How Machine Learning Algorithms are used in Anti Money Laundering (AML)
Knowledge gained from dissecting the article “Detecting Money Laundering” by Arshak Navruzyan
Starting with the Basics
There are two forms of AML monitoring used by financial institutions:
[1] knowledge-based systems and
[2] link analysis
“Structuring” and/or “Smurfing” are small-ish deposits designed to avoid currency reporting requirements, kind of like the scam in Office Space. If you take small unnoticeable bits from large accounts and systematically iterate this extraction a bunch of times, you end up with a large sum of cash.
“FinCEN.gov” — [Financial Crimes Enforcement Network] is one of the US department of Treasury’s top agencies in detecting money laundering.
The mission of the Financial Crimes Enforcement Network is to safeguard the financial system from illicit use and combat money laundering and promote national security through the collection, analysis, and dissemination of financial intelligence and strategic use of financial authorities.
— FinCEN.gov
Currency Transaction Reports (CTR) — are used by FinCEN for processing and analysis of large transactions. These reports are required by FinCEN for every “deposit, withdrawal, exchange of currency, or other payment or transfer” involving a financial institution where the transaction totals over 10,000 USD.
Backing up a bit to [1], knowledge-based AML systems include statistical analysis, machine learning and data visualization. A knowledge-based system, or a KBS, typically refers to “a computer program that reasons and uses a knowledge base to solve complex problems”.
Network Modeling
This is the first technique Navruzyan introduces for techniques applied as a machine learning approach to AML. Quickly read through this infographic and I’ll follow up below with a bit more information on what we’re doing.
Above are network module plots. They are a form of data visualization that can be used to connect pieces of data in order to uncover knowledge or actionability from raw information.
For the last listed relationship, “pagerank”, this reminds me of a sort of cosigner system. Or even less than that, if you have a child account at a bank that struggles to pay its debts while a parent account (that is pseudo-connected) is healthy in its payments, you’ll get more lenience on your transactions because you are “well connected”, in a sense.
Aside: not sure yet what he means about “subgraph relations”. This is related to “community detection”, so I would assume this is something like: different nodes on the graph may not appear related in the present form of the graph, but upon further inspection/more iterations of the graph, maybe you would see that there is an underlying “community” between these nodes; they have a relationship, though this might not appear clearly at first glance.
Clustering
Second tactic, used with interpretation to “identify natural groupings within the data”.
So this would help with AML because you can pick up on clusters which show you what natural data looks like. Then, when something doesn’t fit or looks off, it may be “indicitive of AML activit[y]”.
Impress your friends! Mention something about how in clustering, spectral clustering is totally state of the art and the current name of the game (or the cat’s meow in machine learning). However, some bleeding edge research seems to think some little processes that use deep learning autoencoders are the new way to go. I’m not exactly sure what DLAs are either, but you could just mention them and then step back and let the engineers talk and nod in concentrated approval. That usually works.
For the most part, on a more serious note, deep learning autoencoders (which are a form of “unsupervised learning”) seem to mean pieces of algorithm code that pick up on patterns in a set of data and can then reproduce those patterns on their own. Like if a program watches a video and summarizes it in pixel movements, and then has to type out that summary, but in logical text form. So they are in charge of producing a summary form that humans could read and understand. Maybe there can be a deep learning auto encoder that recognizes dogs and people within the pixels of a video, so that the summary output would say something about the amount of people or dogs in the video. In the reality though, the program’s input was just a bunch of color-changing pixels. Or was it?
Time Series on Graphs
Third technique. Monitor things like “centrality”, “connected components”, “etc. for active nodes”, over time. Look for these things at a set of time signatures, then analyze the data for anomalies. Maybe the anomalies point out iffy money laundering behavior.
As an example, Navruzyan mentions Twitter’s Anomaly Detection, an open source program that was used for detecting time signature anomalies, like why are so many people using Twitter on Christmas, rather than paying attention to their families? Or why do so many people favorite pictures on Halloween? Or of course to detect spammers on days where a bunch of ‘likes’ or ‘favorites’ occur, but there’s no event that logically would have caused the anomaly. Maybe worth checking out. Somebody could be Tweet laundering.
Seriously though, this sort of analysis can have huge implications if you’re doing research into your digital marketing strategy against the objectives or KPIs you’ve defined for yourself. The idea: let’s make sure good anomalies keep occurring! If you can iterate an occurrence, then it’s not really an anomaly anymore.
Learning More
The fourth technique Navruzyan mentions is to learn more. I’m not going to cover this anymore, because learning more is what you’re doing now, so if I discuss it, this whole thing will become metta and make my brain hurt. To dive deeper, check out the article below, or crawl around on Google.com or ResearchGate.com, searching for terms like ‘knowledge-based systems machine learning’, ‘what are deep learning network modules’, etc., or crawling around the documents that are publicly available on FinCEN.gov.
Here’s the full article: Detecting Money Laundering
Don’t forget to tap the heart symbol if you found any of this useful, so others get a chance to see this article too!
As always, thank you for reading. I’ll also plug a couple of my other articles below, in case you’re interested in exploring my work a bit more:
