SCOUR - AI-Enhanced Safety Feature

AI-Enhanced Safety Feature (SCOUR)

Summary

We’re going to increase trust and safety across the Rarible platform and NFT ecosystem using predictive analytics.

Problem

Wash trading is a process whereby a trader buys and sells a security for the express purpose of feeding misleading information to the market.

Wash trading and ill-intentioned behaviours are the leading problems deterring new users from entering the NFT ecosystem and it has been an issue with Rarible too.

Solution

Scour AI agent is our solution that acts as a watchdog to flag the spoofing transactions between the traders that manipulates both volume and price of the assets in the NFT ecosystem thereby protecting the platform (marketplace’s) market mining.

We are building a detailed knowledge graph of the complete NFT transaction history that includes transfers, wallet addresses, and reward token distributions.

Business Model

We would like to provide this tool as a SaaS Model (Software-as-a-Service).

Minimum Viable Product

Based on the interactions with the Rarible team, we have already built a sample static web application that is capable of detecting wash trades, as a MVP to prove our hypothesis in identifying the wash trades.

Payments

  • Phase 1: Both Rarible and our team will be working in a collaborative fashion for a period of first 2 months.
  • Phase 2: Standalone product flagging the wash traded wallets.

The payment comprises of two components : Component A and Component B

  • Component A is a fixed base price ($8000 in USDT) that covers up the infrastructure and operations cost.
  • Component B is a variable price that amounts to 5% of the wash trade that happens on a weekly basis that the internal Rarible team has not identified internally.

The contract runs for 2 months, the Component B will be waived during this period. Deliverables will be at the end of every week. Fees for Component B will be assessed after 2 months.

Payments will be made to the following Ethereum address: 0x283cC2177BC157708a5Ed19a53159185a18DBE78 (ERC20 - USDT)

We consider ourselves independent contractors of the Rarible DAO and will report taxes accordingly. Legal compliant invoices will be issued accordingly before funds get distributed to our team.

Use of funds

A portion of these funds will be used for operations and infrastructure that includes transaction fees and contract deployment fees as well as the procurement of a server that can continuously handle marketplace support prediction requests, data aggregation and model training.

Team Members

  • Vijay Pravin Maharajan: Founder. Vijay is a TEDx Speaker, TU Munich Graduate, Top 20 Data Scientists to be followed on LinkedIn, followed by over 40k+ connections, and has over 7+ years of technical experience in Blockchain, Mobility, Automotive and Telecom Industries. https://www.linkedin.com/in/vijaypravin

  • Saravanan Jaichandaran: Chief Data Scientist. Saravanan is a Data Analytics professional with proven ability in machine learning, deep learning and data science to drive process efficiency and data integrity.
    https://www.linkedin.com/in/saravanan-jaichandaran

  • Ashok Varadarajan: CTO. Ashok has over 9+ years of experience as a Data Scientist in the Life Sciences Industry. He is specialized in the field of multi-omics data analytics, Machine learning and Artificial Intelligence.
    https://www.linkedin.com/in/ashok-varadharajan-52a994a9

  • Ajay Prashanth: Product Manager. Ajay is a certified DevOps Expert
    and has 5+ years of experience in deployment of applications into AWS, Google Cloud, Azure and with CI/CD and development/testing tools for Android and iOS.
    https://www.linkedin.com/in/ajayprashanth

  • Kyle Webster: Community Manager. Experienced Owner with a demonstrated history of working in the media production industry. Skilled in Sales, Design, Marketing, and Data Analysis.
    https://www.linkedin.com/in/kyle-stargarden

The contact person for this project will be Vijay Pravin Maharajan. Contact information is vijay@bitscrunch.com and his Discord handle is @Vijay | bitsCrunch#9613

Our team has been ideating and developing NFT marketplace platform safety features since September of last year. We have put a lot of thought and energy into this solution space.

Here is the image forgery and provenance detection model that we built for the Untitled NFT hackathon last year:

( https://kylewebster.medium.com/n-safety-forgery-detection-model-untitled-nft-hackathon-submission-11fb9c2b01b1 )

You can test out our image forgery detection model right here:

( live demo https://nsafety.bitscrunch.com/ )

Accountability

We’ve developed a number of relationships with Rarible team members including Alex Salnikov, Eric Arsenault and Ethan van Ballegooyen. We will be working hand in hand for milestones eight through ten above. It’s very important that the finished product adheres to Rarible’s technical specifications. We will also work with community members to ensure updates are distributed through the Rarible communication channels.

Why should Rarible DAO fund this?

We’re very interested to continue deploying AI-enhanced safety features in the NFT marketplace space — like the wash-trade detection & image forgery and provenance detection. While NFTs are more popular than ever, the most impactful way to strengthen the platform is with wash-trade detection and flagging models.

In our analysis of 1/3 of confirmed malicious trader wallets; 2.95 million $RARI tokens have been minted to addresses which were involved with malicious trading activity. This represents ~10% of the $RARI token distributions.

Our goal is to reduce $RARI tokens rewarded to malicious actors. This will drastically increase platform trust by detecting and stopping wash trading in a manner that is fast, reliable and scalable. We also eliminate the manual effort of approximately 8 hours per week of the support team.

Useful Links & Media

An example of one of the complex wash trading patterns identified by our model.

3 Likes

I don’t think wash trading is deterring people as much as waiting 2 month just to get a response on verification. that’s what’s really deterring people. seen it many times. After about a month people loose hope and start talking about other platforms. This is the biggest problem.

4 Likes

@EthanVB is working on clearing the backlog… hopefully this does not stay a problem for long!

4 Likes

Component B sounds like a terrible idea. What’s the check to ensure that’s correct rather than an incentive to increase false positives.

To be frank as token holder, I’m thinking about overall value of the platform and price of RARI. And this seems like an absolute negative to the platform.

Wash trading is a catch-22. Rarible’s value is trading volume and taking a fee of that trading volume. Cutting out trading volume is a negative and then giving those tokens instead to SCOUR only to be dumped is not a net positive either. Also those tokens get redistributed to others, doesn’t guarantee those people aren’t dumping RARI.

Of course someone selling 5 ETH NFT for minimal effort and adds nothing to platform should be removed. But your example, no one is thinking “OMG can’t believe that is happening. I don’t trust Rarible because Redlioneye sold to x”. If it’s that person getting RARI or some other artist getting more, for other token holders is probably doesn’t matter.

Ultimately what matters is high quality artists using the platform making high quality pieces, buyers buying and people holding and not dumping tokens.

For me looking at this, I don’t see this is going to increase the value of RARI.

1 Like

Also its absolutely incomplete looking just at the blockchain.

Any model should incorporate digital fingerprinting from the web browser + IP address, etc.

1 Like

We can rework the component B Bryan. Point noted. And also, it’s only the component A that has been given the nod so far.

I really like @bryandough’s idea of using IP… in a certain sense, this might be better for detection.

I also agree that component B needs to be revised to ensure the net is positive for the ecosystem.

1 Like

Most credit card fraud takes IP into account plus useragent, etc. Pretty much all information from whoer.net and anything else to establish a fingerprint of the user.

I believe that can be more concrete than blockchain analysis. Ofcourse everything should be in combination.

Again I’m strongly of the opinion, Rarible should be careful about shooting it self in the foot because you don’t necessary want to eliminate trading volume and fees collected. Its about eliminating those that add no value to the platform. That’s a question of what that is. (Low quality artists, egregious and obvious offenders, token dumpers, etc.).

If the OP model is correct, than that’s 10% less in fees collected, which based on Dune is almost a $1mm lost in fees. To be frank I’ll rather that fee be used to buyback and burn RARI

That’s probably a deeper analysis that needs to be done and talked about. Its like coinbase having that washtrading fine with the SEC before their IPO. To the general public it probably wasn’t noticeable but the fees and volume probably helped establish credibility early on. But I agree first and foremost there should be trust.

Also you give a great example of finding complex wash trading, but the most egregious offenders seem to still get through, see below.

Both accounts received liked 200+ RARI in the past few weeks.


Someone posted this on the discord, he also seems like a legitimate user that got flagged. How are we evaluating the effectiveness of this? @eric

1 Like

The example shown here is one such example of how good and accurate our model is! Of course, scammers can always find their way out. But the more and more we can find, the model can train by itself @bryan

1 Like

Ofcourse, but what positive feedback does the model get to say yes this was actually a wash trader vs false positives.

While its great to identify these complex situations, the actual biggest “drain” on the ecosystem are those washtraders I mentioned above and those seem to be missed

I would agree with @bryandough tbh…
I’m pretty sure I could completely dupe your AI model, without much work. Also, there is no certainty that the graph you are showing is actually wash trading.

This entire space feels like a game of chasing ghosts. I would rather solve this by changing the token incentive model I think, but I still think it is worth the experiment possibly.

One more question: are you also using “money flow” as an input for your model?

For example, in the graphic above… if b8d08 received funding from Rare-Designer for their purchase… I think that’s a critical part of all this… where is the money coming from?

2 Likes

We haven’t integrated our product yet with Rarible. So by the existing manual method, they missed those wash traders you mentioned above.

1 Like

Yes we’ve spent a significant amount of time discussing token flow and how that can be incorporated to affect the overall confidence score. Both different models represent a certain confidence adjustment. To build the confidence score actually requires an ensemble of different approaches. The two primary approaches are to look for circular movements in both “money flows” and NFTs.

1 Like

The most common indicator we see is money flowing into an exchange and then back out of the exchange or into Tornado cash and back out. It’s less about where the funds go and come from and more about similarity in the amounts.

It’s also possible that this model can be extrapolated after it’s incorporation for other use cases. So as an experiment I believe there is significant value.

2 Likes

What about new accounts? The easy form I see is of new accounts (non-verified) selling to other new accounts and repeat. They sell for high prices that completely unrealistic

Those are the accounts that are the biggest drain because that actually add no value to the platform

2 Likes