timely). Timelines is defined by the EvidenceParams: MaxAgeNumBlocks and MaxAgeDuration. In Proof of Stake chains where validators are bonded, evidence age should be less than the unbonding period so validators still can be punished. Given these two propoerties the following initial checks are made.

Has the evidence expired? This is done by taking the height of the Vote within DuplicateVoteEvidence or CommonHeight within LightClientAttakEvidence. The evidence height is then used to retrieve the header and thus the time of the block that corresponds to the evidence. If CurrentHeight - MaxAgeNumBlocks > EvidenceHeight && CurrentTime - MaxAgeDuration > EvidenceTime, the evidence is considered expired and ignored.
Has the evidence already been committed? The evidence pool tracks the hash of all committed evidence and uses this to determine uniqueness. If a new evidence has the same hash as a committed one, the new evidence will be ignored.

DuplicateVoteEvidence

Valid DuplicateVoteEvidence must adhere to the following rules:

Validator Address, Height, Round and Type must be the same for both votes
BlockID must be different for both votes (BlockID can be for a nil block)
Validator must have been in the validator set at that height
Vote signature must be correctly signed. This also uses ChainID so we know that the fault occurred on this chain

LightClientAttackEvidence

Valid Light Client Attack Evidence must adhere to the following rules:

If the header of the light block is invalid, thus indicating a lunatic attack, the node must check that they can use verifySkipping from their header at the common height to the conflicting header
If the header is valid, then the validator sets are the same and this is either a form of equivocation or amnesia. We therefore check that 2/3 of the validator set also signed the conflicting header.
The nodes own header at the same height as the conflicting header must have a different hash to the conflicting header.
If the nodes latest header is less in height to the conflicting header, then the node must check that the conflicting block has a time that is less than this latest header (This is a forward lunatic attack).

Gossiping

If a node verifies evidence it then broadcasts it to all peers, continously sending the same evidence once every 10 seconds until the evidence is seen on chain or expires.

Commiting on Chain

Evidence takes strict priority over regular transactions, thus a block is filled with evidence first and transactions take up the remainder of the space. To mitigate the threat of an already punished node from spamming the network with more evidence, the size of the evidence in a block can be capped by EvidenceParams.MaxBytes. Nodes receiving blocks with evidence will validate the evidence before sending Prevote and Precommit votes. The evidence pool will usually cache verifications so that this process is much quicker.

Sending Evidence to the Application

After evidence is committed, the block is then processed by the block executor which delivers the evidence to the application via EndBlock. Evidence is stripped of the actual proof, split up per faulty validator and only the validator, height, time and evidence type is sent.

enum EvidenceType {
  UNKNOWN             = 0;
  DUPLICATE_VOTE      = 1;
  LIGHT_CLIENT_ATTACK = 2;
}

message Evidence {
  EvidenceType type = 1;
  // The offending validator
  Validator validator = 2 [(gogoproto.nullable) = false];
  // The height when the offense occurred
  int64 height = 3;
  // The corresponding time where the offense occurred
  google.protobuf.Timestamp time = 4 [
    (gogoproto.nullable) = false, (gogoproto.stdtime) = true];
  // Total voting power of the validator set in case the ABCI application does
  // not store historical validators.
  // https://github.com/tendermint/tendermint/issues/4581
  int64 total_voting_power = 5;
}

DuplicateVoteEvidence and LightClientAttackEvidence are self-contained in the sense that the evidence can be used to derive the abci.Evidence that is sent to the application. Because of this, extra fields are necessary:

type DuplicateVoteEvidence struct {
  VoteA *Vote
  VoteB *Vote

  // abci specific information
  TotalVotingPower int64
  ValidatorPower   int64
  Timestamp        time.Time
}

type LightClientAttackEvidence struct {
  ConflictingBlock *LightBlock
  CommonHeight     int64

  // abci specific information
  ByzantineValidators []*Validator
  TotalVotingPower    int64
  Timestamp           time.Time
}

These ABCI specific fields don’t affect validity of the evidence itself but must be consistent amongst nodes and agreed upon on chain. If evidence with the incorrect abci information is sent, a node will create new evidence from it and replace the ABCI fields with the correct information.