SUI NFT Scam Detector

A lightweight detector scam classifier for NFTs on Sui with 95% accuracy. Can run anywhere that webassembly runs: on a server, in a lambda function, and even running entirely in your browser.

Also included is the model training code and data, so you can train and bring your own model if the default model is not performing well.

Feature extraction is done with a combination of on-chain data and OCR using the tesseract.js library. Classification is done with naive bayes and hand-picked labels.

We labeled NFT into 2 categories:

• Scam: NFT that our classifier detected as scammed
• Ham: NFT that our classifier detected as trusted

Installation

• Using npm

npm install nft-scam-detector --save

• Using yarn

yarn add nft-scam-detector

Getting started

import { extractAndClassify } from "nft-scam-detector";

const imageUrl = "YOUR_NFT_IMAGE_URL";
const result = extractAndClassify(imageUrl);

result.then(console.log);
// result here 
// 
// {
//     "classification": "scam",
//     "scam_likelihood": 0.0001565123,
//     "ham_likelihood": 0.0000004214,
// }
// 

Training

1. Input your list of ham and scam in ham_ids.json and scam_ids.json

2. Run training model.json

yarn run train

Testing

Live testing at: https://sui-nft-spam-fe.getnimbus.io/

You can test the accuracy of a model using the code in the /test folder. Make sure that your training set and test set do not overlap. It should spit out a confusion matrix as well as all of the mistakes made with accuracy more than 95%

	scam (predicted)	ham (predicted)
scam (actual)	20	0
ham (actual)	2	18

• avg time: 1930ms
• median time: 1820ms

License

MIT License