System and method for predicting authentication detectability of counterfeit

Abstract

The present invention relates to a computer-implemented method for identifying whether an item is genuine or counterfeit, wherein the method uses a combination of outputs from a predictive machine learning model and from an authenticity detection algorithm. In particular, the present invention relates to a method for identifying counterfeit.

Description

Technical Field

[0001] The present invention relates to machine learning systems, methods, and processes in the field of anti-counterfeiting and authentication of manufactured articles, products, security documents, and banknotes. Background Art

[0002] To determine whether a product item is genuine, various anti-counterfeiting techniques can be used (Anti-Counterfeiting Technology Guide, European Union Intellectual Property Office, 2021). Generally, these techniques add elements that are difficult to replicate or copy onto the item, or they characterize specific physical or chemical features of the item, similar to a fingerprint of the item. The challenges can be technical, such as the replication of holograms, or require products that are difficult to obtain on the market, such as rare isotopes or special inks. Generally, anti-counterfeiting features can be classified into overt technologies (visible, or more generally perceivable by the end user with their own bodily senses without a specific detection device) or covert technologies (invisible / imperceptible, but detectable with a dedicated device). Examples of covert technologies include:

[0003] · Product identification: Technologies such as digital watermarks have been designed to better prevent the forgery of product packaging and security documents by electronic and digital means. As an example of the widespread deployment of such technologies, AlpVision Cryptoglyph exists in two styles, either as a random or pseudo-random pattern of microdots printed with visible ink (WO0225599, WO04028140), or as a cloud of micropores in a varnish layer (WO06087351). The distribution of the microdots or micropores can be controlled using a secret cryptographic key. Authentication can be performed using a conventional imaging device, such as a smartphone or an off-the-shelf office scanner, in combination with dedicated signal processing software. Product identification includes digital identification, chemical identification, holograms, etc.

[0004] · Surface fingerprint: These technologies do not add any security elements to the product, but instead use existing inherent microscopic surface features. For example, the matte surface of an injection-molded product is an ideal candidate for a fingerprint solution. An image of the surface can be acquired during the production process and then compared to a subsequent image captured from the product under inspection. Authentication can be performed using a conventional imaging device, such as a smartphone or an off-the-shelf office scanner, in combination with dedicated signal processing software.

[0005] An example of fingerprint technology described, for example, in US10332247.

[0006] · Invisible ink: These special inks are invisible but become visible when illuminated by an appropriate light source.

[0007] A well-known example is UV (Ultra-Violet) or IR (Infra-Red) ink. A dedicated imaging device integrated with a dedicated lighting device can be used, in combination with image analysis methods, to perform authentication to characterize the presence of special ink on the item to be authenticated (e.g., EP1295263, US6903342).

[0008] · Chemical tagging agents: There are a large number of chemical tagging agents. Generally, these tagging agents are invisible and detectable under laboratory conditions. Digital olfactory sensors and dedicated signal analysis methods can be used to perform authentication to characterize the presence of chemical tagging agents on the item to be authenticated (e.g., WO2014201099, WO2020160377).

[0009] · Micrographic patterns: Artworks can be protected by adding very small graphic elements that are invisible without visual magnification. Microtext is one of the most popular implementations of this security feature (e.g., WO2012131474).

[0010] For any of the above technologies, authenticating an item involves using a detector suitable for the specific authentication technology employed by the item to identify whether the authentication technology can be retrieved by examining the object. In the past two decades, the emergence of digital detection technologies has facilitated the automation of this process and popularized it to non-experts and even potentially the general public, thanks to the use of digital signal processing algorithms that are embedded in software applications, or in detection equipment (such as smartphones), or executed on a computer communicating with the detection equipment via a communication network.

[0011] These digital authentication detection methods generally follow the following steps:

[0012] 1) Capture a digital signal representation of the item to be authenticated (e.g., an image of the item's surface captured by a camera sensor, an RFID signal read by an RFID sensor, a chemical signal from a digital olfactory sensor, etc.) through a sensor;

[0013] 2) Process the digital signal representation with a signal processing algorithm implemented by a computer to characterize the authenticity of the item - for example, by measuring the difference, distance, or signal-to-noise ratio (SNR) between the captured digital signal representation and a template digital signal representation of a reference genuine item; or by extracting mathematical features from the digital signal representation, which are used by a classifier to distinguish between forged and genuine items.

[0014] 3) Depending on the measurement results (e.g., using a predetermined threshold or using a machine learning classifier), classify the item as genuine or counterfeit (using a computer-implemented authentication classification or decision algorithm).

[0015] In the past decade, many authentication technologies have increasingly used consumer electronic devices (such as smartphone cameras) instead of dedicated sensing equipment (such as flatbed scanners and professional cameras with dedicated optics). The diversity of digital signal representations of the capture conditions poses specific challenges that affect the detection of signals in authentication technologies, which in turn affects the reliability of autonomously distinguishing counterfeits from genuine items. Therefore, many authentication algorithms have been adjusted to ensure that genuine items are detected as true positives, but it remains challenging to rely on the expected digital signal representations that fail to detect genuine items to confirm that an item is counterfeit. For example, as a machine learning-based method, WO2015157526 describes the use of a convolutional neural network to classify genuine and counterfeit items. The latter method uses a training set that includes genuine and forged items, combined with data augmentation to facilitate training. The latter method requires the brand owner to collect multiple forged samples that are sufficient to represent the forger's ability to replicate the original product. This places an additional burden on the organization and long-term operation of anti-counterfeiting operations. There remains a risk of misclassifying genuine items as forged items (false negative classification), or more generally, there are too many doubtful cases.

[0016] In an ideal scenario, a digital authentication detection method applied to a perfect digital signal representation of a genuine item would always be able to detect it as genuine. In other words, by applying a digital authentication detection method to a perfect digital signal representation of a genuine item, the authenticity would be "100% detectable" or "always detectable".

[0017] However, under real-world authentication conditions, the detectability of the authenticity of an item depends on the quality of the digital signal representation of the item. This quality itself depends on multiple variable digital signal capture factors; for example, in the case of imaging capture (but not limited to):

[0018] · How the image of the surface of the genuine item is captured:

[0019] - Depending on the position and orientation of the item relative to the sensor position and orientation - especially when using a handheld device camera to capture images instead of a flatbed scanner;

[0020] - Depending on the imaging sensor, such as camera resolution, focal length, aperture, field of view, white balance;

[0021] - Depending on the lighting environment, where shadows, reflections, etc. vary with the surface condition of the item (even possibly second-hand items, such as luxury watches in the second-hand market) and the physical environment of the shot (camera flash on or off, indoor or outdoor ambient lighting, etc.).

[0022] · The aging condition of the item or the way it has been handled during its life, e.g., dirt on the surface, scratches, discoloration due to UV exposure, deformation due to vibration or humid environment, oxidation, etc.

[0023] Historically, most digital detection methods in the past decade have required a large amount of scientific research and experimental development (R&D) as well as testing to adjust the algorithms to process and classify as diverse as possible digital signal representations of the item to be authenticated. Many solutions currently deployed in the authentication market provide users with guidance on optimizing the capture conditions: e.g., visual guidance for correctly orienting smartphone capture and / or capturing video as a series of multiple digital representations and / or capturing repeatedly from different positions until a genuine item can be detected.

[0024] Currently, it is not possible to achieve a perfect detection paradigm in the real world. Therefore, it cannot be determined whether the failure of the authentication technology detection of an item is caused by the item being a forgery or by detection limitations due to the capture conditions of the item, i.e., the item may still be genuine. For example, end users using the Veriscan fingerprint smartphone application technology of Swiss precious metals supplier PAMP to verify the authenticity of PAMP brand cast gold bars can only obtain a "pass" (= detected as genuine) or "no available result" (= cannot be detected as genuine, which may mean it is genuine but cannot be detected from the smartphone image, or cannot be detected because the gold bar is a counterfeit) (https: / / www.pamp.com / veriscan / ).

[0025] Therefore, improved methods and systems are needed to predict the detectability of counterfeits and reliably classify items as counterfeits when using any digital authentication algorithm of the prior art.

[0026] The object of the present invention is to provide improved methods and systems to predict the detectability of counterfeits. Summary of the Invention

[0027] The present invention is based on the discovery that using a predictive machine learning model to predict the authenticity detectability value of an item and combining it with an authenticity detection algorithm (also referred to herein as an authentication algorithm) allows for the identification of genuine and counterfeit items. Additionally, the present invention is based on the development of a specific training protocol for obtaining a predictive machine learning model, wherein the training data consists of a set of digital signal representations of genuine items and their associated authenticity detectability values of the items. The predictive machine learning model allows for the prediction of the authenticity detectability value of an item to be identified, and the authenticity detection algorithm allows for determining whether the authenticity detection algorithm can detect the item as genuine based on the predicted authenticity detectability value of the item. Thus, the predicted detectability value determines whether the digital signal representation of the item to be identified is sufficient for the authenticity detection algorithm to identify the item as genuine. The predicted detectability value is the result of the environment in which the digital signal representation of the item is obtained (captured) by a sensor and can also be defined as the detectability value of the authenticity measurement of the item. In the next step, if the item to be identified can be detected as genuine, then if the authenticity detection algorithm does not identify the item as genuine, it can be determined that the item is counterfeit.

[0028] In one embodiment, a computer-implemented method for training a predictive machine learning model to predict the authenticity detectability value of an item is provided, the method comprising the steps of:

[0029] a) obtaining an authenticity detection algorithm that generates an authenticity detectability value of an item from a digital signal representation of the item;

[0030] wherein the detectability value determines whether the digital signal representation of the item is sufficient to identify the item as genuine;

[0031] b) obtaining one or more genuine items of a type;

[0032] c) obtaining, by a sensor, a set consisting of digital signal representations of each genuine item;

[0033] wherein each digital signal representation of each genuine item is obtained by the operation of the sensor under different capture conditions;

[0034] d) inputting each digital signal representation of each genuine item (obtained in step c) into the authenticity detection algorithm (obtained in step a), and

[0035] outputting the authenticity detectability value of the item,

[0036] such that each digital signal representation of each genuine item has an associated authenticity detectability value of the item;

[0037] e) training a predictive machine learning model to predict the authenticity detectability value of an item

[0038] Use a set composed of digital signal representations of each genuine article and the associated article authenticity detectability value (obtained in step d).

[0039] In another embodiment, there is provided a computer-implemented method for predicting an article authenticity detectability value, the method comprising the following steps:

[0040] a) Obtain the article to be detected;

[0041] b) Obtain a digital signal representation of the article to be detected by a sensor;

[0042] c) Obtain a predictive machine learning model for predicting the article authenticity detectability value, wherein the predictive machine learning model is trained according to the method of the present invention;

[0043] d) Input the digital signal representation of the article to be detected (obtained in step b) into the predictive machine learning model (obtained in step c), and

[0044] Output the predicted article authenticity detectability value for the digital signal representation of the article to be detected.

[0045] In another embodiment, there is provided a computer-implemented method for identifying whether an article is genuine or counterfeit, the method comprising the following steps:

[0046] a) Obtain

[0047] a.1) The article to be identified;

[0048] a.2) A digital signal representation of the article to be identified, by a sensor;

[0049] a.3) An authenticity detection algorithm that returns an authenticity decision as an output based on the digital signal representation of the article to be identified, a.4) A predictive machine learning model for predicting the article authenticity detectability value of the article to be identified using the authenticity detection algorithm (obtained in step a.3), wherein the predictive machine learning model is trained according to the method of the present invention;

[0050] b) Input the digital signal representation of the article to be identified (obtained in step a.2) into the authenticity detection algorithm (obtained in step a.3) and output

[0051] - The article to be identified is identified as genuine,

[0052] Or

[0053] - The article to be identified cannot be identified as genuine, and if the article to be identified cannot be identified as genuine, then

[0054] c) Input the digital signal representation of the item to be identified (obtained in step a.2) into the predictive machine learning model (obtained in step a.4), and

[0055] Output a predicted item authenticity detectability value for the digital signal representation of the item to be identified;

[0056] d) Determine the authenticity detection algorithm (obtained in step a.3) based on the predicted item authenticity detectability value (obtained in step c)

[0057] - capable of detecting the item as genuine and determining the item to be identified as a counterfeit,

[0058] or

[0059] - the item to be identified cannot be recognized as genuine and, optionally, determine the item to be identified as undetectable. Description of the Drawings

[0060] FIG. 1 shows an example of the processing workflow of a prior art authentication algorithm.

[0061] Figure 2 Shows an example of the processing workflow of a method for predicting the item authenticity detectability value using a predictive machine learning model (labeled "ML model" in the figure).

[0062] Panel a) of FIG. 3 shows a robotic arm (330) manipulating an imaging device (310) to capture multiple photos of a genuine product component (320); Panel b) shows a robotic arm (330) manipulating an imaging device (310) to capture multiple pictures as a digital representation of a genuine banknote physical item (340).

[0063] Panel a) of FIG. 4 shows a side view of a system including a fixed imaging device and two robotic arms (410, 420), one robotic arm manipulating a genuine physical item (101) and one robotic arm manipulating a lighting device (430); Panel b) shows a side view of a system including a conveyor (450) for transporting a genuine physical item (101) and two robotic arms (410, 420), one robotic arm manipulating an imaging device and one robotic arm manipulating a lighting device (430).

[0064] Figure 5 Shows an example of the processing workflow of a method for identifying whether an item is genuine or a counterfeit according to one embodiment (Embodiment 1).

[0065] Figure 6 Shows another example of the processing workflow of a method for identifying whether an item is genuine or a counterfeit according to another embodiment (Embodiment 2).

[0066] Figure 7 Shows 13 images of the genuine article described in Example 1. Based on two images (frames 7A and 7B), the article can be identified as genuine, and based on eleven images (frames 7C to 7M), the article cannot be identified as genuine. Detailed implementation

[0067] The term "article" or "item" or "article item" (used interchangeably) refers to a substance that can be sensed by the senses. It can be a man-made article or a hand-made article. Examples of articles include, but are not limited to, security documents, precious metals, banknotes, watches, leather products (such as bags), parts of articles (such as components), labels, packaging, printed surfaces, embossed surfaces, metallized surfaces, etc. Articles with the same characteristics belong to the same "article type" or "article category".

[0068] The term "genuine article" or "real article" or "certified article" (used interchangeably) refers to the article it purports to be, rather than a fake or imitation. In other words, a genuine article is original, real, certified, rather than forged or counterfeited. A genuine article can add or integrate security features that can be detected by a certification algorithm.

[0069] The term "counterfeit article" or "forged article" or "non-certified article" (used interchangeably) refers to an article or item that is made to imitate a genuine article and is intended to be regarded as a genuine article, which is fake or an imitation. In other words, a counterfeit article is a fake, a replica, an imitation, or a forgery.

[0070] "Genuine / Counterfeit" can be described as having a "detectability attribute" that allows identification of whether an item can be detected as genuine / counterfeit. The detectability attribute can be identified or recognized based on the detectability value of the item. In the context of the present invention, the "detectability value of the authenticity of the item" or "authenticity detectability value" is used. The "detectability value of the authenticity of the item" or "detectability value" refers to the value identified from the digital signal representation of the item, where the detectability value determines whether the digital signal representation of the item is sufficient to identify the item as genuine. In the context of the present invention, the predicted detectability value determines whether the digital signal representation of the item is sufficient for an authenticity detection algorithm to identify the item as genuine. "Sufficient" means that the digital signal representation of the item (rather than the item itself) has sufficient signal (or signal quality) for the authenticity detection algorithm to identify the item as genuine. Example 1 shows whether the digital signal representation of the item is sufficient to identify the item as genuine. Thus, the "detectability value of the authenticity of the item" is not related to the item, but rather to the digital signal representation of the item. In other words, it is the "detectability value of the measurement of the authenticity of the item". Therefore, it is not a function of the item, but a function of the environment in which the digital signal representation of the item is obtained (captured) by a sensor. Examples of the detectability value of the authenticity of the item include, but are not limited to: binary tags (e.g., detectable or undetectable), ternary tags (e.g., detectable or undetectable or unknown), or scalar values (e.g., signal-to-noise ratio (SNR) measurements, difference measurements, distance metrics, or similar values known in the art). In one possible embodiment, the tag can be 0 for undetectable and 1 for detectable (see Example 2). In an alternative possible embodiment, the tag can be 1 for undetectable and 0 for detectable. In a possible embodiment, the detectability value can be a scalar value. In a possible embodiment, the detectability value can be a signal processing metric. The signal processing metric can be, for example, the signal-to-noise ratio (SNR) of the cross-correlation of the captured digital signal representation with a template digital signal representation reference of the item to be authenticated. In an alternative possible embodiment, the detectability metric can also be a simple distance measurement (e.g., difference) between an extracted feature from the captured digital signal representation and a matching reference feature from the digital signal representation template. In an alternative possible embodiment, the detectability metric can be a composite distance measurement between a set of features from the captured digital signal representation and a set of matching reference features from the digital signal representation template. Examples of composite distance measurements include L0, L1, L2 norms, and other ways used in statistical modeling to measure the distance between sets of values.

[0071] The term "digital signal representation" or "digital representation" of an article refers to a representation of the article in digital data form. Examples of the digital signal representation of an article include, but are not limited to: binary images, digital recordings, chemical compositions, spectral representations of waves obtained by spectrometer hardware, such as electromagnetic representations in the case of images, or mechanical / pressure representations in the case of sounds or the like, or combinations of the above in the case of multi-modal capture. The digital signal representation of an article can be obtained from signals captured by sensors.

[0072] In one embodiment, the digital signal representation of the article is a binary image.

[0073] In one embodiment, the digital signal representation of the article is obtained (i.e., captured) by a sensor. Each digital signal representation of the article can be obtained by sensor operation under different capture conditions.

[0074] The term "prediction" refers to inferring a detectability value from the digital signal representation of an article using a statistical analysis model or a predictive machine learning model. Prediction can be defined as a method of outputting potential multi-dimensional values from previously unseen potential multi-dimensional input values using a model. The model can come from a set of observed values obtained, or can be an analytical / prior model defined according to a known set of relationships.

[0075] A "machine learning model" refers to a data model or data classifier that has been trained using supervised, semi-supervised, or unsupervised learning techniques known in the field of data science, as opposed to an explicit statistical model. The data input can be represented as a 1D signal (vector), a 2D signal (matrix), or more generally a multi-dimensional array signal (e.g., a tensor, or an RGB color image represented as a 3*2D signal, the 3*2D signal containing red, green, and blue decomposition planes - i.e., 3 matrices), and / or combinations thereof. A multi-dimensional array is mathematically defined as a data structure arranged along at least two dimensions, with each dimension recording more than one value.

[0076] In the case of a deep learning classifier, the data input is further processed through a series of data processing layers to implicitly capture hidden data structures, data signatures, and underlying patterns. Due to the use of multiple data processing layers, deep learning helps to generalize automated data processing to various complex pattern detection and data analysis tasks. Machine learning models can be trained within supervised, semi-supervised, or unsupervised learning frameworks. Within the supervised learning framework, the model learns a function that maps from an input data set to an output result based on pairs of input and matching output examples. Examples of machine learning models for supervised learning include: support vector machines (SVMs), regression analysis, linear regression, logistic regression, naive Bayes, linear discriminant analysis, decision trees, k-nearest neighbor algorithms, random forests, artificial neural networks (ANNs) (such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), fully connected neural networks, long short-term memory (LSTM) models, etc.); and / or combinations thereof. Models trained within the unsupervised learning framework infer a function that identifies the hidden structure of a data set without prior knowledge about the data. Examples of unsupervised machine learning models known in the art include clustering, such as k-means clustering, mixture model clustering, hierarchical clustering; anomaly detection methods; principal component analysis (PCA), independent component analysis (ICA), t-distributed stochastic neighbor embedding (t-SNE); generative models; and / or unsupervised neural networks; autoencoders; and / or combinations thereof. Semi-supervised learning (SSL) is a machine learning framework within which models can be trained using both labeled and unlabeled data. Data augmentation methods can optionally be used to generate artificial data samples from scarce real data sample sets and increase the quantity and diversity of data used for model training. Compared to other frameworks, unlabeled data can significantly improve learning accuracy when combined with a small amount of labeled data. This approach is particularly attractive when only a portion of the available data is labeled data.

[0077] "Convolutional neural network" or "CNN" refers to a machine learning model that uses multiple data processing layers, called convolutional layers, to represent input data in a way that is most suitable for solving classification or regression tasks. During training, the weight parameters of each CNN layer are optimized using optimization algorithms known in the art, such as the backpropagation algorithm, to perform stochastic gradient descent. At runtime, the trained CNN can then very effectively process the input data, for example, in the case of learning a classification task, classifying it into the correct data output label and minimizing false positives and false negatives as much as possible. Convolutional neural networks can also be combined with recurrent neural networks to generate deep learning classifiers.

[0078] In the context of the present invention, the term "authenticity detection algorithm" refers to an algorithm that generates an authenticity value from a digital signal representation of an item and / or also returns an authenticity decision as an output. Thus, in some embodiments, the authenticity value is referred to as an authenticity detectability value. The authenticity decision can be determined, for example, from a predefined threshold of the detectability value or from a set of multiple detectability values calculated from different subsamplings of the digital representation, such as in the case of multiple croppings on the same acquired input image.

[0079] The authenticity detection algorithm can also be referred to as an "authentication algorithm". In the context of the present invention and in the prior art (Figure 1), an "authenticity detection algorithm" or "authentication algorithm" can be any authentication algorithm that takes as input a digital signal representation of an item to be identified / authenticated. It is obvious to those skilled in the art of authentication that the authentication algorithms of the prior art typically return an output that is either an authenticity decision (as a binary label "authenticated as genuine" / "not authenticated / detected as genuine") or a ternary label "authenticated as genuine" / "not authenticated as genuine" / "unknown". However, as shown in Figure 1, the label "not authenticated / detected as genuine" is not necessarily equivalent to the definitive label "fake" (the dashed line in Figure 1). Similarly, the label "not detected" is not equivalent to the definitive label "unknown". The authentication algorithms of the prior art may provide the label "fake" only when the acquisition conditions are fully controlled, as in the case of a flatbed scanner, such that a single acquisition ensures a detectable digital representation of the item, or if they have been trained using representations of forged items or synthetically generated forged items.

[0080] The authentication algorithms of the prior art can also include internal signal processing algorithms to calculate a difference measurement between the digital signal representation and a reference template digital signal representation. When such a measurement is available, it can quantify the authenticity detectability value of the item as a scalar value (e.g., from 0 for no genuine detection to 100 for complete genuine detection). The resulting scalar value can then be further used by the authentication algorithm to classify the authenticity of the item to be authenticated, for example, using a predefined threshold to distinguish measurement values corresponding to the decisions "not authenticated as genuine" (lower range below the threshold) and "authenticated as genuine" (higher range above the threshold), respectively.

[0081] Figure 1 shows an example of the processing workflow of an existing authentication algorithm (or authenticity detection algorithm) (100), which is consistent with authentication methods described, for example, in WO0225599, WO04028140, Micropore Cloud WO06087351, or US10332247. Such an authentication algorithm (100) takes as input one or more digital signal representations of an item, which can be captured by a sensor (such as an image sensor). The authenticity detection algorithm (100) can optionally preprocess the captured digital signal representations, for example, by using geometric transformations (such as scaling, rotation, translation, downsampling, upsampling, cropping, etc.), frequency domain transformations (such as Fourier transform, discrete cosine transform DCT, etc.), filters (such as low-pass filters, high-pass filters, equalizers, etc.), to generate a set consisting of preprocessed digital signal representations of the item, which is suitable for comparative analysis against a reference template (such as the cross-correlation of a cropped region captured by the input image with a matching cropped region of a genuine reference image stored in a template database). Through the comparative analysis, the authenticity detection algorithm (100) can output a scalar value of the authenticity measurement (such as the signal-to-noise ratio SNR scalar value obtained by cross-correlation calculation). The authenticity detection also includes a decision module for determining whether the item can be detected as genuine with high confidence or cannot be detected based on the latter value (such as by comparing it with a predefined threshold). An undetectable event occurs either because the item is actually a counterfeit (but people cannot recognize it), or because the digital signal representation of the item does not enable the authenticity detection algorithm (100) to detect the authenticity of the item with a sufficiently high confidence.

[0082] In the context of the present invention, the term "predictive machine learning model" or "predictive model" or "predictive machine learning model for predicting the authenticity detectability value of an item" refers to a machine learning model that can be trained to predict the authenticity detectability value of an item based on at least one digital signal representation of the item. In one embodiment, the predictive machine learning model is trained according to the method of the present invention.

[0083] The term "preprocessing" refers to a series of digital operations that convert the raw data or signal or raw digital signal representation captured by a sensor into a preprocessed digital signal representation that can be used, for example, by a predictive machine learning algorithm or an authenticity detection algorithm (authentication algorithm). Examples of known preprocessing methods include, but are not limited to, geometric transformations (e.g., scaling, rotation, translation, downsampling, upsampling, cropping, etc.), frequency domain transformations or other domain transformations (e.g., Fourier transform, discrete cosine transform, wavelet transform, etc.), filters (e.g., low-pass filter, high-pass filter, equalizer, etc.), etc. Background suppression algorithms or image registration algorithms can be applied as preprocessing steps. Fully convolutional neural networks, UNet networks, and spatial transformation networks can even be trained to optimize the correlation between reference images. High dynamic range preprocessing combines multiple digital representations.

[0084] Method for training a predictive machine learning model and use of the predictive machine learning model

[0085] In one embodiment, a computer-implemented method is provided for training a predictive machine learning (ML) model to predict an authenticity detectability value of an item based on at least one digital signal representation of at least one genuine article.

[0086] Training using digital signal representations of counterfeits may not be interesting because there is no exact quantity of counterfeits and new counterfeits can be produced, leading to the algorithm becoming obsolete and requiring continuous updates. Thus, according to the method of the present invention, training is performed based on genuine articles to allow for reliable further classification of items. In other words, the method of the present invention is based on positive detection.

[0087] In another embodiment, training can be performed using genuine articles and in combination with a small number of genuine articles that do not have security / authentication features embedded therein, and these genuine articles without security / authentication features embedded therein can thus serve as potential representations of counterfeits. These representations are positioned at exactly the same viewing positions as the similar genuine articles and are mapped to the authenticity detectability values observed for the corresponding genuine articles. This may improve the learning efficiency.

[0088] Thus, the computer-implemented method for training a predictive machine learning model according to the present invention allows for obtaining a predictive machine learning model to predict an authenticity detectability value of an item.

[0089] A computer-implemented method for training a predictive machine learning model according to the present invention includes the step of obtaining a training data set, which includes a set composed of digital signal representations of each genuine article used, wherein each digital signal representation has an associated item authenticity detectability value. It should be understood that the set of digital signal representations can be obtained by using sensors to capture the digital signal representations. Therefore, the captured digital signal representations of the articles can be obtained.

[0090] In one embodiment, there is provided a computer-implemented method for training a predictive machine learning model to predict the authenticity detectability value of an item, the method including the following steps:

[0091] a) Obtain an authenticity detection algorithm that generates the authenticity detectability value of an item from the digital signal representation of the item;

[0092] b) Obtain one or more genuine articles of this type;

[0093] c) Obtain a set composed of digital signal representations of each genuine article;

[0094] d) Input each digital signal representation of each genuine article (obtained in step c) into the authenticity detection algorithm (obtained in step a), and

[0095] Output the authenticity detectability value of the item,

[0096] such that each digital signal representation of each genuine article has an associated item authenticity detectability value;

[0097] e) Use the set composed of digital signal representations of each genuine article and the associated item authenticity detectability values (obtained in step d) to train a predictive machine learning model to predict the authenticity detectability value of an item.

[0098] In one embodiment, the detectability value determines (or identifies) whether the digital signal representation of an item is sufficient to identify (or characterize) the item as a genuine article. However, the detectability value alone cannot determine (or identify) whether the digital signal representation of an item represents a forged item.

[0099] In one embodiment, there is provided a computer-implemented method for training a predictive machine learning model, wherein the authenticity detection algorithm is selected from a surface fingerprint detector and a product identification detector.

[0100] In one embodiment, the set composed of digital signal representations of each genuine article is obtained by a sensor. In another embodiment, the set composed of digital signal representations of each genuine article is obtained by a sensor, wherein each digital signal representation of each genuine article is obtained by the operation of the sensor under different capture conditions.

[0101] In one embodiment, a predictive machine learning model trained according to the method of the present invention is used to identify whether an item is genuine or counterfeit in a computer-implemented method.

[0102] In one embodiment, a computer-implemented method for training a predictive machine learning model is provided, wherein the selected authenticity detection algorithm and the predictive machine learning model to be trained are selected to be capable of processing digital signal representations of a specific type of genuine item. The selected authenticity detection algorithm can be any suitable algorithm known in the art, for example selected from a surface fingerprint detector and a product identification detector. In particular, AlpVision fingerprint detector, AlpVision cryptoglyph detector, taggant detector, Scantrust security graphic detector, SICPA security ink detector, etc.

[0103] In one embodiment, the method for training a predictive machine learning model according to the present invention uses an authenticity detection algorithm (e.g., obtained in step a)) that produces an authenticity detectability value of an item, where the value can be a scalar value, a label, a hash value, a vector, a multi-dimensional vector / tensor, an image, a matrix, a distribution curve, etc., preferably a scalar value or a label.

[0104] In one embodiment, the method for training a predictive machine learning model according to the present invention uses an authenticity detection algorithm (e.g., obtained in step a)) that generates an authenticity detectability value of an item, where the value is a scalar value, such as selected from a signal-to-noise ratio (SNR) measurement result, a difference measurement result, and a distance metric. In a preferred embodiment, the authenticity detectability value of the item is a scalar value, preferably SNR, which quantifies the strength of the security feature signal. Using a continuous variable as the target of the predictive machine learning model has the advantage of removing hyperparameters compared to using a categorical variable. When using a categorical variable, a threshold representing the boundary between detectable and non-detectable predictions needs to be explicitly selected before training.

[0105] In an alternative embodiment, the method for training a predictive machine learning model according to the present invention uses an authenticity detection algorithm (e.g., obtained in step a)) that generates an authenticity detectability value of an item, where the value is a label, such as selected from a binary label (e.g., detectable or non-detectable; or e.g., pass or no result available), and a ternary label (e.g., detectable or non-detectable or unknown), or any classification of a continuous variable, such as one-hot encoding of values for each integer step.

[0106] In one embodiment, the method for training a predictive machine learning model according to the present invention is based on a selected type of article (e.g., in step b). Thus, different types of articles may require different predictive machine learning models, which are trained separately considering the differences in the attributes of the article types.

[0107] In one embodiment, the method for training a predictive machine learning model according to the present invention is based on 1, at least 1, at least 2, at least 10, at least 50, at least 100, at least 200, or at least 500 genuine articles (e.g., in step b), preferably at least 100, and more preferably at least 10. The effect of using more than one genuine article is to increase the size of the training set, which in turn allows for an improvement in the predictive ability of the obtained predictive machine learning model of the present invention. All the articles used in the training method belong to one class or type of article.

[0108] In one embodiment, the method for training a predictive machine learning model according to the present invention uses a set composed of digital signal representations of each genuine article (e.g., in step c), where the set composed of digital signal representations of each genuine article may contain or include 1, at least 1, at least 10, at least 100, at least 200, at least 1000, at least 1000, at least 10000, at least 20000, or at least 40000 digital signal representations, preferably at least 10 or at least 100 digital signal representations. The effect of increasing the number of digital signal representations used (e.g., at least 200 for each genuine article) is to increase the size of the training set, which in turn allows for an improvement in the predictive ability of the obtained predictive machine learning model of the present invention. The effect of reducing the number of digital signal representations used (e.g., at least 10 for each genuine article) is to obtain a training set of sufficient size, which in turn allows the obtained predictive machine learning model of the present invention to have good predictive ability. Those skilled in the art will use common methods to determine the appropriate training set size in order to obtain the best training results.

[0109] It should be understood that the training thus includes multiple sets composed of digital signal representations of each genuine article. For example, in one training set, each of the 10 genuine articles has at least 200 digital signal representations, for a total of at least 2000 digital signal representations, or for example, in one training set, each of the 100 genuine articles has at least 200 digital signal representations, for a total of at least 20000 digital signal representations. In a preferred embodiment, a method for training a predictive machine learning model according to the present invention uses at least 200 digital signal representations for each genuine article of at least 50 articles.

[0110] In one embodiment, a method for training a predictive machine learning model according to the present invention uses a training set consisting of digital signal representations of each genuine article, and associated article authenticity detectability values, where all detectability values are labeled as detectable, or where all detectability values are labeled as non-detectable, or where some detectability values are labeled as detectable and the remaining detectability values are labeled as non-detectable. Preferably, the ratio of the number of digital signal representations in the set with detectability values to the number of digital signal representations in the set with non-detectability values is selected from 50:50, 70:30, 80:20, 90:10 of detectable and non-detectable. The effect of using a training set with a mixture of associated detectability values (e.g., detectable and non-detectable) is generally to improve the predictive ability of the predictive machine learning model obtained by the present invention.

[0111] In one embodiment, a computer-implemented method for training a predictive machine learning model according to the present invention is provided, where each digital signal representation of each genuine article is obtained from acquiring a signal captured by a sensor, such as selected from an image sensor, a digital olfactory sensor, a digital chemical sensor, a microphone, a micro-text reader, a barcode reader, a QR code reader, a laser-based sensor, a code reader, an RFID reader, an infrared sensor, a UV sensor, a digital camera, and a smartphone camera. In one embodiment, at least one sensor is used. In alternative embodiments, at least two, at least three, or at least five different sensors are used. In one embodiment, the sensor is a camera.

[0112] In one embodiment, the signal captured by the sensor can optionally further undergo the step of signal preprocessing with a signal preprocessing algorithm to obtain a digital signal representation.

[0113] In one embodiment, a computer-implemented method for training a predictive machine learning model according to the present invention is provided, where obtaining a set consisting of digital signal representations of each genuine article further includes using a signal preprocessing method to convert the acquired digital signal representations into digital signal representations suitable for input into an authenticity detection algorithm.

[0114] In one embodiment, the preprocessing of the signals captured by the sensor may include the step of preprocessing by means of an authenticity detection algorithm (authentication algorithm) according to known methods using a known system as described herein. Examples of preprocessing methods include, but are not limited to: geometric transformations such as scaling, rotation, translation, downsampling, upsampling, cropping, etc.; frequency domain transformations such as Fourier transform, Discrete Cosine Transform (DCT), etc.; and filters such as low-pass filters, high-pass filters, equalizers, etc.

[0115] In another embodiment, a computer-implemented method for training a predictive machine learning model according to the present invention may use two sets composed of digital signal representations of each genuine article, where one set is the input to the authenticity detection algorithm and the other set is the input to the predictive machine learning model. These different sets may be obtained based on different preprocessing steps.

[0116] In an exemplary embodiment, a computer-implemented method for training a predictive machine learning model according to the present invention may use an AlpVision Cryptoglyph detector as the selected authenticity detection algorithm, where the signals captured by the sensor are cropped to a field of view larger than the field of view used for the cryptographic detector and downsampled in order to obtain a digital signal representation suitable for further processing.

[0117] In an alternative exemplary embodiment, a computer-implemented method for training a predictive machine learning model according to the present invention may use a surface fingerprint detector as the selected authenticity detection algorithm, where the surface fingerprint detector does not perform downsampling on the signals captured by the sensor and obtains a digital signal representation suitable for further processing. Since there are also microstructures with a similar distribution on counterfeits, this does not prevent the model from recognizing the digital representation of counterfeits as detectable while improving the ability to exclude irrelevant items.

[0118] In one embodiment, a computer-implemented method for training a predictive machine learning model is provided according to the present invention, where each digital signal representation of each genuine article is acquired by the sensor at different sensor positions and / or orientations. The different sensor positions and / or orientations are related to the genuine article. In one embodiment, the different sensor positions and / or orientations are selected from a series of possible sensor positions and / or orientations. In one embodiment, the different sensor positions and / or orientations are predetermined. In one embodiment, for each genuine article, the different sensor positions and / or orientations are the same.

[0119] In one embodiment, a computer-implemented method for training a predictive machine learning model is provided according to the present invention, wherein the relative position and / or orientation of the sensor and the genuine article are different to produce different sensor capture conditions.

[0120] In one embodiment, the different sensor positions and / or orientations are controlled (or set) by a robotic arm. In embodiments using two sensors, at least one or at least two robotic arms can be used. Examples of using robotic arms can be seen in FIGS. 3 and 4.

[0121] In another embodiment, the different sensor positions and / or orientations are controlled (set) by an operator positioning the sensor. The operator can position the sensor manually or using a suitable device.

[0122] In one embodiment, a computer-implemented method for training a predictive machine learning model is provided according to the present invention, wherein each digital signal representation of each genuine article is acquired at different sensor positions and / or orientations. The different genuine article positions and / or orientations are related to the sensor. In one embodiment, the different article positions and / or orientations are selected from a range of possible genuine article positions and / or orientations. In one embodiment, the different genuine article positions and / or orientations are predetermined. In one embodiment, for each genuine article, the different genuine article positions and / or orientations are the same.

[0123] In one embodiment, the different genuine article positions and / or orientations are controlled (set) by a robotic arm positioning the genuine article. FIGS. 3 and 4 show examples of using robotic arms.

[0124] In one embodiment, the different genuine article positions and / or orientations are controlled (set) by a conveyor positioning the genuine article. An example of using a conveyor can be seen in Figure 4b )

[0125] In one embodiment, the different genuine article positions and / or orientations are controlled (set) by an operator positioning the genuine article. The operator can position the genuine article manually or using a suitable device.

[0126] In one embodiment, a computer-implemented method for training a predictive machine learning model is provided according to the present invention, wherein each digital signal representation of each genuine article is acquired at at least one predetermined physical environment parameter value, such that the physical environment parameter value changes the digital signal representation of each genuine article at a predetermined article position and orientation and / or at a predetermined sensor position and orientation. In one embodiment, the at least one predetermined physical environment parameter value is selected from a range of possible values.

[0127] Therefore, it should be understood that a set consisting of digital signal representations of each genuine article is obtained by a sensor, and each digital signal representation of each genuine article is obtained by the operation of the sensor under different capture conditions. Such different capture conditions are understood as different values of physical environment parameters, or are characterized by at least one physical environment parameter around the genuine article.

[0128] In one embodiment, a computer-implemented method for training a predictive machine learning model is provided according to the present invention, wherein each digital signal representation of each genuine article is obtained under capture conditions characterized by at least one physical environment parameter around the genuine article.

[0129] In one embodiment, a computer-implemented method for training a predictive machine learning model is provided according to the present invention, wherein each digital signal representation of each genuine article is obtained under capture conditions characterized by a plurality of physical environment parameters around the genuine article. In one embodiment, at least one physical environment parameter (value) varies around the article. This may be due to changes in the relative position and / or orientation of the sensor and the genuine article. This results in each digital signal representation of each genuine article being obtained by the operation of the sensor under different capture conditions. In other words, the capture conditions are different from each other because at least one physical environment parameter value is different when capturing each digital signal representation of each genuine article by the sensor.

[0130] In the most general case, there are three different variables in the physical environment parameters that can affect the digital representation signal of a physical item captured by at least one sensor:

[0131] - The position and orientation of the item in three-dimensional space, which can be modeled with up to 6 variables ("6DOF"), corresponding to the translational and angular degrees of freedom of the rigid item body; for flexible items (such as banknotes, textiles, etc...), the position and orientation can be further modeled according to the position and / or orientation of the vertices on the deformation grid.

[0132] - The position and orientation of the sensor in three-dimensional space, which can be modeled with up to 6 variables ("6DOF"), corresponding to the translational and angular degrees of freedom of the rigid item body, either as absolute values in space or as differences relative to the item to be sensed; this includes, for example, the tilt, zoom, rotation, and translation of a camera sensor.

[0133] - Certain changes in the physical environment that may directly affect the sensor capture, such as the lighting conditions (position, intensity, orientation, spectrum of each light, ambient light, presence of smoke, etc.) of an image sensor or the ambient noise conditions of an audio sensor.

[0134] In one embodiment, in both the training method and the recognition method, the capture conditions are further characterized by the parameters or states of the sensors. In the case where the sensor is a camera, examples of sensor parameterization include camera resolution, autofocus, aperture, focal length, field of view, white balance, etc. In the case where the sensor is a camera, examples of sensor states include the camera temperature (which may cause CMOS thermal noise), aging of electronic components, or the presence of dust, fingerprints, and the like on the camera. In a specific example, the capture conditions are characterized by the presence of dust and / or fingerprints on the camera lens.

[0135] In one embodiment, in both the training method and the recognition method, the capture conditions are further characterized by the quality or state of the item. The quality or state of the item may depend on its aging condition or the way it is handled during its life, for example, dirt on the surface, scratches, color fading due to ultraviolet radiation, deformation due to vibration or a humid environment, oxidation, etc. Examples of such items include items having dirt on the surface, scratches, color fading due to ultraviolet radiation, deformation due to vibration or a humid environment, oxidation, etc.

[0136] According to the actual application, various system settings can be used to utilize the item locator to control at least one variable position and / or orientation parameter of the training item during capture. In a possible embodiment, a mechanical setting with automated control can be used, such as a robotic arm or a conveyor with its software controller, to precisely manipulate the training item and control the variable position and / or orientation parameters of the training item. In another possible embodiment, the training item can be placed in a fixed position and orientation, and at least one other variable in the physical environment around the training item (such as sensor position or parameterization) can be varied.

[0137] According to the actual application, various system settings can be used to utilize the sensor locator to control at least one variable position and / or orientation parameter of the sensor during capture. In a possible embodiment, a mechanical setting with automated control can be used, such as a robotic arm. In another possible embodiment, a group of multiple sensors can be placed at different fixed positions and orientations around the training item, and each sensor can be sequentially controlled to take different captures of the training item, each capture corresponding to a different position and orientation of the training item relative to the fixed training item. According to the actual application, at least one physical environment parameter can be automatically set by a physical environment parameter controller. For example, when using a smartphone for sensor capture, a dedicated smartphone application can be developed that controls the smartphone to illuminate the item, such as using the smartphone flashlight in the flashlight mode. More generally, the physical environment around the item to be captured for training can be adapted to at least one physical environment control device, such as lights, speakers, and / or mechanical actuators.

[0138] Examples of variable characteristics of the physical environment light include, but are not limited to: color temperature, polarization, emission spectrum, intensity, beam shape, pulse shape, light orientation, distance of the light towards the item, etc. Examples of variable characteristics of the actuator include, but are not limited to: volume of water or air projected towards the item; force applied to brush, tap, shear, bend it; temperature for heating or cooling it; distance to a magnet approaching it; variable placement of a movable light reflector, blocker, diffuser or filter (such as a Wratten filter or an interference filter); and so on.

[0139] In a possible embodiment, a series of digital representations of training items can be captured as a series of image acquisitions over time with an imaging device under an orientable neon light with variable intensity, then under an orientable LED light with variable intensity, and then under direct sunlight at different times of the day. Using this scheme, the set of digital representations will inherently represent multiple light sources with different spectra, variable intensities, and variable positions, as input to a machine learning classifier generation engine.

[0140] In a possible embodiment, a series of digital representations of training items can be captured as a series of image acquisitions over time with an imaging device parameterized by different sensors.

[0141] In another possible embodiment, by independently controlling two flashlight LEDs, a series of digital representations of training items can be captured with an Apple iPhone version iPhone 7 or higher, each flashlight LED having a different, adjustable, variable color temperature. It will be obvious to those skilled in the imaging art that the resulting set of digital representations will inherently represent a variable spectral reflectance, transmittance, and emissivity environment, as input to a machine learning classifier generation engine.

[0142] In one embodiment, there is provided a computer-implemented method for training a machine learning model, wherein each digital signal representation of each genuine article is a digital image representation of the genuine article, wherein at least one physical environment parameter around the genuine article is the illumination of the genuine article by an illumination device, and wherein the illumination of the genuine article is controlled by a robotic arm, a conveyor, or an operator, which positions the illumination device relative to the genuine article. In one embodiment, at least one physical environment parameter varies around the item.

[0143] In another embodiment, there is provided a computer-implemented method for training a machine learning model, wherein the sensor is a smartphone camera and the illumination device is a smartphone flash.

[0144] The above embodiments can also be combined to provide the widest possible input digital representation space for training items. In the most general case, each of the three different variables in the physical environment can be gradually changed within different possible ranges to generate multiple digital representations for each input item, where the three different variables can affect the digital representation signal of the physical item captured by at least one sensor, such that machine learning can better predict the diversity of the end-user physical environment that the generated classifier solution will encounter during detection:

[0145] - The position and orientation of an item in three-dimensional space can be gradually changed along any position axis or orientation axis, for example, with a translational increment of 1 mm and an Euler angle increment of 1°;

[0146] - The position and orientation of a sensor in three-dimensional space can be gradually changed along any position axis or orientation axis, for example, with a translational increment of 1 mm and an Euler angle increment of 1°;

[0147] - Various physical environment parameters can simply change with on / off, or change with a gradual increment of any variable associated with the underlying physical component, for example, with a light intensity increment of 10 lux, or a sound intensity increment of 1 dB, or a frequency intensity increment of 1 Hz, etc.;

[0148] - Various sensor parameter settings can simply change with the enabling of the sensor controller API, for example, by changing the resolution, focal length, aperture, field of view, white balance, or other optical parameters of the camera sensor on a smartphone.

[0149] In one embodiment, the trained predictive machine learning model can be a machine learning classifier. In an alternative embodiment, the trained predictive machine learning model can be a machine learning regressor.

[0150] In one embodiment, the trained predictive machine learning model can be an artificial neural network tool (ANN), such as a deep learning model, and a convolutional neural network (CNN) or any equivalent model, preferably a CNN model. In a possible embodiment, a pre-trained convolutional neural network (CNN), such as AlexNet, VGG, GoogleNet, UNet, Vnet, ResNet, or other networks, can be used to further train the predictive machine learning model, but other embodiments are also possible. In one embodiment, the trained predictive machine learning model is a supervised machine learning algorithm, where the training set includes or consists of the digital signal representation of each genuine item, and each digital signal representation has an associated item authenticity detectability value.

[0151] In one embodiment, a computer-implemented method for predicting an authenticity detectability value of an item is provided, the method comprising the steps of:

[0152] a) Obtaining the item to be detected;

[0153] b) Obtaining a digital signal representation of the item to be detected;

[0154] c) Obtaining a predictive machine learning model to predict the authenticity detectability value of the item;

[0155] d) Inputting the digital signal representation of the item to be detected (obtained in step b) into the predictive machine learning model (obtained in step c), and

[0156] Outputting a predicted authenticity detectability value for the digital signal representation of the item to be detected.

[0157] In one embodiment, a computer-implemented method for predicting an authenticity detectability value of an item is provided, the method comprising the steps of:

[0158] a) Obtaining the item to be detected;

[0159] b) Obtaining a digital signal representation of the item to be detected;

[0160] c) Obtaining a predictive machine learning model to predict the authenticity detectability value of the item, wherein the predictive machine learning model is trained according to the method of the present invention;

[0161] d) Inputting the digital signal representation of the item to be detected (obtained in step b) into the predictive machine learning model (obtained in step c), and

[0162] Outputting a predicted authenticity detectability value for the digital signal representation of the item to be detected.

[0163] In one embodiment, a computer-implemented method for predicting an authenticity detectability value of an item is provided, wherein the detectability value can further be used to determine (or identify) whether the digital signal representation of an object is sufficient to identify (or characterize) the item as genuine.

[0164] In one embodiment, a computer-implemented method for predicting an authenticity detectability value of an item is provided, wherein a digital signal representation of the item to be detected is obtained (or captured) by a sensor. In another embodiment, a digital signal representation of the item to be detected is obtained by a sensor, wherein the digital signal representation of the item is obtained from the operation of the sensor under given capture conditions. The given (or one) capture condition is understood to be a given value of a physical environment parameter, and / or characterized by at least one physical environment parameter around the item, and / or characterized by the parameterization or state of the sensor, and / or characterized by the quality or state of the item. The sensor can be the same sensor as the sensor used for training the model, or a similar sensor.

[0165] In one embodiment, a computer-implemented method is provided for predicting an authenticity detectability value of an item using a predictive machine learning model, wherein the method can include an algorithm referred to herein as the detectability prediction algorithm (200)( Figure 2 ).

[0166] In one embodiment, a computer-implemented method is provided for predicting an authenticity detectability value of an item based on at least one digital signal representation of such item to be detected. In one embodiment, a computer-implemented method for predicting an authenticity detectability value of an item uses more than one digital signal representation of the item to be detected, wherein these digital signal representations are obtained and processed sequentially or in parallel. This improves the credibility of the prediction. For example, if the predictions of different digital signal representations are considered independent, then if N predictions exceeding a given detectability prediction threshold are required to consider the item detectable, the false detectability prediction rate is actually divided by N. In one embodiment, a computer-implemented method is provided for predicting an authenticity detectability value of an item based on at least two, at least 10, at least 25, at least 50, at least 100, or at least 250 digital signal representations of such item to be detected, preferably at least 50.

[0167] In one embodiment, a computer-implemented method for predicting an authenticity detectability value of an item is used to identify whether the item is genuine or counterfeit in a computer-implemented method.

[0168] In one embodiment, a computer-implemented method for predicting an authenticity detectability value of an item is provided, wherein the selected predictive machine learning model is selected to be capable of processing a specific type of digital signal representation of the item to be detected. It should be understood that in the method for predicting an authenticity detectability value of an item, the predictive machine learning model is trained based on items of the same type as the item to be identified.

[0169] In one embodiment, a computer-implemented method for predicting the authenticity detectability value of an item outputs a predicted authenticity detectability value of the item to be detected, where the value can be a continuous value, a categorical value, a hash value, a vector, a tensor, an image, a matrix, etc.

[0170] In one embodiment, a computer-implemented method for predicting the authenticity detectability value of an item outputs a predicted authenticity detectability value of the item to be detected, where the value is a scalar value, such as selected from signal-to-noise ratio (SNR) measurement results, difference measurement results, and distance metrics. In a preferred embodiment, the authenticity detectability value of the item is a continuous scalar value, such as SNR. Using a continuous value as the target of a predictive machine learning model has the advantage of removing hyperparameters compared to using a categorical variable. When using a categorical variable, a threshold representing the boundary between detectable and non-detectable predictions needs to be explicitly selected before training.

[0171] In another embodiment, a computer-implemented method for predicting the authenticity detectability value of an item outputs a predicted authenticity detectability value of the item to be detected, where the value is a label, such as selected from binary labels (e.g., detectable or non-detectable) and ternary labels (e.g., detectable or non-detectable or unknown).

[0172] In one embodiment, a computer-implemented method for predicting the authenticity detectability value of an item is applicable to one type of item (e.g., in step a). Thus, different types of items may require different methods to predict the detectability value, and these methods can use different predictive machine learning models that are trained separately taking into account the attribute differences of the item types.

[0173] In one embodiment, a computer-implemented method for predicting the authenticity detectability value of an item is provided, where the digital signal representation of the item to be detected is obtained by acquiring a signal captured by a sensor, and the sensor is, for example, selected from an image sensor, a digital olfactory sensor, a digital chemical sensor, a microphone, a microtext reader, a barcode reader, a QR code reader, a laser-based sensor, a code reader, an RFID reader, an infrared sensor, a UV sensor, a digital camera, and a smartphone camera. In one embodiment, at least one sensor is used. In an alternative embodiment, at least two, at least three, or at least five different sensors are used.

[0174] In one embodiment, the signal captured by the sensor can optionally further undergo a step of signal preprocessing with a signal preprocessing algorithm to obtain a digital signal representation. This preprocessing step can be included in the detectability prediction algorithm (200).

[0175] In one embodiment, a computer-implemented method for predicting an authenticity detectability value of an item is provided, wherein obtaining a digital signal representation of the item to be identified further includes converting the obtained digital signal representation into a digital signal representation suitable for input to a predictive machine learning model using a signal preprocessing method. Similarly, this step may be included in the detectability prediction algorithm (200).

[0176] In one embodiment, preprocessing of the signal captured by the sensor may include preprocessing steps by a predictive machine learning model according to known methods using known systems as described herein.

[0177] Figure 2 An example of a processing workflow of a method for predicting an authenticity detectability value of an item using a predictive machine learning model (labeled "ML model" in the figure) is shown, wherein the method includes an algorithm that may be referred to herein as the detectability prediction algorithm (200). Figure 2 An example detectability prediction algorithm 200 according to certain embodiments of the present disclosure is shown. Such a detectability prediction algorithm (200) takes as input one or more digital signal representations of an item, which may be captured by a sensor (e.g., an image sensor). The detectability prediction algorithm (200) may optionally preprocess the captured digital signal representations, for example, by using geometric transformations (e.g., scaling, rotation, translation, downsampling, upsampling, cropping, etc.), frequency domain transformations (e.g., Fourier transform, discrete cosine transform DCT, etc.), filters (e.g., low-pass filter, high-pass filter, equalizer, etc.), to generate a set of digital signal representations of the item suitable as input to a machine learning model that has been trained to predict an authenticity detectability value of the same type of item, using an authenticity detection algorithm (100) suitable for determining the authenticity of the same type of item from its captured digital signal representation. Based on the prediction of the machine learning model, the detectability prediction algorithm (200) may then output a predicted authenticity detectability value of the item based on its captured digital signal representation.

[0178] Method for identifying whether an item is genuine or counterfeit and its uses

[0179] In one embodiment, a computer-implemented method for identifying whether an item is genuine or counterfeit is provided, the method comprising the steps of:

[0180] a) Obtaining

[0181] a.1) The item to be identified;

[0182] a.2) A digital signal representation of the item to be identified;

[0183] a.3) An authenticity detection algorithm that returns an authenticity decision as an output based on the digital signal representation of the item to be identified.

[0184] a.4) A predictive machine learning model to predict the authenticity detectability value of the item to be identified using the authenticity detection algorithm (obtained in step a.3), where the predictive machine learning model is trained according to the method of the present invention.

[0185] b) Input the digital signal representation of the item to be identified (obtained in step a.2) into

[0186] b.1.) The predictive machine learning model (obtained in step a.4), and

[0187] Output the predicted item authenticity detectability value for the digital signal representation of the item to be identified;

[0188] And determine whether the authenticity detection algorithm (obtained in step a.3) can detect the item as genuine based on the predicted item authenticity detectability value;

[0189] b.2) Input the digital signal representation of the item to be identified (obtained in step a.2) into the authenticity detection algorithm (obtained in step a.3) and output that the item to be identified is recognized as genuine, or

[0190] The item to be identified cannot be recognized as genuine,

[0191] c) Provide a decision based on the outputs of the predictive machine learning model (obtained in step b.1) and the authenticity detection algorithm (obtained in step b.2), where

[0192] c.1) Optionally, if the item can be detected as genuine (step b.1)) and the item is recognized as genuine (step b.2)), provide an output of genuine;

[0193] c.2) If the item cannot be detected as genuine (step b.1)) and optionally the item is recognized as genuine (step

[0194] b.2)), provide an output of non-detectable item, and optionally repeat steps a.2), b) and c);

[0195] c.3) If the item can be detected as genuine (step b.1)) and the item is not recognized as genuine (step b.2)), provide an output of fake;

[0196] c.4) If the item cannot be detected as genuine (step b.1)), and optionally the item is not identified as genuine (step b.2)), then provide an output for the non-detectable item, and optionally repeat steps a.2), b), and c).

[0197] In one embodiment, a computer-implemented method for identifying whether an item is genuine or counterfeit is provided, wherein a digital signal representation of the item to be identified is obtained (or captured) by a sensor. In another embodiment, a digital signal representation of the item to be identified is obtained by a sensor, wherein the digital signal representation of the item is obtained by operating the sensor under given capture conditions. The given (or one) capture condition is understood to be a given value of physical environment parameters, and / or characterized by at least one physical environment parameter around the item, and / or characterized by the parameterization or state of the sensor, and / or characterized by the quality or state of the item. Physical environment parameters are defined as above.

[0198] In one embodiment, a computer-implemented method for identifying whether an item is genuine or counterfeit is provided, wherein a predicted detectability value (e.g., obtained in step b.1) of the method) identifies whether the digital signal representation of the item is sufficient to identify the item as genuine. It should be understood that this means whether the digital signal representation of the item is sufficient to cause a authenticity detection algorithm (e.g., obtained in step a.3) of the method) to identify the item as genuine. However, it cannot be determined (or identified) whether the digital signal representation of the item represents a forged item solely based on the predicted detectability value.

[0199] In another embodiment, a computer-implemented method for identifying whether an item is genuine or counterfeit is provided according to the present invention, wherein in order to identify the item as genuine

[0200] - the item to be identified is detected as genuine by an authenticity detection algorithm (step b.2)); or

[0201] - the item to be identified is detected as genuine by an authenticity detection algorithm (step b.2)), and it is determined that the item can be detected as genuine according to the predicted detectability value of the item's authenticity (step b.1)).

[0202] In another embodiment, a computer-implemented method for identifying whether an item is counterfeit is provided according to the present invention, wherein in order to identify the item as counterfeit

[0203] - the item can be detected as genuine according to the predicted detectability value of the item's authenticity (step b.1)), and the authenticity detection algorithm does not identify the item as genuine (step b.2)).

[0204] It should be understood that security features are present in the item to be identified so that, based on the predicted item authenticity detectability value, it can be determined whether the authenticity detection algorithm can detect the item as genuine. In the case where it is not known whether security features are present, the method of the present invention allows for determining whether the authenticity detection algorithm can detect the item as genuine based on the predicted item authenticity detectability value.

[0205] It should be understood that a computer-implemented method for identifying whether an item is genuine or counterfeit according to the present invention uses a combination of the output of a predictive machine learning model and the output of an authenticity detection algorithm, where these two algorithms can be connected in series, in parallel, or can also be combined according to any logical operation, such as addition, subtraction, multiplication, division, convolution, AND, OR, XOR, NOT, comparison above, comparison below, equality comparison, and inequality comparison.

[0206] An example embodiment of algorithms connected in series, i.e., an example embodiment where the result of the upstream algorithm is passed to the downstream algorithm, is presented herein as Embodiment 1 ( Figure 5 ) or Embodiment 2 ( Figure 6 ), where these are not limiting examples.

[0207] In one embodiment, a method for identifying whether an item is genuine or counterfeit is provided according to the present invention, which uses a predictive machine learning model (e.g., within the authenticity detection algorithm (100)) and a detectability prediction algorithm (200), where the combination of these two algorithms causes the method to include an algorithm referred to herein as the "forgery authentication algorithm" ( Figure 5 in (500) or Figure 6 in (600)).

[0208] It should be understood that the method and system of the present invention allow for the detection of counterfeits due to the combined use of an authenticity detection algorithm and a predictive machine learning model trained according to the method of the present invention.

[0209] In one embodiment, a method for identifying whether an item is genuine or counterfeit is provided, where the selected authenticity detection algorithm and predictive machine learning model are selected to be capable of processing a digital signal representation of a specific type of item. The selected authenticity detection algorithm can be any suitable algorithm known in the art, such as selected from a surface fingerprint detector and a product identification detector. In particular, AlpVision fingerprint detector, AlpVision cryptoglyph detector, taggant detector, Scantrust security graphic detector, SICPA security ink detector, etc.

[0210] It should be understood that any authentication algorithm outputs a label "genuine" that may include true positive or false positive cases, or a label "counterfeit" that may include true negative or false negative cases.

[0211] In another embodiment, by using or combining further statistical methods and parameters in the authenticity detection algorithm, the effect of reducing the number of false negatives in the "counterfeit" label is achieved. For example, at runtime, multiple digital representations of the same item can be predicted for detectability multiple times to form a distribution and select a meaningful quantity, such as the maximum or average prediction, or a similar quantity known in the art such as a confidence interval.

[0212] In one embodiment, a computer-implemented method is provided for identifying whether an item is genuine or counterfeit based on at least one digital signal representation of the item to be detected. In one embodiment, a computer-implemented method for identifying whether an item is genuine or counterfeit uses more than one digital signal representation of the item to be detected, where these digital signal representations are obtained and processed sequentially or in parallel. This improves the confidence of the prediction. For example, if the predictions of different digital signal representations are considered independent, then if N predictions exceeding a given detectable prediction threshold are required to consider the item detectable, the false detectable prediction rate is actually divided by N.

[0213] In another embodiment, a computer-implemented method is provided for identifying whether an item is genuine or counterfeit based on at least two, at least 10, at least 25, at least 50, at least 100, or at least 250 digital signal representations of such item to be detected, preferably at least 50. Using multiple digital signal representations of the item to be detected allows providing, for example, a statistical distribution of these representations. This can reduce the number of false negative cases within the "counterfeit" label and increase the confidence level of true negative cases within the "counterfeit" label.

[0214] At least two digital signal representations of the item to be detected can be obtained, for example, by capturing with a camera sensor for approximately 5 - 15 seconds, such as approximately 10 seconds, and at least approximately 25 - 150 frames, such as approximately 50 frames.

[0215] It should be understood that using at least two digital signal representations can further reduce the number of false negative cases within the "counterfeit" label.

[0216] It should be understood that in the method for identifying whether an item is genuine or counterfeit according to the present invention, the predictive machine learning model is trained based on items of the same type as the item to be identified. An illustration is provided in Example 2.

[0217] In one embodiment, in a method for identifying whether an item is genuine or counterfeit, in one step, the method outputs a predicted authenticity detectability value of the item to be detected, where the value can be any classification of a scalar value, a label, groundtruth, a regressor, a hash value, a vector, a multi-dimensional vector, an image, a matrix, or a continuous variable, such as one-hot encoding of values for each integer step, etc., preferably a scalar value or a label.

[0218] In one embodiment, in a method for identifying whether an item is genuine or counterfeit, in one step, the method outputs a predicted authenticity detectability value of the item to be detected, where the value is a scalar value, such as selected from a signal-to-noise ratio (SNR) measurement result, a difference measurement result, and a distance metric. In a preferred embodiment, the authenticity detectability value of the item is a scalar value, such as SNR. Using a continuous variable as the target of a predictive machine learning model has the advantage of removing hyperparameters compared to using a categorical variable. When using a categorical variable, a threshold representing the boundary between detectable and undetectable predictions needs to be explicitly selected before training.

[0219] In one embodiment, in a method for identifying whether an object is genuine or counterfeit, in one step, the method outputs a predicted authenticity detectability value of the object to be detected, where the value is a label, such as selected from a binary label (e.g., detectable or undetectable) and a ternary label (e.g., detectable or undetectable or unknown).

[0220] In one embodiment, a method for identifying whether an item is genuine or counterfeit according to the present invention is applicable to one type of item (e.g., in step a). Therefore, different types of items may require different methods to identify whether the item is genuine or counterfeit, and these methods can use different predictive machine learning models, which are separately trained considering the attribute differences of the item types.

[0221] In one embodiment, a method for identifying whether an item is genuine or counterfeit is provided according to the present invention, where the digital signal representation of the item to be detected is obtained by acquiring a signal captured by a sensor, and the sensor is, for example, selected from an image sensor, a digital olfactory sensor, a digital chemical sensor, a microphone, a micro-text reader, a barcode reader, a QR code reader, a laser-based sensor, a code reader, an RFID reader, an infrared sensor, a UV sensor, a digital camera, and a smartphone camera. In one embodiment, at least one sensor is used. In an alternative embodiment, at least two, at least three, or at least five different sensors are used. In another embodiment, a method for identifying whether an item is genuine or counterfeit is provided according to the present invention, where the sensor is a smartphone camera and the lighting device is a smartphone flash.

[0222] In one embodiment, the signal captured by the sensor can optionally further undergo the step of signal preprocessing with a signal preprocessing algorithm to obtain a digital signal representation. This preprocessing step can be included in the authenticity detection algorithm (100) and / or the detectability prediction algorithm (200).

[0223] In one embodiment, according to the present invention, a method for identifying whether an item is genuine or counterfeit is provided, wherein obtaining a digital signal representation of the item to be identified further includes converting the obtained digital signal representation into a digital signal representation suitable for input into a predictive machine learning model and / or an authenticity detection algorithm using a signal preprocessing method. Similarly, this step can be included in the authenticity detection algorithm (100) and / or the detectability prediction algorithm (200). In one embodiment, the preprocessing of the signal captured by the sensor can include the step of preprocessing through a predictive machine learning model according to a known method using a known system as described herein. In one embodiment, the preprocessing of the signal captured by the sensor can include the step of preprocessing through an authenticity detection algorithm according to a known method using a known system as described herein.

[0224] In another embodiment, a method for identifying whether an item is genuine or counterfeit according to the present invention can use two sets composed of digital signal representations of each genuine item, where one set is the input of the authenticity detection algorithm and the other set is the input of the predictive machine learning model. These different sets can be obtained based on different preprocessing steps.

[0225] In an exemplary embodiment, a method for identifying whether an item is genuine or counterfeit according to the present invention can use a Cryptoglyph detector as the selected authenticity detection algorithm, wherein the signal captured by the sensor is cropped to a field of view larger than the field of view used for this Cryptoglyph detector.

[0226] In an alternative exemplary embodiment, a method for identifying whether an item is genuine or counterfeit according to the present invention can use a surface fingerprint detector as the selected authenticity detection algorithm, wherein the surface fingerprint detector does not perform downsampling on the signal captured by the sensor and obtains a digital signal representation suitable for further processing. Since there are also microstructures with a similar distribution on counterfeits, this does not prevent the model from identifying the digital representation of a counterfeit as detectable while improving the ability to exclude irrelevant items.

[0227] In one embodiment (here is exemplary embodiment 1), a computer-implemented method for identifying whether an item is genuine or counterfeit is provided, the method comprising the following steps:

[0228] a) Obtaining

[0229] a.1) An item to be identified;

[0230] a.2) A digital signal representation of the item to be identified, via a sensor;

[0231] a.3) An authenticity detection algorithm that returns an authenticity decision as output based on the digital signal representation of the item to be identified, a.4) A predictive machine learning model to predict an authenticity detectability value of the item to be identified using the authenticity detection algorithm (obtained in step a.3), wherein the predictive machine learning model is trained according to the method of the present invention;

[0232] b) Input the digital signal representation of the item to be identified (obtained in step a.2) into the authenticity detection algorithm (obtained in step a.3) and output

[0233] - The item to be identified is identified as genuine,

[0234] or

[0235] - The item to be identified cannot be identified as genuine, and if the item to be identified cannot be identified as genuine, then

[0236] c) Input the digital signal representation of the item to be identified (obtained in step a.2) into the predictive machine learning model (obtained in step a.4), and

[0237] Output a predicted item authenticity detectability value for the digital signal representation of the item to be identified;

[0238] d) Determine the authenticity detection algorithm (obtained in step a.3) based on the predicted item authenticity detectability value (obtained in step c)

[0239] - Can detect the item as genuine and determine the item to be identified as a fake,

[0240] or

[0241] - The item to be identified cannot be identified as genuine, and optionally determine the item to be identified as undetectable.

[0242] In one embodiment (here is exemplary embodiment 1), a computer-implemented method for identifying fakes is provided, the method comprising the steps of:

[0243] a) Obtain

[0244] a.1) An item to be identified;

[0245] a.2) A digital signal representation of the item to be identified, via a sensor;

[0246] a.3) An authenticity detection algorithm that returns an authenticity decision as an output based on the digital signal representation of the item to be identified, a.4) A predictive machine learning model to predict the authenticity detectability value of the item to be identified using the authenticity detection algorithm (obtained in step a.3), wherein the predictive machine learning model is trained according to the method of the present invention;

[0247] b) Input the digital signal representation of the item to be identified (obtained in step a.2) into the authenticity detection algorithm (obtained in step a.3) and output

[0248] - Optionally, the item to be identified is identified as genuine,

[0249] Or

[0250] - The item to be identified cannot be identified as genuine, and if the item to be identified cannot be identified as genuine, then

[0251] c) Input the digital signal representation of the item to be identified (obtained in step a.2) into the predictive machine learning model (obtained in step a.4), and

[0252] Output the predicted item authenticity detectability value for the digital signal representation of the item to be identified;

[0253] d) Determine the authenticity detection algorithm (obtained in step a.3) based on the predicted item authenticity detectability value (obtained in step c)

[0254] - Can detect the item as genuine and determine that the item to be identified is a counterfeit.

[0255] In one embodiment (this is Example Embodiment 1 herein), a computer-implemented method for identifying whether an item is genuine or counterfeit is provided, the method comprising the following steps:

[0256] a) Obtain

[0257] a.1) The item to be identified;

[0258] a.2) The digital signal representation of the item to be identified, through a sensor;

[0259] a.3) An authenticity detection algorithm that returns an authenticity decision as an output based on the digital signal representation of the item to be identified, a.4) A predictive machine learning model to predict the authenticity detectability value of the item to be identified using the authenticity detection algorithm (obtained in step a.3), wherein the predictive machine learning model is trained according to the method of the present invention;

[0260] b) Input the digital signal representation of the item to be identified (obtained in step a.2) into the authenticity detection algorithm (obtained in step a.3) and output

[0261] - The item to be identified is identified as genuine,

[0262] or

[0263] - The item to be identified cannot be identified as genuine, then

[0264] c) Input the digital signal representation of the item to be identified (obtained in step a.2) into the predictive machine learning model (obtained in step a.4), and

[0265] Output a predicted item authenticity detectability value for the digital signal representation of the item to be identified;

[0266] d) Determine the authenticity detection algorithm (obtained in step a.3) based on the predicted item authenticity detectability value (obtained in step c)

[0267] - Can detect the item as genuine and determine the item to be identified as a fake,

[0268] or

[0269] - The item to be identified cannot be identified as genuine and determine the item to be identified as undetectable, if the item is determined to be undetectable (in step d), then

[0270] e) Obtain another digital signal representation of the item to be identified through a sensor;

[0271] f) Input the other digital signal representation of the item to be identified (obtained in step e) into the authenticity detection algorithm (obtained in step a.3) and output

[0272] - The item to be identified is identified as genuine,

[0273] or

[0274] - The item to be identified cannot be identified as genuine, and if the item to be identified cannot be identified as genuine, then

[0275] g) Input the digital signal representation of the item to be identified (obtained in step e) into the predictive machine learning model (obtained in step a.4), and

[0276] Output another predicted detectability of the item authenticity for the digital signal representation of the item to be identified;

[0277] h) Determine the authenticity detection algorithm (obtained in step a.3) based on the predicted item authenticity detectability value (obtained in step g).

[0278] - Be able to detect the item as genuine and determine the item to be identified as a counterfeit

[0279] Or

[0280] - The item to be identified cannot be identified as genuine, and optionally determine that the item to be identified is undetectable, and repeat steps e) to h) at least once.

[0281] Figure 5 Fig. shows the processing workflow of a method for identifying whether an item is genuine or a counterfeit according to Example Embodiment 1.

[0282] In one embodiment (here is Example Embodiment 2), a computer-implemented method for identifying whether an item is genuine or a counterfeit is provided, and the method includes the following steps:

[0283] a) Obtain

[0284] a.1) The item to be identified;

[0285] a.2) The digital signal representation of the item to be identified, through a sensor;

[0286] a.3) An authenticity detection algorithm that returns an authenticity decision as an output based on the digital signal representation of the item to be identified, a.4) A predictive machine learning model to predict the authenticity detectability value of the item to be identified using the authenticity detection algorithm (obtained in step a.3), where the predictive machine learning model is trained according to the method of the present invention;

[0287] b) Input the digital signal representation of the item to be identified (obtained in step a.2) into the predictive machine learning model (obtained in step a.4), and

[0288] Output the predicted item authenticity detectability value for the digital signal representation of the item to be identified;

[0289] c) Determine based on the predicted item authenticity detectability value (obtained in step b) that the authenticity detection algorithm (obtained in step a.3) can detect the item as genuine or

[0290] Optionally, the item to be identified cannot be identified as genuine, and further optionally determine that the item to be identified is undetectable,

[0291] d) Input the digital signal representation of the item to be identified (obtained in step a.2) into the authenticity detection algorithm (obtained in step a.3) and output

[0292] - The item to be identified is identified as genuine,

[0293] or

[0294] - The item to be identified cannot be identified as genuine, and it is determined that the item to be identified is a counterfeit.

[0295] In one embodiment (here it is Example Embodiment 2), a computer-implemented method for identifying counterfeits is provided, and the method includes the following steps:

[0296] a) Obtain

[0297] a.1) The item to be identified;

[0298] a.2) A digital signal representation of the item to be identified, via a sensor;

[0299] a.3) An authenticity detection algorithm that returns an authenticity decision as an output based on the digital signal representation of the item to be identified, a.4) A predictive machine learning model to predict an authenticity detectability value of the item to be identified using the authenticity detection algorithm (obtained in step a.3), where the predictive machine learning model is trained according to the method of the present invention;

[0300] b) Input the digital signal representation of the item to be identified (obtained in step a.2) into the predictive machine learning model (obtained in step a.4), and

[0301] Output a predicted item authenticity detectability value for the digital signal representation of the item to be identified;

[0302] c) Determine, based on the predicted item authenticity detectability value (obtained in step b), whether the authenticity detection algorithm (obtained in step a.3) can detect the item as genuine or

[0303] the item to be identified cannot be identified as genuine, and optionally determine that the item to be identified is undetectable,

[0304] If the item to be identified can be identified as genuine, then

[0305] d) Input the digital signal representation of the item to be identified (obtained in step a.2) into the authenticity detection algorithm (obtained in step a.3) and output

[0306] - Optionally, the item to be identified is identified as genuine,

[0307] or

[0308] - The item to be identified cannot be identified as genuine, and it is determined that the item to be identified is a counterfeit.

[0309] In one embodiment (Example Embodiment 2 herein), a computer implements a method for identifying whether an item is genuine or counterfeit, the method comprising the steps of:

[0310] a) Obtain

[0311] a.1) The item to be identified;

[0312] a.2) A digital signal representation of the item to be identified, via a sensor;

[0313] a.3) An authenticity detection algorithm that returns an authenticity decision as output based on the digital signal representation of the item to be identified, a.4) A predictive machine learning model to predict an authenticity detectability value of the item to be identified using the authenticity detection algorithm (obtained in step a.3), wherein the predictive machine learning model is trained according to the method of the present invention;

[0314] b) Input the digital signal representation of the item to be identified (obtained in step a.2) into the predictive machine learning model (obtained in step a.4), and

[0315] Output a predicted item authenticity detectability value for the digital signal representation of the item to be identified;

[0316] c) Determine the authenticity detection algorithm (obtained in step a.3) according to the predicted item authenticity detectability value (obtained in step b)

[0317] - Is able to detect the item as genuine, or

[0318] Is not able to detect the item as genuine and determine that the item to be identified is undetectable, and if the item is identified as undetectable (in step c), then

[0319] e) Obtain another digital signal representation of the item to be identified via a sensor;

[0320] f) Input the another digital signal representation of the item to be identified (obtained in step e) into the predictive machine learning model (obtained in step a.4), and

[0321] Output another predicted item authenticity detectability value for the digital signal representation of the item to be identified;

[0322] g) Determine that the authenticity detection algorithm (obtained in step a.3) is able to detect the item as genuine according to the predicted item authenticity detectability value (obtained in step f), and then perform step h)

[0323] Or

[0324] - The item to be identified cannot be identified as genuine, and optionally it is determined that the item to be identified is undetectable, and steps e) to g) are repeated at least once;

[0325] h) Input another digital signal representation of the item to be identified (obtained in step e)) into the authenticity detection algorithm (obtained in step a.3)) and output

[0326] - The item to be identified is identified as genuine,

[0327] Or

[0328] - The item to be identified cannot be identified as genuine, and it is determined that the item to be identified is a fake.

[0329] Figure 6 Fig. shows a processing workflow of a method for identifying whether an item is genuine or fake according to Example Embodiment 2.

[0330] It should be understood that multiple example embodiments of computer-implemented methods for identifying whether an item is genuine or fake are provided, wherein the predicted detectability value obtained for one or another digital signal representation of the item to be identified identifies whether the digital signal representation of the item is sufficient to identify the item as genuine. It should be understood that this means whether the digital signal representation of the item is sufficient to cause the authenticity detection algorithm (obtained, for example, in step a.3) of the method) to identify the item as genuine. In Example Embodiment 1 (shown by Figure 5 Fig.) and Example Embodiment 2 (shown by Figure 6 Fig.), when the digital signal representation of the item is sufficient to cause the authenticity detection algorithm (obtained, for example, in step a.3) of the method) to identify the item as genuine, the authenticity detection algorithm returns an authenticity decision as output according to the digital signal representation of the item to be identified.

[0331] It is obvious to those skilled in the field of authenticity detection that the system architecture selection in various Example Embodiments 1 or 2 can depend on application-related requirements, such as the respective computational performances of the authenticity detection algorithm (100) and the detectability prediction (200), and / or the ratio of genuine items to fake items to be processed by the system, and / or the degree of exposure to reverse engineering of the authenticity detection algorithm (100). (Therefore, it is preferably to use Embodiment 2 - Figure 6 ).

[0332] In a possible embodiment, the detectability prediction model (200) and the authenticity detection algorithm (100) can run on the same device (e.g., as a smartphone application or on a remote server connected to a local device that operates sensors and captures digital signal representations). In another possible embodiment, the detectability prediction model (200) and the authenticity detection algorithm (100) can run on different devices. For example, in one embodiment, the detectability prediction model can first run as a front-end application on a local device, and the detection algorithm can then run as a back-end application on a remote server, which is only invoked when the predicted item authenticity detectability value from the front-end indicates that the authenticity detection algorithm can detect the item as genuine (i.e., following Figure 6 's method steps). This approach is particularly advantageous when the authenticity detection algorithm requires complex computations and the processing speed of the user device is not fast enough, or when the brand owner needs server-side control over the authentication application. It is also advantageous when it is necessary to isolate the authenticity detection algorithm from hacking attempts, as access is only permitted when the predicted item authenticity detectability value indicates that the authenticity detection algorithm can detect the item as genuine. Conversely, in an alternative embodiment, the authenticity detection algorithm model can first run as a front-end application on a local device, and the detectability prediction model can then run as a back-end application on a remote server, which is only invoked when the front-end authenticity detection algorithm fails to detect the item as genuine (i.e., following Figure 5 's method steps).

[0333] In one embodiment, in a computer-implemented method for identifying whether an item is genuine or counterfeit, where the item to be identified is marked as undetectable, at least 2, at least 5, at least 10, or at least 50 additional (or another) digital signal representations of the item to be identified can be obtained, preferably at least 50, and processed again by the method (as Figure 5 and Figure 6 shown). The method for identifying whether an item is genuine or counterfeit repeatedly processes the additional digital signal representations of the item to be identified until a counterfeit is identified or until a maximum allowed number of cycles (timeout) has been reached. In one embodiment, the maximum allowed number of cycles is reached when the item is predicted to be detectable in at least 20% of the trials.

[0334] In one embodiment, the method for identifying whether an item is genuine or counterfeit repeatedly processes the additional digital signal representations of the item to be identified until a genuine item is identified or until a maximum allowed number of cycles is reached. The maximum allowed number of cycles is reached when the item is predicted to be detectable in at least 20% of the trials.

[0335] In one embodiment, a data processing device is provided that includes means for performing the method of the present invention described herein.

[0336] In one embodiment, a data processing device including instructions is provided that, when executed by the device, cause the device to perform the method of the present invention described herein.

[0337] In one embodiment, a computer-readable medium including instructions is provided that, when executed by a computer, cause the computer to perform the method of the present invention described herein.

[0338] It should be understood that the method of the present invention is a computer-implemented or computer-based method.

[0339] Examples

[0340] Example 1 - Different Signal Representations of the Same Item under Different Capture Conditions

[0341] Illustrated is the preparation of a training set for training a predictive machine learning model according to the method of the present invention, where an item can be used for training and testing algorithms.

[0342] Thirteen images (i.e., digital signal representations) of a genuine item (here, an electronic circuit) have been obtained (i.e., captured) by a sensor (here, a smartphone camera). These images are input into an authenticity detection algorithm (authentication algorithm).

[0343] In this example, the output authenticity detectability value of the item is a binary label, such as detectable or not detectable as genuine.

[0344] Figure 7 Shown are the 13 images of the genuine item, i.e., a single set, i.e., an item has only its 13 digital signal representations. In two images (image references 7A and 7B), the item can be recognized as genuine (i.e., the relevant item authenticity detectability value is detectable). In the other eleven images, the item cannot be recognized as genuine (i.e., the relevant item authenticity detectability value is not detectable) (image references 7C to 7M). Examples of images where the item cannot be recognized as genuine include images where the item is incompatible with the image background ( Figure 7 C), images where the item is blurred ( Figure 7 D), images where the item is blurred and the item in the image is cropped ( Figure 7 E), images where color distortion is introduced making the item unclear ( Figure 7 F), images where the item in the image is cropped ( Figure 7 G), images where there is interfering light making the item unclear ( Figure 7 H), images where the item is overexposed ( Figure 7I), an image of an article that is small in the image ( Figure 7 J), an image of an article that is in an inclined position in the image ( Figure 7 K), an image of an article with insufficient exposure ( Figure 7 L), an image in which only a part of the article is present and a magnified portion of the article is shown ( Figure 7 M).

[0345] Example 2 - A training method and a testing (recognition) method using different signal representations of multiple articles under different capture conditions

[0346] Illustrates, within the framework of a method for identifying whether an article is genuine or counterfeit, a computer-implemented method for preparing a training set (Example 2A) in a method for training a predictive machine learning model and a test set (Example 2B) for testing the trained predictive machine learning model (from 2A). Multiple articles of one type can be used for training and testing the algorithm.

[0347] 2A) - The training set used in the method for training a predictive machine learning model

[0348] Four genuine articles of the same type were obtained (labeled 1, 2, 3, and 4). Four images (i.e., digital signal representations) of genuine article 1 were obtained by the sensor (labeled (object.capture)1.1, 1.2, 1.3, 1.4; Table 1, column 1). Four images of genuine article 2 were obtained by the sensor (labeled 2.1, 2.2, 2.3). Four images of genuine article 3 were obtained by the sensor (labeled 3.1, 3.2, 3.3, 3.4). Four images of genuine article 4 were obtained by the sensor (labeled 4.1, 4.2, 4.3, 4.4). Thus, four sets of digital signal representations of each genuine article (all of the same type) were obtained, forming a training set consisting of 15 digital signal representations (4 + 3 + 4 + 4 = 15; Table 1, column 2). It should be understood that each image was captured under given capture conditions. These images were input into an authenticity detection algorithm (authentication algorithm), which returned an authentication decision of "pass" or "no available result" (i.e., outputting the authenticity detectability value of each digital signal representation of each genuine article. The output is the authenticity detectability value of the article in the form of a binary label, which can be detected as genuine, meaning authentication is successful ("pass"), or not detected as genuine ("no available result"). As a reminder, "detectable" and "detected" are different. In this way, the relevant authenticity detectability values of the article were generated from 4 sets of 15 digital signal representations (4 sets for each of the 4 genuine articles used for training) (Table 1, column 4). These sets consisting of the digital signal representations of each genuine article and the relevant authenticity detectability values (i.e., 4 sets of images; 1 set for each of the 4 genuine articles; a total of 15 images as the training set) were used to train a predictive machine learning model to predict the authenticity detectability value of an article based on the digital signal representation of the article.

[0349] Table 1

[0350]

[0351] 2B) - Test set used in a method for identifying whether an article is genuine or counterfeit

[0352] Illustrated is a computer-implemented method for identifying whether an article is genuine or counterfeit according to the present invention, in which a predictive machine learning model for predicting the authenticity detectability value of an article to be identified is trained according to Example 2A, and in which the article to be identified is of the same type as the articles used for training purposes in Example 2A.

[0353] Five articles with known properties were used in the test, namely three genuine articles marked 5, 7, and 8; and two forged articles marked 6 and 9 (Table 2, column 2), as the articles to be authenticated / identified. One or more images (i.e., digital signal representations) of these articles (marked (object.capture) 5.1, 5.2, etc.) (Table 2, column 1) have been obtained by the sensor. It should be understood that each image was captured under given capture conditions.

[0354] According to Figure 5 Example 1 shown, each digital signal representation of the article to be identified is input into the same algorithm as the authenticity detection algorithm used during training. The output is that the article to be identified is identified as genuine ("passed"), or the article to be identified cannot be identified as genuine ("no available result") (Table 2, column 3). If the article cannot be identified as genuine, then each digital signal representation of the article to be tested is input into a predictive machine learning model trained according to the method of Example 2A. The output is the predicted article authenticity detectability value of the digital signal representation of the article to be identified (Table 2, column 4). This allows determining whether the authenticity detection algorithm (for authentication) can or cannot detect the article as genuine based on the predicted article authenticity detectability value. It should be noted that if the article can be identified as genuine by the authenticity detection algorithm in the first step, then there is no need to apply detectability prediction, because there is no need to distinguish between forgeries and non-detectable articles.

[0355] Finally, the final result of article authentication is provided (Table 2, column 5), which is a decision based on the output from the predictive machine learning model (Table 2, column 4) and the output from the authenticity detection algorithm (Table 2, column 3).

[0356] As summarized in Table 2:

[0357] - Digital signal representations 5.1, 7.1, 7.2, 7.3, 7.4, 8.2, 8.3, 8.4: If the article is detected as genuine by the authenticity detection algorithm (100) (column 3), then it can be concluded that the article is detectable as genuine (without detectability value prediction (200), column 4), and then an output of genuine is provided (column 5);

[0358] - Digital signal representations 6.1, 9.1: If the article is not identified as genuine by the authenticity detection algorithm (column 3), and the predictive machine learning model indicates that the article is detectable as genuine (column 4), then an output of counterfeit (forged) article is provided (column 5);

[0359] - Digital signal representations 5.2, 8.1, 9.2: If the item is not recognized as genuine by the authenticity detection algorithm (column 3), and the predictive machine learning model indicates that the item cannot be detected as genuine (column 4), then an output of "not authenticated" item is provided (column 5).

[0360] Table 2

[0361]

[0362] More applications

[0363] When testing the authentication limitations of euro banknotes with the ValiCash application available in the Apple App Store as of July 2023, the advantages of the proposed invention can be further illustrated. This application is an example of an authenticity detection algorithm that generates an authenticity detectability value for euro banknotes based on digital image representations of the euro banknote artwork captured using a smartphone camera. The application is very sensitive to capture conditions; due to the overlay on the screen to align with the banknote pattern, this forces the end user to manually geometrically position the smartphone relative to the banknote correctly (e.g., ensure the banknote is flat). However, even with correct geometric positioning, especially under certain capture conditions (light reflection, dirty camera), the algorithm cannot detect the authenticity of the euro banknote, even if it is genuine. Then, the application suggests manual inspection steps to the end user (visually inspect the intaglio printing, manual serial number, watermark). But the end user will still be suspicious whether the banknote is fake or genuine and is likely to consider it fake. In contrast, the proposed invention can mainly detect whether the capture conditions are suitable for the underlying authenticity detection algorithm, thus enabling the redirection of the user to manually check the capture conditions rather than the banknote features. For example, the application can suggest cleaning or drying the sensor lens in a foggy or dusty environment, and / or avoiding placing the item under a transparent plastic film or glass sheet to keep it flat, as this may cause flash reflections on the plastic or glass surface, which may interfere with the digital image representation signal processing for authenticity detection.

[0364] Other examples of applications that can be improved due to the proposed invention include:

[0365] - Gold bar security features that use a high-resolution image processing algorithm to verify the authenticity of gold bars according to the independent authoritative standards of the London Bullion Market Association (https: / / www.lbma.org.uk / good-delivery / gold-bar-security-features#-);

[0366] - A hidden anti-counterfeiting technology that uses digital image processing algorithms to detect counterfeit goods, such as products manufactured in the tobacco industry; drugs that comply with the European False Medicines Directive (FMD) regulations and / or the anti-counterfeiting protocol of the US Drug Quality and Security Act (DQSA); and any goods or security documents that typically develop digital anti-counterfeiting technologies based on digital signal processing algorithms.

Claims

1. A computer-implemented method for training a predictive machine learning model to predict the authenticity detectability value of an item, the method comprising: a) obtaining an authenticity detection algorithm that generates a detectability value of the authenticity of an item type based on a digital signal representation of the item, wherein the detectability value determines whether the digital signal representation of the item is sufficient to identify the item as genuine; b) obtaining one or more genuine items of the same type; c) obtaining, via a sensor, a set of digital signal representations of each of the one or more genuine items; wherein each digital signal representation of each genuine item is obtained by operating the sensor under different capture conditions; d) inputting each digital signal representation of each genuine item (obtained in step c) into the authenticity detection algorithm (obtained in step a), and outputting an item authenticity detectability value for each digital signal representation of each genuine item; e) using the set of digital signal representations of each of the one or more genuine items, and the associated item authenticity detectability values (obtained in step d), to train a predictive machine learning model to predict the authenticity detectability value of the item based on the digital signal representation of the item.

2. The computer-implemented method for training a predictive machine learning model according to claim 1, wherein, The authenticity detection algorithm is selected from a surface fingerprint detector and a product identification detector.

3. A computer-implemented method for training a predictive machine learning model according to any one of claims 1-2, wherein, A computer-implemented method using the trained predictive machine learning model to identify whether an item is genuine or counterfeit.

4. A computer-implemented method for training a predictive machine learning model according to any one of claims 1-2, wherein, The authenticity detection algorithm generates an authenticity detectability value of the item, wherein the detectability value is a scalar value or a label.

5. The computer-implemented method for training a predictive machine learning model according to any one of claims 1-4, wherein, Each digital signal representation of each of the one or more genuine items is obtained based on a signal captured by a sensor, such as an image sensor, a digital olfactory sensor, a digital chemical sensor, a microphone, a microtext reader, a barcode reader, a QR code reader, a laser-based sensor, a code reader, an RFID reader, an infrared sensor, a UV sensor, a digital camera, or a smartphone camera.

6. The computer-implemented method for training a machine learning model according to any one of claims 1-5, wherein, The relative position and / or orientation of the sensor and the genuine item are different to produce different sensor capture conditions.

7. The computer-implemented method for training a machine learning model according to claim 6, wherein, The different sensor positions and / or orientations are controlled by a robotic arm or an operator who positions the sensor relative to the genuine item.

8. The computer-implemented method for training a machine learning model according to claim 6, wherein, The different genuine item positions and / or orientations are controlled by a robotic arm, a conveyor, or an operator who positions the genuine item relative to the sensor.

9. A computer-implemented method for training a machine learning model according to any one of claims 1-8, wherein, Each digital signal representation of each of the one or more genuine items is obtained under capture conditions characterized by physical environmental parameters around the genuine item.

10. The computer-implemented method for training a machine learning model according to claim 9, wherein, Each digital signal representation of each of the one or more genuine items is a digital image representation of the genuine item, wherein at least one physical environmental parameter around the genuine item varies between different capture conditions, wherein the parameter is the illumination of the genuine item by an illumination device, and wherein the illumination of the genuine item is controlled by a robotic arm, a conveyor, or an operator who positions the illumination device relative to the genuine item.

11. The computer-implemented method for training a machine learning model according to claim 10, wherein, The sensor is a smartphone camera, and the lighting device is a smartphone flash.

12. A computer-implemented method for predicting an authenticity detectability value of an item, the method comprising: a) obtaining the item to be detected; b) obtaining, via a sensor, a digital signal representation of the item to be detected; c) obtaining a predictive machine learning model to predict the authenticity detectability value of the item, wherein the predictive machine learning model is trained according to the method of any one of claims 1-11; d) inputting the digital signal representation of the item to be detected (obtained in step b) into the predictive machine learning model (obtained in step c), and outputting, for the digital signal representation of the item to be detected, a predicted authenticity detectability value of the item.

13. A computer-implemented method for identifying whether an item is genuine or counterfeit, the method comprising: a) obtaining a.1) the item to be identified; a.2) the digital signal representation of the item to be identified, via a sensor; a.3) an authenticity detection algorithm that returns an authenticity decision as an output based on the digital signal representation of the item to be identified; a.4) a predictive machine learning model to predict the authenticity detectability value of the item to be identified using the authenticity detection algorithm (obtained in step a.3), wherein the predictive machine learning model is trained according to the method of any one of claims 1-11; b) inputting the digital signal representation of the item to be identified (obtained in step a.2) into the authenticity detection algorithm (obtained in step a.3) and outputting - the item to be identified is identified as genuine, or - the item to be identified cannot be identified as genuine, and if the item to be identified cannot be identified as genuine, then c) inputting the digital signal representation of the item to be identified (obtained in step a.2) into the predictive machine learning model (obtained in step a.4), and outputting, for the digital signal representation of the item to be identified, a predicted authenticity detectability value of the item; d) determining, based on the predicted authenticity detectability value of the item (obtained in step c), the authenticity detection algorithm (obtained in step a.3) - to be able to detect the item as genuine and determine that the item to be identified is counterfeit, or - the item to be identified cannot be identified as genuine, and optionally determining that the item to be identified is undetectable.

14. The computer-implemented method for identifying whether an item is genuine or counterfeit according to claim 13, wherein, In step d), the item to be identified cannot be identified as genuine, and the item to be identified is determined to be undetectable, wherein the method further comprises the step of: e) obtaining, via a sensor, another digital signal representation of the item to be identified; f) inputting the another digital signal representation of the item to be identified (obtained in step e) into the authenticity detection algorithm (obtained in step a.3) and outputting - the item to be identified is identified as genuine, or - the item to be identified cannot be identified as genuine, and if the item to be identified cannot be identified as genuine, then g) Input the digital signal representation of the item to be identified (obtained in step e)) into the predicted machine learning model (obtained in step a.4)), and output another predicted item authenticity detectability value for the digital signal representation of the item to be identified; h) Determine the authenticity detection algorithm (obtained in step a.3)) based on the predicted item authenticity detectability value (obtained in step g)) - capable of detecting the item as genuine and determining the item to be identified as a counterfeit, or - the item to be identified cannot be identified as genuine, and optionally determine the item to be identified as undetectable, and repeat steps e) to h) at least once.

15. A computer-implemented method for identifying whether an item is genuine or a counterfeit, the method comprising: a) Obtain a.1) The item to be identified; a.2) The digital signal representation of the item to be identified, through a sensor; a.3) An authenticity detection algorithm that returns an authenticity decision as an output from the digital signal representation of the item to be identified, a.4) A predicted machine learning model to predict the authenticity detectability value of the item to be identified using the authenticity detection algorithm (obtained in step a.3)), wherein the predicted machine learning model is trained according to the method of any one of claims 1-11; b) Input the digital signal representation of the item to be identified (obtained in step a.2)) into the predicted machine learning model (obtained in step a.4)), and output a predicted item authenticity detectability value for the digital signal representation of the item to be identified; c) Based on the predicted item authenticity detectability value (obtained in step b)), determine that the authenticity detection algorithm (obtained in step a.3)) can detect the item as genuine; d) Input the digital signal representation of the item to be identified (obtained in step a.2)) into the authenticity detection algorithm (obtained in step a.3)) and output - the item to be identified is identified as genuine, or - the item to be identified cannot be identified as genuine, and determine the item to be identified as a counterfeit.

16. The computer-implemented method for identifying whether an item is genuine or counterfeit according to claim 15, wherein, In step c), the item to be identified cannot be identified as genuine, and the item to be identified is determined to be undetectable, wherein the method further comprises the steps: e) Obtain another digital signal representation of the item to be identified through a sensor; f) Input the another digital signal representation of the item to be identified (obtained in step e)) into the predicted machine learning model (obtained in step a.4)), and output another predicted item authenticity detectability value for the digital signal representation of the item to be identified; g) Based on the predicted item authenticity detectability value (obtained in step f)), determine that the authenticity detection algorithm (obtained in step a.3)) can detect the item as genuine, and then perform step h) or - The item to be identified cannot be identified as genuine, and optionally determine that the item to be identified is undetectable, and repeat steps e) to g) at least once; h) Input the other digital signal representation of the item to be identified (obtained in step e)) into the authenticity detection algorithm (obtained in step a.3)) and output - The item to be identified is identified as genuine, or - The item to be identified cannot be identified as genuine, and determine that the item to be identified is a fake.

17. A computer-implemented method for identifying whether an item is genuine or counterfeit according to any one of claims 13-16, wherein, The authenticity detection algorithm is selected from a surface fingerprint detector and a product identification detector.

Images