Two factor wallet with one-time-passwords
4 messages, 27 Jul to 8 Aug 2013. Started by Peter Todd.
Peter Todd (3), John Dillon
Gavin Andresen recently suggested a design for a wallet protected by two-factor authentication via one-time-passwords with the aid of a third-party service to counter-sign 2-of-2 protected wallets.(1) The design is useful when the user can't sign transactions on a second device, such as a phone, but can provide one-time-passwords. (possibly generated on a smart phone or stored on paper) However involving a third-party has privacy and availability risks. Here is an alternate design, also using one-time-passwords, that removes the requirement for a third-party, along with other advantages and disadvantages.
User experience ===============
The user has a wallet with a separate balances for savings and a smaller day-to-day spending amount. Transactions spending the day-to-day balance do not need two-factor authorization, while spending the savings balance does. As the day-to-day balance becomes low the user is able to top it up by authorizing the movement of discrete multiples of some amount from savings to spending. That authorization requires one one-time-password per multiple being moved.
Implementation ==============
Savings use P2SH outputs matching the following scriptPubKey form:
HASH160 <H(nonce_i)> EQUALVERIFY <pubkey> CHECKSIG
spent with:
<sig> <nonce_i>
The way the pubkey/seckey is generated is unimportant, although some kind of deterministic scheme is preferable. Nonces on the other hand are generated deterministically using a counter-based one-time-password scheme that takes some secret seed and an integer i. A large number of H(nonce_n) are generated in advance and moved to the computer holding the wallet. (generating them on that computer is also possible, but obviously risks the secret seed being compromised)
A brute-force attack to spend a signed txout requires the attacker to find a preimage, thus the security level is the number of bits for the nonce; 64 bits is sufficient. (remember the birthday attack doesn't apply here) Unfortunately the most popular one-time-password scheme, the RFC6238 used in Google Authenticator, only outputs six digits numbers, well below the security level required. (Google Auth is generally used in a time-mode, but also has a counter mode)
The older RFC2289 however turns the passwords into six words from a 2048 entry wordlist, giving a 64-bit nonce with 2-bits of checksum. RFC2289 implementations are also well suited to paper print-outs and generally make it easy to do so. RFC2289 as written uses SHA1, however the suspected vulnerabilities in SHA1 are partial-preimage collisions, not relevant in this application.
In a sense the user is now acting as an oracle answering the question of whether or not funds should be allowed to move from savings to spending, without being responsible for where those funds are allowed to go. As described in (2) it is easy to create a whole range of conditions by using multiple nonces if the use-case demanded. For instance a corporate environment may want multiple parties to be required to authorize the funds to move, possible with multiple nonces.
It's interesting to note how in some cases it may be preferable that the authorization is simply authorization to spend without any other involvement. Here the party acting as an oracle not only doesn't need to know where funds are going but can even authorize the spend in advance without two-way communication - possibly even prior to the funds being received in the first place. This authorization can be easily given manually, for instance over the phone, and the accounting to keep track of the total amount authorized can be easily done with pen and paper - something not possible with CHECKMULTISIG wallets.
Funding the wallet ==================
As with any multi-party wallet receiving funds must also be handled carefully to ensure an attacker can't fool the user into giving the sender the wrong address. This requires the involvement of all parties required to authorize an outgoing payment. In addition here the protection only works if funds sent to the wallet are split up into the discrete authorization amounts the user wishes. (possibly with more than one amount level)
There hasn't been as much thought put into these systems as there has been on payment protocols between a customer and a merchant, but the basic idea is to have more than one device participate in the generation of payment request signed somehow. For fund splitting the request can be that the funds are paid to multiple txouts in one go. For recurring payments the request could have some mechanism for multiple addresses to be specified for future use. Fall-back to a standard multi-signature wallet is possible as well.
More research is needed.
1) https://gist.github.com/gavinandresen/5616606 2) https://bitcointalk.org/index.php?topic=260898.msg2804469#msg2804469
On Sat, Jul 27, 2013 at 07:49:18PM -0400, Peter Todd wrote:
Quoted text (10 lines)
> Implementation > ============== > > Savings use P2SH outputs matching the following scriptPubKey form: > > HASH160 <H(nonce_i)> EQUALVERIFY <pubkey> CHECKSIG > > spent with: > > <sig> <nonce_i>
FWIW with some minor scripting language additions such as access to txin and txout contents, along with merklized abstract syntax tree (MAST) support, we can even implement a version where scriptPubKey's can be reused:
<pubkey> CHECKSIGVERIFY
// Verify we aren't spending more than the maximum spend amount
0 GET-TXIN-VALUE // relative indexing
0 GET-TXOUT-VALUE
SUB
<max-spend-amount>
LESSTHAN
VERIFY
// If the txout is greater than the maximum spend amount force it to
// also follow these same rules.
0 GET-TXOUT-VALUE
<max-spend-amount>
LESSTHAN
IFNOT
GET-THIS-SCRIPT
MAST-HASH
<serialized script "MAST-HASH MAST-EVAL">
CAT
GET-TXOUT-SCRIPT
EQUALVERIFY
ENDIF
// Hash the provided oracle nonce, saving original for later.
DUP
HASH160
// Use the txid:vout nonce as an index to a table, embedded with MAST
// script compression.
0 GET-TXIN-TXID
0 GET-TXIN-VOUT
CAT
HASH160
// The table, n=64 levels deep, not all levels shown for brevity.
DUP
1
AND
IF
1
RSHIFT
DUP
1
AND
IF
1
RSHIFT
DUP
1
AND
IF
<MAST digest, not executed>
ELSE
1
RSHIFT
DUP
1
AND
IF
// Lowest level contains the following pushdata,
// with 0 <= i < 2^64
<HASH160(HASH160(nonce-secret + i))>
ELSE
<MAST digest, not executed>
ENDIF
ELSE
<MAST digest, not executed>
ENDIF
ELSE
<MAST digest, not executed>
ENDIF
// Drop the txid:vout nonce
SWAP
DROP
// Verify that the hash of the nonce and the pre-committed value in
// the H(nonce) table match.
EQUALVERIFY
// Stack now only contains the nonce preceeded by a merkle path linking
// that nonce to the tip of a merkle tree over all nonces.
//
// Verify that path.
SWAP // Move direction flag to the top
IF
SWAP
ENDIF
HASH160
(repeat above five lines 63 more times)
<nonce-merkle-tree-tip-digest>
EQUAL
The scriptPubKey is spent by the following scriptSig:
<nonce-merkle-path-0>...<nonce-merkle-path-63>
<nonce>
<signature>
<serialized-script>
(note that I've left off a number of possible optimizations for clarity)
Now when the user wishes to spend a txout greather than their spending limit their wallet software will first give them a short 6 word string calculated from the last 64-bits of H(txid:vout). They simply enter this string into their phone, ideally via convenient qr-code or voice/thought recognition, and their phone provides a second short 6 word string to enter into the wallet software on their computer, authorizing the payment. If they opt for a paper-based one-time-password table they simply use the 6 word string as an index to their pre-printed OTP encyclopedia set.
Like the previously described version the security level is still a healthy 2^64 - again the attacker needs to find a 64-bit pre-image, considered to be a highly difficult task for any attacker unable to count from 0 to 2^64 or store a table containing 2^64 values.
There is the disadvantage of the large storage requirements for both wallets, however because of the double hashed construction, H(H(nonce-secret+i)), neither table needs to be kept secret. Thus without loss of security both tables can be easily stored in a distributed hash table in the cloud and queried as needed.
On Sun, Jul 28, 2013 at 1:20 AM, Peter Todd <pete@petertodd•org> wrote:
Quoted text (4 lines)
> FWIW with some minor scripting language additions such as access to txin > and txout contents, along with merklized abstract syntax tree (MAST) > support, we can even implement a version where scriptPubKey's can be > reused:
<snip>
Quoted text (8 lines)
> // Stack now only contains the nonce preceeded by a merkle path linking > // that nonce to the tip of a merkle tree over all nonces. > // > // Verify that path. > SWAP // Move direction flag to the top > IF > SWAP > ENDIF
You missed a 'CAT' opcode here.
Quoted text (2 lines)
> HASH160 > (repeat above five lines 63 more times)
<snip>
Quoted text (3 lines)
> payment. If they opt for a paper-based one-time-password table they > simply use the 6 word string as an index to their pre-printed OTP > encyclopedia set.
I think you should disclose whether or not you have any ties to the pulp and paper business... By my calculations the production of a single OTP table would consume roughly half of all the forest biomass on this planet.
Quoted text (10 lines)
> Like the previously described version the security level is still a > healthy 2^64 - again the attacker needs to find a 64-bit pre-image, > considered to be a highly difficult task for any attacker unable to > count from 0 to 2^64 or store a table containing 2^64 values. > > There is the disadvantage of the large storage requirements for both > wallets, however because of the double hashed construction, > H(H(nonce-secret+i)), neither table needs to be kept secret. Thus > without loss of security both tables can be easily stored in a > distributed hash table in the cloud and queried as needed.
ROTFL!
Your idea is better than you realize, you are just too paranoid for your own good. The thing is the attacker isn't going to be someone paying you funds over your minimum spending limit, which means the size of the table deriving which H(nonce) is selected for a given txid:vout can be significantly smaller. For instance if you want to have 256 total payments before a 50:50 chance of any pair using the same nonce, you only need a table with ~2^16 elements or with 20 byte hashes just a megabyte of data. It is the 16 level merkle proofs that are the problem, 16*21=336 bytes of data in the scriptSig. Then again, that's only 4.5x the size of a single signature, not unreasonable.
Also your nested IF statements, while a lovely and hilarious use of MAST, can be replaced by simply creating the merkle tree over the tuples [i,H(nonce_i)] and proving that the nonce_i you provided matched the precommitted tree. Now you only need to provide one merkle proof, not two.
But don't let me discourage you, rarely do I see elaborate jokes that also meet the criteria to be a least publishable unit. :)
On Sat, Jul 27, 2013 at 07:49:18PM -0400, Peter Todd wrote:
Quoted text (10 lines)
> Funding the wallet > ================== > > As with any multi-party wallet receiving funds must also be handled > carefully to ensure an attacker can't fool the user into giving the > sender the wrong address. This requires the involvement of all parties > required to authorize an outgoing payment. In addition here the > protection only works if funds sent to the wallet are split up into the > discrete authorization amounts the user wishes. (possibly with more than > one amount level)
Quick note for patent prior-art/my own memory - did a talk yesterday about multifactor wallets, one time passwords and hash digest based oracles. Someone getting involved in the business of selling bitcoins pointed out that legally it can actually be desirable to give the bitcoins to the customer by giving them a physical private key, perhaps on a sheet of paper in a mailed envelope. Obviously the customer would be wise to sweep the funds. Of course, the advantage of doing it with paper is the legal system has a long history of dealing with the concept of a secret on a piece of paper. (your customers won't have handy PKI to use after all)
With multi-factor wallets you can have the customer provide one or more keys, and you give them one final key on a sheet of paper, with instructions to scan it on their phone via QR-code or something. Now the transfer is absolute on your end - you can't get the funds back. If it's a large amount you may want to split it up among multiple addresses, and deliver the keys to the customer in a way that makes it obvious when they are revealed. (scratch off for instance)
Finally, one-time-passwords do much the same thing, but they don't require the second device, and the sheet of paper the customer is dealing with can be much shorter. Similarly the final approval could just be done over the phone by telling the customer the ~6-8 magic words that unlock their funds - legally it could be useful to record that phone call. Similarly for a large transfer, make it clear how much each scratched off text field is unlocking to defend yourself in court.
Of course, in both there is still the risk of the funds ending up locked due to a mistake, but at least there isn't financial incentive to make that event happen. (usually)
I'll admit I hadn't thought of any of this stuff until I talked to an actual business with real problems, worth doing.
Finally it's too bad we didn't get OP_EVAL; the customer could have provided a P2SH script with, well, anything in it, and the unlock could could have easily been a "bolt-on":
HASH160 <digest> EQUALVERIFY DUP HASH160 <p2sh-code-digest> EQUALVERIFY EVAL
Oh well, MAST support can do the same thing one day.