MEVCheck
The AI way of tracking MEV Transactions! We are using Graph Neural Networks to ensure better MEV tracking for fair reward distribution among Validators & for comprehensive statistical analysis.
Project Description
We use Graph Neural Networks to track transactions happening on-chain.
The project is a comprehensive one & involves 3 separately working parts linked to each other privately & securely using Lit Actions. The first part of the project is the dataset generation. We utilise Powerloom Snapshotters for this. The snapshotters take snap of the epoch from block 15615328 to block 15643187. The snapshotters generate transaction receipts of each transaction that was part of these blocks, the transactions are then sampled into a CSV and labelled as MEV & Non-MEV according to the algo mentioned here.
The dataset is then archived onto filecoin using the py-ipfs-client library provided by Powerloom. We made several improvements to the library such as a documentation to use it, added 7 more tests, optimised the code, added error handling & retrying mechanisms & also optimised exceptions handling & logging. Then Lassie can be used by Clients to retrieve the data from Filecoin for training the GNN model.
The second part of MEVSPy involves training the GNN using the dataset generated from Part 1. The clients using LIt Actions train securely on their own end the model & sends the results to the server who aggregates them & keeps updating the global parameters. The approach of Federated Learning is utilised here for decentralised training of GNN using Lit Actions & Filecoin.
How it's Made
Used- Powerloom Snapshotter Lit Actions Filecoin Storage Galadriel Devnet
The project is a comprehensive one & involves 3 separately working parts linked to each other privately & securely using Lit Actions. The first part of the project is the dataset generation. We utilise Powerloom Snapshotters for this. The snapshotters take snap of the epoch from block 15615328 to block 15643187. The snapshotters generate transaction receipts of each transaction that was part of these blocks, the transactions are then sampled into a CSV and labelled as MEV & Non-MEV according to the algo mentioned here.
The dataset is then archived onto filecoin using the py-ipfs-client library provided by Powerloom. We made several improvements to the library such as a documentation to use it, added 7 more tests, optimised the code, added error handling & retrying mechanisms & also optimised exceptions handling & logging. Then Lassie can be used by Clients to retrieve the data from Filecoin for training the GNN model.
The second part of MEVSPy involves training the GNN using the dataset generated from Part 1. The clients using LIt Actions train securely on their own end the model & sends the results to the server who aggregates them & keeps updating the global parameters. The approach of Federated Learning is utilised here for decentralised training of GNN using Lit Actions & Filecoin.
