[bitcoin-dev] Building Blocks of the State Machine Approach to Consensus

[Peter Todd]

In light of Ethereum's recent problems with its imperative, account-based,
programming model, I thought I'd do a quick writeup outlining the building
blocks of the state-machine approach to so-called "smart contract" systems, an
extension of Bitcoin's own design that I personally have been developing for a
number of years now as my Proofchains/Dex research work.


# Deterministic Code / Deterministic Expressions

We need to be able to run code on different computers and get identical
results; without this consensus is impossible and we might as well just use a
central authoritative database. Traditional languages and surrounding
frameworks make determinism difficult to achieve, as they tend to be filled
with undefined and underspecified behavior, ranging from signed integer
overflow in C/C++ to non-deterministic behavior in databases. While some
successful systems like Bitcoin are based on such languages, their success is
attributable to heroic efforts by their developers.

Deterministic expression systems such as Bitcoin's scripting system and the
author's Dex project improve on this by allowing expressions to be precisely
specified by hash digest, and executed against an environment with
deterministic results. In the case of Bitcoin's script, the expression is a
Forth-like stack-based program; in Dex the expression takes the form of a
lambda calculus expression.


## Proofs

So far the most common use for deterministic expressions is to specify
conditions upon which funds can be spent, as seen in Bitcoin (particularly
P2SH, and the upcoming Segwit). But we can generalize their use to precisely
defining consensus protocols in terms of state machines, with each state
defined in terms of a deterministic expression that must return true for the
state to have been reached. The data that causes a given expression to return
true is then a "proof", and that proof can be passed from one party to another
to prove desired states in the system have been reached.

An important implication of this model is that we need deterministic, and
efficient, serialization of proof data.


## Pruning

Often the evaluation of an expression against a proof doesn't require all all
data in the proof. For example, to prove to a lite client that a given block
contains a transaction, we only need the merkle path from the transaction to
the block header. Systems like Proofchains and Dex generalize this process -
called "pruning" - with built-in support to both keep track of what data is
accessed by what operations, as well as support in their underlying
serialization schemes for unneeded data to be elided and replaced by the hash
digest of the pruned data.


# Transactions

A common type of state machine is the transaction. A transaction history is a
directed acyclic graph of transactions, with one or more genesis transactions
having no inputs (ancestors), and one or more outputs, and zero or more
non-genesis transactions with one or more inputs, and zero or more outputs. The
edges of the graph connect inputs to outputs, with every input connected to
exactly one output. Outputs with an associated input are known as spent
outputs; outputs with out an associated input are unspent.

Outputs have conditions attached to them (e.g. a pubkey for which a valid
signature must be produced), and may also be associated with other values such
as "# of coins". We consider a transaction valid if we have a set of proofs,
one per input, that satisfy the conditions associated with each output.
Secondly, validity may also require additional constraints to be true, such as
requiring the coins spent to be >= the coins created on the outputs. Input
proofs also must uniquely commit to the transaction itself to be secure - if
they don't the proofs can be reused in a replay attack.

A non-genesis transaction is valid if:

1. Any protocol-specific rules such as coins spent >= coins output are
   followed.

2. For every input a valid proof exists.

3. Every input transaction is itself valid.

A practical implementation of the above for value-transfer systems like Bitcoin
could use two merkle-sum trees, one for the inputs, and one for the outputs,
with inputs simply committing to the previous transaction's txid and output #
(outpoint), and outputs committing to a scriptPubKey and output amount.
Witnesses can be provided separately, and would sign a signature committing to
the transaction or optionally, a subset of of inputs and/or outputs (with
merkle trees we can easily avoid the exponential signature validation problems
bitcoin currently has).

As so long as all genesis transactions are unique, and our hash function is
secure, all transaction outputs can be uniquely identified (prior to BIP34 the
Bitcoin protocol actually failed at this!).


## Proof Distribution

How does Alice convince Bob that she has done a transaction that puts the
system into the state that Bob wanted? The obvious answer is she gives Bob data
proving that the system is now in the desired state; in a transactional system
that proof is some or all of the transaction history. Systems like Bitcoin
provide a generic flood-fill messaging layer where all participants have the
opportunity to get a copy of all proofs in the system, however we can also
implement more fine grained solutions based on peer-to-peer message passing -
one could imagine Alice proving to Bob that she transferred title to her house
to him by giving him a series of proofs, not unlike the same way that property
title transfer can be demonstrated by providing the buyer with a series of deed
documents (though note the double-spend problem!).


# Uniqueness and Single-Use Seals

In addition to knowing that a given transaction history is valid, we also want
to know if it's unique. By that we mean that every spent output in the
transaction history is associated with exactly one input, and no other valid
spends exist; we want to ensure no output has been double-spent.

Bitcoin (and pretty much every other cryptocurrency like it) achieves this goal
by defining a method of achieving consensus over the set of all (valid)
transactions, and then defining that consensus as valid if and only if no
output is spent more than once.

A more general approach is to introduce the idea of a cryptographic Single-Use
Seal, analogous to the tamper-evidence single-use seals commonly used for
protecting goods during shipment and storage. Each individual seals is
associated with a globally unique identifier, and has two states, open and
closed. A secure seal can be closed exactly once, producing a proof that the
seal was closed.

All practical single-use seals will be associated with some kind of condition,
such as a pubkey, or deterministic expression, that needs to be satisfied for
the seal to be closed. Secondly, the contents of the proof will be able to
commit to new data, such as the transaction spending the output associated with
the seal.

Additionally some implementations of single-use seals may be able to also
generate a proof that a seal was _not_ closed as of a certain
time/block-height/etc.


## Implementations

### Transactional Blockchains

A transaction output on a system like Bitcoin can be used as a single-use seal.
In this implementation, the outpoint (txid:vout #) is the seal's identifier,
the authorization mechanism is the scriptPubKey of the output, and the proof
is the transaction spending the output. The proof can commit to additional
data as needed in a variety of ways, such as an OP_RETURN output, or
unspendable output.

This implementation approach is resistant to miner censorship if the seal's
identifier isn't made public, and the protocol (optionally) allows for the
proof transaction to commit to the sealed contents with unspendable outputs;
unspendable outputs can't be distinguished from transactions that move funds.


### Unbounded Oracles

A trusted oracle P can maintain a set of closed seals, and produce signed
messages attesting to the fact that a seal was closed. Specifically, the seal
is identified by the tuple (P, q), with q being the per-seal authorization
expression that must be satisfied for the seal to be closed. The first time the
oracle is given a valid signature for the seal, it adds that signature and seal
ID to its closed seal set, and makes available a signed message attesting to
the fact that the seal has been closed. The proof is that message (and
possibly the signature, or a second message signed by it).

The oracle can publish the set of all closed seals for transparency/auditing
purposes. A good way to do this is to make a merkelized key:value set, with the
seal identifiers as keys, and the value being the proofs, and in turn create a
signed certificate transparency log of that set over time. Merkle-paths from
this log can also serve as the closed seal proof, and for that matter, as
proof of the fact that a seal has not been closed.


### Bounded Oracles

The above has the problem of unbounded storage requirements as the closed seal
set grows without bound. We can fix that problem by requiring users of the
oracle to allocate seals in advance, analogous to the UTXO set in Bitcoin.

To allocate a seal the user provides the oracle P with the authorization
expression q. The oracle then generates a nonce n and adds (q,n) to the set of
unclosed seals, and tells the user that nonce. The seal is then uniquely
identified by (P, q, n)

To close a seal, the user provides the oracle with a valid signature over (P,
q, n). If the open seal set contains that seal, the seal is removed from the
set and the oracle provides the user with a signed message attesting to the
valid close.

A practical implementation would be to have the oracle publish a transparency
log, with each entry in the log committing to the set of all open seals with a
merkle set, as well as any seals closed during that entry. Again, merkle paths
for this log can serve as proofs to the open or closed state of a seal.

Note how with (U)TXO commitments, Bitcoin itself is a bounded oracle
implementation that can produce compact proofs.


### Group Seals

Multiple seals can be combined into one, by having the open seal commit to a
set of sub-seals, and then closing the seal over a second set of closed seal
proofs. Seals that didn't need to be closed can be closed over a special
re-delegation message, re-delegating the seal to a new open seal.

Since the closed sub-seal proof can additionally include a proof of
authorization, we have a protcol where the entity with authorization to close
the master seal has the ability to DoS attack sub-seals owners, but not the
ability to fraudulently close the seals over contents of their choosing. This
may be useful in cases where actions on the master seal is expensive - such as
seals implemented on top of decentralized blockchains - by amortising the cost
over all sub-seals.


## Atomicity

Often protocols will require multiple seals to be closed for a transaction to
be valid. If a single entity controls all seals, this is no problem: the
transaction simply isn't valid until the last seal is closed.

However if multiple parties control the seals, a party could attack another
party by failing to go through with the transaction, after another party has
closed their seal, leaving the victim with an invalid transaction that they
can't reverse.

We have a few options to resolve this problem:

### Use a single oracle

The oracle can additionally guarantee that a seal will be closed iff some other
set of seals are also closed; seals implemented with Bitcoin can provide this
guarantee. If the parties to a transaction aren't already all on the same
oracle, they can add an additional transaction reassigning their outputs to a
common oracle.

Equally, a temporary consensus between multiple mutually trusting oracles can
be created with a consensus protocol they share; this option doesn't need to
change the proof verification implementation.


### Two-phase Timeouts

If a proof to the fact that a seal is open can be generated, even under
adversarial conditions, we can make the seal protocol allow a close to be
undone after a timeout if evidence can be provided that the other seal(s) were
not also closed (in the specified way).

Depending on the implementation - especially in decentralized systems - the
next time the seal is closed, the proof it has been closed may in turn provide
proof that a previous close was in fact invalid.


# Proof-of-Publication and Proof-of-Non-Publication

Often we need to be able to prove that a specified audience was able to receive
a specific message. For example, the author's PayPub protocol[^paypub],
Todd/Taaki's timelock encryption protocol[^timelock], Zero-Knowledge Contingent
Payments[^zkcp], and Lightning, among others work by requiring a secret key to
be published publicly in the Bitcoin blockchain as a condition of collecting a
payment. At a much smaller scale - in terms of audience - in certain FinTech
applications for regulated environments a transaction may be considered invalid
unless it was provably published to a regulatory agency.  Another example is
Certificate Transparency, where we consider a SSL certificate to be invalid
unless it has been provably published to a transparency log maintained by a
third-party.

Secondly, many proof-of-publication schemes also can prove that a message was
_not_ published to a specific audience. With this type of proof single-use
seals can be implemented, by having the proof consist of proof that a specified
message was not published between the time the seal was created, and the time
it was closed (a proof-of-publication of the message).

## Implementations

### Decentralized Blockchains

Here the audience is all participants in the system. However miner censorship
can be a problem, and compact proofs of non-publication aren't yet available
(requires (U)TXO commitments).

The authors treechains proposal is a particularly generic and scalable
implementation, with the ability to make trade offs between the size of
audience (security) and publication cost.

### Centralized Public Logs

Certificate Transparency works this way, with trusted (but auditable) logs run
by well known parties acting as the publication medium, who promise to allow
anyone to obtain copies of the logs.

The logs themselves may be indexed in a variety of ways; CT simply indexes logs
by time, however more efficient schemes are possible by having the operator
commit to a key:value mapping of "topics", to allow publication (and
non-publication) proofs to be created for specified topics or topic prefixes.

Auditing the logs is done by verifying that queries to the state of the log
return the same state at the same time for different requesters.

### Receipt Oracles

Finally publication can be proven by a receipt proof by the oracle, attesting
to the fact that the oracle has successfully received the message. This is
particularly appropriate in cases where the required audience is the oracle
itself, as in the FinTech regulator case.


# Validity Oracles

As transaction histories grow longer, they may become impractical to move from
one party to another. Validity oracles can solve this problem by attesting to
the validity of transactions, allowing history prior to the attested
transactions to be discarded.

A particularly generic validity oracle can be created using deterministic
expressions systems. The user gives the oracle an expression, and the oracle
returns a signed message attesting to the validity of the expression.
Optionally, the expression may be incomplete, with parts of the expression
replaced by previously generated attestations. For example, an expression that
returns true if a transaction is valid could in turn depend on the previous
transaction also being valid - a recursive call of itself - and that recursive
call can be proven with a prior attestation.

## Implementations

### Proof-of-Work Decentralized Consensus

Miners in decentralized consensus systems act as a type of validity oracle, in
that the economic incentives in the system are (supposed to be) designed to
encourage only the mining of valid blocks; a user who trusts the majority of
hashing power can trust that any transaction with a valid merkle path to a
block header in the most-work chain is valid. Existing decentralized consensus
systems like Bitcoin and Ethereum conflate the roles of validity oracle and
single-use seal/anti-replay oracle, however in principle that need not be true.


### Trusted Oracles

As the name suggests. Remote-attestation-capable trusted hardware is a
particularly powerful implementation - a conspiracy theory is that the reason
why essentially zero secure true remote attestation implementations exist is
because they'd immediately make untraceable digital currency systems easy to
implement (Finney's RPOW[^rpow] is a rare counter-example).

Note how a single-use seal oracle that supports a generic deterministic
expressions scheme for seal authorization can be easily extended to provide a
validity oracle service as well. The auditing mechanisms for a single-use seal
oracle can also be applied to validity oracles.


# Fraud Proofs

Protocols specified with deterministic expressions can easily generate "fraud
proofs", showing that claimed states/proof in the system are actually invalid.
Additionally many protocols can be specified with expressions of k*log2(n)
depth, allowing these fraud proofs to be compact.

A simple example is proving fraud in merkle-sum tree, where the validity
expression would be something like:

    (defun valid? (node)
        (or (== node.type leaf)
            (and (== node.sum (+ node.left.sum node.right.sum))
                 (and (valid? node.left)
                      (valid? node.right)))))

To prove the above expression evaluates to true, we'll need the entire contents
of the tree. However, to prove that it evaluates to false, we only need a
subset of the tree as proving an and expression evaluates to false only
requires one side, and requires log2(n) data. Secondly, with pruning, the
deterministic expressions evaluator can automatically keep track of exactly
what data was needed to prove that result, and prune all other data when
serializing the proof.


## Validity Challenges

However how do you guarantee it will be possible to prove fraud in the first
place? If pruning is allowed, you may simply not have access to the data
proving fraud - an especially severe problem in transactional systems where a
single fraudulent transaction can counterfeit arbitrary amounts of value out of
thin air.

A possible approach is the validity challenge: a subset of proof data, with
part of the data marked as "potentially fraudulent". The challenge can be
satisfied by providing the marked data and showing that the proof in question
is in fact valid; if the challenge is unmet participants in the system can
choose to take action, such as refusing to accept additional transactions.

Of course, this raises a whole host of so-far unsolved issues, such as DoS
attacks and lost data.


# Probabilistic Validation

Protocols that can tolerate some fraud can make use of probabilistic
verification techniques to prove that the percentage of undetected fraud within
the system is less than a certain amount, with a specified probability.

A common way to do this is the Fiat-Shamir transform, which repeatedly samples
a data structure deterministically, using the data's own hash digest as a seed
for a PRNG. Let's apply this technique to our merkle-sum tree example. We'll
first need a recursive function to check a sample, weighted by value:

    (defun prefix-valid? (node nonce)
        (or (== node.type leaf)
            (and (and (== node.sum (+ node.left.sum node.right.sum))
                      (> 0 node.sum)) ; mod by 0 is invalid, just like division by zero
                                      ; also could guarantee this with a type system
                 (and (if (< node.left.sum (mod nonce node.sum))
                          (prefix-valid? node.right (hash nonce))
                          (prefix-valid? node.left (hash nonce)))))))

Now we can combine multiple invocations of the above, in this case 256
invocations:

    (defun prob-valid? (node)
        (and (and (and .... (prefix-valid? node (digest (cons (digest node) 0)))
             (and (and ....
                            (prefix-valid? node (digest (cons (digest node) 255)))

As an exercise for a reader: generalize the above with a macro, or a suitable
types/generics system.

If we assume our attacker can grind up to 128 bits, that leaves us with 128
random samples that they can't control. If the (value weighted) probability of
a given node is fraudulent q, then the chance of the attacker getting away with
fraud is (1-q)^128 - for q=5% that works out to 0.1%

(Note that the above analysis isn't particularly well done - do a better
analysis before implementing this in production!)


## Random Beacons and Transaction History Linearization

The Fiat-Shamir transform requires a significant number of samples to defeat
grinding attacks; if we have a random beacon available we can significantly
reduce the size of our probabilistic proofs. PoW blockchains can themselves act
as random beacons, as it is provably expensive for miners to manipulate the
hash digests of blocks they produce - to do so requires discarding otherwise
valid blocks.

An example where this capability is essential is the author's transaction
history linearization technique. In value transfer systems such as Bitcoin, the
history of any given coin grows quasi-exponentially as coins are mixed across
the entire economy. We can linearize the growth of history proofs by redefining
coin validity to be probabilistic.

Suppose we have a transaction with n inputs. Of those inputs, the total value
of real inputs is p, and the total claimed value of fake inputs is q. The
transaction commits to all inputs in a merkle sum tree, and we define the
transaction as valid if a randomly chosen input - weighted by value - can
itself be proven valid. Finally, we assume that creating a genuine input is a
irrevocable action which irrevocable commits to the set of all inputs, real and
fake.

If all inputs are real, 100% of the time the transaction will be valid; if all
inputs are fake, 100% of the time the transaction will be invalid. In the case
where some inputs are real and some are fake the probability that the fraud
will be detected is:

    q / (q + p)

The expected value of the fake inputs is then the sum of the potential upside -
the fraud goes detected - and the potential downside - the fraud is detected
and the real inputs are destroyed:

    E = q(1 - q/(q + p)) - p(q/(q + p)
      = q(p/(q + p)) - p(q/(q + p)
      = (q - q)(p/(q + p))
      = 0

Thus so long as the random beacon is truly unpredictable, there's no economic
advantage to creating fake inputs, and it is sufficient for validity to only
require one input to be proven, giving us O(n) scaling for transaction history
proofs.


### Inflationary O(1) History Proofs

We can further improve our transaction history proof scalability by taking
advantage of inflation. We do this by occasionally allowing a transaction proof
to be considered valid without validating _any_ of the inputs; every time a
transaction is allowed without proving any inputs the size of the transaction
history proof is reset. Of course, this can be a source of inflation, but
provided the probability of this happening can be limited we can limit the
maximum rate of inflation to the chosen value.

For example, in Bitcoin as of writing every block inflates the currency supply
by 25BTC, and contains a maximum of 1MB of transaction data, 0.025BTC/KB. If we
check the prior input proof with probability p, then the expected value of a
transaction claiming to spend x BTC is:

    E = x(1-p)

We can rewrite that in terms of the block reward per-byte R, and the transaction size l:

    lR = x(1-p)

And solving for p:

    p = 1 - lR/x

For example, for a 1KB transaction proof claiming to spending 10BTC we can omit
checking the input 0.25% of the time without allowing more monetary inflation
than the block reward already does. Secondly, this means that after n
transactions, the probability that proof shortening will _not_ happen is p^n,
which reaches 1% after 1840 transactions.

In a system like Bitcoin where miners are expected to validate, a transaction
proof could consist of just a single merkle path showing that a single-use seal
was closed in some kind of TXO commitment - probably under 10KB of data. That
gives us a history proof less than 18.4MB in size, 99% of the time, and less
than 9.2MB in size 90% of the time.

An interesting outcome of thing kind of design is that we can institutionalize
inflation fraud: the entire block reward can be replaced by miners rolling the
dice, attempting to create valid "fake" transactions. However, such a pure
implementation would put a floor on the lowest transaction fee possible, so
better to allow both transaction fee and subsidy collection at the same time.


# References

[^paypub] https://github.com/unsystem/paypub
[^timelock] https://github.com/petertodd/timelock
[^zkcp] https://bitcoincore.org/en/2016/02/26/zero-knowledge-contingent-payments-announcement/
[^rpow] https://cryptome.org/rpow.htm

-- 
https://petertodd.org 'peter'[:-1]@petertodd.org
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 455 bytes
Desc: Digital signature
URL: <http://lists.linuxfoundation.org/pipermail/bitcoin-dev/attachments/20160620/73f11845/attachment.sig>