Graph For A Better Token Economy

Make Crypto & Virtual Assets safer than paper currency - using Graph-based AI

Comment

Virtual Assets (Crypto coins, NFTs, ...) enable under represented population

... But the platforms are not as secure as they need to be. There should be a better way to enable the masses and usher the new improved kinder-gentler Web 3.0 & crypto currency platform

Problem Statement

  • Immutable, Distributed, Unanimous & Secure DLT & Blockchain stack for Crypto currency & Web 3.0. While the other properties are well implemented, security is still a challenge
  • Specifically, this project addresses the detection of Fraud Rings (e.g., Money Laundering) & Camouflaged Fraud (e.g., Fake Product Reviews)

Hypothesis

  • Graph representation & algorithms are well suited to extract fraud rings of arbitrary complexity

Approach

  • Tackle one of the complex fraud pattern in a scalable, extensible way

Challenges

  • Payment network is very different from social or other networks
    • Temporal, monotonic time
    • Not symmetric
    • Other attributes (like amount) matters
    • Many of the concepts like connected components and page rank do not mean much
  • Money Laundering Layering is not just one cycle
    • But multiple ordered cycles with a fraud actor at the helm
  • Detecting fraud ring is literally finding the proverbial needle in a haystack
    • Class imbalance, Long Fraud Chains

Approach a.k.a How we built it

  1. Layered approach with well defined pipeline stages
    • Overlay fraud rings progressively
    • Narrow down and organize the vertices as we progress
    • Always keep time sequencing in the processing
  2. Customize relevant gsql graph algorithms
    • e.g., Rocha-Thatte Cycle Detection, but it is unordered
    • Also need to combine cycles and find the fraud actor at the helm
  3. Add runtime attributes to vertices
    • That will help the processing downstream the pipeline 4, Start with a simple schema and add more elements as required
  4. Stay in TigerGraph as much as possible (more later)

Accomplishments that we're proud of

  • Constructing multi-cycle fraud ring using graph algorithms
  • The possibilities the platform provides - and we have a lot more ideas !

What we learned

  1. Graph representation is appropriate for DLT/Blockchain Fraud Detection
    • TigerGraph is a feature rich, flexible & scalable Graph Database well suited for this problem
    • But it requires disciplined thinking, at times different than what we are used to !
  2. Spend time thinking about & understanding the problem
  3. Make simplified assumptions and relax them as you progress (Layered approach)
  4. Think Graphs & more specifically Parallel Graphs – It will take a little time to get used to that concept
    • Read, write & learn the patterns from the GSQL Graph Algorithms and the GSQL code
  5. Draw graph diagrams to visualize the problem - Draw the happy path 1st & then edge cases
  6. Create datasets depicting multiple scenarios – 1st use a small dataset to test the algorithms
  7. Do as much in TigerGraph as possible, staying true to the platform
    • It is tempting to process a list outside (say in python), after a quick GSQL algorithm; don’t stop there, persist (using GSQL) until you have exhausted all graph ideas

What's next for Graph For A Better Token Economy

This is only the beginning !

  1. Scale ! Load Bitcoin/Ethereum blockchains and apply the algorithms
    • Cross-Ledger tracking of fraud rings
  2. Fine grained temporals
    • There could be many such rings by the same actors, so need to separate the rings by time
    • solution : time tree “If you need to filter use vertices” – TigerGraph pragma
    • Opportunity for Payment Networks in TigerGraph Graph Algorithms
  3. More expressive Graph schema with derived runtime attributes, especially to track cross-ledger behaviors
  4. Explore Graph Motif extraction, Weighted Graphs
  5. Graph Neural Networks leveraging the extended dynamic attributes
  6. Entity Resolution
    • Need to understand heavy spans & differentiate between Exchanges, Tumblers, Mixers - Add Vertex type based customized logic
    • Probably via highest measure of eigenvalue centrality

Thanks for the opportunity, Enjoyed the journey a lot !!

Built With

Share this project:

Updates