DSS 2026 / Day 2 / Amber Hall
Fund-Flow Modeling by Solving Large Linear Algebra Systems
Yannis Smaragdakis
Source: YouTube automatic caption track · Exported 2026-11-16. Timestamps link to the source video. Cue end times were not supplied.
Transcript
0:00Thank you very much.
0:02Okay.
0:03So
0:05fund flow modeling by solving large
0:07linear algebra systems. And if that
0:09title didn't turn you off, I don't know
0:11what will cuz like it's about as boring
0:14as it can be, right? Large linear
0:16algebra systems. What's what's this all
0:18about? It's actually going to be real
0:20fun.
0:21But the the slides are a bit heavy, so
0:24I'll try to depart a bit from the slides
0:26and try to give you
0:28a bit of an intuitive understanding of
0:30what I'm proposing here. And we can
0:33discuss whether you agree that this is
0:35interesting. I think it's pretty
0:36interesting. It's been my pet project
0:40for a very large part of the past year,
0:42let's say. And the the the result is
0:45deployed on our site. So if you go to
0:48app.dedup.com, you you can see the
0:50results of this algorithm running every
0:52two weeks. And I will tell you what are
0:55the drawbacks and what alternative
0:57algorithms we have for the same problem.
0:58But what's the problem first of all?
1:01So
1:02fund flow.
1:04Every time there's a hack,
1:06everyone asks, where did the money go?
1:09There was recently a certain hack you
1:11may have heard in the past three days.
1:13And everyone is asking, where did the
1:14money go? And they will be asking that
1:16for years to come.
1:18And everyone, meaning exchanges,
1:20investigators, regulators, everyone
1:23wants to know who got funded from the
1:26proceeds of the hack. And the standard
1:28answer is to do taint analysis. And
1:31there's a whole bunch of algorithms
1:33there. And I'm not sure if I can point
1:35at the screen. There's poison. Poison
1:38means
1:39you you get poisoned by some tainted
1:41funds. You transfer the poison to
1:43everyone else. You are marked forever.
1:45Or haircut. You proportionally assign
1:49what portion of your funds is tainted,
1:51etc., etc.
1:52I'm not going to go into all the
1:54alternative algorithms that everyone is
1:56using to to do fund flow. But the main
1:58point is is that everyone else is doing
2:02fund transfer tracking. So you need to
2:05know when money changes hands. You need
2:08to know when there are transfers.
2:11what's the problem with that? It's that
2:14we don't really know all the time how
2:16money changes hands in DeFi.
2:19Uh so in DeFi value can move without
2:23transfers and that's the principle of
2:25money laundering. That's how people
2:27launder money after they get money that
2:30they shouldn't have gotten.
2:32Okay, so examples and this is not an
2:35exhaustive list, but and and you may
2:37have seen many of these. You can deposit
2:39something on behalf of someone else.
2:40Let's say I'm a hacker H and I want to
2:43give value to some other party C. I want
2:47to
2:47give value to C, but not do a token
2:50transfer to C directly. So what
2:53technique can I use? Of course, I can
2:55deposit with any DeFi protocol on their
2:57behalf. That's the most stupid thing.
3:00I can deposit and get an NFT that looks
3:03valueless.
3:05So it looks like I got nothing in
3:06return. I just deposited to a well-known
3:09protocol which is not tainted by itself.
3:11The the well-known protocol is innocent.
3:15And but I got an NFT that's value zero
3:17that has value zero and then I pass it
3:20on. So I passed value zero.
3:22So the other person should not be
3:23tainted, but the other person can now
3:25withdraw from the main protocol.
3:28I can do all sorts of hiding off-chain.
3:31I can have an off-chain order and some
3:35innocent party can jump in and fill that
3:38order
3:39and indirectly the proceeds go to C, the
3:43party that I want to benefit.
3:45So C gets the funds, but I'm never doing
3:49a transaction with C. I'm signing I'm
3:52signing an off-chain order.
3:54And there's lots more.
3:57I can do value injection.
4:00C holds a worthless token and I give it
4:02value. How do I give it value? Through
4:04any combination of liquidity provision
4:08or whatever. I can buy it at an inflated
4:11price. I can
4:13I can enable liquidity in another wide
4:15otherwise worthless pool, whatever. The
4:18point is DeFi lets us create lots of
4:22complex financial mechanisms and all
4:25these financial mechanisms allow value
4:27transfer without a direct token
4:30transfer.
4:32So we all believe this?
4:34Are we okay
4:35with the principle so far? Okay.
4:38So none of this is malicious.
4:41It's very hard to keep up with that.
4:43Obviously not impossible. There are lots
4:45of people who are trying to do it, but
4:47it's very very hard to have a catalog of
4:51all the DeFi protocols and all the
4:52patterns they can use to obscure value
4:55transfer. So what are we trying what are
4:59examples of this kind of valueless
5:01transfer? This is actually in parallel
5:03with the work I was doing. I saw this
5:05paper some seven months ago by people
5:08who were then at Eigen Fi and it said
5:10exactly the same thing. Basically here's
5:13an example of how back a few million
5:17blocks there was 38 million of US 38
5:20million of funds laundered through a
5:23transaction that nobody could track
5:26because it was an off-chain order that
5:27some MEV bot filled in and indirectly
5:31there was value transfer from P to B and
5:34everyone could see the transactions, but
5:36nobody could tell that there was value
5:38transfer there
5:39because there was no direct transfer of
5:42tokens, no direct taint.
5:44So what's the principle that we're
5:46trying to explore here
5:48to exploit here? The principle is what's
5:50in yellow. So this is the entire
5:51presentation.
5:53The principle is we can have an
5:56innocuous DeFi protocol and it can
5:59obscure that I'm passing some value to
6:02someone else.
6:04But if it's an innocuous protocol, what
6:06it will not obscure is that I'm
6:09mentioning that someone else that I'm
6:11passing the value to.
6:12There has to be a transaction where I
6:14participate in
6:16and somehow my address and their address
6:20appear in that same one transaction.
6:24And we have to link that transaction
6:26with others perhaps, but at the very
6:28least, whoever is sending the money, the
6:31who is very supplying the value and
6:33whoever is receiving it, they are named
6:35in one transaction.
6:37Obviously, there's exceptions. The
6:40exceptions are privacy protocols that
6:41obscure the recipient addresses, Tornado
6:44Cash, Railgun. We'll talk about those at
6:46the very end.
6:47But if you take out privacy protocols
6:50and you just keep regular DeFi, which is
6:53used for the
6:54lion's share of money laundering, this
6:56principle holds.
7:00Are we clear? And I want some feedback
7:02because if we haven't followed this,
7:03we're not going to go very far. Are we
7:05clear on what I'm I'm saying here?
7:07Yes.
7:08Excellent. So
7:10I'm trying to link the taint of
7:14co-participants of every transaction,
7:17meaning I'm going to go into all
7:19transactions ever that ever appeared on
7:22chain. I'm going to take the call diet
7:25data for all of them. I'm going to parse
7:27it and find addresses,
7:29send the recipient but anyone who's also
7:31mentioned.
7:32And if they appear anywhere in the same
7:35transaction, I'm going to link all those
7:37addresses in a very large linear system.
7:40And then our problem will be scaling.
7:42Our problem will be solving this very
7:44large linear system. So let's look at a
7:47couple of technical details and then we
7:50can get to the high level again.
7:52So
7:53the model
7:54pretty much says I will give a value to
7:58the taint of every participant, of every
8:02addresses,
8:03of every address. What's the taint? It's
8:05a value that
8:07how to what percentage, let's say, it's
8:10a it's a let's say a fraction that will
8:14give some metric to the taintedness of
8:17an address.
8:19I will take as input the participants of
8:21every transaction, like I said at the
8:22top, and I will take as input some
8:25amounts that will help me with weighing
8:27what transactions are important. And and
8:30then there will be some extra
8:32assumptions here. So I'll start with a
8:34list of known hackers. There are roughly
8:3810,000 addresses that are directly
8:40involved in past hacks so far in
8:42Ethereum. So we start with this
8:44information, and then we we're going to
8:46check how these addresses relate to
8:49every other address that ever
8:51participated in transactions on mainnet.
8:54And the simplifying assumptions here is
8:57that any address that has many thousands
8:59of transactions, I'm going to exclude it
9:01just to make the system a little more
9:03scalable. It would most likely end up
9:05being zero tainted anyway or very close
9:08to zero, but it also helps us with
9:10scalability here to just do it upfront.
9:12So any well-known DeFi protocol will go
9:15to zero right away based on this
9:17criterion.
9:18And the
9:21high-level description of the math is
9:23the sentence up there.
9:25An address is as tainted as the average
9:28of its co-participants in each
9:30transaction.
9:31And then the transaction the
9:34transactions are averaged
9:36compared to other transactions on the
9:38same address based on how much value
9:40that address gets.
9:42And if that address is completely
9:44worthless, they get an arbitrary value
9:46of epsilon, same for every one of their
9:48transactions.
9:50So again, what I'm saying up here is you
9:54participate in a transaction with
9:56someone else. Your taint value is
9:58averaged and contributes to their
10:01taint, and their taint value is averaged
10:04and contributes to your taint.
10:06And the this average is also averaged in
10:10proportion to how much the transaction
10:12contributed to your total worth.
10:14So if we actually
10:16do this in math, it comes to a linear
10:19equation that looks like this and I know
10:21it's it's going to be very hard to
10:23parse, but let's it's it's not really
10:26that hard if you spend two minutes
10:28looking at it. So very quickly I'll just
10:30go over it.
10:32This is how much the transaction matters
10:34to address A.
10:36What does that mean? The address A, if
10:38we take the absolute values of every
10:42funds transferred to or from it, sum
10:44over all of those and weigh the
10:47transaction based on how much value that
10:49transaction contributes, that's this
10:51factor. And then everything else is just
10:54averaging over co-participants in the
10:56same transaction.
10:58So
11:00hopefully this scared you enough, but in
11:02intuition it's very simple. We are just
11:05averaging
11:07all participants in a transaction in a
11:09large linear system. It can be
11:11recursive. My taint contributes to
11:13yours, yours contributes to mine
11:16because we participate in the same
11:17transaction and then we solve that
11:19linear
11:20equation, that large linear system to
11:23get a score. This is very much like
11:25PageRank, but over all participants in
11:28all transactions.
11:31Okay, so
11:33large linear system,
11:35very
11:36very standard, but here are some
11:38intuitive features.
11:41The result of this is not a probability,
11:44it's not a fraction of stolen funds.
11:46It's just a metric. So if I get a 0.7,
11:49it doesn't mean anything other than
11:52probably I'm more tainted than someone
11:54who got a 0.6.
11:56I mean there will be a threshold. Let's
11:59say everything above 0.3 is most likely
12:02very involved in some money laundering,
12:06but
12:07it's not an absolute threshold. It's not
12:09a confidence. It's not a percentage.
12:12Second, everything is symmetric. I
12:15funded a hacker is the same as a hacker
12:18funded me.
12:19It's going to give me the same score.
12:21So it's a completely symmetric system
12:24and it goes back in time. So I funded a
12:27hacker 3 million blocks ago, it's still
12:30going to give me a high tainted score
12:32the way this system is being solved.
12:35Okay, so with all this intuitive
12:37understanding of the system,
12:40you can see results from that system
12:43live if you log in right now as a free
12:46user on app.ddub.com
12:48and you find you start from any address
12:50or you actually look at the list that's
12:52called hack fish related.
12:54That's the name of the list here. You
12:57will get graphs like this. You can go
12:59and explore all the way to taint sources
13:01and when you get to 100% I cannot give a
13:03live demo here. These are frozen slides,
13:06but if you go all the way to taint
13:08sources to 100% and you click on that,
13:11you actually get what incident gave us
13:13the confidence originally that this is a
13:15hacker address. But you can also trace
13:17back and say why does this address have
13:20a 74%?
13:21Oh,
13:23I have at least two transactions that
13:25give me this confidence. There may be
13:26many more and these two transactions
13:29contribute in the linear system and it's
13:31because this address got linked to that
13:34known hacker.
13:36So go and explore and if you want to ask
13:39me about the Binance hack, obviously,
13:42like I said, this is a huge system.
13:44We solve this every two weeks. In two
13:47weeks, in 15 days from now, you will see
13:49how funds moved from the Binance hack.
13:52But right now we only have for the
13:54Binance hack, we only have internal
13:56systems like this one.
13:59So we have some internal data, but on
14:01the public system you cannot see it just
14:03yet. You will see the results in 15 days
14:06though.
14:07Okay, so
14:10just to give you an idea of the scaling
14:12of this system, it's it's huge. We
14:14actually have to solve it at ranges of 1
14:17million blocks, but very, very quickly
14:20right now on the full chain, just
14:23Ethereum. Let's not even go to base,
14:25Binance, anything with higher
14:27transaction rates.
14:29It right now takes on CPU
14:31like 3 days, 2 and 1/2 days to solve the
14:34thing.
14:35Like just on this range of blocks, it
14:37takes 22 hours on an even faster machine
14:39than our production server. But
14:42if this is to scale, if we find like
14:45really good business cases for that, we
14:47can speed it up quite a bit. So on a on
14:49a really expensive server,
14:51like a $150,000 server with 4 GPUs, we
14:55can actually get it to be more than 25
14:57times faster. So we can solve this in an
14:59hour
15:01if we really if it really made a lot of
15:03sense for us, if we had the right
15:06business value for it. So that's kind of
15:08the exercise on how fast we can get in
15:11solving these systems. But again, this
15:13is a very heavy system. This is a very
15:15large system and it's not something that
15:18one should do casually.
15:20Okay, so it does it give good results?
15:23It gives you really good results.
15:25Informally, you can go and explore and
15:27let us know if you don't like what you
15:29see, but it gives pretty good results.
15:32Just an example from like addresses
15:35that were humanly found tainted from the
15:38Bybit hack by ZachXBT.
15:40Automatically, the first time you run
15:42this system, it finds virtually all of
15:45them. Like
15:47from the 52 addresses, 42 of the 52 have
15:51taint scores of over 0.7. All of them
15:54have some taint score that not zero, but
15:57at least 42 of the 52 automatically are
15:59very highly flagged in the taint scores.
16:03And overall, it flags many thousands of
16:06addresses. Like of course, there are 350
16:09million active Ethereum addresses
16:11historically, so 150k is not that many,
16:15but that's exactly what you expect. How
16:17many addresses are in some way tainted?
16:20That's roughly what you would expect
16:22overall.
16:23Like I said, we also have alternative
16:25models that are taint-based, not so
16:28scalable, but this is our most general
16:31model and it's also publicly exported.
16:33You can play with it. The results are
16:35visible with the caveat that they're
16:37only updated every 15 days, but you can
16:40actually go and run this or not run, see
16:43the results
16:45on our site as a free user.
16:48Okay, so
16:50what's next? There's lots of ways. This
16:53is this is good, interesting an
16:55interesting model I think and there's
16:57lots of ways
16:58where this can be improved. I think the
17:00main one is this.
17:03Can we start the large linear system
17:05with a proportional taint
17:07doing the same thing as a human
17:09investigator will do for a Tornado Cash
17:11or a railgun?
17:13Like can I start with a taint of 0.2 for
17:16that address because I have 0.2 or 20%
17:20confidence
17:22that this address really got out of
17:23Tornado Cash, that deposit that came
17:26from a hack based on sizes or based on
17:29timing or whatever? So can we replicate
17:32what human investigators do and
17:34integrate that information in the linear
17:36system and extrapolate from that? That's
17:39I think the highest promise vector.
17:41So that's all I had, but I think
17:45there may be interesting questions and
17:47we'll let me say a bit more. Thank you
17:49very much.
Source and accuracy details
- Names, organizations, technical terms, code and numbers may be misrecognized.
- Source wording and disfluencies are preserved; no editorial corrections were applied.
- Export contains whole-second cue starts only; cue ends and subsecond timing are unavailable.
- A full caption-track export does not establish that every spoken word is represented.
- Automatic YouTube captions; not human-verified against the audio.
Source export SHA256: 4a335fe6c98691089c43cfc39d18a572a7714ae94d59b559cdfa222e60e6b01c