Fwd: [Lightning-dev] AMP: Atomic Multi-Path Payments over Lightning
(too old to reply)
Bryan Bishop via bitcoin-dev
2018-02-08 17:49:23 UTC
---------- Forwarded message ----------
From: Olaoluwa Osuntokun <***@gmail.com>
Date: Mon, Feb 5, 2018 at 11:26 PM
Subject: [Lightning-dev] AMP: Atomic Multi-Path Payments over Lightning
To: lightning-dev <lightning-***@lists.linuxfoundation.org>

Hi Y'all,

A common question I've seen concerning Lightning is: "I have five $2
channels, is it possible for me to *atomically* send $6 to fulfill a
payment?". The answer to this question is "yes", provided that the receiver
waits to pull all HTLC's until the sum matches their invoice. Typically, one
assumes that the receiver will supply a payment hash, and the sender will
re-use the payment hash for all streams. This has the downside of payment
hash re-use across *multiple* payments (which can already easily be
correlated), and also has a failure mode where if the sender fails to
actually satisfy all the payment flows, then the receiver can still just
pull the monies (and possibly not disperse a service, or w/e).

Conner Fromknecht and I have come up with a way to achieve this over
Lightning while (1) not re-using any payment hashes across all payment
flows, and (2) adding a *strong* guarantee that the receiver won't be paid
until *all* partial payment flows are extended. We call this scheme AMP
(Atomic Multi-path Payments). It can be experimented with on Lightning
*today* with the addition of a new feature bit to gate this new
feature. The beauty of the scheme is that it requires no fundamental changes
to the protocol as is now, as the negotiation is strictly *end-to-end*
between sender and receiver.

TL;DR: we repurpose some unused space in the onion per-hop payload of the
onion blob to signal our protocol (and deliver some protocol-specific data),
then use additive secret sharing to ensure that the receiver can't pull the
payment until they have enough shares to reconstruct the original pre-image.

Protocol Goals
1. Atomicity: The logical transaction should either succeed or fail in
entirety. Naturally, this implies that the receiver should not be unable to
settle *any* of the partial payments, until all of them have arrived.

2. Avoid Payment Hash Reuse: The payment preimages validated by the
consensus layer should be distinct for each partial payment. Primarily,
this helps avoid correlation of the partial payments, and ensures that
malicious intermediaries straddling partial payments cannot steal funds.

3. Order Invariance: The protocol should be forgiving to the order in which
partial payments arrive at the destination, adding robustness in the face of
delays or routing failures.

4. Non-interactive Setup: It should be possible for the sender to perform an
AMP without directly coordinating with the receiving node. Predominantly,
this means that the *sender* is able to determine the number of partial
payments to use for a particular AMP, which makes sense since they will be
the one fronting the fees for the cost of this parameter. Plus, we can
always turn a non-interactive protocol into an interactive one for the
purposes of invoicing.

Protocol Benefits

Sending pay payments predominantly over an AMP-like protocol has several
clear benefits:

- Eliminates the constraint that a single path from sender to receiver
with sufficient directional capacity. This reduces the pressure to have
larger channels in order to support larger payment flows. As a result,
the payment graph be very diffused, without sacrificing payment

- Reduces strain from larger payments on individual paths, and allows the
liquidity imbalances to be more diffuse. We expect this to have a
non-negligible impact on channel longevity. This is due to the fact that
with usage of AMP, payment flows are typically *smaller* meaning that
each payment will unbalance a channel to a lesser degree that
with one giant flow.

- Potential fee savings for larger payments, contingent on there being a
super-linear component to routed fees. It's possible that with
modifications to the fee schedule, it's actually *cheaper* to send
payments over multiple flows rather than one giant flow.

- Allows for logical payments larger than the current maximum value of an
individual payment. Atm we have a (temporarily) limit on the max payment
size. With AMP, this can be side stepped as each flow can be up the max
size, with the sum of all flows exceeding the max.

- Given sufficient path diversity, AMPs may improve the privacy of LN
Intermediaries are now unaware to how much of the total payment they are
forwarding, or even if they are forwarding a partial payment at all.

- Using smaller payments increases the set of possible paths a partial
payment could have taken, which reduces the effectiveness of static
analysis techniques involving channel capacities and the plaintext
values being forwarded.

Protocol Overview
This design can be seen as a generalization of the single, non-interactive
payment scheme, that uses decoding of extra onion blobs (EOBs?) to encode
extra data for the receiver. In that design, the extra data includes a
payment preimage that the receiver can use to settle back the payment. EOBs
and some method of parsing them are really the only requirement for this
protocol to work. Thus, only the sender and receiver need to implement this
feature in order for it to function, which can be announced using a feature

First, let's review the current format of the per-hop payload for each node
described in BOLT-0004.

│Realm (1 byte) │Next Addr (8 bytes)│Amount (8 bytes)│Outgoing CLTV (4
bytes)│Unused (12 bytes)│ HMAC (32 bytes) │
│65 Bytes Per Hop │

Currently, *each* node gets a 65-byte payload. We use this payload to give
each node instructions on *how* to forward a payment. We tell each node: the
realm (or chain to forward on), then next node to forward to, the amount to
forward (this is where fees are extracted by forwarding out less than in),
the outgoing CLTV (allows verification that the prior node didn't modify any
values), and finally an HMAC over the entire thing.

Two important points:
1. We have 12 bytes for each hop that are currently unpurposed and can be
used by application protocols to signal new interpretation of bytes and
also deliver additional encrypted+authenticated data to *each* hop.

2. The protocol currently has a hard limit of 20-hops. With this feature
we ensure that the packet stays fixed sized during processing in order to
avoid leaking positional information. Typically most payments won't use
all 20 hops, as a result, we can use the remaining hops to stuff in *even
more* data.

Protocol Description
The solution we propose is Atomic Multi-path Payments (AMPs). At a high
level, this leverages EOBs to deliver additive shares of a base preimage,
from which the payment preimages of partial payments can be derived. The
receiver can only construct this value after having received all of the
partial payments, satisfying the atomicity constraint.

The basic protocol:

Let H be a CRH function.
Let || denote concatenation.
Let ^ denote xor.

Sender Requirements
The parameters to the sending procedure are a random identifier ID, the
number of partial payments n, and the total payment value V. Assume the
sender has some way of dividing V such that V = v_1 + 
 + v_n.

To begin, the sender builds the base preimage BP, from which n partial
preimages will be derived. Next, the sender samples n additive shares s_1,

, s_n, and takes the sum to compute BP = s_1 ^ 
 ^ s_n.

With the base preimage created, the sender now moves on to constructing the
n partial payments. For each i in [1,n], the sender deterministically
computes the partial preimage r_i = H(BP || i), by concatenating the
sequence number i to the base preimage and hashing the result. Afterwards,
it applies H to determine the payment hash to use in the i’th partial
payment as h_i = H(r_i). Note that that with this preimage derivation
scheme, once the payments are pulled each pre-image is distinct and
indistinguishable from any other.

With all of the pieces in place, the sender initiates the i’th payment by
constructing a route to the destination with value v_i and payment hash h_i.
The tuple (ID, n, s_i) is included in the EOB to be opened by the receiver.

In order to include the three tuple within the per-hop payload for the final
destination, we repurpose the _first_ byte of the un-used padding bytes in
the payload to signal version 0x01 of the AMP protocol (note this is a PoC
outline, we would need to standardize signalling of these 12 bytes to
support other protocols). Typically this byte isn't set, so the existence of
this means that we're (1) using AMP, and (2) the receiver should consume the
_next_ hop as well. So if the payment length is actually 5, the sender tacks
on an additional dummy 6th hop, encrypted with the _same_ shared secret for
that hop to deliver the e2e encrypted data.

Note, the sender can retry partial payments just as they would normal
payments, since they are order invariant, and would be indistinguishable
from regular payments to intermediaries in the network.

Receiver Requirements

Upon the arrival of each partial payment, the receiver will iteratively
reconstruct BP, and do some bookkeeping to figure out when to settle the
partial payments. During this reconstruction process, the receiver does not
need to be aware of the order in which the payments were sent, and in fact
nothing about the incoming partial payments reveals this information to the
receiver, though this can be learned after reconstructing BP.

Each EOB is decoded to retrieve (ID, n, s_i), where i is the unique but
unknown index of the incoming partial payment. The receiver has access to
persistent key-value store DB that maps ID to (n, c*, BP*), where c*
represents the number of partial payments received, BP* is the sum of the
received additive shares, and the superscript * denotes that the value is
being updated iteratively. c* and BP* both have initial values of 0.

In the basic protocol, the receiver cache’s the first n it sees, and
verifies that all incoming partial payments have the same n. The receiver
should reject all partial payments if any EOB deviates. Next, the we update
our persistent store with DB[ID] = (n, c* + 1, BP* ^ s_i), advancing the
reconstruction by one step.

If c* + 1 < n, there are still more packets in flight, so we sit tight.
Otherwise, the receiver assumes all partial payments have arrived, and can
being settling them back. Using the base preimage BP = BP* ^ s_i from our
final iteration, the receiver can re-derive all n partial preimages and
payment hashes, using r_i = H(BP || i) and h_i = H(r_i) simply through
knowledge of n and BP.

Finally, the receiver settles back any outstanding payments that include
payment hash h_i using the partial preimage r_i. Each r_i will appear random
due to the nature of H, as will it’s corresponding h_i. Thus, each partial
payment should appear uncorrelated, and does not reveal that it is part of
an AMP nor the number of partial payments used.

Non-interactive to Interactive AMPs

Sender simply receives an ID and amount from the receiver in an invoice
before initiating the protocol. The receiver should only consider the
invoice settled if the total amount received in partial payments containing
ID matches or exceeds the amount specified in the invoice. With this
variant, the receiver is able to map all partial payments to a pre-generated
invoice statement.

Additive Shares vs Threshold-Shares

The biggest reason to use additive shares seems to be atomicity. Threshold
shares open the door to some partial payments being settled, even if others
are left in flight. Haven’t yet come up with a good reason for using
threshold schemes, but there seem to be plenty against it.

Reconstruction of additive shares can be done iteratively, and is win for
the storage and computation requirements on the receiving end. If the sender
decides to use fewer than n partial payments, the remaining shares could be
included in the EOB of the final partial payment to allow the sender to
reconstruct sooner. Sender could also optimistically do partial
reconstruction on this last aggregate value.

Adaptive AMPs

The sender may not always be aware of how many partial payments they wish to
send at the time of the first partial payment, at which point the simplified
protocol would require n to be chosen. To accommodate, the above scheme can
be adapted to handle a dynamically chosen n by iteratively constructing the
shared secrets as follows.

Starting with a base preimage BP, the key trick is that the sender remember
the difference between the base preimage and the sum of all partial
preimages used so far. The relation is described using the following

X_0 = 0
X_i = X_{i-1} ^ s_i
X_n = BP ^ X_{n-1}

where if n=1, X_1 = BP, implying that this is in fact a generalization of
the single, non-interactive payment scheme mentioned above. For i=1, ...,
n-1, the sender sends s_i in the EOB, and X_n for the n-th share.

Iteratively reconstructing s_1 ^ 
. ^ s_{n-1} ^ X_n = BP, allows the
receiver to compute all relevant r_i = H(BP || i) and h_i = H(r_i). Lastly,
the final number of partial payments n could be signaled in the final EOB,
which would also serve as a sentinel value for signaling completion. In
response to DOS vectors stemming from unknown values of n, implementations
could consider advertising a maximum value for n, or adopting some sort of
framing pattern for conveying that more partial payments are on the way.

We can further modify our usage of the per-hop payloads to send (H(BP),
s_i) to
consume most of the EOB sent from sender to receiver. In this scenario, we'd
repurpose the 11-bytes *after* our signalling byte in the unused byte
to store the payment ID (which should be unique for each payment). In the
of a non-interactive payment, this will be unused. While for interactive
payments, this will be the ID within the invoice. To deliver this slimmer
2-tuple, we'll use 32-bytes for the hash of the BP, and 32-bytes for the
partial pre-image share, leaving an un-used byte in the payload.

Cross-Chain AMPs

AMPs can be used to pay a receiver in multiple currencies atomically...which
is pretty cool :D

Open Research Questions

The above is a protocol sketch to achieve atomic multi-path payments over
Lightning. The details concerning onion blob usage serves as a template that
future protocols can draw upon in order to deliver additional data to *any*
hop in the route. However, there are still a few open questions before
something like this can be feasibly deployed.

1. How does the sender decide how many chunked payments to send, and the
size of each payment?

- Upon a closer examination, this seems to overlap with the task of
congestion control within TCP. The sender may be able to utilize
inspired heuristics to gauge: (1) how large the initial payment should
and (2) how many subsequent payments may be required. Note that if the
first payment succeeds, then the exchange is over in a signal round.

2. How can AMP and HORNET be composed?

- If we eventually integrate HORNET, then a distinct communications
sessions can be established to allow the sender+receiver to exchange
up-to-date partial payment information. This may allow the sender to
accurately size each partial payment.

3. Can the sender's initial strategy be governed by an instance of the
Push-relabel max flow algo?

4. How does this mesh with the current max HTLC limit on a commitment?

- ATM, we have a max limit on the number of active HTLC's on a particular
commitment transaction. We do this, as otherwise it's possible that the
transaction is too large, and exceeds standardness w.r.t transaction
size. In a world where most payments use an AMP-like protocol, then
overall ant any given instance there will be several pending HTLC's on
commitments network wise.

This may incentivize nodes to open more channels in order to support
the increased commitment space utilization.


We've presented a design outline of how to integrate atomic multi-path
payments (AMP) into Lightning. The existence of such a construct allows a
sender to atomically split a payment flow amongst several individual payment
flows. As a result, larger channels aren't as important as it's possible to
utilize one total outbound payment bandwidth to send several channels.
Additionally, in order to support the increased load, internal routing nodes
are incensed have more active channels. The existence of AMP-like payments
may also increase the longevity of channels as there'll be smaller, more
numerous payment flows, making it unlikely that a single payment comes
across unbalances a channel entirely. We've also showed how one can utilize
the current onion packet format to deliver additional data from a sender to
receiver, that's still e2e authenticated.

-- Conner && Laolu

Lightning-dev mailing list
- Bryan
1 512 203 0507