Here's my draft. I don't claim this to be "official", but I think this
should represent the consensus we've come to at the DarkWallet
Hackathon; I'm writing it up as an email in part to preserve a record of
that consensus.


* General Principles

We believe in decentralization, user-defined privacy, education as
opposed to "magic", and security based on openness rather than just
trust. We consider users who are individuals as well as the needs of
businesses such as merchants and exchanges. We recognize that often the
more people who use privacy protection technologies such as CoinJoin,
the better protected we all are.


* Privacy

Bitcoin inherently makes the flow of money visible, however it does not
require that flow to have real-world identities attached, or even for
that matter, pseudonyms. We see this as an unfortunate flaw in the
Bitcoin protocol that is to be corrected; the Satoshi whitepaper itself
included one such correction by outlining how avoiding address re-use
helped preserve privacy.


** Threat model

We assume a worst-case sophisticated state-level attacker with the goal
of deanonymizing and otherwise subverting Bitcoin users. Such an
attacker can be assumed to control up to 100% of the Bitcoin relay
network as well as have the ability to wiretap up to 100% of the
node-to-node traffic. (for nodes that they do not control) Such
attackers are however constrained by politics and budget. We assume they
use their ability to conduct MITM attacks sparingly - for instance by
subverting certificate authorities - due to the risk of detection. (note
how such attackers may be more willing to use detectable attacks in the
future now that their activities are public knowledge) We also assume
that while their budgets for individual targets may be very large, the
equally large number of targets required for en-mass survailance leads
to relatively low budgets per target. In particular note how the 51%
assumption assumes that the overall "economic value" of Bitcoin to its
participants is greater than the attacker's budget by some margin.


** Address re-use

Wallet software SHALL avoid address re-use. New addresses will be used
for all change and users will be encouraged through the user-interface
and other measures to use new addresses for every payment to the wallet.


** CoinJoin

With CoinJoin the more users that make use of it, the larger the
anonymity set and the better protected user privacy is. Thus we
encourage wallet software to agressively make trade-offs between
absolute privacy and usability; compromise is not a dirty word.

Wallet software SHALL implement basic two-party mix functionality and
MAY implement more sophisticated CoinJoin functionality such as n-party
mixes. CoinJoin SHALL be the default way that all transactions are sent.
Wallet authors are cautioned that more sophisticated functionality may
be more secure in theory, but if users do not use it the functionality
is wasted; focus on the general case first and only then try to improve.


*** Two-Party Mixes

The most basic form of CoinJoin is for two parties to author a
transaction. A key distinction between a 2 party mix and an n-party mix
is that in the two party case both parties automatically learn the
other's inputs and outputs by simple elimination; sophisticated
cryptographic blinding protocols are useless. To an external attacker
each transaction doubles the size of the anonymity set: the coins may
have come from one party or the other and the attacker has no way of
knowing which. (modulo value analysis, which will be discussed later)


*** n-party Mixes and Blinding

If two parties can jointly author a transaction, n parties can too.
Without special effort this has the disadvantage of revealing the input
to output mapping to all parties. Various cryptographic blinding schemes
have been proposed to fix this problem, either with multi-party
computational techniques, or by making use of multiple communication
channels along with a robust anti-DoS scheme. In either case, for now we
reject such schemes as complex and inconvenient and prioritize robust
two-party mixing. However we do take the existance of such schemes into
account; note how a n-party mix can act as a single party in a two-party
mix scheme.


*** Four-stage two-party mix protocol

<on the wiki>


*** Defeating value analysis

Attackers can make good guesses at the mapping of inputs to outputs
based on value. For instance with two inputs of 3 and 5, and fours
outputs of 1.4, 1.6, 2.4 and 2.6 the attacker can easily map inputs to
outputs based on what values match up, in this case 3 split into 1.6 and
1.4, and 5 split into 2.4 and 2.6


**** Value Matching

Not all CoinJoin users need their transactions to have specific output
amounts; some users simply need to move money from one place to another
but do not need a specific amount moved or at a specific time. These
users can assist users with more stringent requirements by matching the
input or output values they request. As a general principle wallets
SHOULD make these anonymity optimizations possible by allowing users to
schedule transactions to complete by a specific time and/or allow users
to specify that they do not wish the transaction to happen unless
CoinJoin is used.

With four-stage two-party mixes the Alice, who advertised a desire to do
a transaction first, can easily do ths by picking the transaction output
amounts only after Bob replies with his desired inputs and outputs, and
picking those amounts so they match Bob's. (or some combination of Bob's
outputs)


**** Merge Avoidance

Merge avoidance is the practice of avoiding the merging of multiple
transaction inputs into a single new transaction output, thus implying a
common relationship between those inputs. The most primitive form of
merge avoidance is to create multiple individual transactions, each
moving single transaction input to an output. (or perhaps some small
number) This is of course inefficient, and appears to have only been
proposed as a means to still allow for coin blacklists to function while
preserving some financial privacy.

Combined with CoinJoin however merge avoidance becomes much more
powerful. For instance even in its most simple form multiple parties can
combine their merge-avoiding transaction sets, giving even transactions
without actual counterparties a useful measure of plausible deniability.

In addition the underlyng features that make merge-avoidance possible -
the ability of a recipient to designate they are willing to receive
payments made to multiple addresses - synergisticly make very
sophisticated value matching strategies possible.


***** Cut-thru payments

Related to merge avoidance the idea of a cut-thru payment is that if an
intermediary is both a debitor and a creditor, with sophisticated
payment protocols they can request incoming payments to directly pay
outgoing liabilities, skipping them as an intermediary. While premature
to implement this feature now, it is worth thinking about for the future.


** Tor

While Tor isn't perfect there is real-world evidence - specifically the
Snowden leaks - that it works well enough to be considered a worthy
adversary by state-level attackers. Wallets MUST suppoort the basic
proxy options that allow the Tor proxy - or some other similar
technology - to be used for privacy enhancement and SHOULD make use of
Tor-specific features such as hidden services.


* Decentralization

** Fees

In a decentralized system distinguishing DoS attackers from legitimate
users is at best difficult, at worst impossible. Wallets that do not
provide users with the ability to set fees, both when a transaction is
created initially and after initial broadcast, do their users a
disservice by taking away a vital method of responding to an attack:
outspending the attacker.

Wallets MUST give users the ability to set the fee per KB they are
willing to pay for their transactions. Wallets SHOULD allow users to
change that fee after the fact via transction replacement. Wallets MAY
additionally implement fee estimation techniques, such as watching what
transactions in the mempool are finally mined, or using estimates
provided by miners. However it must be recognized that such data is
inherently unreliable, and this may become a problem in practice in the
future; giving users robust ways to alter fees after the fact will make
lying about fee data - perhaps to push fees upwards - less advantageous.

Note that the current direction of the Bitcoin Foundation maintained
reference implementation is weakly towards a pure estimation scheme;
deployment of full nodes supporting replacement and support from miners
is a precondition to doing things correctly.


*** Fees and privacy

Where there is a trade-off between fees and privacy - such as with merge
avoidance strategies - users should be given options to specify how much
extra they are willing to pay for extra privacy. Wallets SHOULD default
to being willing to pay some extra, perhaps 25% more over the basic fee.


** SPV, full nodes and partial nodes

Wallet software SHOULD whenever possible blur the distinctions between
full UTXO set nodes, SPV nodes, and partial UTXO set nodes. In addition
to those three basic categories there is also the question of whether or
not a node stores archival blockchain data, something that all three
categories of nodes can participate in.

Instead how a node contributes back to the health of the network should
be a function of what resources it has available to it. Of course in
some cases, like a phone wallet, that won't be very much, but for
desktop or business usage the resources available can be significant in
both bandwidth and storage capacity.


*** Relaying data

**** Blocks and blockheaders

Any node can safely relay blocks and block headers, where "safely" is
defined as SPV-level security. Our threat model implies that we don't
trust random peers on the network, thus we are not relying on them for
block validity; as a SPV node we are relying on miners to do validity
checking for us. In short feel free to relay data that you yourself
would trust.


**** Transactions

Remember that relaying transactions has a DoS-attack risk; the Bitcoin
model relies entirely on mining fees and/or priority as the limited
resource to prevent DoS attacks. Thus at present nodes SHOULD NOT relay
transactions if they do not have an up-to-date copy of the relevant
parts of the UTXO set spent by the transaction. (relaying transactions
spending only inputs in a partial UTXO set is acceptable):


**** Block-header diversity

Wallet software MUST make it possible to get block-header information
from a diverse set of sources. These sources SHOULD comprise more than
just peers on a single P2P network. Remember that it is acceptable to
use even centralized sources in addition to decentralized ones for
blockheader data - knowing that a miner did the work required to create
a block header is always valuable information. (for reasonable amounts
of work) For instance the author's block headers over twitter project -
while an April Fools joke - is equally a perfectly reasonable backup
source of blockheader data.


** Updating wallets from blockchain data

In an ideal world wallets wouldn't need to sync their state with
blockchain data at all: pervasive use of payment protocols would have
senders send txout proofs directly to recipients. But that's not the
case. Instead wallet implementations sync themselves from the
blockchain, and when bandwidth limited this becomes a tradeoff between
bandwidth and privacy: your transactions hide in the anonymity set of
the false positives matched by the filter.


*** Bloom filters

The current implementation for SPV nodes is to simply give peers a bloom
filter; the false-positives make the anonymity set. For n peers this has
O(n) cost when a new block comes in; Bloom filters are cheap to test
against and this system works reasonably well.

However, for archival blockchain data bloom filters are seriously
flawed: every block has to be read from disk in full, the bloom filter
matched, and some (potentially very small!) subset sent to the peer. n
peers. The result is high IO load on the node relative to the client,
enabling easy DoS attacks.

Wallet software SHOULD NOT implement only Bloom filters, however using
them when availalbe is acceptable. Note how the Bloom filter design has
at best O(n^2) scaling ruling it out for large-blocksize future
scenarios.


*** Prefix filters

TXO or UTXO data can be easily indexed by in radix trees with log2(k)
lookup cost per query. We can take advantage of the fact that the query
keys need not be provided in full by only providing partial keys.
Because scriptPubKeys are randomly distributed a prefix n bits long has
an anonymity set of roughly 1/2^n * # of transactions in total.

Wallet software SHOULD implement prefix filters and SHOULD use them in
preference to bloom filters whenever available. Wallet software that
currently uses full-key filtering - e.g. Electrum - MUST be upgraded to
support prefix filters in the future.

Wallet software MUST NOT assume that matching anyting other than
H(scriptPubkey) is possible. This applies to bloom filter matches as
well.

In the future miners may commit to either the TXO set in conjunction
with per-block lookup trees, or possibly the full UTXO set. In either
case many of the leading designs may be implemented with only
H(scriptPubKey) lookup capability for reasons of scalability.


* Security

Bitcoin wallet software is unprecedented in how they provide attackers
targets that are highly profitable to attack and highly liquid. (note
the irony here!) A succesfull attack that injects malicious theft
routines into either sourcecode or binaries can steal thousands of
Bitcoins in one go, and the attacks target is you and your team.
Following basic good practices for robust code is a start, but it's far
from enough.


** Source-code integrity

Sourcecode MUST be maintained using a revision control system that
provides strong integrity guarantees; git is recommended.

Sourcecode MUST be PGP signed on a regular basis. Releases MUST be
signed - in git this is accomplished by signing the release tag.
Individual commits SHOULD be signed, particularly if source-code used in
day-to-day development is kept on an untrusted server, e.g. github.
Recovering from a server compromise is made significantly easier if
every commit is securely signed.


** Binary integrity

All things being equal it is better to use an interpreted language
rather than a compiled one; auditing the former is significantly easier
than the latter. Similarly, all things being equal, do not distribute
binaries of your software - have end-users compile binaries themselves.

Of course all things are not equal, and frequently compiled languages
and distributing binaries is the correct choice. If that is the case
deterministic build systems MUST be used when possible; if using them is
not possible take great care with the process by which binaries are
created and try to create long-term plans to move to a deterministic
build system in the future.


** PGP

Developers of wallet software MUST make use of PGP and participate in
the web-of-trust. Developers MUST advertise their PGP fingerprint
widely, for instance on personal websites, forum profiles, business
cards etc. simultaneously. Multiple paths by which someone can find a
fingerprint claimed to be of some developer make subterfuge easier to
detect and more costly to carry out. When possible it is highly
recommended to attach these advertisements to real-world, physical,
actions. For instance the author has included his PGP fingerprint in
highly public, videotaped, talks he has given at conferences. He has
also created a videotaped statement of his PGP key that was timestamped
in the Bitcoin blockchain. While it certainly is possible for such
artifacts to be faked, doing so convincingly is expensive, labour
intensive, and error prone.

Developers SHOULD sign as much communication as practical. Sourcecode is
one form; your emails to development lists and between each other are
another. Signing both leaves a large body of widely distributed work,
all tied to your identity. (it's highly unfortunate that nearly all
publicly available mail archives do not make mail accessible to the
public in such a way as to allow verification of PGP signatures; often
even inline signatures are broken for various reasons)


*** Increasing adoption of PGP

Keep in mind that end-users very rarely verify PGP fingerprints at all,
let alone correctly - the above advice with regard to PGP is currently
mostly useful in allowing *other developers* the tools they need to
verify the integrity of your code. For instance, in the event of a
suspected compromise consistantly signed code allows anyone competent in
the use of PGP to quickly evaluate the situation, and if warrented,
inform less sophisticated users through traditional measures such as the
media.

While this is somewhat out of scope for this document the "DarkWallet
effort" should include work to make PGP more user-friendly and a better
experience. But that does *not* have to mean "making PGP easier for
grama", right now "making PGP easier for Joe Wallet Developer" is a
laudable goal. For instance understanding and using the web-of-trust
sucks right now. How can we make that experience better for a user who
understands the basics of cryptography?


** SSL/Certificate authorties

While certificate authorities are notoriously bad at the job they are
supposed to be doing the CA system is still better than nothing - use it
where appropriate. For instance if you have a website advertising your
software, use https rather than http.


** Multi-factor spend authorization, AKA multisig wallets

<mainly discussed at the conference in terms of multiple individuals
controlling funds, which is out of scope for this document>

Assuming any individual device is uncompromised is risky; wallet
software SHOULD support some form of multi-factor protection of some or
all wallet funds. Note that this is a weak "should"; mainly we want to
ensure that users have options to keep large quantities of funds secure;
accepting higher risks for smaller quantities is an acceptable
trade-off.

FIXME: go into more detail.


*** P2SH

Wallet software MUST support paying funds to P2SH addresses.


** Payment Integrity

Multi-factor wallets protect your funds from being spent without your
knowledge, but they provide no assurance about where those funds went; a
Bitcoin address is not an identity. A payment protocol, such as BIP70,
is needed.

Wallet software SHOULD support BIP70. Yes, there are (justified)
concerns about its current dependence on the certificate authority
system, but those concerns should be addressed by a combination of
fixing that system, and extending BIP70 with different and better
identity verification options.

However, remember that in the absense of multi-factor wallets the "know
who you are paying" protections of BIP70 are actually pretty much
useless; malware that defeats the payment protocol locally is not much
different than malware that silently rewrites Bitcoin addresses. There
are other motivations for the BIP70 version of the payment protocol, but
whether or not they are actually interesting for users is an open
question; it was not designed by user-experience experts. Thus wallet
authors should consider supporting a low priority for now.


--
'peter'[:-1]@petertodd.org
000000000000000f9102d27cfd61ea9e8bb324593593ca3ce6ba53153ff251b3