Systems and methods for authenticating the scarcity of digital assets

Abstract

A system and method for authenticating the fingerprinting of assets and the scarcity of digital assets. This solution can be implemented on a peer-to-peer decentralized platform for securely registering, trading, and collecting non-fungible tokens associated with digital visual artworks and recorded on a blockchain. The system comprises a fingerprinting engine (FE) and a relative scarcity engine (RRE). The FE processes digital assets (e.g., NFT images) using a trained neural network model that generates a digital fingerprint vector representation of the image. The RRE compares the digital fingerprint vector to a dataset of registered digital fingerprint vectors using multiple correlation measures. The correlation results are analyzed and combined to generate a relative scarcity score representing the scarcity of the NFT. In addition, the RRE can assess the scarcity of an image compared to images on the internet.

Description

Technical Field

[0001] The present disclosure relates to a system and method for authenticating digital works recorded in a distributed ledger or blockchain. In one particular configuration, the present disclosure describes a system and method for fingerprinting digital visual assets and assessing the scarcity of digital assets associated with non-fungible tokens stored on a blockchain.

Background Art

[0002] In the case of conventional artworks such as paintings, drawings, sculptures, and limited edition prints, copies are inferior to the originals. As long as it is possible to distinguish the genuine work from the imitation, no reasonable buyer would pay the same amount for a copy as for the genuine (original) one. However, what should be thought about digital artworks from the beginning, that is, artworks intended by the producer to be presented in a digital format that is literally a specific sequence of 0 and 1 (for example, in contrast to a digital photograph of a physical painting)? Aren't all the same sequences of numbers the same, just as all the atoms of gold are the same? Put another way, the physical embodiment of a digital artwork from the beginning seems to be basically secondary to its essence. Would anyone care whether it is stored on a DVD or on a USB flash memory drive? Wouldn't future art historians be more concerned with the data itself, which is nothing more than information without a physical form?

[0003] An analogy is useful to understand the different dynamics of digital artworks. Imagine the following two scenarios. In the first scenario, someone owns a brand-new, first edition of the original Spider-Man comic book. In the second scenario, someone owns a high-resolution PDF file of the same comic book. Both people can read and enjoy the artwork contained in their respective objects, but one is worth thousands of dollars on the market, while the other would be considered worthless by most people. The reason is obvious: the physical edition is rare. When Spider-Man was first conceived, no one knew that it would later gain iconic cultural status. Only a very small number were produced, and of those produced, the vast majority were discarded, lost, or destroyed by the children and adults who bought them.

[0004] Therefore, the de facto exclusivity associated with the ownership of a physical object (i.e., it is in one's own home and therefore cannot simultaneously be in someone else's home) is the most important element in the value of traditional works of art. This attribute is necessarily present in physical works of art due to the inherent nature of space and matter, whereas it is completely absent in the digital realm.

[0005] The use of blockchain or distributed ledger technology as a registry for digital artworks is one method that provides a degree of exclusivity for digital artworks. However, no existing project has solved the specific challenges that arise when developing a digital asset registry system that can function truly decentralized, reliably, and securely.

[0006] The emergence of non-fungible tokens (NFTs) has increased both awareness and demand for scarce digital assets. The term NFT is commonly used to describe blockchain-based crypto tokens created for digital assets and stored on the blockchain. Although NFTs are crypto tokens, unlike cryptocurrencies such as Bitcoin and many network tokens or utility tokens, which are interchangeable (i.e., fungible), each NFT is verifiable unique (i.e., non-fungible). Typically, NFTs are stored on the blockchain, and the metadata contained in the token (e.g., URL) refers to the corresponding digital asset stored elsewhere. Therefore, NFTs can be created for a wide range of digital assets, including digital artwork, images, videos, audio, and more.

[0007] While NFTs can provide official authentication of the authenticity or ownership of the token itself, the legal rights conveyed by NFTs can be uncertain. Furthermore, since NFTs are typically linked to stored digital files, NFTs alone do not restrict the sharing or copying of the associated digital files, nor do they prevent the creation of NFTs with related digital works that are identical or near-identical copies.

[0008] NFTs do indeed offer certain advantages to both creators and speculators of digital assets, with interoperability across ecosystems improving tradability and liquidity, while token standards like ERC721 guarantee verifiable scarcity. However, NFT platforms face some challenges. For one, no existing NFT platform offers a system that can provide the network advancements necessary to detect quasi-reproductions of digital artwork.

[0009] Effective detection of quasi-replica images is an unresolved challenge in computer vision, and it assumes that visual data is extremely high-dimensional. Even a relatively small 100kb JPEG file can contain well over 500,000 pixels, each with red, green, and blue components. Furthermore, someone might edit the JPEG file in Photoshop so that its visual appearance is easily recognizable to a human observer as a simple derivative of the original image, but even then, every pixel is likely altered in a complex way, leaving little of the original structure unchanged at the pixel level.

[0010] These and other considerations are presented herein. [Overview of the project] [Means for solving the problem]

[0011] One aspect of this disclosure discloses a method for authenticating the scarcity of a digital asset. The method includes the step of generating a first digital fingerprint of a digital asset by running one or more deep learning models on a first visual appearance of the digital asset. The generated first digital fingerprint is a first vector representing the first visual appearance. In addition, one or more deep learning models are trained to process the visual appearance of the digital asset and generate corresponding first vectors suitable for determining the similarity of the visual appearances. The method also includes the step of evaluating the similarity of the first digital fingerprint to a registry of registered digital fingerprints by calculating the corresponding dot product of the first digital fingerprint and the registered digital fingerprints. Similar to the first digital fingerprint, the registry of registered digital fingerprints is obtained by running one or more deep learning models on the corresponding visual appearances of registered digital assets. Furthermore, the method includes the step of determining a corresponding first measure of statistical dependence for registered digital fingerprints by normalizing the calculated dot product to a range having a first end corresponding to no similarity to the first visual appearance and a second end corresponding to identical similarity to the first visual appearance. Furthermore, the method includes the step of selecting a fixed number N registered digital fingerprints whose corresponding N normalized dot products are closest to the second end of the normalized range. In addition, the method includes the step of determining a corresponding N second measure of statistical dependence by applying a population measure test of deviation from independence to N combinations of the first digital fingerprint and the selected N digital fingerprints. Finally, the method includes the step of authenticating the rarity of the digital asset using the determined first and second measures of statistical dependence.

[0012] In a further aspect of this disclosure, an automated system for authenticating the scarcity of digital assets is disclosed. The system comprises a processing circuit and a non-temporary storage medium storing a registry of registered digital fingerprints and machine learning models. Instructions are also stored in the storage medium, which, when executed by the processing circuit, configure the processing circuit to generate a first digital fingerprint of a digital asset by executing one or more deep learning models of the stored machine learning models on a first visual appearance of the digital asset. In detail, the generated first digital fingerprint is a first vector representing a first visual appearance, and the deep learning models are trained to process the visual appearance of the digital asset and generate corresponding first vectors suitable for determining the similarity of the visual appearances. The instructions further configure the processing circuit to evaluate the similarity of the first digital fingerprint to registered digital fingerprints in a registry by calculating the corresponding dot product of the first digital fingerprint and the registered digital fingerprints. The registry of registered digital fingerprints is obtained by executing the deep learning models on the corresponding visual appearances of the registered digital assets.

[0013] The instruction further configures a processing circuit to determine a corresponding first measure of statistical dependence for registered digital fingerprints by normalizing the calculated dot product into a range having a first end corresponding to no similarity to the first visual appearance and a second end corresponding to identical similarity to the first visual appearance. In addition, the instruction configures a processing circuit to select a fixed number N registered digital fingerprints whose corresponding N normalized dot products are closest to the second end of the normalized range, and to determine a corresponding N second measure of statistical dependence by applying a population measurement test of deviation from independence to the N combinations of the first digital fingerprint and the selected N digital fingerprints. Furthermore, the instruction configures a processing circuit to authenticate the rarity of the digital asset using the determined first and second measures of statistical dependence.

[0014] In a further embodiment, a non-temporary computer-readable medium (CRM) storing computer instructions, wherein the computer instructions, when executed by a processing circuit, cause the processing circuit to execute an automated process for authenticating the rarity of a digital asset. The process includes the step of generating a first digital fingerprint of a digital asset by running one or more deep learning models on a first visual appearance of the digital asset. The generated first digital fingerprint is a first vector representing the first visual appearance. In addition, one or more deep learning models are trained to process the visual appearance of the digital asset and generate corresponding first vectors suitable for determining the similarity of the visual appearances. The process also includes the step of evaluating the similarity of the first digital fingerprint to a registry of registered digital fingerprints by calculating the corresponding dot product of the first digital fingerprint and the registered digital fingerprints. Similar to the first digital fingerprint, the registry of registered digital fingerprints is obtained by running one or more deep learning models on the corresponding visual appearances of registered digital assets. Furthermore, the process includes the step of determining a corresponding first measure of statistical dependence for registered digital fingerprints by normalizing the calculated dot product to a range having a first end corresponding to no similarity to the first visual appearance and a second end corresponding to identical similarity to the first visual appearance. Furthermore, the process includes the step of selecting a fixed number N registered digital fingerprints whose corresponding N normalized dot products are closest to the second end of the normalized range. In addition, the process includes the step of determining a corresponding N second measure of statistical dependence by applying a population measurement test for deviation from independence to the N combinations of the first digital fingerprint and the selected N digital fingerprints. Finally, the process includes the step of authenticating the rarity of the digital asset using the determined first and second measures of statistical dependence.

[0015] These and other embodiments, features, and advantages can be understood from the description of the specific embodiments of the present disclosure, as well as from the accompanying drawings and claims. [Brief explanation of the drawing]

[0016] The above-mentioned and other features of the configuration of this disclosure will become more readily apparent from the following detailed description and drawings of one or more exemplary embodiments contained herein. [Figure 1A] This figure shows an example of a computer network environment equipped with a system for fingerprinting digital assets and authenticating the scarcity of digital assets, based on the principles of this disclosure. [Figure 1B] This is a conceptual block diagram showing an exemplary configuration of the processor used in the system of Figure 1A, based on the principles of this disclosure. [Figure 2] This is a process flowchart illustrating an exemplary routine for generating a digital fingerprint vector based on the principles of this disclosure. [Figure 3] This is a process flowchart illustrating an exemplary routine for evaluating the relative scarcity of NFTs based on the principles of this disclosure. [Figure 4] This figure shows three sets of images, including a base image and multiple transformations of the base image, based on the principles of this disclosure. [Figure 5] This figure shows an image and a converted version of that image, based on the principles of this disclosure. [Figure 6] This figure shows an image and a converted version of that image, based on the principles of this disclosure. [Figure 7] This figure shows an image and a converted version of that image, based on the principles of this disclosure. [Figure 8A] This is a diagram illustrating an example image. [Figure 8B] This figure shows a set of images that are visually similar to the image in Figure 8A, based on the principles of this disclosure. [Figure 9] This figure includes the text of an exemplary output file generated by the principles of this disclosure. [Figure 10A]~ [Figure 10C] FIG. 3 is a diagram showing an original art image and a converted art image for training machine learning techniques to authenticate the scarcity of digital assets according to one embodiment. [Figure 11] FIG. 4 is a flowchart of an exemplary automated method for authenticating the scarcity of digital assets according to one embodiment.

DETAILED DESCRIPTION OF THE INVENTION

[0017] It should be noted that the drawings are for illustrative purposes and are not necessarily to scale, and that the same or similar features have the same or similar reference numerals throughout the figures.

[0018] The present disclosure and details of various features and advantages will be more fully described by referring to the non-limiting embodiments and examples illustrated in the accompanying drawings and described in detail in the following description. It should be noted that the features shown in the drawings are not necessarily to scale, and that, as will be appreciated by those skilled in the art, the features of one embodiment may be used in conjunction with those of other embodiments even if not explicitly stated. Descriptions of well-known components and processing techniques may be omitted for the sake of brevity. Those examples are intended only to facilitate an understanding of the manner in which the present disclosure may be implemented and to enable those skilled in the art to implement embodiments of the present disclosure. Therefore, the examples and embodiments should not be construed as limiting the scope of the present disclosure. Further, it should be noted that like reference numerals represent like parts throughout several views of the drawings.

[0019] As an overview and introduction, this application describes a system and method for fingerprinting and scarcity assessment of digital assets. In non-restrictive embodiments, the system and method are described herein in relation to certain types of digital assets, namely digital visual works (e.g., digital images, works of art, etc.) having a corresponding visual appearance associated with NFTs. However, it should be understood that the principles of this disclosure are not limited to those exemplary types of digital assets. In addition, it should be understood that the term NFT as used herein is intended to refer to a digital asset associated with a non-fungible token stored on a blockchain. Furthermore, it should be understood that the terms NFT, image, digital asset, and visual appearance are used interchangeably herein, and their actual meanings are clear from the context.

[0020] In one exemplary practical application, the system and method for fingerprinting and scarcity assessment of digital assets can be implemented in digital art registry and marketplace platforms, such as Pastel Network, a peer-to-peer decentralized platform for securely registering, trading, and collecting NFTs. As can be understood, it may be preferable for a digital art registry or marketplace to allow only works of sufficient originality to be registered on the network, or at least to assign scarcity scores that users can consider when making purchase decisions. That is, it may be preferable to determine whether an image is a quasi-copy of another image previously registered on the network, for example, to prevent the registration of quasi-copyings or to quantify the scarcity / rarity of a digital asset in other ways. Therefore, systems and methods for fingerprinting and rarity assessment of digital assets can be configured to implement secure cryptographic digital signature schemes and robust quasi-replica image detection techniques (for example, detecting similarity despite various modifications that may be made to the original asset, without generating excessive false negatives or false positives, i.e., having a wide area under the accuracy-recall curve), thereby providing digital asset collectors with a high degree of confidence when determining the rarity, authenticity, and origin of specific works of art registered in the system.

[0021] If the sole concern is to detect an exact bitwise copy of the original image file, the system can simply use file hashes and determine that files with different hashes are unique. However, file hashes are unstable because changing just one pixel of an existing registered image changes the entire hash. Therefore, the System and Method for Fingerprinting and Rarity Assessment of Digital Assets is configured to generate a robust image fingerprint that remains stable even when superficial changes occur. In other words, the System and Method for Fingerprinting and Rarity Assessment of Digital Assets is configured to identify or characterize an image (and its corresponding visual appearance) and generate a digital fingerprint that is robust against various transformations of the original image, such transformations including, but not limited to, cropping, scaling, or rotating the image; adjusting the color, contrast, brightness, or curves of the image; adding random noise or dots to the image; and applying any type of image filter, such as those included in the Adobe Photoshop software package (e.g., blur / sharpen, edge detection, invert image, or nonlinear image distortion filters such as Spherize or Twist).

[0022] Detecting quasi-replica images remains an unresolved challenge in computer vision, but previous solutions have been inefficient given the extremely high dimensionality of visual data. Even a relatively small 100kb JPEG file can contain well over 500,000 pixels, each with red, green, and blue components. In addition, JPEG files can be easily edited with photo editing software in a way that makes it easily recognizable to a human observer that the visual appearance is a simple derivative of the original image, but even then, every pixel is altered in a very complex way, perhaps leaving little of the original structure unchanged at the individual pixel level, thereby making it difficult for existing quasi-replica image detection techniques to identify the edited image as a quasi-replica of the original. Furthermore, as described herein, systems and methods for fingerprinting and rarity assessment of digital assets can be configured to react in the same way that a human observer can when determining whether two images are related; that is, if the average person can reliably determine that the visual appearance of a given image is an over-imitation of an existing registered image, then an automated system can reliably reach the same conclusion. Preferably, the system rejects a high percentage of works that are true reproductions, while allowing the majority of works that are truly original. The greatest challenge lies with works of art on the borderline, namely works that resemble existing works of art but are different enough that they are clearly not reproductions according to selected criteria.

[0023] Therefore, to address these and other challenges and objectives, the Systems and Methods for Fingerprinting and Scarcity Assessment of Digital Assets introduce an innovative fingerprinting and quasi-replica detection framework, which leverages advances in machine learning techniques as well as unique applications of classical statistical techniques, as further described herein.

[0024] In one embodiment, a system for fingerprinting and rarity assessment of digital assets includes a fingerprinting engine (FE) component configured to generate a digital fingerprint of a digital asset, and a relative rarity engine (RRE) component for assessing the relative rarity of each digital fingerprint in a dataset. Generating an NFT fingerprint involves generating a compressed representation of the NFT in a way that dramatically reduces the required dimensionality while maintaining a high level of structural content of the input image data. The compressed representation becomes a digital fingerprint, which is a list of numbers relative to the original pixel data and is robust to various transformations. The relative rarity engine for assessing rarity is configured to compare the digital fingerprint with the digital fingerprints of an underlying dataset (e.g., a registry), quantify how rare a given NFT is compared to all NFTs in the underlying dataset, and generate a relative rarity score representing the uniqueness / rarity (and therefore similarity) of that NFT.

[0025] In this way, even when the fingerprint of the original NFT is compared to a candidate NFT, which is, for example, a known NFT transformed using random noise, it will appear suspiciously similar to the fingerprint of the original NFT. By quantifying this similarity, systems and methods for fingerprinting and rarity assessment of digital assets can generate a measure that can be used as a relative known rarity score. In one embodiment, this score is a number between 0% (i.e., the NFT is exactly the same as an existing NFT) and 100% (i.e., the NFT does not resemble any known NFT).

[0026] A system and method for fingerprinting and rarity assessment of digital assets makes it possible to recognize even the most subtle similarities between two digital assets, even if one of them has been transformed. This protocol surpasses other digital fingerprinting techniques that determine the rarity of NFTs, particularly by evaluating the rarity of pixel patterns in the data.

[0027] Overall system architecture Figure 1A shows a non-restrictive example of a computer network environment 1 with a technical solution for fingerprinting and scarcity valuation of digital assets according to the principles of this disclosure. More specifically, Figure 1A shows an exemplary architecture of the Pastel Network environment 1 described above, which is a peer-to-peer decentralized platform for securely registering, trading, and collecting NFTs. As shown in Figure 1A, the computer network environment 1 includes a supernode 5 having a server 10 configured to implement a technical solution for fingerprinting and scarcity valuation of digital assets. More specifically, the server 10 may include, among other things, a processor 100 (shown in Figure 1B) having a fingerprinting engine 150 and a relative scarcity engine 155.

[0028] Figure 1B shows a non-restrictive embodiment of a processor 100 that can be implemented in server 10 (shown in Figure 1A). Exemplary configurations of the processor components are described in detail below, but in the relevant parts, the processor 100 may comprise a computer processor 110 such as a computer processing unit (CPU) and one or more modules having the fingerprinting engine 150 and relative scarcity engine 155, a machine learning module 160, and an image processor module 145 as described above.

[0029] Fingerprinting engine Figure 2 is a flowchart of a hybrid system and process illustrating a method 200 for fingerprinting an NFT according to the principles of this disclosure. One or more steps of routine 200 may be performed using a processor 100, more specifically, a fingerprinting engine 150.

[0030] In one embodiment, the fingerprinting engine is configured to utilize various well-trained deep neural network models, thereby achieving exceptional results for complex data classification tasks. More specifically, in step 205, each model is provided with image data relating to a given NFT250. Each model is configured to generate a list of N numbers in a specific order that characterize the contents of the image, which are referred to as the digital fingerprint vectors of the given image and the model, respectively.

[0031] An analogy for how neural network models generate vectors is as follows: A human subject's brain is scanned in real time to determine precisely which nerve cells are active at any given moment and how each cell is activated. The subject is then shown candidate images, and the results of the brain's activation patterns are recorded as a series of numbers. Similar to how the human brain works, a deep neural network model may contain tens of millions of artificial neurons, and what a given neural network model sees (i.e., the vector "embeddings" it generates depending on its programming and training) is not merely a precise mechanical description of pixels, but a high-level description of the pixel's features. The ability of a neural network model to generate a high-level abstract representation of an image's content makes the output representation powerful for characterizing the unique features of an image and thus evaluating its relative rarity.

[0032] In one embodiment, to construct a digital fingerprint vector, the fingerprinting engine is configured to utilize multiple well-defined neural network models, for example, four neural network models. Each model may require a unique preprocessing pipeline applied to the image. The preprocessing pipeline may include various image processing operations performed on the image, such as resizing and pixel representation.

[0033] The fingerprinting engine is configured to acquire the respective fingerprint vectors output from each model and combine them in step 210 to generate a composite digital fingerprint vector 255.

[0034] In one embodiment, individual fingerprint vectors are concatenated to define a composite digital fingerprint vector. For example, a single digital fingerprint vector may consist of exactly 10,048 decimal digits. However, longer or shorter fingerprint vectors may be used depending on the application.

[0035] In light of the foregoing, the fingerprint engine implementing the digital asset fingerprinting method 200 can be recognized as effectively converting input data into a unique digital fingerprint vector of a given image, which is a compressed representation of the image that dramatically reduces the required dimensions while maintaining a high level of structural content of the image data. Furthermore, tests have shown that this fingerprinting process can be completed in less than a few seconds.

[0036] Relative scarcity engine The Relative Rarity Engine (RRE) utilizes digital fingerprint vectors, which serve as the representation of NFT image data, to evaluate the relative rarity of each digital fingerprint within a dataset more accurately than conventional techniques.

[0037] Figure 3 is a process flowchart illustrating a method 300 for calculating the relative rareness score (RRS) of a digital fingerprint for a dataset of digital fingerprints, based on the principles of this disclosure. One or more steps of routine 300 may be performed using a processor 100, more specifically a relative rareness engine 155.

[0038] In step 305, RRE155 compares the digital fingerprint vector with the digital fingerprint vector of a previously registered NFT in the database. Based on this comparison, RRE calculates a relative rarity score in step 310. In one embodiment, this score is a number between 0% (i.e., the NFT is exactly the same as an existing NFT) and 100% (i.e., the NFT is not even similar to any known NFT). In step 315, RRE determines, based on RRS, whether the NFT is a quasi-replica of a previously registered NFT.

[0039] For this solution to generate digital fingerprint vectors, the digital fingerprint vectors are robust to simple transformations and describe similarities that allow for a deeper level of fingerprint investigation.

[0040] In one embodiment, the RRE is configured to measure the rarity of a given digital fingerprint relative to a database of digital fingerprints (e.g., database 170) and generate an RRS by utilizing different correlation and statistical dependency measures. Specifically, in step 305, the RRE may be configured to compare a candidate digital fingerprint vector with digital fingerprint vectors of all previously registered NFTs in the system using each of several different correlation measures. For each correlation measure, a corresponding correlation value can be calculated, and a list of correlation values ​​can be output by the RRE. From each correlation value, in step 310, the RRE can further calculate the RRS. Tests have shown that the RRE can calculate the correlation between a candidate digital fingerprint and an entire database of hundreds of thousands or millions of NFTs in just a few seconds.

[0041] To reliably calculate RRS and identify quasi-replicating NFTs with reasonable confidence intervals, RRE is configured to leverage various functions and correlation measures, some of which are quite sophisticated and computationally intensive. For example, without limitation, various correlation measures may include Pearson's R correlation, Spearman's Rho correlation, Kendall's Tau correlation, Hoeffding's D dependency, Mutual Information, Hilbert-Schmidt Independence, and XG Boost Feature Importance. For instance, RRE is configured to utilize correlation measures based on the rank of the data rather than the data values ​​themselves, and similarity measures of statistical dependency. Essentially, these measures tell RRE how suspiciously similar two fingerprint vectors are. In other words, they allow RRE to measure how unlikely it would find such a particular pattern between fingerprints if it were actually looking at random or unrelated data.

[0042] RRE is designed to more accurately measure relative rarity using several differentiated, diverse, and powerful similarity measures, and to address the problems of false-negative and false-positive quasi-replica detection results.

[0043] Optimization technology RRE can be configured to further optimize system performance and minimize false-negative and false-positive quasi-replica detection results using additional technologies.

[0044] In one embodiment, the RRE is configured to evaluate all Pearson correlation scores of candidate NFTs compared to all registered NFTs, and to compare the value of the highest correlation of a registered NFT to the correlation at the 99.99th percentile across all registered NFTs. The percentage increase (i.e., Pearson gain) between the highest correlation and the 99.99th correlation can provide some useful information if it is sufficiently large. For example, suppose there are 10,000 registered fingerprints and 10,000 correlation scores sorted in descending order. If the RRE is configured to compare the highest value to the 99.99th percentile score, and the highest score is 86.00% and the second highest score is 65.00%, then it suggests a Pearson gain of 86.00% / 65.00%-1=32.3%. This means that exactly one had a much higher correlation than the rest of the dataset. Extending this across the entire dataset, the RRE can identify correlations across broad clusters of NFT data objects. Implementing this requirement can dramatically improve the confidence threshold within the system.

[0045] In one embodiment, the RRE is configured to quantify similarity scores with high accuracy in a range of 0.00% to 100.00% in a way that resembles human intuition, rather than as binary 0-1. The RRE may be configured to combine the results of the processes described above to generate various subscores that can be converted into a single number between 0.00% and 100.00%. One subscore sums up various similarity measures and compares the sum to the maximum value that would have been obtained if the NFTs were identical, essentially averaging the results of different similarity measures within the available range. The RRE is further configured to combine the subscores across each methodology to calculate an overall relative rarity score.

[0046] Machine Learning This solution for fingerprinting and scarcity valuation of digital assets may be further configured to further optimize the system and method for calculating RRS and detecting quasi-replica images using a parallel approach with machine learning.

[0047] In one embodiment, the processor 100 includes a machine learning module 160 which includes one or more supervised machine learning systems and / or one or more unsupervised machine learning systems. The machine learning module includes, for example, Word2vec deep neural networks, Convolutional Architecture for Fast Feature Embedding (CAFFE), Artificial Immune Systems (AIS), Artificial Neural Networks (ANN), Convolutional Neural Networks (CNN), Deep Convolutional Neural Networks (DCNN), Region-Based Convolutional Neural Networks (R-CNN), You-Only-Look-Once (YOLO) method, Mask-R-CNN, Deep Convolutional Encoder-Decoder (DCED), and Recurrent Neural Networks. This may include Work (RNN), Neural Turing Machine (NTM), Differential Neural Computer (DNC), Support Vector Machine (SVM), Deep Learning Neural Network (DLNN), Naive Bayes, Decision Tree, Logistic Model Tree Induction (LMT), NBTree Classifier, Case-based, Linear Regression, Q-Learning, Time Difference (TD), Deep Adversarial Network, Fuzzy Logic, K-Nearest Neighbors, Clustering, Random Forest, Rough Sets, or any other machine intelligence platform capable of supervised or unsupervised learning.

[0048] In one embodiment, the processor 100 can access a large number of known NFT files by accessing, for example, the open data of OpenSea, a platform for creating and trading NFTs. In addition, the processor can isolate a certain percentage of the data to define a subset of registered NFTs and calculate their digital fingerprint vectors, which are stored in a database 170 of registered NFTs. The remaining NFT files in the dataset are isolated as a subset of unknown true original NFTs, i.e., their digital fingerprint vectors are not calculated, and it is known that none of this subset of NFTs are in the database. Finally, the processor 100 may be configured to generate a large corpus of artificially generated quasi-replica NFTs through transformation techniques applied to the NFTs of the subset of registered NFTs, as shown and explained in the example below. For example, Figure 4 shows three sets of images 405, 410, and 415, each containing the original image (the leftmost image) and quasi-replica NFTs generated by various transformation techniques (the remaining images).

[0049] In this case, the processor 100 may be configured to apply digital asset fingerprinting and rarity assessment protocols to these transformations stored, for example, in a database 170. Specifically, known quasi-replica NFTs are selected from a corpus of artificially generated quasi-replica NFTs, and their digital fingerprint vectors are calculated (e.g., according to Method 200). In addition, a correlation scale funnel is applied to compare the digital fingerprint vectors with all registered NFTs in the database (e.g., according to Method 300). Next, original NFTs are selected from a subset (it is known that the selected original NFTs cannot be identified as quasi-replicas of any registered NFT in the database), and the same fingerprinting and rarity assessment routines are applied to these original NFTs. For each of these, the processor is configured to observe how many registered fingerprints reach the final stage of the funnel. Instead of tracking an overall relative rarity score, the processor may assign a binary label of 1 to artificial quasi-replica NFTs and 0 to true originals. The processor can then model the input data in light of the various similarity measures and subscores calculated for each image.

[0050] This methodology enables the processor 100, more specifically the machine learning module 160 (Figure 1B), to generate a predictive model for determining whether an NFT is a duplicate or an original, using machine learning training or supervised learning. More specifically, given a column of data representing the maximum correlation score obtained by comparing a candidate NFT with all registered digital fingerprints, the predictive model is configured to predict whether the label is 1 (i.e., a duplicate) or 0 (i.e., an original) using various techniques. In one embodiment, the predictive model may include a trained random forest classifier configured to predict labels from input data via XGBoost using a set of decision trees. In another embodiment, the predictive model may comprise a deep neural network classifier built using a Keras application and configured to predict labels from input data. Each of these models is subtly differentiated and provides a gradient of different degrees. Thus, the system can combine each score to produce a final overall average score, which is more accurate than any individual score and maps closer to human intuition.

[0051] Example image transformation The following description illustrates exemplary techniques that may be configured to be implemented by a processor 100 to generate quasi-replica images and corresponding results from systems and methods for fingerprinting and scarcity assessment of digital assets performed on exemplary NFTs taken from a test corpus of quasi-replica NFTs artificially generated by each conversion technique.

[0052] In one embodiment, the processor 100 may be configured to create a quasi-replica NFT by modifying a known NFT by adding random noise. Preferably, the noise is added in amounts sufficient to prevent conventional image matching solutions, such as Google's reverse image search, from finding any matches at all. For example, Figure 5 shows two images: the original NFT image 505 and a modified version of image 510 modified according to the technique described above. During testing, the system classified image 510 as a quasi-replica and assigned it a fairly low overall average rarity score of ~28.82%. The system was also able to accurately identify the original exact registered image from which the candidate image was derived.

[0053] As a further example, a quasi-replica can be generated by stretching a registered image. For example, Figure 6 shows two images: the original NFT image 605 and a modified version, image 610, which has been altered by stretching. During testing, the system classified the modified image 610 as a quasi-replica and assigned it a rarity score of ~0.087, which is quite low considering that identical images would receive a score close to 0.0.

[0054] As a further example, quasi-replicas can be generated by applying a contour or edge detection filter to a registered image. For example, Figure 7 shows two images: the original NFT image 705 and a modified version image 710, which has been altered by applying a contour or edge detection filter. During testing, the system classified the modified image 710 as a quasi-replica and assigned it an overall mean rarity score of only ~0.15, despite its dramatically different appearance compared to the original.

[0055] Rarity on the Internet In one embodiment, the processor 100 may be further configured to crawl and index websites and evaluate the rarity of NFTs compared to those known on the internet, leveraging existing open resource functions such as Google's reverse image search. For example, in response to a user's attempt to register an NFT, such as image 805 shown in Figure 8A, the processor 100 may be configured to present the NFT to such a reverse image search tool, which can return results described as visually similar data. For example, based on image 805, the reverse image search tool may return a set of visually similar images 810 on the internet, as shown in Figure 8B. The images shown in 810 are indexed by various websites, and the reverse image search tool can find the exact image. The same applies to a collection of NFTs, where there are tens or even thousands of extremely similar images created as part of a larger group of collectible NFTs.

[0056] Therefore, when a user attempts to register an NFT within the Pastel Network environment 1, in addition to determining its rarity and detecting quasi-replicas by comparing it to previously registered digital assets within the network as described above, the processor 100 may be further configured to evaluate the rarity of that NFT compared to the internet. In one embodiment, the rarity of the NFT relative to the internet may be performed entirely in a distributed manner by each node within the network environment 1. Specifically, multiple randomly selected supernodes 5 (each including a server 10 and a processor 100) may be configured to perform the same search and ensure that all their results are exactly the same (for example, RRS scores compared to registered NFTs are independently calculated by multiple supernodes, and the consistency of the results is checked).

[0057] If an artist creates a completely new image and has never shared it online before, the artist can initially register it within the Pastel Network environment 1. The resulting NFT is therefore rare within the Pastel Network environment 1 and also rare on the internet, allowing processor 100 to assign it the highest level of certified rarity available on the system. In addition, both of these rarity indicators can be written to an immutable NFT registration ticket, which is recorded as part of the Pastel blockchain (not shown). If the image is later shared on social media or other websites, the NFT's rarity score does not change. In other words, the most important thing is how rare the image was when the NFT was registered with Pastel.

[0058] To make it understandable, Exemplary Network Environment 1, which implements a system and method for fingerprinting and scarcity assessment of digital assets, promotes a much stronger concept of what it means for a digital image to be scarce. This solution can not only verify authenticity and origin using the creator's digital signature (as in all NFT systems in use), but also, going further, actually assess how scarce the underlying pixel pattern of an image is, both within Pastel Network Environment 1 itself and on the broader internet. Since value is determined primarily by scarcity / rarity, the additional layer of scarcity authentication provided by this exemplary solution can provide a more favorable assessment for NFT creators. After all, even if another creator makes a similar NFT in the future, that creator can register it with Pastel Network 1, but it will not be assigned a scarcity score close to that of the original image.

[0059] Furthermore, even if the original creator attempts to create another similar or identical NFT in the future, the later-created NFT will not have the same rarity score as the creator's original NFT. This protects NFT buyers from inflation caused by the creator, something that cannot be achieved with NFT systems that rely solely on digital signature verification. This is because a second or third very similar or identical NFT could appear genuine even if it is a plagiarism of the original, simply because the artist has correctly signed it.

[0060] Supernode collaborative operation In one embodiment, as shown in Figure 1A, the Pastel Network environment 1 includes multiple supernodes 5, from which multiple Pastel supernodes can be randomly selected to calculate the RRS of NFTs and / or evaluate their rarity relative to the internet, and the results can be compared for consistency. The coordinated and distributed configuration of the supernodes can provide redundancy, reliability, and scalability.

[0061] More specifically, in one embodiment, each supernode 5 is provided with several support directories for local machine presence. Thousands of digital fingerprints are seeded into a file in a database 170 of digital fingerprints (for example, SQLite can be used as the database system). In addition, there are pre-trained classifiers such as XGBoost or Keras, which are downloaded and trained by each processor 100.

[0062] Each supernode's processor runs a software process (e.g., a Python file) which is configured to loop and monitor the input file folder 108 for new NFT files or for duplicate detection. The input file folder 108 (and output file folder 118) are shown in Figure 1 as independent file repositories accessible by multiple supernodes 5, although such folders could be provided locally to each supernode. When new data from a user (e.g., using client device 106) is uploaded to this folder, the processor is configured (e.g., by code) to wait for the file upload to finish, and then automatically proceeds to analysis. More specifically, it first calculates the RRS of the NFT, and then calculates the Internet scarcity score.

[0063] In addition, the processor may be configured (by code, for example) to compute a series of perceptual hashes. Perceptual hashes are used to scale the system when a candidate NFT is an almost exact copy of an already registered NFT. In fact, if the hash-based system finds a copy, the processor can skip the computationally intensive process and proceed directly to assign a scarcity score of zero to the image. In this regard, the processor may be configured to use various image hash generation algorithms, including, but not limited to, Facebook Research's pdq algorithm and Apple's neural hash algorithm, as well as other conventional methods.

[0064] Once the above process is complete, the processor may be configured (e.g., by a program) to generate a JSON file to be output to a folder, for example, a duplicate detection output file folder 118. In one embodiment, the file name is the first 10 characters of the SHA3-256 hash of the image file, for example, 00d201498d.json. The contents of this file may include various information about the NFT and the prominent indicators generated by the system. Figure 9 shows the text contents of an exemplary output file 905, which includes the hash of the candidate input file "hash_of_candidate_image_file", the duplicate detection result ("is_likely_dupe": 0), the average rarity score value ("overall_average_rareness_score"), the internet rarity score value ("is_rare_on_internet"), the number of matches found on the first page of the internet search results ("matches_found_on_first_page": 3), and the number of matches returned by the internet search results. This includes the number of pages ("number_of_pages_of_results": 2), the URL of the first match in the internet search ("url_of_first_match_in_page"), a score indicating whether the image contains explicit material ("open_nsfw_score" and "alternative_nsfw_scores"), the image hash ("image_hashes"), and the digital fingerprint generated for the candidate image file ("image_fingerprint_of_candidate_image_file", abbreviated for brevity).

[0065] The above-described architecture of an exemplary Pastel computer network environment 1, equipped with technical solutions for fingerprinting and scarcity assessment of digital assets, can scale to hundreds of thousands or more images. In one embodiment, the machine executing the copy detection code preferably maintains a table of all registered images in memory, which can be a limiting factor. However, one way to scale the system to millions of images is to use a technique known as sharding. More specifically, each supernode (SN) 5 has an identifier used as its name in the Pastel Network environment 1, and using the concept of XOR distance, each SN may be assigned a particular subset of the space of all previously registered images. These associations are not disparate; that is, the same image is assigned to at least three SNs so that they can be compared to ensure that each of their results matches.

[0066] Each SN is responsible for calculating correlation / dependency scores for candidate images compared to a subset of all registered images assigned to that SN. After this, these correlation scores are shared with other SNs in the network, and consistency is checked by comparing the results from SNs assigned the same subset of images. All verified results from each group of SNs are sent to the top three supernodes, which combine these results and then terminate the calculation (e.g., scarcity score, internet scarcity score, etc.). This avoids the aforementioned limiting factor, namely, the complete table of registered fingerprints existing in the memory of a single machine at once. In addition, the results of the overall calculations from each of the top three SNs can be compared by all SNs to check for consistency, and if they match, the results can be written to the blockchain.

[0067] Exemplary Embodiments The following describes further embodiments of a system for fingerprinting and scarcity assessment of digital assets, which can be implemented using a processor 100 comprising processing circuits or modules such as an image processor 145, a fingerprinting engine (FE) 150, a relative scarcity engine (RRE) 155, and a machine learning module 160. More specifically, given a candidate digital asset (having a corresponding visual appearance), the system is configured to detect whether the digital asset is a copy of another (e.g., previously encountered) digital asset. The system can further assess or authenticate the scarcity of the digital asset based on its visual appearance against a registry of other digital assets (or equivalent discriminatory information of the visual appearance, such as a digital fingerprint of those other digital assets). While the exemplary fingerprinting and scarcity assessment technique applies to a digital image having a relevant visual appearance, the technique can be generalized to other types of digital works. For simplicity of explanation, the terms digital asset, NFT, visual appearance, input image, and digital fingerprint may be used interchangeably, and their actual meanings are clear from the context.

[0068] This technology begins with generating a digital fingerprint of a digital asset. For example, the digital fingerprint may be generated by a fingerprinting engine 150. The digital fingerprint preferably satisfies a variety of properties, including (1) being relatively easy to represent, such as a vector of numbers where each number represents one aspect of the visual appearance of the digital asset; (2) being easy to handle, such as a sufficient number to characterize the visual appearance; (3) being discriminative, such as different visual appearances correspondingly having different digital fingerprints; and (4) capturing similarity, such that the closer the visual appearances, the closer the corresponding digital fingerprint. These and other required properties often conflict, and due to the vastness of the similarity of visual appearances and the need for computationally usable digital fingerprints to detect duplicates or establish rarity against potentially millions of other visual appearances, this problem is well suited to being solved by machine learning.

[0069] Therefore, in an exemplary embodiment, deep learning is used by FE150, for example, using the ResNet50 deep learning model to generate fingerprint vectors. ResNet50 is a 50-layer deep convolutional neural network that takes an input image and is trained to "embed" image features into vectors, such as a 2048-dimensional vector of values ​​(e.g., floating-point numbers) that characterize the content of the image. While this vector embedding of the image can serve as a vector representation of the image, it may require further processing (such as additional learning steps) to obtain the desired properties of the digital fingerprint. In addition, other embodiments of FE are described as processing the image using multiple different deep learning models (e.g., four), each generating a vector, and the resulting output vectors being concatenated into one larger vector, but the resulting vector may have redundant numbers, making the image embedding inefficient. The number of elements that make up a vector is sometimes called a dimension, and it should be noted that although these vectors have only one dimension in a mathematical sense, a vector with N elements can also be considered to represent a point in an N-dimensional space.

[0070] Therefore, in exemplary embodiments, an additional deep neural network may be used by FE150 after the initial (e.g., Resnet50) deep learning network. This additional neural network is trained to take vectors from the initial (e.g., Resnet50) neural network as input and transform those vectors into new vectors (e.g., 2,048-dimensional vectors) optimized for the specific task of determining image similarity. This can be done using the machine learning module 160 by training a model on a large number of (millions or millions of) image pairs, where the first element of the pair is the original image and the second element is the transformed image (e.g., a modified image that visually looks the same as the original image). The transformations may include (but are not limited to) various operations such as cropping, flipping, rotating, adding a rectangular overlay (e.g., partial covering), and deleting parts of the image. For example, the image transformation may be performed by the image processor 145.

[0071] More specifically, the model is trained on the concept that, since the related images are known beforehand to be simple transformations of each other, they should be mapped (or embedded) into similar vectors by the model. This additional training is significant because running ResNet50 through the original image and a transformed image (for example, the same image but with randomly placed black rectangles filling 25% of the original image) can produce vectors that look different. However, this additional training phase brings the related images closer together in the embedding space, allowing the model to see beyond superficial changes such as simple transformations and to consider such transformed images as similar to the original image. Thus, while the first deep learning model learns to characterize image content into simple vectors of properties, the second deep learning model is trained to refine or optimize its characterization to determine image similarity by recognizing simple transformations of one image as similar to one another. This results in much higher accuracy using a single vector than is possible with a single deep learning model.

[0072] As already mentioned, the output of the second deep learning model can have the same number of dimensions as the output of the first deep learning model, for example, 2048 dimensions (or numbers). However, the second deep learning model may generate redundant numbers in the output vector (e.g., numbers that behave similarly to linear combinations of 1 or more other numbers in the vector). Therefore, in one embodiment, these extra numbers (dimensions) in the output vector can be removed using dimensionality reduction techniques such as principal component analysis (PCA). For example, PCA (a type of unsupervised machine learning) can be performed by the machine learning module 160 on hundreds of thousands of sample outputs of the second deep learning model. This compression of numbers into a more orthogonal (linearly independent) set can remove more than a quarter of the numbers, resulting in a final "optimized" fingerprint vector with, for example, about 1500 numbers (dimensions).

[0073] These final vectors are discriminative and optimized for the task of measuring image similarity, from which the answer (NFT similarity) can be obtained relatively easily. To compare two images (visual appearances), the dot product (also known as cosine similarity) of the two vectors is calculated, which measures the degree to which those vectors make the same angle with respect to the origin in a 1,500-dimensional space. For example, these dot products can be calculated by the relative rarity engine 155. This is a calculation that can be performed simply, quickly, and efficiently, provided that the original images (visual appearances) are aggregated into a digital fingerprint consisting of 1,500 consistent numbers as described above. The result of the dot product is a single number representing how similar the vectors are, with a larger number indicating higher similarity (and vice versa for smaller dot products).

[0074] These results are then normalized using normalization techniques to generate numbers within a specified range (-1 to 1, or 0 to 1, etc.) or distribution, for example, positive values ​​representing a high probability that two images are quasi-replicas, and negative values ​​representing a high probability that the images are not replicas. For example, this normalization can be performed by the relative scarcity engine 155. Furthermore, within the normalized range, the magnitude plays a role in similarity, with larger positive values ​​indicating a higher probability of replicas, and smaller negative values ​​indicating a higher degree of dissimilarity. For simplicity of explanation, unless otherwise specified, the normalized range is assumed to be 0 to 1 (including 0 and 1).

[0075] The technique described above generates a digital fingerprint vector consisting of 1500 numbers, with little correlation between the numbers across diverse sets of visual appearances. Therefore, as the first step in determining the rarity of a digital asset relative to a known or registered set of visual appearances (or registry), FE150 can calculate the corresponding digital fingerprint for each registered digital asset and store it in the registry. The fingerprinting engine 150 then uses the same technique to calculate the digital fingerprint of the visual appearance of the digital asset. The corresponding dot product of the registered digital fingerprint and the digital fingerprint of the digital asset is then calculated and normalized (e.g., by the relative rarity engine 155). These are efficient calculations that can be performed even on large registries (millions of visual appearances, or more precisely, their digital fingerprints).

[0076] In one embodiment, the RRE155 is configured to use the aforementioned dot product to narrow down the vast set of registered digital fingerprints to a small number N (e.g., 10) of the most similar (i.e., most likely to be quasi-replicas of candidate digital fingerprints) possible digital fingerprints.

[0077] Assuming N=10 for simplicity of explanation, the above technique efficiently identifies the 10 registered visual appearances that are closest (or most similar) to the visual appearance of the digital asset whose rarity is to be determined. The 10 digital fingerprints selected are potentially taken from a database of millions of candidate fingerprints. However, subtle differences between two similar visual appearances can be difficult to determine from a single digital fingerprint (and corresponding dot product or normalized dot product), especially when trying to determine if they are duplicates or measure similarity. Therefore, RRE can be configured to implement additional comparison techniques. Specifically, RRE can apply additional measures of statistical dependence to compare candidate digital fingerprints to each of the top 10 most similar registered image fingerprints.

[0078] In one embodiment, this additional measure of statistical dependence consists of a population measure of deviation from independence, e.g., Heffding's D, or Heffding's D, and the Hilbert-Schmidt Independence Criteria, or HSIC. For example, this additional measure of statistical dependence can be calculated by the relative rarity engine 155. These measures are computationally more complex than cosine similarity (e.g., the simple dot product of digital fingerprints), but they provide a measure of more subtle differences in how two digital fingerprints being compared relate to each other. These two additional measures of similarity (e.g., dependence, independence, or rarity) help reduce the possibility of spurious matches that occur by chance with cosine similarity, which can be a significant risk when a database of registered images contains millions of image fingerprints and can lead to a large number of false positive matches. Therefore, in one embodiment, Hefting's D and HSIC, or both, are used to clearly measure the similarity of a candidate fingerprint to an existing fingerprint that shows a high cosine similarity to the candidate fingerprint.

[0079] Using one or both of these specific similarity measures to further refine the closest matches generated by cosine similarity significantly reduces the frequency of false positive matches. Heffding's D and HSIC are computationally intensive, making it impractical to apply them to all registered image fingerprints, especially for millions of registered fingerprints. Narrowing the list of possible matches to just 10 initially makes these additional measures more manageable. While Heffding's D and HSIC are specifically cited as additional measures of statistical dependence, cosine similarity can be supplemented with any additional measure of statistical dependence capable of reducing false positive matches from cosine similarity (or other simple measures such as Pearson correlation), particularly when dealing with large databases of registered digital fingerprints (e.g., millions). Here again, for simplicity of explanation, we assume N=10, but N could be other values ​​such as 8 or 12 in other embodiments.

[0080] In another embodiment, in addition to these three measures of statistical dependence (normalized cosine similarity, Heffding's D, and HSIC), the RRE155 may be configured to evaluate the “gain” of these measures for multiple digital fingerprints in order to further assess similarity. More specifically, cosine similarity can be used to establish a similarity ranking between the top 10 most similar registered digital fingerprints and a candidate digital fingerprint for which rarity is to be determined. The RRE can also calculate additional statistical dependences (Heffding's D and HSIC) for multiple digital fingerprints with consecutive ranks, and compare those measures to determine the “gain” of those measures. In other words, the gain quantifies how much each measure of different statistical dependence differs with respect to its own cosine similarity rank. As will be further explained, these gain / difference values ​​can be used to more favorably assess how similar the top 10 matches are to (and to some extent to) the candidate digital fingerprint. This allows for a more informed selection regarding whether one of the top ten is a duplicate, and enables the determination of how rare (or different) a candidate digital fingerprint (and therefore the corresponding digital asset) is within the database of registered digital fingerprints (and their corresponding visual appearance or image).

[0081] For more details, consider, as an example, the top 10 registered fingerprints measured by cosine similarity to the candidate digital fingerprint (or candidate digital fingerprint). These 10 digital fingerprints can be sorted from 1st, 2nd to 10th according to their respective cosine similarities to the candidate digital fingerprint. RRE can then calculate two other measures of statistical dependence on the candidate digital fingerprint (Hefting's D and HSIC) for multiple most similar fingerprints, e.g., the 1st most similar fingerprint and the 2nd most similar fingerprint. The difference (difference, gain) between each measure for digital fingerprints with consecutive ranks can then be calculated.

[0082] For example, suppose the three scores for the top-ranked, most similar fingerprint are (normalized cosine similarity: 0.25; Hefting's D: 0.15; HSIC: 0.09). Then, suppose the three scores for the second-ranked, most similar fingerprint are (normalized cosine similarity: 0.21; Hefting's D: 0.09; HSIC: 0.07). The "gain" score can be calculated as the percentage (percentage) that the top score exceeds the second-ranked score, i.e., as follows: Gain of normalized cosine similarity: 0.25 / 0.21-1=~0.190; The gain of D in Hefting's equation: 0.15 / 0.09-1=~0.666; HSIC gain: 0.09 / 0.06-1=0.5.

[0083] This can be done similarly for the second and third digital fingerprints, the third and fourth digital fingerprints, and up to the top 10 digital fingerprints. This gain can be generalized to all of the top 10 results by comparing the Nth most similar fingerprint with the (N-1)th most similar fingerprint (which requires scoring the 11th most similar fingerprint).

[0084] It should be noted that differences, disparities, or gains can be expressed in alternative forms. For example, a gain could be the absolute difference or percentage increase of a higher-ranked digital fingerprint compared to a lower-ranked digital fingerprint.

[0085] In summary, gain measures the degree to which the statistical dependence of a registered fingerprint on a candidate image fingerprint changes as the similarity ranking of the registered fingerprint changes. Here again, this measure provides additional insight into how similar a candidate digital fingerprint is to a registered digital fingerprint. For example, if there are several registered fingerprints that are similar to a candidate image fingerprint, but the ranked 1st digital fingerprint has a larger gain than the ranked 2nd digital fingerprint, this provides additional evidence that the 1st fingerprint is somewhat unique and likely to be a duplicate, while the 2nd, 3rd, and other similar registered digital fingerprints are merely visually similar and do not reach the level of a duplicate.

[0086] In an exemplary embodiment, multiple measures of statistical dependence between registered fingerprints with consecutive ranks, and their corresponding gains (a total of six numerical inputs in the exemplary configuration of three measures and three gains), are determined for each of the top ten registered digital fingerprints. These six numerical scores can then be input into a machine learning statistical model. This final model can use logistic regression to process these six input numbers and generate a single numerical output, i.e., a number between 0.0 and 1.0 representing the probability that a candidate image is a copy of a particular registered image. For example, this statistical model can be trained in machine learning module 160. This machine learning model is trained by taking a large corpus of images and generating transformed images as quasi-replicas that exhibit advanced transformations that make image detection difficult.

[0087] For example, in one embodiment, 100,000 images are selected from various sources, such as a publicly available dataset of artwork images like Wikiart. From these, tens of thousands of complex quasi-replica images are generated through an automated process using various complex transformations. This is further explained with reference to Figures 10A-10C.

[0088] Figures 10A–10C show original and transformed artwork images for training a machine learning technique (such as a logistic regression model) to authenticate the rarity of digital assets, according to one embodiment. Figures 10A–10C illustrate several examples of complex transformations. Figure 10A shows the original artwork image, and Figures 10B–10C show two different complex transformations of the original artwork image in Figure 10A. A human observer would likely be able to immediately recognize that Figures 10B–10C are transformed copies of Figure 10A, but such a determination is more difficult for a computer algorithm. This is because almost every pixel has been changed, resulting in complex structural changes that significantly affect the overall pattern in the transformed image.

[0089] In a more specific embodiment, training images can be divided into three general categories: (1) registered original images (e.g., 80,000 out of 100,000 original images), (2) generated copies of registered original images (e.g., 50,000 images based on some of the 80,000 images in the first category), and (3) true original images (e.g., the remaining 20,000 out of 100,000 original images). Registered original images form a registry of protected visual appearances (already seen and registered), while generated copies are used to check for false negatives (i.e., copies that are determined to be originals) and false positives (i.e., copies that are determined to be copies but are based on incorrect registered original images). In addition, true originals are used to check for false positives (i.e., originals that are determined to be copies).

[0090] In addition to the training data, the logistic regression model is trained using a loss function that generates a single number to measure how well the model is performing. In one embodiment, the Area Under the Precision-Recall Curve (AUC-PR or AUPRC) is used. AUPRC is a model performance metric for binary responses such as rare events (e.g., the detection of a duplicate). For example, AUPRC may be computed by the machine learning module 160 as part of training the logistic regression model. Here, the logistic regression model learns to work with six input statistics for the top 10 cosine similarity registered fingerprints and generates a single registered fingerprint with the highest probability of being a duplicate among the other top 10 fingerprints. If this probability is above a threshold probability (e.g., 50% or 75%), the logistic regression model determines that a duplicate has been found and outputs 1 for the highest registered fingerprint and 0 for all other registered fingerprints. If this probability is below the threshold probability, the logistic regression model determines that no duplicate has been found and outputs 0 for all registered fingerprints.

[0091] The logistic regression model is trained by randomly selecting images from generated copies and true (unregistered) originals, generating corresponding measures (and their gains) of their statistical dependence on digital fingerprints and registered images, and running the logistic regression model on the measures and gains for the top 10 registered fingerprints. Generated copies are passed through the system to detect false negatives and false positives, and true originals are passed through the system to detect false positives. Ideally, none of these exist (i.e., the logistic regression model is perfect), but in reality, there is a trade-off between precision (precision in the sense of minimizing false positives, i.e., saying something is a copy when something is a true original, and false positives, i.e., saying something is a copy of a registered image when something is a copy of another registered image) and recall (recall in the sense of minimizing the number of false negatives, i.e., saying something is a true original when something is actually a copy). Precision and recall depend on the selected cutoff threshold probability. The AUPRC index captures both recall and precision across all possible thresholds into a single number, and is therefore an ideal measure.

[0092] In one embodiment, logistic regression is trained using AUPRC as the loss function, and given six numerical inputs for each of the top 10 registered fingerprints, the coefficients that best optimize AUPRC are found. That is, AUPRC can determine the optimal coefficients (or weights) for each of the three scales and their gains. The six coefficients (weights) obtained with an exemplary AUPRC are 11.023235 (cosine scale), 247.833809 (Hefting's D scale), -154.621053 (HSIC scale), 26.973047 (cosine gain), -254.253315 (Hefting's gain), and 132.568845 (HSIC gain).

[0093] In summary, in an exemplary embodiment, all six measures and difference / gain of each of the top 10 registered fingerprints are input to a logistic regression, which outputs a value representing the probability of replication. For example, the probability determination may be performed by a relative scarcity engine 155 running the trained logistic regression model. In addition, in one such embodiment, AUPRC is used as an intrinsic loss function for training the logistic regression, which provides the best overall balance between the dual problem of capturing as many replications as possible while minimizing the blame for false positives of true originals.

[0094] In an exemplary technique for authenticating the rarity of a digital asset (or more precisely, its visual appearance or image) against a registered set of visual appearances or images, a rarity score between 0 and 1 is returned, where 0 is not rarity (i.e., it is a copy of one of the registered images), and 1 is extremely rarity (it has no similarity to any of the registered images). In other words, rarity defines a measure of novelty of previously registered images to some specified corpus, as opposed to being identical to a registered image or a simple transformation of one or two of the registered images. Thus, in some embodiments, rarity is defined as 1 minus the highest probability of it being a copy of any of the previously registered images. For example, the calculation of rarity may be performed by a relative rarity engine 155. For example, if the above technique is used for a candidate image against a registry of known images, the top 10 closest registered images are identified. Of these, the most likely copy image is selected, and the probability that the candidate image is a copy of the selected image is determined (e.g., 0.7 or 70%). Finally, the rarity is calculated by subtracting this from 1 (e.g., 1-0.7=0.3 or 30%).

[0095] In summary, the scarcity score is calculated by subtracting the probability of the candidate image being the largest copy of any previously registered image from 1.0. Therefore, if there are no images similar to the candidate image among the already registered images, none of them have a high probability of being copies, and thus the candidate image's scarcity score will be high. Conversely, if a registered image has a high probability of being one or more copies of the registered image, the candidate image's scarcity score will be close to zero. Thus, the scarcity score measures the novelty or rarity of one or more images for the purpose of quantifying the potential value of digital image products (such as NFTs or other digital assets).

[0096] Figure 11 is a flowchart of an exemplary automated method 1100 for authenticating the scarcity of a digital asset, according to one embodiment. Method 1100 is automated under the control of an electronic circuit (such as a microprocessor) configured to perform the steps of Method 1100 (e.g., by code, such as being programmed, by custom logic such as configurable logic gates, or similar).

[0097] Parts or all of Method 1100 may be carried out using the components and techniques shown in Figures 1A to 10C. In addition, parts of this method and other methods or processes disclosed herein may be carried out in or using simulation logic, such as a custom or pre-programmed control logic device, circuit, or processor, computer, software, or other circuit configured by code or logic to perform an assigned task (e.g., ASIC, FPGA), such as a programmable logic circuit (PLC). The device, circuit, or processor may be, for example, a dedicated or shared hardware device (such as a laptop, single-board computer (SBC), workstation, tablet, smartphone, part of a server, or dedicated hardware circuit such as an FPGA or ASIC), or a computer server, or part of a server or computer system. The device, circuit, or processor may include a non-temporary computer-readable medium (such as read-only memory (ROM), flash drive, or disk drive, or other CRM) that stores instructions causing a part of Method 1100 (or other disclosed methods or processes) to be carried out when executed by one or more processors. It should be noted that in other embodiments, the order of operations can be different, and some operations can be omitted. Parts of Method 1100 can also be performed using logic, circuits, or processors located in, or communicating with, processing circuits configured by code to perform those parts of Method 1100.

[0098] In Method 1100, the process begins with a step (1110) in which a first digital fingerprint of a digital asset is generated (by a fingerprinting engine 150, etc.) by running a deep learning model on the first visual appearance of the digital asset (e.g., a digital artwork). The generated first digital fingerprint is a first vector consisting of 1,000 to 2,000 numbers (e.g., 1,500 numbers) for determining visual appearance similarity (e.g., 1,500 numbers that quantify the aspects of the visual appearance that make it look similar to other visual appearances). The deep learning model is trained (by a machine learning module 160, etc.) on at least 100,000 (e.g., 1 million) visual appearances to generate the corresponding first vector for determining visual appearance similarity.

[0099] Method 1100 further includes a step (1120) of evaluating the similarity of a first digital fingerprint to a registry of digital fingerprints (by a relative rarity engine 155, etc.) by calculating the corresponding dot product (for cosine similarity, etc.) of the first digital fingerprint and the registered digital fingerprints. Here, the registry of digital fingerprints functions as a database of registered images, where the rarity of a digital asset is based on how dissimilar its visual appearance is to other such visual appearances measured by its corresponding digital fingerprint. A large dot product indicates strong similarity, and a small dot product indicates weak similarity. The registry of digital fingerprints is obtained by identifying visual appearances to establish rarity (e.g., building a database of visual appearances), and then running a deep learning model on the identified visual appearances to generate corresponding digital fingerprints.

[0100] In addition, Method 1100 includes a step (1130) of determining a corresponding first measure of the statistical dependence of registered digital fingerprints (by relative rarity engine 155, etc.) by normalizing the calculated dot product. The normalized dot product takes a range (0 to 1, or -1 to 1, etc.) having a first end (0 or -1, etc.) corresponding to no similarity to the first visual appearance and a second end (1, etc.) corresponding to identical similarity to the first visual appearance. Method 1100 also includes a step (1140) of selecting a fixed number N (10, etc.) registered digital fingerprints (10 digital fingerprints whose corresponding visual appearance is most similar to the visual appearance of the digital asset) (by relative rarity engine 155, etc.) whose corresponding N normalized dot products are closest to the second end of the normalized range.

[0101] Furthermore, Method 1100 includes a step (1150) of determining a second measure of corresponding N statistical dependencies (by relative scarcity engine 155, etc.) by applying Heffding's Dependency Scale (Heffding's D) or the Hilbert-Schmidt Independence Criterion (HSIC) to N combinations of a first digital fingerprint and N selected digital fingerprints. These are more computationally based measures of dependency (similarity) and can be used to further distinguish the most similar digital fingerprints. Finally, Method 1100 includes a step (1160) of authenticating the scarcity of the digital assets (by relative scarcity engine 155, etc.) using the determined first and second measures of statistical dependency (by adding them together, etc., assuming they are normalized to the same, equivalent, or comparable ranges).

[0102] In one embodiment, running a deep learning model on a first visual appearance includes running a first deep learning model (such as a ResNet50 deep learning model) on the first visual appearance to generate a second vector consisting of 2,000 to 4,000 numbers (such as 2,048 numbers) that characterize the first visual appearance. The first deep learning model is trained (by machine learning module 160, for example) on at least 100,000 visual appearances (such as 1 million visual appearances) to generate the corresponding second vector that characterizes the visual appearance. Running a deep learning model further includes running a second deep learning model (such as an additional deep neural network) on the generated second vector to generate a third vector of similar size to the second vector (such as 2,048 numbers) that is optimized for determining the similarity of visual appearances.

[0103] More specifically, the second deep learning model is trained on at least one million pairs of second vectors (by machine learning module 160, etc.). Each pair of second vectors is generated by running the first deep learning model on the corresponding original visual appearance and the transformed visual appearance obtained by applying one of several transformations to the original visual appearance (by image processor 145, etc.). In addition, the second deep learning model brings the generated pairs of second vectors closer together in embedding space to generate corresponding third vectors optimized for determining the similarity of the visual appearances. Running the deep learning model further includes compressing the generated third vectors into the first vectors by removing redundant information, such as by using principal component analysis (PCA), in order to generate a first digital fingerprint. In one such embodiment, applying one of the transformations to each original visual appearance and the corresponding transformed visual appearance includes performing one of the following: cropping the original visual appearance, inverting the original visual appearance, rotating the original visual appearance, adding one or more rectangular overlays to the original visual appearance, and removing one or more portions from the original visual appearance.

[0104] In one embodiment, the step (1150) of determining corresponding N second measures of statistical dependence includes applying Heffding's D to N digital fingerprint combinations. Method 1100 further includes the step of determining corresponding N third measures of statistical dependence by applying HSIC to N digital fingerprint combinations. In addition, the step (1160) of authenticating the scarcity of the digital asset includes further using the determined third measures of statistical dependence.

[0105] In a further embodiment thereof, method 1100 further includes the steps of: ranking N digital fingerprints based on a first measure of their respective corresponding statistical dependence (such as cosine similarity); calculating the difference / gain (such as a difference) of the first measure of statistical dependence between consecutively ranked digital fingerprints; calculating the difference / gain of a second measure of statistical dependence between consecutively ranked digital fingerprints; and calculating the difference / gain of a third measure of statistical dependence between consecutively ranked digital fingerprints. In addition, the step of authenticating the scarcity of the digital asset (1160) further includes using the calculated differences / gains of the first, second, and third measures of statistical dependence.

[0106] In a further embodiment thereof, the step (1160) of authenticating the scarcity of a digital asset includes running a machine learning model that uses logistic regression to process, for each of N digital fingerprints, a measure of statistical dependence measured for that digital fingerprint and the difference / gain, and to produce a single output representing the probability that the corresponding visual appearance of the digital fingerprint is a copy of a first visual appearance, where the six inputs include first, second, and third measures of statistical dependence of the digital fingerprint and the difference of the first, second, and third measures of statistical dependence of the digital fingerprint.

[0107] In further embodiments, the logistic regression model is trained (by machine learning module 160, for example) with tens of thousands of visual appearances and their corresponding digital fingerprints, as well as combinations of six inputs and a single output, and with tens of thousands of complex transformations of visual appearances and their corresponding digital fingerprints, as well as combinations of six inputs and a single output.

[0108] In a further embodiment of this, the logistic regression model is further trained (by machine learning module 160, for example) using the area under the precision-recall curve (AUPRC) as a loss function to measure how well the model is performing.

[0109] In a further embodiment of the logistic regression model, method 1100 further includes the steps of determining the second of N digital fingerprints which has the highest probability of being a copy of the first digital fingerprint among the probabilities output by the logistic regression model, and determining a scarcity score of the digital asset by subtracting the highest output probability from 1, where the step of authenticating the scarcity of the digital asset (1160) further includes using the determined scarcity score.

[0110] Any of the methods described herein may be compiled into a non-temporary computer-readable medium (CRM) that stores computer instructions causing the processing circuit to execute an automated process for performing each method when executed by the processing circuit in the corresponding embodiment.

[0111] The methods described herein may be executed in whole or in part by machine-readable software or firmware located on a tangible (e.g., non-temporary) storage medium. For example, the software or firmware may be in the form of a computer program containing computer program code adapted to execute any part of the methods described herein when the program is executed on a computer or a suitable hardware device (e.g., an FPGA), and the computer program may be embodied on a computer-readable medium. Examples of tangible storage mediums include computer storage devices having computer-readable media such as disks, thumb drives, and flash memory, but do not include propagating signals. Propagating signals may reside in a tangible storage medium, but propagating signals themselves are not examples of tangible storage mediums. The software may be suitable for execution on parallel or serial processors so that the steps of the method can be executed in any suitable order or simultaneously.

[0112] Returning to Figure 1B, which is a conceptual block diagram showing an unrestricted embodiment of a processor 100 that may be included in the server 10 of the supernode 5 (shown in Figure 1A). As seen in Figure 1B, the processor 100 may include a computer processor 110 such as a central processing unit (CPU), read-only memory (ROM) 115, random access memory (RAM) 120, a disk drive (DD) 125, a network interface 130, an input / output (I / O) interface 135, and a database (DB) 170. Although the database 170 is shown as part of the processor in Figure 1B, the database 170 is not limited in that way and may include one or more databases located outside the processor 100, as shown in Figure 1A. Various components within the processor 110 may be connected to the bus 105 via one or more communication links. The processor 100 can receive or otherwise access NFT data for processing, for example, via the network interface 130, I / O interface 135, DD125, or DB170.

[0113] The system bus 105 may be any of several bus structures that can be further interconnected with a memory bus (which may or may not have a memory controller), a peripheral bus, and a local bus using any of the various commercially available bus architectures.

[0114] CPU110 may be any of the various commercially available processors. Dual microprocessors and other multiprocessor architectures may also be used as CPU110. CPU110 may also be a graphics processing unit (GPU).

[0115] The processor 100 includes a computer-readable medium capable of holding executable or interpretable computer code (or instructions) that, when executed by the CPU 110, causes the described steps, processor, and method to be executed. The computer-readable medium may be provided as a ROM 115, RAM 120, DD 125, DB 170, or as an external computer-readable medium connected to the processor 100 via a network interface 130 or I / O interface 135. The computer-readable medium may include functional modules, for example, portions of computer code that, when executed by the CPU 110, cause the workflow steps and all other process steps described or intended in the description to be executed.

[0116] The Basic Input / Output System (BIOS) may be stored in non-volatile memory within the processor 100, such as ROM 115. ROM 115 may include ROM, erasable programmable read-only memory (EPROM), or electrically erasable programmable read-only memory (EEPROM). The BIOS may include basic routines to help transfer information between components within the processor 100, such as during startup. RAM 120 may include high-speed RAM, such as static RAM, for caching data.

[0117] The disk drive (DD) 125 may include a hard drive, such as an EIDE (enhanced integrated drive electronics) drive or a SATA (serial advanced technology attachments) drive. The DD125 may include an optical disc drive capable of reading and writing to and from a compact disc read-only memory (CD-ROM) disc (not shown), or reading or writing to other high-capacity optical media such as a digital video disc (DVD). The DD125 may be configured for external use in a suitable enclosure (not shown). The DD125 may be connected to the system bus 105 by a hard disk drive interface (not shown) and an optical drive interface (not shown), respectively. The hard disk drive interface (not shown) may include a Universal Serial Bus (USB) (not shown) or an IEEE 1394 interface (not shown) for external applications.

[0118] The DD125 and associated computer-readable media can provide non-volatile storage of data, data structures, or computer executable instructions. The DD125 can accommodate the storage of any data in a suitable digital format. The DD125 may include one or more applications used to implement embodiments of the architecture described herein.

[0119] Several program modules, including an operating system (not shown), one or more application programs (not shown), other program modules (not shown), and program data (not shown), may be stored in DD125, ROM115, or RAM120. Any (or all) of the operating system, application programs, program modules, and program data may be cached in RAM120 as executable portions of computer code.

[0120] The network interface 130 can communicately connect the processor to a communication network. The network interface 130 may include a wired or wireless communication network interface (not shown) or a modem (not shown). When used in a data communication network, the processor 100 may be connected to a LAN network via a wired or wireless communication network interface, and when used in a wide area network (WAN), the processor 100 may be connected to a WAN network via a modem. The modem (not shown) may be internal or external and wired or wireless. The modem may be connected to the system bus 105 via, for example, a serial port interface (not shown).

[0121] The I / O interface 135 can receive commands and data from an operator via a user interface device (not shown), such as a keyboard (not shown), mouse (not shown), pointer (not shown), microphone (not shown), speaker (not shown), or display (not shown). The received commands and data can be transferred from the I / O interface 135 to the CPU 110 as instruction and data signals via the bus 105.

[0122] The image processor 145 may include an image processing device, a graphics driver (not shown), a video adapter (not shown), or any other device necessary for processing, generating, and drawing image signals onto a display device, such as a display device of the I / O interface 135 or a display device in a communication device connected thereto.

[0123] As further shown in Figure 1B and as described above, the processor 100 may include a fingerprinting engine 150, a relative scarcity engine 155, and a machine learning module 160. It should be understood that the fingerprinting engine, relative scarcity engine, and machine learning module shown in Figure 1B, as well as one or more of other such hardware and / or software-based components, can be integrated with the CPU 110 or provided separately as seen in Figure 1B. Furthermore, such modules may be stored in one of the computer-readable storage media and may take the form of instructions executable by the CPU. It should be further understood that one or more of the hardware and / or software-based components shown in Figure 1B can be integrated with the processor or provided separately / externally from the processor.

[0124] As used in this disclosure, the terms “a,” “an,” and “the” mean “one or more” unless expressly specified otherwise.

[0125] As used in this disclosure, the term "backbone" means a transmission medium that interconnects one or more computing devices or communication devices and provides a path for transmitting data signals and command signals between one or more computing devices or communication devices. A backbone may include a bus or a network. A backbone may include Ethernet TCP / IP. A backbone may include a distributed backbone, a collapsed backbone, a parallel backbone, or a serial backbone.

[0126] As used in this disclosure, the term "bus" means any of several bus structures that can be further interconnected to a memory bus (which may or may not have a memory controller), a peripheral bus, or a local bus using any of the various commercially available bus architectures. The term "bus" may include a backbone.

[0127] As used in this disclosure, the terms “communicating device” and “communication device” mean any hardware, firmware, or software capable of transmitting or receiving data packets, command signals, data signals, or radio frequency signals over a communication link. A device may include a computer or a server. A device may be portable or stationary.

[0128] As used in this disclosure, the term “communication link” means a wired or wireless medium that transmits data or information between at least two points. Wired or wireless mediums may include, for example, metal conductor links, radio frequency (RF) communication links, infrared (IR) communication links, or optical communication links. RF communication links may include, for example, Wi-Fi, WiMAX, IEEE 802.11, DECT, 0G, 1G, 2G, 3G, or 4G cellular standards, or Bluetooth.

[0129] As used in this disclosure, the terms “computer,” “computing device,” or “processor” mean any machine, device, circuit, component, or module, or any system of machines, devices, circuits, components, or modules, that are capable of manipulating data according to one or more instructions. The terms “computer,” “computing device,” or “processor” can include, for example, but are not limited to, communication devices, computing resources, processors, microprocessors (μPs), microcontrollers (μCs), central processing units (CPUs), graphics processing units (GPUs), application-specific integrated circuits (ASICs), general-purpose computers, supercomputers, personal computers, laptop computers, palmtop computers, note computers, desktop computers, workstation computers, servers, server farms, computer clouds, or arrays or systems of processors, μCs, CPUs, GPUs, ASICs, general-purpose computers, supercomputers, personal computers, laptop computers, palmtop computers, note computers, desktop computers, workstation computers, or servers.

[0130] As used in this disclosure, the terms “computing resource” or “computer resource” mean software, software applications, web applications, web pages, computer applications, computer programs, computer code, machine-executable instructions, firmware, or processes that can be configured to run as one or more processes on a computing device.

[0131] As used in this disclosure, “computer-readable medium” means any storage medium relating to providing data (e.g., instructions) that can be read by a computer. Such mediums can take many forms, including non-volatile and volatile media. Non-volatile media may include, for example, optical or magnetic disks and other persistent memory. Volatile media may include dynamic random-access memory (DRAM). Common forms of computer-readable media include, for example, floppy disks, flexible disks, hard disks, magnetic tapes, any other magnetic media, CD-ROMs, DVDs, any other optical media, punch cards, paper tapes, any other physical media with perforation patterns, RAM, PROMs, EPROMs, FLASH-EEPROMs, any other memory chips or cartridges, carriers, or any other media that can be read by a computer. Computer-readable media may include “clouds,” which include a large number (e.g., thousands) of memory caches or the distribution of files across a large number (e.g., thousands) of computers.

[0132] Various computer-readable media may be involved in transporting sequences of instructions to a computer. For example, a sequence of instructions may be (i) transmitted from RAM to a processor, (ii) transported via a wireless transmission medium, or (iii) formatted according to a number of formats, standards, or protocols, including, for example, Wi-Fi, WiMAX, IEEE 802.11, DECT, 0G, 1G, 2G, 3G, 4G, or 5G cellular standards, or Bluetooth.

[0133] As used in this disclosure, the term “database” means any combination of software or hardware, including at least one application or at least one computer. A database may include a structured collection of records or data organized according to a database model, for example, at least one of the relational model, a hierarchical model, or a network model, for example, but not limited to these. A database may include a database management system application (DBMS) as known in the art. At least one application may include, for example, an application program that can accept connections to service requests from clients by sending responses back to the client. A database may be configured to run at least one application unattended and with minimal human intervention over long periods of time, often under heavy workloads.

[0134] As used in this disclosure, the terms “including,” “equipped with,” and variations thereof mean “including, but not limited to,” unless expressly specified otherwise.

[0135] As used in this disclosure, the term “Network” means, but is not limited to, a personal area network (PAN), local area network (LAN), wireless local area network (WLAN), campus area network (CAN), metropolitan area network (MAN), wide area network (WAN), global area network (GAN), broadband area network (BAN), cellular network, storage area network (SAN), system area network, passive optical local area network (POLAN), enterprise private network (EPN), virtual private network (VPN), the Internet, or any combination thereof, and any of these can be configured to transmit data over wireless and / or wired communication media. These networks can implement a variety of protocols, but are not limited to, Ethernet, IP, IPX, TCP, UDP, SPX, IP, IRC, HTTP, FTP, Telnet, SMTP, DNS, ARP, and ICMP.

[0136] As used in this disclosure, the term "server" means any combination of software or hardware, including at least one application or at least one computer that performs services to connected clients as part of a client-server architecture. The at least one server application may include, but is not limited to, an application program that can accept connections to service requests from clients by sending responses back to the clients. A server may be configured to run at least one application unattended and with minimal human intervention for extended periods, often under heavy workloads. A server may include multiple configured computers, with at least one application being distributed among them depending on the workload. For example, under light loads, at least one application may run on a single computer; however, under heavy loads, multiple computers may be required to run at least one application. The server or either of its computers may also be used as a workstation.

[0137] As used in this disclosure, the terms “send,” “sent,” “transmit,” or “transmit” mean the carrying of data, data packets, computer instructions, or any other digital or analog information via electrical, acoustic, light waves, or other electromagnetic radiation, such as those generated in communications in the radio frequency (RF) or infrared (IR) spectrum. The transmission medium for such transmissions may include coaxial cables, copper wires, and optical fibers, including wires with a system bus connected to a processor.

[0138] Devices communicating with each other do not need to communicate with each other continuously unless explicitly specified otherwise. In addition, devices communicating with each other may communicate directly or indirectly through one or more intermediaries.

[0139] Process steps, method steps, or algorithms may be described in a sequential or parallel order, but such processes, methods, and algorithms may be configured to function in an alternative order. In other words, any order or sequence of steps that may be described in a sequential order does not necessarily mean that the steps must be performed in that order, and some steps may be performed concurrently. Similarly, if the order or sequence of steps is described in a balanced (or concurrent) order, such steps may be performed in a sequential order. The steps of the processes, methods, or algorithms described herein may be performed in any order that is practical.

[0140] Where a single device or article is described, it will be readily apparent that two or more devices or articles may be used in place of that single device or article. Similarly, where two or more devices or articles are described, it will be readily apparent that a single device or article may be used in place of those two or more devices or articles. The function or features of a device may be embodied by one or more other devices that are not explicitly stated to have such a function or features.

[0141] The subject matter described above is provided for illustrative purposes only and should not be construed as restrictive. Various modifications or changes can be made to the subject matter described herein without following the illustrated and described exemplary embodiments and applications, and without departing from the true spirit and scope of the invention as encompassed by this disclosure. The spirit and scope of the invention are defined by the following set of claims and by equivalent structures and functions or steps.

Claims

1. An automated method for verifying the scarcity of digital assets, A step of generating a first digital fingerprint of a digital asset by running one or more deep learning models on the first visual appearance of the digital asset, wherein the generated first digital fingerprint is a first vector representing the first visual appearance, and the one or more deep learning models are trained to generate corresponding first vectors suitable for processing the visual appearance of the digital asset and determining the similarity of the visual appearances, A step of evaluating the similarity of the first digital fingerprint to a registry of registered digital fingerprints by calculating the corresponding dot product of the first digital fingerprint and the registered digital fingerprint, wherein the registry of registered digital fingerprints is obtained by running the one or more deep learning models on the corresponding visual appearance of the registered digital assets, The steps include determining a corresponding first measure of the statistical dependence of the registered digital fingerprint by normalizing the calculated dot product to a range having a first end corresponding to no similarity with the first visual appearance and a second end corresponding to identical similarity with the first visual appearance, A step of selecting a fixed number N of registered digital fingerprints whose corresponding N normalized dot product is closest to the second end of the normalized range, The steps include determining N corresponding second measures of statistical dependence by applying a population measurement test for deviation from independence to N combinations of the first digital fingerprint and the selected N digital fingerprints, A step of authenticating the scarcity of the digital asset using the first and second measures of statistical dependence determined above. The method, including the method described above.

2. The step of running the one or more deep learning models on the first visual appearance is: A step of running a first deep learning model on the first visual appearance to generate a second vector containing a plurality of values ​​characterizing the first visual appearance, wherein the first deep learning model is trained on the plurality of visual appearances to generate a corresponding second vector characterizing each of the plurality of visual appearances, and the second vector contains a plurality of values ​​characterizing the first visual appearance. A step of generating a third vector optimized for determining visual appearance similarity, wherein the second deep learning model is run on the generated second vector to generate a third vector optimized for determining visual appearance similarity, the second deep learning model is trained on a pair of second vectors, each pair of second vectors is generated by running the first deep learning model on a corresponding original visual appearance and a transformed visual appearance obtained by applying one of a plurality of transformations to the original visual appearance, and the second deep learning model brings the generated pairs of second vectors closer to each other in an embedding space in order to generate a corresponding third vector optimized for determining visual appearance similarity, The method according to claim 1, comprising the step of compressing the generated third vector into the first vector by removing redundant information in order to generate the first digital fingerprint.

3. The method according to claim 2, wherein the first deep learning model is a convolutional neural network.

4. The method according to claim 2, wherein the step of applying one of the transformations to each original visual appearance and the corresponding transformed visual appearance includes performing one of the following: trimming the original visual appearance, inverting the original visual appearance, rotating the original visual appearance, adding one or more rectangular overlays to the original visual appearance, and removing one or more portions from the original visual appearance.

5. The method according to claim 2, wherein removing the aforementioned redundant information includes using principal component analysis (PCA).

6. The method according to claim 1, wherein the population measurement test for deviation from independence is selected from the group consisting of Hefding's Dependence Scale (Hefding's D) and the Hilbert-Schmidt Independence Criterion (HSIC).

7. The step of determining the corresponding N second measures of statistical dependence includes the step of applying a first population measurement test for deviation from independence to the N digital fingerprint combinations, The method further includes the step of determining a corresponding N third measure of statistical dependence by applying a second population measurement test for deviation from independence to the N digital fingerprint combinations, The method according to claim 1, wherein the step of authenticating the scarcity of the digital asset is performed using a third measure of the determined statistical dependency.

8. The steps include ranking the N digital fingerprints based on a first measure of their corresponding statistical dependence, A step of calculating the difference of the first measure of statistical dependence between consecutively ranked digital fingerprints, A step of calculating the difference of the second measure of the statistical dependence between the sequentially ranked digital fingerprints, A step of calculating the difference of the third measure of the statistical dependence between the sequentially ranked digital fingerprints. It further includes, The method according to claim 7, wherein the step of authenticating the scarcity of the digital asset is performed using the calculated differences of the first, second, and third measures of statistical dependence.

9. The method according to claim 8, wherein the step of authenticating the scarcity of the digital asset includes running a machine learning model that uses logistic regression to process a plurality of inputs of the digital fingerprint for each of the N digital fingerprints and generates a single output representing the probability that the corresponding visual appearance of the digital fingerprint is a copy of the first visual appearance, wherein the plurality of inputs include the first, second, and third measures of the statistical dependence of the digital fingerprint and the difference between the first, second, and third measures of the statistical dependence of the digital fingerprint.

10. The method according to claim 9, wherein the logistic regression model is trained on tens of thousands of visual appearances and their corresponding digital fingerprints and combinations of six inputs and a single output, and tens of thousands of complex transformations of the visual appearances and their corresponding digital fingerprints and combinations of six inputs and a single output.

11. The method according to claim 10, wherein the logistic regression model is further trained using a model performance index for a binary response as a loss function to measure how well the logistic model is performing.

12. The step of determining the second fingerprint among the N digital fingerprints which has the highest probability of being a copy of the first digital fingerprint among the probabilities output by the logistic regression model, The steps include:

1. Determining the scarcity score of the digital asset by subtracting the highest output probability from the above-mentioned probability; It further includes, The method according to claim 9, wherein the step of authenticating the scarcity of the digital asset is based on the scarcity score.

13. An automated system for authenticating the scarcity of digital assets, Processing circuit and A non-temporary storage medium that stores a registry of registered digital fingerprints, a machine learning model, and instructions. Equipped with, When the instruction is executed by the processing circuit, Generating a first digital fingerprint of a digital asset by running one or more deep learning models of the stored machine learning models on the first visual appearance of the digital asset, wherein the generated first digital fingerprint is a first vector representing the first visual appearance, and the deep learning models are trained to generate corresponding first vectors suitable for processing the visual appearance of the digital asset and determining the similarity of the visual appearances. The similarity of the first digital fingerprint to the registry of registered digital fingerprints is evaluated by calculating the corresponding dot product of the first digital fingerprint and the registered digital fingerprint, wherein the registry of registered digital fingerprints is obtained by running the deep learning model on the corresponding visual appearance of the registered digital assets. The calculation of the dot product is normalized to a range having a first end corresponding to no similarity to the first visual appearance and a second end corresponding to identical similarity to the first visual appearance, thereby determining the corresponding first measure of the statistical dependence of the registered digital fingerprint. Selecting a fixed number N of registered digital fingerprints whose corresponding N normalized dot product is closest to the second end of the normalized range, To determine the corresponding N second measures of statistical dependence by applying a population measurement test for deviation from independence to the N combinations of the first digital fingerprint and the N selected digital fingerprints, The scarcity of the digital asset is authenticated using the first and second measures of statistical dependence determined above. The system comprising the processing circuit configured to perform the above.

14. When the aforementioned instruction is executed by the processing circuit, The method involves running a first deep learning model on the first visual appearance to generate a second vector of values ​​characterizing the first visual appearance, wherein the first deep learning model is trained on the plurality of visual appearances to generate corresponding second vectors characterizing the plurality of visual appearances, The method involves running a second deep learning model on the generated second vectors to generate a third vector of similar size to the second vectors and optimized for determining visual appearance similarity, wherein the second deep learning model is trained on pairs of second vectors, each pair of second vectors is generated by running the first deep learning model on a corresponding original visual appearance and a transformed visual appearance obtained by applying one of a plurality of transformations to the original visual appearance, and the second deep learning model generates the optimized third vector by bringing the generated pairs of second vectors closer together in the embedding space in order to generate a corresponding third vector optimized for determining visual appearance similarity. In order to generate the first digital fingerprint, the generated third vector is compressed into the first vector by removing redundant information, The system according to claim 13, wherein the processing circuit is further instructed to run the deep learning model on the first visual appearance.

15. The system according to claim 14, wherein applying one of the transformations to each original visual appearance and the corresponding transformed visual appearance includes performing one of the following: trimming the original visual appearance, inverting the original visual appearance, rotating the original visual appearance, adding one or more rectangular overlays to the original visual appearance, and removing one or more portions from the original visual appearance.

16. Determining the statistical dependence of the corresponding N second measures includes applying a first population measurement test for deviation from independence to the N digital fingerprint combinations, When the instruction is executed by the processing circuit, the processing circuit is further instructed to determine the corresponding N third measures of statistical dependence by applying a second population measurement test for deviation from independence to the N digital fingerprint combinations. The system according to claim 13, wherein the scarcity of the digital asset is authenticated using a third measure of the determined statistical dependency.

17. When the instruction is executed by the processing circuit, The N digital fingerprints are ranked based on a first measure of their corresponding statistical dependence. Calculating the difference of the first measure of statistical dependence between consecutively ranked digital fingerprints, Calculating the difference of the second measure of the statistical dependence between the sequentially ranked digital fingerprints, Calculating the difference of the third measure of the statistical dependence between the sequentially ranked digital fingerprints, The processing circuit is then further instructed to perform the following: The system according to claim 16, wherein the scarcity of the digital asset is authenticated using the calculated difference of the first, second, and third measures of statistical dependence.

18. The system according to claim 17, wherein authenticating the rarity of the digital asset involves running another machine learning model among the stored machine learning models, the other machine learning model using logistic regression to process a plurality of inputs of the digital fingerprint for each of the N digital fingerprints, and to produce a single output representing the probability that the corresponding visual appearance of the digital fingerprint is a copy of the first visual appearance, the plurality of inputs comprising the first, second, and third measures of the statistical dependence of the digital fingerprint and the difference between the first, second, and third measures of the statistical dependence of the digital fingerprint.

19. The system according to claim 18, wherein the logistic regression model is trained on tens of thousands of visual appearances and their corresponding digital fingerprints and combinations of six inputs and a single output, and tens of thousands of complex transformations of the visual appearances and their corresponding digital fingerprints and combinations of six inputs and a single output.

20. When the aforementioned instruction is executed by the processing circuit, the processing circuit further: To determine the second fingerprint among the N digital fingerprints that has the highest probability of being a copy of the first digital fingerprint among the probabilities output by the logistic regression model, The scarcity score of the digital asset is determined by subtracting the highest probability output from 1. Have them do it, The system according to claim 18, wherein the scarcity of the digital asset is authenticated using the scarcity score.

21. A non-temporary computer-readable medium (CRM) storing computer instructions, wherein when the computer instructions are executed by a processing circuit, the processing circuit executes an automated process for authenticating the rarity of digital assets, and the process A step of generating a first digital fingerprint of a digital asset by running one or more deep learning models on the first visual appearance of the digital asset, wherein the generated first digital fingerprint is a first vector of values ​​representing the first visual appearance, and the one or more deep learning models are trained to generate corresponding first vectors suitable for processing the visual appearance of the digital asset and determining the similarity of the visual appearances, A step of evaluating the similarity of the first digital fingerprint to a registry of registered digital fingerprints by calculating the corresponding dot product of the first digital fingerprint and the registered digital fingerprint, wherein the registry of registered digital fingerprints is obtained by running the deep learning model on the corresponding visual appearance of the registered digital asset, The steps include determining a corresponding first measure of the statistical dependence of the registered digital fingerprint by normalizing the calculated dot product to a range having a first end corresponding to no similarity with the first visual appearance and a second end corresponding to identical similarity with the first visual appearance, A step of selecting a fixed number N of registered digital fingerprints whose corresponding N normalized dot product is closest to the second end of the normalized range, The steps include determining N corresponding second measures of statistical dependence by applying a population measurement test for deviation from independence to N combinations of the first digital fingerprint and the selected N digital fingerprints, A step of authenticating the scarcity of the digital asset using the first and second measures of statistical dependence determined above. The non-temporary computer-readable medium (CRM), including the said non-temporary computer-readable medium.

22. The step of running the deep learning model on the first visual appearance is, A step of running a first deep learning model on the first visual appearance to generate a second vector containing a plurality of values ​​characterizing the first visual appearance, wherein the first deep learning model is trained on the plurality of visual appearances to generate corresponding second vectors characterizing the plurality of visual appearances, and the second vector containing a plurality of values ​​characterizing the first visual appearance is generated. A step of generating a third vector optimized for determining visual appearance similarity, wherein the second deep learning model is run on the generated second vector to generate a third vector optimized for determining visual appearance similarity, the second deep learning model is trained on a second pair of vectors, each pair of second vectors is generated by running the first deep learning model on a corresponding original visual appearance and a transformed visual appearance obtained by applying one of a plurality of transformations to the original visual appearance, and the second deep learning model brings the generated pairs of second vectors closer to each other in embedding space in order to generate a corresponding third vector optimized for determining visual appearance similarity, The CRM according to claim 21, comprising the step of compressing the generated third vector into the first vector by removing redundant information in order to generate the first digital fingerprint.

23. The CRM according to claim 22, wherein the step of applying one of the transformations to each original visual appearance and the corresponding transformed visual appearance includes performing one of the following: cropping the original visual appearance, inverting the original visual appearance, rotating the original visual appearance, adding one or more rectangular overlays to the original visual appearance, and removing one or more portions from the original visual appearance.

24. The step of determining the corresponding N second measures of statistical dependence includes the step of applying a first population measurement test for deviation from independence to the N digital fingerprint combinations, The process further includes the step of determining a corresponding N third measure of statistical dependence by applying a second population measurement test for deviation from independence to the N digital fingerprint combinations, The CRM according to claim 21, wherein the step of authenticating the scarcity of the digital asset further includes using a third measure of the determined statistical dependency.

25. The aforementioned process, The steps include ranking the N digital fingerprints based on a first measure of their corresponding statistical dependence, A step of calculating the difference of the first measure of statistical dependence between consecutively ranked digital fingerprints, A step of calculating the difference of the second measure of the statistical dependence between the sequentially ranked digital fingerprints, A step of calculating the difference of the third measure of the statistical dependence between the sequentially ranked digital fingerprints. It further includes, The CRM according to claim 24, wherein the step of authenticating the scarcity of the digital asset further includes using the calculated differences of the first, second, and third measures of statistical dependence.

26. The CRM according to claim 25, wherein the step of authenticating the scarcity of the digital asset includes running a machine learning model that uses logistic regression to process a plurality of inputs of the digital fingerprint for each of the N digital fingerprints and generates a single output representing the probability that the corresponding visual appearance of the digital fingerprint is a copy of the first visual appearance, wherein the plurality of inputs include the first, second, and third measures of the statistical dependence of the digital fingerprint and the difference between the first, second, and third measures of the statistical dependence of the digital fingerprint.

27. The CRM according to claim 26, wherein the logistic regression model is trained on tens of thousands of visual appearances and their corresponding digital fingerprints and combinations of six inputs and a single output, and tens of thousands of complex transformations of the visual appearances and their corresponding digital fingerprints and combinations of six inputs and a single output.

28. The aforementioned process, The step of determining the second fingerprint among the N digital fingerprints which has the highest probability of being a copy of the first digital fingerprint among the probabilities output by the logistic regression model, The steps include:

1. Determining the scarcity score of the digital asset by subtracting the highest output probability from the above-mentioned probability; It further includes, The CRM according to claim 26, wherein the step of authenticating the scarcity of the digital asset further includes using the scarcity score.

Images