Date: Mon, 7 Nov 2022 21:46:28 +1000
From: Anthony Towns <aj@erisian.com.au>
To: Antoine Riard <antoine.riard@gmail.com>,
 Bitcoin Protocol Discussion <bitcoin-dev@lists.linuxfoundation.org>
Message-ID: <Y2jwFOeVep6WefKb@erisian.com.au>
References: <Y2HpaWnAC5dIGkx+@erisian.com.au>
 <CALZpt+G7SCTtsHUjdxdYYxFoQW4k7_BEbygVx2ksgZpTQZBUTw@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <CALZpt+G7SCTtsHUjdxdYYxFoQW4k7_BEbygVx2ksgZpTQZBUTw@mail.gmail.com>
Subject: Re: [bitcoin-dev] Preventing/detecting pinning of jointly funded txs
Precedence: list

On Sun, Nov 06, 2022 at 06:22:08PM -0500, Antoine Riard via bitcoin-dev wrote:
> Adding a few more thoughts here on what coinjoins/splicing/dual-funded
> folks can do to solve this DoS issue in an opt-in RBF world only.
> 
> I'm converging that deploying a distributed monitoring of the network
> mempools in the same fashion as zeroconf people is one solution, as you can
> detect a conflicting spend of your multi-party transaction.

> Let's say you
> have a web of well-connected full-nodes, each reporting all their incoming
> mempool transactions to some reconciliation layer.
> 
> This "mempools watchdog" infrastructure isn't exempt from mempools
> partitioning attacks by an adversary,

A mempool partitioning attack requires the adversary to identify your
nodes. If they're just monitoring and not being used as the initial
broadcaster of your txs, that should be difficult. (And I think it would
make sense to do things that make it more difficult to successfully
partition a node for tx relay, even if you can identify it)

> where the goal is to control your
> local node mempool view. A partitioning trick is somehow as simple as
> policy changes across versions (e.g allowing Taproot Segwit v0.1 spends) or
> two same-feerate transactions. The partitioning attack can target at least
> two meaningful subsets.

An even easier way to partition the network is to create two conflicting
txs at the same fee/fee rate and to announce them to many peers
simultaneously. That way neither will replace the other, and you can
build from there.

In order to attack you, the partition would need to be broad enough
to capture all your monitoring nodes on one side (to avoid detection),
and all the mining nodes on the other side (to prevent your target tx
from being confirmed). If that seems likely, maybe it indicates that
it's too easy to identify nodes that feed txs to mining pools...

> Either the miner mempools only, by conflicting all
> the reachable nodes in as many subsets with a "tainted" transaction (e.g
> set a special nSequence value for each), and looking on corresponding
> issued block. Or targeting the "watchdog" mempools only, where the
> adversary observation mechanism is the multi-party blame assignment round
> itself.

I think there's a few cases like that: you can find out what txs mining
pools have seen by looking at their blocks, what explorers have seen by
looking at their website...

Being able to use that information to identify your node(s) -- rather
than just your standardness policy, which you hopefully share with many
other nodes -- seems like something we should be working to fix...

> There is an open question on how many "divide-and-conquer" rounds
> from an adversary viewpoint you need to efficiently identify all the
> complete set of "mempools watchdog". If the transaction-relay topology is
> highly dynamic thanks to outbound transaction-relay peers rotation, the
> hardness bar is increased.

I'm not sure outbound rotation is sufficient? In today's world, if
you're a listening node, an attacker can just connect directly, announce
the conflicting tx to you, and other txs to everyone else.

For a non-listening node, outbound rotation might be more harmful than
helpful, as it increases the chances a node will peer with a given
attacker at some point.

> Though ultimately, the rough mental model I'm thinking on, this is a
> "cat-and-mouse" game between the victims and the attacker, where the latter
> try to find the blind spots of the former. I would say there is a strong
> advantage to the attacker, in mapping the mempools can be batched against
> multiple sets of victims. While the victims have no entry barriers to
> deploy "mempools watchdog" there is a scarce resource in contest, namely
> the inbound connection slots (at least the miners ones).

Anytime you deploy a new listening node, you use up 10 inbound
connections, but provide capacity for 115 new ones. Worst case (if
it's too hard to prevent identifying a listening node if you use it for
monitoring), you setup all your monitoring nodes as non-listening nodes,
and also set enough listening nodes in different IP ranges to compenasate
for the number of outbound connections your monitoring nodes are making,
and just ignore the mempools of those listening nodes.

> Victims could batch their defense costs, in outsourcing the monitoring to
> dedicated entities (akin to LN watchtower). However, there is a belief in
> lack of a compensation mechanism, you will have only a low number of public
> ones (see number of BIP157 signaling nodes, or even Electrum ones).

Seems like a pretty simple service to pay for, if there's real demand:
costs are pretty trivial, and if it's a LN service, you can pay for
it over lightning... Fairly easy to test if you're getting what you're
paying for too, by simultaneously announcing two conflicting txs paying
yourself, and checking you get an appropriate alert.

> Assuming we can solve them, there is still the issue of assigning blame
> reliably among a set of trust-minimized joint funding protocol
> participating UTXOs. Indeed, you're running quickly into issues like *two*
> double-spend from two sybilling participants, aiming to halt the assignment
> process.

I don't see how that's a problem: anyone who wants to continue as part of
the group never signs a conflicting tx; anyone who does sign a conflicting
tx is saying "I don't want to be part of this group anymore", whether
that conflicting tx arrives via normal channels or p2p. If you want a way
of saying "I want to abort this coinjoin, but stay a part of the group",
you need to get 51% of the rest of the group to sign off on that claim
(outside of the blockchain/mempool), before getting control of your
utxo back. If you can't get 51% of the group to sign off on that in a
reasonable time, then you should just exit the group.

I don't think it makes much sense for "group A said Bob's a cheater,
so therefore we won't let Bob into group C either".

> for each round of blame assignment. Assuming 255 inputs (current LN's
> interactive construction protocol limit) and a transaction propagation
> delay of 2min (30s ?) on the p2p network, an attacker controlling all the
> inputs minus 1 might be able to DoS for ~50 blocks (do we have other
> factors to think of in the design of the blame assignment process ?).

Hmm, in messing around with dandelion simulations, I've found 10s to 30s
seems more reasonable for propagation to 95% of the network, though that's
modelled without delays due to latency/low bandwidth, block propagation,
or competition with large/higher value txs.

Shouldn't be hard to estimate that actually: if you setup a node with
-debug=net and -datacarrier=0, then check for blocks with txs with
OP_RETURN outputs, figure out the w/txid and grep your logs for how
far apart. Because you set -datacarrier=0 you won't accept those txs
into your mempool, or announce them to your peers, but your peers will
announce them to you when they receive them, so the time difference
between first and last announcement should give you a decent sampling
of info that you could use to estimate the overall time it takes a tx
to flood through the network.

> On the other hand, you have a full-rbf world, where instead to deploy or
> gain access to "mempools watchdog" and proceed to a timevalue-expensive
> blame assignment protocol, any participant should be able to fee-bump the
> joint transaction (assuming multiple pre-signed feerate version of the
> transactions, or ephemeral, nversion=3 and package-relay to do unilateral
> CPFP).

A full-rbf world doesn't get you that: you can do low fee rate pinning
via other mechanisms than the opt-out flag. In a world where nodes don't
mostly have a consistent mempool policy you introduce new pinning vectors
too; eg, if one of your coinjoin outputs is bare multisig, then that
will pass core's standardness checks, but be rejected by fullrbf knots
nodes that haven't specifically enabled that configuration option.

Even if you assume the attacker doesn't know about those other methods,
or we somehow come up with an eventually consistent mempool/relay policy,
where pinning isn't possible that gets near universal adoption; if you're
only monitoring the blockchain and not the mempool, your hypothetical
attacker can immediately RBF your tx with a higher fee spend of one of
their inputs -- you'll find out about that in the next block, but at one
input per block, that's going to allow you to keep thinking that maybe
one of your peers is honest for 254 blocks instead of 50.

This seems like an argument not to do coinjoins with a ridiculously
large number of untrusted counterparties...

If you really want to salvage the opportunity to have really large
anonymous groups, seems better to quickly say "oops, someone cheated,
let's exclude them, and split the remainder into two groups" until you
either succeed or are down to a more reasonable group size of perhaps
10 or 20.

> [0] "Of course, even assuming full-rbf, propagation of the multi-party
> funded
> transactions can still be interfered with by an attacker, simply
> broadcasting a double-spend with a feerate equivalent to the honest
> transaction. However, it tightens the attack scenario to a scorched earth
> approach, where the attacker has to commit equivalent fee-bumping reserve
> to maintain the pinning and might lose the "competing" fees to miners."

We now know that that isn't correct though: if it really did tighten
the attack scenario like that, that would be great -- but it doesn't:
the attacker can still do low feerate pinning that's not recoverable by
fee bumping the alternative, as Suhas explained elsewhere.

Cheers,
aj