Return-Path: Received: from smtp1.linuxfoundation.org (smtp1.linux-foundation.org [172.17.192.35]) by mail.linuxfoundation.org (Postfix) with ESMTPS id 7D3E5956 for ; Mon, 20 Jun 2016 13:33:29 +0000 (UTC) X-Greylist: delayed 00:06:41 by SQLgrey-1.7.6 Received: from tupac2.dyne.org (tupac2.dyne.org [178.62.188.7]) by smtp1.linuxfoundation.org (Postfix) with ESMTPS id 2D6AF172 for ; Mon, 20 Jun 2016 13:33:27 +0000 (UTC) Received: from [127.0.0.1] (localhost [127.0.0.1]) (Authenticated sender: policeterror@dyne.org) with ESMTPSA id 140DF181A76 DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=dyne.org; s=mail; t=1466429204; bh=aUmB+MOnSWYmq5CNS1wQ/jAmYUGhBtXyRWcFuafmAGU=; h=Subject:To:References:From:Date:In-Reply-To:From; b=WYOGh1DA0CpGwuaiaszNeGLpg8/+HlL0HCrdWZiKh8M9L5AerGP0Pmt1C1mEojThq f0/xkvKbEher9c1KmZLNz8su35BytDr25Df93GOp4y8/SbCv5BdKDoOqF4D9h34eiK 8NRh+bMKG2DWX8eDZgesPQ7ZJxHud0VpabV+fwsk= To: bitcoin-dev@lists.linuxfoundation.org References: <20160620085649.GA29964@fedora-21-dvm> From: Police Terror Message-ID: <5767EEFE.7060103@dyne.org> Date: Mon, 20 Jun 2016 13:26:22 +0000 MIME-Version: 1.0 In-Reply-To: <20160620085649.GA29964@fedora-21-dvm> Content-Type: text/plain; charset=windows-1252 Content-Transfer-Encoding: 7bit X-Spam-Status: No, score=-3.4 required=5.0 tests=BAYES_00,DKIM_SIGNED, DKIM_VALID, DKIM_VALID_AU, RP_MATCHES_RCVD, UNPARSEABLE_RELAY autolearn=ham version=3.3.1 X-Spam-Checker-Version: SpamAssassin 3.3.1 (2010-03-16) on smtp1.linux-foundation.org X-Mailman-Approved-At: Mon, 20 Jun 2016 14:14:48 +0000 Subject: Re: [bitcoin-dev] Building Blocks of the State Machine Approach to Consensus X-BeenThere: bitcoin-dev@lists.linuxfoundation.org X-Mailman-Version: 2.1.12 Precedence: list List-Id: Bitcoin Protocol Discussion List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Mon, 20 Jun 2016 13:33:29 -0000 Bitcoin could embed a lisp interpreter such as Scheme, reverse engineer the current protocol into lisp (inside C++), run this alternative engine alongside the current one as an option for some years (only for fine tuning) then eventually fade this lisp written validation code instead of the current one. Scheme is small and minimal, and embeds easily in C++. This could be a better option than the libconsensus library - validation in a functional scripting language. That doesn't mean people can't program the validation code in other languages (maybe they'd want to optimize), but this code would be the standard. It's really good how you are thinking deeply how Bitcoin can be used, and the implications of everything. Also there's a lot of magical utopic thinking in Ethereum, which is transhumanist nonsense that is life denying. Bitcoin really speaks to me because it is real and a great tool following the UNIX principle. I wouldn't be so quick to deride good engineering over systematic provable systems for all domains. Bitcoin being written in C++ is not a defect. It's actually a strong language for what it does. Especially when used correctly (which is not often and takes years to master). With the seals idea- am I understand this correctly?: Every transaction has a number (essentially the index starting from 0 upwards) depending on where it is in the blockchain. Then there is an array (probably an on disk array mapping transaction indexes to hashes). Each hash entry in the array must be unique (the hashes) otherwise the transaction will be denied. This is a great idea to solve transaction hash collisions and simple to implement. Probabilistic validation is a good idea, although the real difficulty now seems to be writing and indexing all the blockchain data for lookups. And validation is disabled for most of the blocks. Pruning is only a stop gap measure (which loses data) that doesn't solve the issue of continually growing resource consumption. Hardware and implementation can only mitigate this so much. If only there was a way to simplify the underlying protocol to make it more resource efficient... Peter Todd via bitcoin-dev: > 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 > > > > _______________________________________________ > bitcoin-dev mailing list > bitcoin-dev@lists.linuxfoundation.org > https://lists.linuxfoundation.org/mailman/listinfo/bitcoin-dev >