In this 2-part post, we lay out Pastel Network’s “dupe detection” system — arguably the single most innovative part of our software. Before we jump into the catalyst for such a feature and dive into its inner workings, we thought a quick description of it would be appropriate.
Pastel’s Duplicate Image Detection leverages 7 deep-learning models using Tensorflow with Keras Applications and transforms each NFT into a vector of 8,000 numbers (the NFT ‘fingerprint’). We then assess the correlation between a given fingerprint and the database of all existing fingerprint vectors on Pastel, competitive NFT platforms like OpenSea, and even open databases of images on the entire internet (e.g. Google). We then assign the NFT a ranking in terms of rareness (e.g. ‘Rare to Pastel’ or ‘Highly Rare’).
Part 1 — The Need
The need for a “dupe detector” is fairly obvious, but no other company seems to view it as essential to their platform. For example, when you buy an NFT on any other layer-2 marketplace (Nifty Gateway, OpenSea, Rarible) you are just buying any particular file uploaded by the creator itself — with no concern for the authenticity or originality of the data. Most marketplaces do not have their own native, underlying blockchain and rely on existing infrastructure like Ethereum instead. As a result, they cannot incorporate the bespoke technical architecture and software required to determine whether a digital asset is truly a duplicate or not.
This seems counterintuitive as the whole idea of natively digital artwork being “rare” hinges on the assurance that this rareness continues and cannot be subverted without undermining the artwork’s entire value proposition.
Of course, one approach would be to say: “Who cares if someone else who is not the artist registers a similar artwork? It’s up to the collector to determine if a given artwork is genuine.”
Well, that won’t work for us. We believe that such a stance is counter-productive, since it deflects the problem onto users and exposes both creators and collectors to a variety of scams. Thus, in theory, “dupe detection” is a no-brainer; however, in practice, duplicate image detection is, perhaps, the most technically challenging feature required in any decentralized art registry.
The idea is simple: if a particular, original image is registered on the network, no other user can register a duplicate image and have it be scored as rare. However, using a straightforward file hash will not be sufficient. A file hash is brittle; if you take an existing registered image and change only the upper left pixel, the entire hash will change.
Instead, we want a robust image fingerprint — one that does not change materially when the same file is uploaded or even if an image is modified and re-uploaded with superficial changes.
To do this, we need a duplicate image detection system that can react similarly to the way a human observer could in determining if two images are “related.” For example, if an average person could reliably determine that a given image is similar to another, then we want our automated system to reach the same conclusion. The greatest challenge, though, are those artworks on the boundary line — similar yet different enough to not be considered as clear “duplicates” according to our chosen criteria.
For these problems, Pastel has developed a novel and innovative duplicate detection system, which leverages advances in machine learning technology as well as the creative application of classical statistical techniques.
Check out Part 2 to see how our founders and team of world-class developers built it.
About Pastel
Pastel is the world’s first fully dedicated, decentralized NFT platform allowing users to register, store, and trade ‘provably rare’ assets. The platform gives creators a way to securely connect with fans and sell unique, limited edition digital assets –without the high fees or storage constraints of other crypto projects. Pastel also allows for the development of third-party applications to sit on top of the Network, enabling developers to enjoy the scalable registration features, storage processes, and security of the broader network. By leveraging a purpose-built native blockchain, Pastel deploys a number of key features such as a deep-learning based duplicate detection system to ensure proven authenticity.
The network is a fork from Z-cash, running the Proof-of-Work (PoW) algorithm Equihash and also supports Supernodes, which provide computational resources to the network to support asset registration, distribution, and permanent storage. Pastel is managed by world-class developers, cryptographers, and technologists, supported alongside an experienced and extensive network of marketers, influencers, and third-party agencies. Pastel is backed by key stakeholders including Innovating Capital, a prominent venture fund.
Website | Telegram | Twitter | Instagram |Github | Wiki | Medium
