Re: [Bitcoin-development] Bitcoin at POS using BIP70, NFC and offline payments - implementer feedback

[Eric Voskuil <eric@voskuil.org>]

[-- Attachment #1: Type: text/plain, Size: 22437 bytes --]

On 02/24/2015 11:49 AM, Andy Schroder wrote:
> Hello,
> 
> I think were talking about a lot of the same things. There is one key
> piece of information that I was not thinking about until you made it
> clear. Why the payee needs to identify the payer. In my fuel pump
> application, they really don't, so I wasn't thinking closely about these
> other situations. With my fuel pump, it won't even let you do anything
> until you sign a transaction and submit it. So, the payment request
> contains no personal information, it's just a request for payment, and
> not for anything specific. It doesn't know or care what you are going to
> buy until you make a prepayment, because there is no use in trying to
> start doing business without a signed transaction. This approach
> minimizes risk because once you dispense a fuel, or anything else, it's
> not like you can easily give it back if you happened to not have any
> funds. It also makes it a higher chance for a confirmation before the
> customer leaves. Other transactions have similar post payment
> traditions, like a restaurant (not fast food), where the seller doesn't
> know if you actually have enough money until you've already consumed the
> food, but this work flow seems to be a culturally driven one rather than
> risk driven.
> 
> In the discussion about an https website, there are many websites where
> no login or authentication by the client required to have a private
> connection. With a shopping website though, the customer can identify
> themselves without logging in by saying (in band) what they are
> intending to buy after the private connection has been established. At a
> cash register in person the items being purchased have no tie to the
> customer. The items being purchased were communicated to the seller out
> of band (in real life) and insecurely to establish that link. You are
> trying to make a tie between that list of items and the buyer
> separately, and that is why some unique identifier needs to be
> transmitted via NFC.
> 
> Stepping a bit back: I guess I'm wondering if it would be useful to
> encourage an opposite work flow where a micro payment channel is setup
> for most purchases. For example, if I go to the grocery store, it
> usually takes a minute or so to check out. If I immediately tap and open
> up a payment channel with the store when I start checkout, rather than
> finish, there can be more network propagation of that transaction while
> they are scanning all the items. They'll see the channel is open and
> start adding all the items I want to buy to that micro payment channel.
> I'm identified because I made a payment, not because I've transmitted a
> unique resource or used a unique public key to encrypt communication. A
> payment terminal would not allow for new payment channels to be open
> until the currently active one is closed. If I don't have enough funds
> left in the payment channel, they just stop scanning items. There may be
> some additional privacy implications of setting up micro payment
> channels all the time for everyday purchases. It also may not work for
> every sales case, so we may still need some way to authenticate the
> payer with a unique identifier. So, maybe we don't want to do this, but
> it is just a thought to consider.

It's an interesting thought. As you say, it may be more of a cultural
than technical issue.

> So, unless someone thinks what I am proposing in my previous paragraph
> has any potential (as a complete solution, not a complement to
> solutions), the plan is the following:
> 
>   * Get rid of the "h=" parameter

Agree.

>   * Add a "s=" parameter that uses a unique public key for each session.
>     This public key identifies the payee to the payer and payer to the
>     payee.

This would be the simple model, which just tacks on another parameter to
the bitcoin URL:

bitcoin:[address]?bt=<mac>&s=<key>

But we should also look at the more flexible "r#" approach from your
existing TBIPs, which would yield:

bitcoin:[address]?r=bt:<mac>/<key>

and incorporate the "payment_url" list.

>   * Use a base58 encoding to save space and reduce the character set
>     slightly.

:)

>   * Get rid of the resource? If a terminal is accepting payment from
>     multiple customers simultaneously, it should be smart enough to
>     distinguish between customers based on the public key they are
>     encrypting the data with. Is this approach feasible?

Yes, it is not necessary on the URL. But an id is useful in helping the
BT terminal identify the session without having to try all of its
outstanding keys until it finds one that works.

I proposed that the resource name ("session id" may be a better name) be
deterministically derived from the session key. Given the design change
to pass an EC public key it would need to be derived from that key (not
from the session key because the receiver would not have a copy before
decrypting the first BT message). So any function on the public key that
reduces it to a smaller length, fixed width should be fine. Hashing it
first may be better as is prevents disclosure of any bits of the public
key, which should be treated as a secret during the session.

>   * When you said a new public key for each tap, do you see that as
>     every single tap, or do you consider multiple taps from the same
>     customer the same tap?

Yes, since there would be no other way to distinguish between customers
in some scenarios and this is the safest approach. We certainly won't
run out of numbers, and unused sessions can be discarded based on any
number of criteria, including discarding all but the most recent. That
may may be sufficient for your vending machines given there's little if
any call for parallelism.

e

> On 02/24/2015 06:28 AM, Eric Voskuil wrote:
>> On 02/23/2015 09:53 PM, Andy Schroder wrote:
>>> I was saying provide a public key via NFC (or a public key fingerprint
>>> and then send the full public key over bluetooth). Instead of providing
>>> a new public key on each tap, why can't the payee just stop accepting
>>> connections from new parties on that "resource" after a session key has
>>> been received from the first person?
>> Because the presumption was that there was not an additional secret in
>> the URI. If the public key is reused then anyone can spoof a payer and
>> obtain payment requests.
>>
>> Adding a secret to the URI can resolve this, as long as it is encrypted
>> with the public key before being transmitted back to BT. Otherwise the
>> secret can be intercepted and replayed to the terminal, encrypted with
>> the well-known public key.
>>
>> So if you want to treat the "resource" as a secret this would work.
>> However the resource was designed as a public session identifier,
>> leading the byte stream. This changes it to private session identifier,
>> which loses some utility.
>>
>> Also, reuse of the public key introduces a forward secrecy problem and
>> the potential for persistent seller impersonation in the case of
>> undiscovered key compromise.
>>
>> But there's really no benefit to reusing the key. An ephemeral key
>> resolves these issues and can also seed the public resource name.
>>
>>> If the person decides to have there
>>> friend or family pay for them instead and cancel the payment, they could
>>> just hit cancel on the POS or something (on my fuel pump I have a switch
>>> that needs to be turned, the purpose of this is to avoid wasting too
>>> many addresses)
>> Don't you have someone stop by the pump once a week and empty out the
>> addresses? :)
>>
>>> and/or do another NFC tap (if you're providing QR codes
>>> you'd still need a button of some kind though so it knows to refresh
>>> it), or the POS can just provide a completely new payment request to any
>>> new connections on that same "resource" which use a different session key.
>>>
>>> I feel like the authentication of the payer to the payee in any future
>>> connections after they receive the session key from them (which was
>>> encrypted with the payees public key), comes from the fact that they are
>>> sending responses back that are encrypted using the session key they
>>> gave to the payee. The way I am seeing it is that the NFC tap or QR code
>>> scan is acting in addition to the visual name check on the signature
>>> verification in the wallet.
>> With a secure channel that identifies the parties by proximity, the
>> reason for the payment request signature is for the payer to obtain a
>> non-repudiation guarantee. But it also serves as a defense-in-depth
>> solution to a compromise of the channel (though does not offer a benefit
>> in the case of seller terminal/cert compromise).
>>
>>> If the certificate used isn't signed by a CA
>>> (self signed), it may be fine as long as you heard about it via NFC or
>>> QR code. I don't think it will require PKI and should still work
>>> wallet-to-wallet.
>> In that case the cert provides no benefit. A self-signed cert can be
>> repudiated and if the channel is compromised anyone can sign the payment
>> request.
>>
>>> It sounds like you are saying I'm proposing the customer is going to
>>> need a certificate signed by CA? If so, why? 
>> This was not a serious proposal, it was to point out what would become
>> necessary if the payer could not be identified by proximity.
>>
>> In the case where a public key is reused, any payer can contact the BT
>> terminal and obtain the payment request. If the merchant can't rely on
>> proximity (i.e. can't trust the integrity of the NFC connection) then he
>> would have to fall back on some other means of identifying the payer. A
>> mutual verbal/visual confirmation could work, but the point of of NFC+BT
>> is elimination of that hassle.
>>
>> Yes, it sounds a bit wild, but I have seen on this list a serious
>> proposal to have people broadcast their photo, having the merchant
>> select them and push to them the payment request. Of course anyone can
>> spoof another's image, so at some point your image would need to be
>> certified, and hence a CA.
>>
>> I wouldn't go there, but was just making the point.
>>
>>> I don't need this for any https website I visit.
>> When you go to a web site you first establish a private communication.
>> The site doesn't know who you are (hopefully). Then you log on with your
>> secret, or proof of it, establishing who you are. Customer identity
>> problem solved.
>>
>> Or you create an account, providing your relevant identity information
>> which effectively becomes who you are to the site.
>>
>> Or you shop anonymously and when you go to check out they know that if
>> you pay, you get permission to direct the product shipment. And only you
>> can see the bill. This because your session binds your shopping to your
>> bill and payment.
>>
>> However when you go to the local adult shop to pick up some love toys,
>> the person at the counter has no idea who's asking their terminal for a
>> payment request. You having the shop's public cert doesn't help them
>> with that problem (nor does some anonymous signal sending them a photo
>> of you). Protecting your privacy ironically requires that they know who
>> you are - electronically. That means some sort of crazy consumer cert
>> (not sure that would fly in the love shop)... or trust in
>> (electronically anonymous) proximity.
>>
>>> It's not like the payee is sending anything to
>>> the payer that is private. The payment request only becomes private if
>>> something is actually received to it, otherwise, it is just discarded
>>> and it doesn't matter.
>> The payment request is private. It's a (potentially signed) proposal to
>> contract. It can contain interesting information.
>>
>>> Those bitcoin addresses are never used. It's just
>>> like a shopping cart on a website where someone aborts payment and
>>> cancels the order.
>> Very much so, but in that case your neighbors can't read your potential
>> transactions because your session is secured.
>>
>>> At one point I was thinking we could do something similar to Mike
>>> Hearn's suggestion in another recent e-mail where we re-use some
>>> existing part of the bitcoin URI to bootstrap some trust in a public key
>>> that the payee next sends via bluetooth after the NFC connection. Now
>>> that I'm reviewing my notes though, I can't see how this will work with
>>> a watching only wallet or if no backwards compatible (to BIP21) bitcoin
>>> address is presented in the URI (as Mike said).
>> It can work, but you just end up putting an additional value on the URI
>> (for watchers), requiring legacy addresses (for non-watchers), adding
>> P2SH scripts to the BT broadcast of the public key, and adding another
>> BT round trip to obtain a public key before establishing the session.
>>
>> A few bytes on the NFC tap is a non-issue, especially in comparison to
>> the additional complexity and BT traffic. Those choices are really all
>> based on providing private offline transaction support originating from
>> generally not private QR code scanning. But QR+BT is not the same as NFC+BT.
>>
>> Honestly I think it would be reasonable to use the technique with QR+BT,
>> accepting the limitations for the legacy system while not unduly
>> burdening NFC+BT just for an unachievable cross-consistency goal. Always
>> passing the key on the URL for NFC but giving a non-NFC wallet the
>> option to ask a BT terminal for a public key seems not just reasonable
>> but optimal if we want to support the QR+BT scenario.
>>
>> Note also that the BT-only scenario is different as well (see recent
>> discussion on Airbitz BLE wallet, resulting in the RedPhone-based
>> proposal). And finally, QR-only and NFC-only are also different. The
>> URIs can be consistent, but the communication protocol will vary.
>>
>>> What I was saying above about how you can stop accepting connections on
>>> that "resource" after a session key has been received by the first
>>> person could be problematic though. An evil person could just start
>>> making connections to every device they can, just to be mean, which
>>> would not allow the POS operator to receive payments from their real
>>> customers. If you do the other option I proposed, which is to just keep
>>> giving out new payment requests, you have other problems (on top of
>>> wasting addresses), which are that you can still have mean people giving
>>> you a denial of service attach on your hardware, or you could have an
>>> unusual situation where two people pay (don't know why they would do
>>> this though), so that is why I'm suggesting a manual tap or button press
>>> or switch turn being required.
>> Yes, but even with a manual button you could have these problems. The
>> data transfer needs to be proximate as well.
>>
>>> I guess as more of a abuse filter, a new "resource" could be given
>>> instead with each tap, and the POS would just ignore all requests to an
>>> inactive resource. You may say, why not send a new public key (as you
>>> suggested) instead of a new "resource" with each tap (or button press if
>>> using QR codes), and then you can skip the sending of a static public
>>> key (or public key fingerprint), and ignore any data that is not
>>> encrypted with that public key. Maybe that is a better idea because it
>>> will shorten the bitcoin URI. However, I don't think its required from a
>>> privacy standpoint, it primarily just aids in combining the public key
>>> fingerprint with the changing "resource" name used to filter abuse. Or,
>>> am I missing something?
>> I think this question is covered above.
>>
>>> So, after thinking through the abuse scenarios I mentioned above, I
>>> think I am agreeing with you, but the reason I'm writing all this is to
>>> hopefully just get some feedback on my logic to learn something from
>>> this discussion. I do think sending a unique public key over NFC has to
>>> be better than a unique session key. It adds one more step, but seems to
>>> help. 
>> It doesn't actually add another step to the protocol, just some
>> different but simple code on each end. The only downside is that it
>> extends the NFC URL about 23 characters.
>>
>>> If we do this, can we then safely get rid of the h= parameter?
>> Absolutely, and I believe Mike ack'd this on a previous post today.
>>
>>> That should make Mike Hearn happy, and also may alleviate the base64url
>>> debate?
>> Others may not be aware of the encoding squabble (not sure if it
>> qualifies as debate). In the proposed URL, it affects the mac address
>> and the key:
>>
>> bitcoin:[address]?bt:<mac>/<key>
>>
>> base58:
>> bitcoin:?r=bt:12rAs9mM/234KF8hPkXq5pa6pT1wnJC3hVH7W6yB2Wtw24ztzGtBc4
>>
>> base64url:
>> bitcoin:?r=bt:ABBgss5s/A3xWlhB1GI_t2NMR9Zq9E47hZOzmZ6eZTS8sbq-liugh
>>
>> I prefer base58 because it's available to all bitcoin libraries, nearly
>> as compact as base64 (+1 byte in our example) and better standardized.
>> Some embedded device people might care about having to incorporate
>> base64 as well as base58.
>>
>> It's also better looking (no - or _ characters) and more consistent in
>> the proposed URL (all three values would be base58, as opposed to one
>> base58 and two base64url). There may be some idea that base58 is just
>> for bitcoin addresses (not true) or designed for humans... that's sort
>> of the point, but it's also good for URLs.
>>
>> e
>>
>>> On 02/23/2015 09:55 PM, Eric Voskuil wrote:
>>>> Andy, adding to my previous post below:
>>>>
>>>> On 02/23/2015 01:40 AM, Eric Voskuil wrote:
>>>>> On 02/22/2015 11:36 PM, Andy Schroder wrote:
>>>> ...
>>>>>> It's possible a really sophisticated modification could be done where
>>>>>> the attacker encrypts and decrypts the communication and then relays to
>>>>>> each party (without them knowing or any glitches detected), but I guess
>>>>>> I'm not sure how easy that would be on such a close proximity device?
>>>>> If the NFC tap is sufficiently private, privacy is easy to achieve for
>>>>> the subsequent communication. If it is not, privacy can be completely
>>>>> compromised. The question is only how much more difficult is the attack.
>>>>>
>>>>> With the public cert tap, the level of difficulty is much lower for
>>>>> capturing selected payment requests. The interloper no longer needs to
>>>>> invade the space of the NFC terminal and can instead impersonate the
>>>>> payer from a safe distance. Nobody gets paid, but privacy is
>>>>> compromised.
>>>> This problem in the preceding paragraph can be resolved by sending a
>>>> unique public key on each NFC tap. In that case an attacker would need
>>>> to monitor the NFC communication.
>>>>
>>>> The talk of wrapping the connection in SSL led me to believe you were
>>>> talking about a static public certificate. However that's not a
>>>> necessary assumption here and may not be what you intended.
>>>>
>>>>> The level of difficulty in the case where the interloper wants to taint
>>>>> transactions may appear lower, but it is not:
>>>>>
>>>>> With the session key tap the interloper must compromise the NFC location
>>>>> and then monitor the BT traffic. Monitoring BT traffic without being
>>>>> party to the connection is presumably not rocket surgery, but not
>>>>> standard BT design either.
>>>>>
>>>>> With the public cert tap the interloper must also compromise the NFC
>>>>> location and communicate over BT. Therefore the hardware and physical
>>>>> attack requirements are similar. The only added difficulty is that the
>>>>> attack on the NFC terminal attack is active (modifying the MAC address
>>>>> directing the payer to the BT service).
>>>> I believe your central claim was that the difference in the two
>>>> bootstrapping approaches (public key vs. session key) is that by using a
>>>> unique public key per tap, the attack requires an active vs. passive
>>>> attack on the NFC terminal. I just wanted to make clear here that I
>>>> agree with that assessment.
>>>>
>>>> The symmetric key approach is based on the idea that these attacks are
>>>> comparable in difficulty and otherwise identical in privacy loss.
>>>>
>>>> However, the difference in implementation amounts to about +23
>>>> additional encoded characters for the BT/LE URL, assuming use of the
>>>> secp256k1 curve for DHE. This is really not a material issue in the case
>>>> of the NFC tap. The entire URI+URL could be as small as:
>>>>
>>>> bitcoin:?r=bt:12rAs9mM/79bq48xJaMgqR9YNxnWhqHHM1JB52nxn6VFXBHTP2zrP
>>>>
>>>> In comparison to a symmetric key:
>>>>
>>>> bitcoin:?r=bt:12rAs9mM/12drXXUifSrRnXLGbXg8E
>>>>
>>>> It also does not change the protocol design or complexity at all - it
>>>> would just swap out an AES key for a secp256k1 public key.
>>>>
>>>> bitcoin:[address]?bt:<mac>/<key>
>>>>
>>>> If that gets us aligned I'm all for it.
>>>>
>>>>> However impersonating the payer is just a matter of software - no more
>>>>> difficult than the session key attack. In fact it may be much easier to
>>>>> implement, as the attack can use supported BT features because the
>>>>> attacker has directed the payer to connect to him and is connecting to
>>>>> the receiver as if he was a payer.
>>>>>
>>>>> But it gets worse for the public cert tap, since a more sophisticated
>>>>> attacker can set himself up in the same position without subverting the
>>>>> NFC terminal at all. By broadcasting a more powerful BT service on the
>>>>> same advertised MAC address, the attacker can capture traffic and relay
>>>>> it to the intended service.
>>>> I'm retracting the last paragraph, since the interloper, without
>>>> invading the NFC connection (by substituting the public cert), could not
>>>> read the relayed traffic. It was getting late :/
>>>>
>>>>> So in sum, reliance on a public cert makes the communication less
>>>>> private under the same physical set of constraints. The difference
>>>>> results from the receiver allowing non-proximate payers to impersonate
>>>>> proximate payers from a distance by generating their own session keys
>>>>> and submitting them over BT.
>>>> e
>>>>
>>>
> 


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 473 bytes --]