owned this note
owned this note
Published
Linked with GitHub
# SSZ Transaction Signature Schemes
Roberto Bayardo
May 1, 2023
## Background
[EIP-4844](https://eips.ethereum.org/EIPS/eip-4844) introduces a new blob transaction type for which [SSZ](https://ethereum.org/en/developers/docs/data-structures-and-encoding/ssz/) encoding was chosen for its network & block representations in place of [RLP](https://ethereum.org/en/developers/docs/data-structures-and-encoding/rlp/). The choice of SSZ kicks off a longstanding desire to migrate the Ethereum execution layer (EL) to this more modern and featureful encoding format, which is expected to yield multiple benefits over time. In the short term, once there are enough SSZ transaction types to cover all transaction use cases, new L2s/zk-EVMs can avoid implementing the legacy RLP transaction support entirely, eliminating significant technical debt.
Transactions in the execution layer have two perpetual identifiers:
* `sig_hash`, a hash of the unsigned transaction data which the signature is based on, and
* `tx_hash`, the signed transaction's hash, used as a unique identifier to refer to the transaction.
We currently lack consensus on how to compute sig_hash and tx_hash for SSZ-encoded transactions. This document attempts to summarize the current state, the various other options being considered, and the concerns surrounding each one.
## Current State
For RLP ethereum transaction types, the sig_hash is computed as:
```
keccak([tx_type_byte] + rlp_encode(tx))
```
And its tx_hash as:
```
keccak([tx_type_byte] + rlp_encode(signed_tx))
```
EIP-4844 (as of May 1, 2023) defines sig_hash and tx_hash for a blob transaction as follows
sig_hash:
```
keccak([0x03] + ssz.serialize(tx.message))
```
tx_hash:
```
keccak([0x03] + ssz.serialize(tx))
```
This approach borrows heavily from the existing RLP transaction signature scheme, providing the following advantages:
* conceptually simple to understand & implement without much new code since it follows the previous EL signing idiom
* requires bringing only SSZ serialization/deserialization into the EL at this time instead of a full blown implementation that includes merkelization
* Note: we will need full-blown merkelization in the EL eventually in support of broader SSZ adoption
This simplicity however is countered by the following drawbacks:
* it is non-idiomatic with other SSZ-based signature schemes which are over hash_tree_root of the SSZ object instead of a hash of the flat serialized encoding
* it has some risk of causing hash collisions with txs from other chains that may choose to introduce their own type-3 tx type that is instead RLP encoded
* Note: This could potentially be narrowly mitigated by adding an extra byte or bytes after the type prefix to guarantee collision resistance with any type-3 RLP encoded tx.
* it forgoes any of the potential benefits of SSZ merkelization, such as ability to use the signature to provide efficient proofs over specific elements of a transaction.
* Note: because signing hash is typically ephemeral and only appears during tx validation, it's arguable whether merkelization in this context would prove useful. [EIP-6404](https://eips.ethereum.org/EIPS/eip-6404) describes how such benefits might be unlocked that explicitly treats transaction signatures as opaque.
## Option 1: EIP-6493
[EIP-6493](https://eips.ethereum.org/EIPS/eip-6493) proposes an SSZ transaction signature scheme consistent with how SSZ objects are signed in the Ethereum consensus layer, involving idiomatic usage of the hash_tree_root combined with a domain that prevents collisions with signatures from other domains (see [compute_signing_root](https://github.com/ethereum/consensus-specs/blob/dev/specs/phase0/beacon-chain.md#compute_signing_root)).
This approach resolves the concerns with the current state, though not without new drawbacks. The main concern seems to be:
* ~~**t brings significant complexity into the EL that may have made sense in the CL but may not provide value for this purpose (e.g. forkid as part of the domain)~~
_Update: As of May 3 2023, EIP-6493 has been updated and now largely resembles option 2 below._
## Option 2: EL-Specific Idiomatic SSZ Tx Signature Scheme
We can imagine a new idiomatic ssz signature scheme for transactions that more narrowly attempts to address the specific concerns of the current state without bringing in a full CL-inspired signature scheme. For example, consider:
```
class TxSigningData(Container):
tx_type: Bytes1
object_root: Root
sig_hash(BlobTransaction: tx):
return hash_tree_root(TxSigningData(0x03, hash_tree_root(tx.message))
tx_hash(BlobTransaction: tx):
return hash_tree_root(tx)
```
Because the top-level hash function here is no longer keccak, the probability of collision with any keccak-based hash is negligible. We preserve usage of the tx-type byte as a sort of execution layer transaction domain prefix for preventing collisions over future SSZ transaction types. Cross-chain replays are prevented by the fact that blob-tx type has chain_id as its first element.
If we imagine a future where we move entirely to SSZ type transactions, this approach allows removing of all legacy RLP cruft pertaining to transaction encoding/decoding. While this may not benefit ethereum mainnet any time soon since we expect old transactions should be forever interpretable, it could benefit, say, L2s which are bootstrapped at a point where SSZ transactions provide all the necessary functionality.
## Option 3: RLP-based Signature Scheme
Another option would be to retain use of SSZ for network & block representations of a blob transaction, but to serialize the blob transaction's contents to RLP purely for the purpose of signing & tx id computation.
The obvious drawback having to continue to rely on legacy RLP serialization (but not deserialization) in order to validate the signature of an SSZ transaction type. If we imagine a future where we deprecate older transactions and move entirely to SSZ based types, this approach prevents us from removing legacy RLP cruft pertaining to transaction encoding/decoding. While this may not affect ethereum mainnet any time soon since we expect old transactions to be forever interpretable, it limits the benefit noted in our introduction around post-SSZ bootstrapped L2s being able to avoid significant RLP related technical debt.
### Aside #1: Move chain_id from the tx itself to the signing envelope?
Right now chain_id, whose purpose is *(only? predominantly?)* to prevent tx replay across chains, is within the blob_tx itself. Should this field instead be moved to the "signing envelope? As examples, consider the current state. If we removed chain-id from the blob-tx, we'd instead want to sign:
```
keccak([0x3] + uint_to_bytes(uint32(chain_id)) + ssz.serialize(tx.message))
```
Note that if all other chains were to adopt this new style signing prefix, there would be replay protection across chains regardless of the tx-format chosen. (The scheme would still be vulnerable to bad actors who fork without introducing a new chain_id however.)
For option 2 we'd extend TxSigningData as:
```
class TxSigningData(Container):
tx_type: Bytes1
chain_id: Bytes32
object_root: Root
```
Potential concerns:
* Moving chain_id to the ephemeral signing envelope would prohibit validating a transaction's signature of known type without knowing the chain & the specific block in the chain that it comes from. (Forked chains can have multiple chain_ids depending on block height.)
### Aside #2: Opaque 65-byte signature for the signed blob transaction?
The signed blob tx still has a signature composed of sub-elements v,s, and r which are (de)serialized individually even though their only use is to re-compact them back into a 65-byte array expected by the signature validation / address recovery function. This approach has been abused to some extent to encode chain_id into older transaction types by hacking it into the "V" value's higher order bits, but is no longer relevant for newer tx types where chain_id is encoded explicitly (whether within the tx itself or within the signing envelope).
We could consider treating the signature instead as an opaque 65 byte array with respect to its SSZ encoding, and the implementation could then choose a representation that precisely matches the expectations of the signature verification / address recovery API.