Bitcoin, Tokenomics, Fraud Rings and Graph AI !

How’s that for a SEO keyword clickbait ? Believe it or not, I can make a coherent case tying all four together ! A little oblique, but the connection is definitely there.

BTW, this was the topic of my talk at the TigerGraph Graph+AI summit earlier this week. The 30-minute video is published here.

In this blog I will give a quick highlight for the attention challenged and the time compressed souls among us (which is probably all of us!)

Problem Statement

• The Distributed Ledger Technologies (DLT), in general and the blockchain, specifically offers a variety of interesting opportunities from inclusive financing for underserved communities to being the platform for a scalable web3.
• Security from all forms of fraud is vital for the future of any DLT platform — crypto currency and Web 3.0. Securing the blockchain involves understanding a special class of non-reciprocal, sequenced, temporal networks, that break many of the assumptions on traditional network algorithms.
• We will start with a few fantastic algorithms (and how to find them!) and conclude with the hard and (yet) unsolved challenges and how TigerGraph addresses them — from graph algorithms to Graph Neural Networks.

YouTube Video (Here)

Approach

• Tackle one of the complex fraud pattern in a scalable, extensible way
• Specifically, this talk addresses the detection of Fraud Rings (e.g., Money Laundering), Camouflaged Fraud (e.g., Fake Product Reviews) and other similar activities

First, what are these Non-Reciprocal, Non-Transitive, Sequenced, Temporal Networks, anyway ?

Payment networks have slightly different characteristics from traditional networks like computer networks and social networks

1. Payments networks are not reciprocal or symmetric — If A is a friend of B, usually B is also a friend of A. But if you are a customer to Amazon doesn’t mean they are your customers !
2. Payment networks are not transitive — If B is a friend of A and C is a friend of B, usually A & C might be friends. But if you pay Amazon for some goods and Amazon pays Joe as a refund of something, it doesn’t mean we can infer any kind of relationship between you and Joe
3. Payment networks are sequenced and temporal— There is a sequence to payments/edges, especially if we want to find fraud rings et al. If A pays B and B pays after that payment this could be interesting. But the B-C payment occurred long ago, “these are not the payments you are looking for” !

Examples of payments networks which exhibit the above properties include Blockchain Forensics, Fake Reviews and Fake News.

Implications

This dichotomy results in interesting challenges.The major one being, they break some of the assumptions in traditional graph algorithms and as such we need to tweak the traditional graph algorithms.

• PageRank doesn’t make sense per se. If we apply page rank to a payment network, any payment from amazon will make those nodes more important ! Just because amzn paid someone doesn’t make them any more important
• Community detection or k-clique are not that relevant; connected component is not that informative either
• In a payment network, edges rule. Algorithmically, nodes (while important) have a lesser role

Implementation : Graph based blockchain forensics to detect fraud rings

Now, let us get into the details of implementation. The context is Money Laundering, specifically the Fan-In/Fan-Out Fraud Rings.

• Money Laundering goes through 3 stages viz. Placement, Layering and Integration
• Layering is the most complex — the money is piped through a complex web of strategic transactions, structured layered payments that diffuse into a network — mixers, tumblers, long chain of crypto transactions and cross currency/cross ledger, obscuring the money trail. AN example below:

In Graph #1, a person #1 diffuses the money into 5 entities who in turn give to another 5 entities and so on until all the money comes back to #1. In graph #3, the bad actor #201 gives to a tumbler network (#200) that handles the diffusion and lands the money back to #201. These can be of arbitrary depth, can span entities like crypto exchanges as well as cross-ledger, cross coins like BitCoin to Ethereum to other coins and back. Our goal is to unroll this network, using graph primitives

I implemented this using TigerGraph. The questions to the graph were:

And it worked out very well ! As you can see the output below:

Lessons Learned

Things to try