WO2024042240A1

WO2024042240A1 - Systems and methods for predicting the authentication detectability of counterfeited items

Info

Publication number: WO2024042240A1
Application number: PCT/EP2023/073442
Authority: WO
Inventors: Yves Delacretaz; Frederic Jordan; Martin Kutter
Original assignee: Alpvision S.A.
Priority date: 2022-08-26
Filing date: 2023-08-25
Publication date: 2024-02-29
Also published as: EP4328879A1

Abstract

The present invention is directed to a computer-implemented method for identifying if an object is a genuine or a counterfeited object, wherein the method uses a combination of output from a predictive machine learning model and from a genuineness detection algorithm. In particular, the invention relates to methods for identifying a counterfeited object.

Description

Systems and methods for predicting the authentication detectability of counterfeited items

Field

The present invention relates to the machine learning systems, methods and processes in the field of anti-counterfeiting and authentication of manufactured items, products, security documents and banknotes.

Background

In order to determine whether product items are genuine, various anti-counterfeiting technologies may be used (Anticounterfeiting technology guide, European Union Intellectual Property Office, 2021). In general, these technologies add an element onto the item that is difficult to duplicate, or copy, or they characterize a specific physical or chemical feature of the item, similar to a fingerprint of the item. The challenge may be either technical, for instance on the reproduction of holograms, or require products which are not readily available on the market, such as rare isotopes or special inks. In general, anti-counterfeiting features may be classified as overt technologies (visible, or more generally perceptible by the end user with his own body senses, without the need for specific detection equipment) or covert technologies (invisible/imperceptible, but detectable with a dedicated equipment). Example of covert technologies include:

• Product markings: Technologies such as digital watermarks have been designed to better prevent the counterfeiting of product packages and security documents by electronic and digital means. As a widely deployed example of such a technology, the AlpVision Cryptoglyph exists in two flavors, either as a random or pseudo-random pattern of microdots printed with visible ink (WO0225599, W004028140), or as a distributed cloud of micro-holes in the varnish layer (W006087351). The distribution of microdots or micro-holes can be controlled with a secret cryptographic key. Authentication can be performed using conventional imaging devices, such as smartphones or off-the-shelf office scanners, in combination with dedicated signal processing software. Product markings comprise digital markings, chemical markings, hologram and the like.

• Surface fingerprints: These technologies do not add any security element on the product, but rather use existing, intrinsic microscopic surface characteristics. For instance, a matte surface of a plastic injected product is an ideal candidate for the fingerprint solution. An image of the surface may be acquired during production and then compared to a later image capture from the product under inspection. Authentication can be performed using conventional imaging devices, such as smartphones or off-the-shelf office scanners, in combination with dedicated signal processing software. Examples of fingerprinting technologies are described for instance in US10332247.

• Invisible Ink: These special inks are invisible but become visible when exposed to the appropriate light source. A well-known example is UV (Ultra-Violet) or IR (Infra-Red) inks. Authentication can be performed using specialized imaging devices incorporating dedicated lighting means, in combination with image analysis methods, to characterize the presence of special inks on the object to be authenticated (e.g., EP1295263, US6903342).

• Chemical Taggants: A large number of chemical taggants exist. In general, these taggants are invisible and detectable in laboratory conditions. Authentication can be performed using a digital olfactory sensor and dedicated signal analysis methods to characterize the presence of chemical tags on the object to be authenticated (e.g.,W02014201099, W02020160377).

• Micro-graphics: The artwork can be secured by adding very small graphical elements which are not visible without visual magnification. Micro-text is one of the most popular realizations of this security feature (e.g., WO2012131474).

The authentication of an object for any of the above technologies consists then in identifying, with a detector adapted to the particular authentication technology employed for this object, whether the authentication technology can be retrieved from inspecting the object. In the past two decades, digital detection technologies have emerged which have facilitated the automation of this process and its generalization to non-specialized personnel, possibly also the general public, thanks to the use of digital signal processing algorithms embedded into software applications either embedded into the detector equipment (e.g., a smartphone) or executed on a computer in communication with the detector equipment through a communication network.

These digital authentication detection methods generally employ the following steps:

1) capture (with a sensor) a digital signal representation of the object to be authenticated (for instance, an image of the object surface by a camera sensor, an RFID signal reading by an RFID sensor, a chemical signal from a digital olfactory sensor, etc.);

2) process (with computer-implemented signal processing algorithms) the digital signal representation to characterize the object genuineness - for instance, by measuring a difference, a distance or a signal-to-noise ratio (SNR) between the captured digital signal representation and a template digital signal representation of a reference genuine object; or by extracting mathematical features from the digital signal representation which can be used by a classifier to discriminate between a fake and a genuine object.

3) classify (with computer-implemented authentication classification or decision-making algorithms) the object as genuine or fake depending on the measurement (for instance, using a predetermined threshold or using a machine learning classifier).

In the past decade, a number of authentication technologies have less and less used dedicated sensing equipment (such as flatbed scanners and professional cameras with dedicated optics) to make use of consumer electronics equipment (such as smartphone cameras, for instance). The diversity of the digital signal representation capture conditions raises specific challenges in terms of detecting the authentication technology signals, which impacts the reliability of the automated discrimination between counterfeited and genuine objects. Accordingly, many authentication algorithms have been fine-tuned to ensure the detection of a genuine object as a true positive classification, but it remains challenging to rely upon the failed detection of the expected digital signal representation of a genuine object to confirm that an object is counterfeited. For instance, as an approach based on machine learning, WO2015157526 describes the use of convolutional neural networks to classify genuine vs counterfeited objects. The latter method employs training sets comprising both genuine objects and fake objects, in combination with data augmentation to facilitate the training. The latter method requires the brand owner to collect multiple fake samples which are representative enough of the ability of counterfeiters to reproduce the original products. This creates additional burden to organize and maintain in anticounterfeiting long-term operations. There is still a risk of wrong classifications of genuine objects as fake ones (false negative classification) or more generally, too many doubtful cases.

In an ideal world, a digital authentication detection method applied to a perfect digital signal representation of a genuine object will always enable to detect it as genuine. In other words, the genuineness would be “100% detectable” or “always detectable” by applying the digital authentication detection method on a perfect digital signal representation of a genuine object. However, in the real -world authentication conditions, the detectability of the object genuineness depends on the quality of the digital signal representation of the object. This quality depends itself on multiple, variable digital signal capture factors; for instance, in the case of an imaging capture (but not limited to):

• how an image of the genuine object surface has been captured: - depending on the object position and orientation relative to the sensor position and orientation - especially when acquiring images with a handheld device camera, as opposed to a flatbed scanner;

- depending on the imaging sensor, for instance the camera resolution, focal length, aperture, field of view, white balance;

- depending on the lighting environment, with shades, reflections etc varying according to the state of surface of the object (which may even be a used object, for instance a 2^nd market luxury watch) as well as the physical environment of the capture (camera flash on or off, ambient lighting indoor or outdoor, etc.),

• how the object has aged or how it has been handled over its lifetime, e.g., dirt surfaces, scratches, colour loss due to UV exposure, deformation due to shocks or wet environment, oxidation, etc.

Historically, most of the digital detection methods of the past decade have required significant R&D and testing to fine-tune the algorithms to process and classify as diverse as possible digital signal representations of the objects to be authenticated. A number of solutions currently deployed in the authentication market propose guidelines for the user to optimize the capture conditions: for instance, visual guides for properly orienting the smartphone capture and/or capturing a video as a series of multiple digital representations and/or repeating the capture from different positions until one enables to detect the object as genuine.

Currently, it is still impossible to reach a perfect detection paradigm in the real world. It is therefore not possible to determine whether the failure of detection of the authentication technology on an object is caused by the object being a fake or caused by the limit of detection due to the capture conditions of the object, which may still be a genuine one. For instance, end users using the Veriscan fingerprint smartphone app technology from the Swiss precious metals supplier PAMP to verify the genuineness of a PAMP -branded minted gold bullion only get the result of either “pass” (=detected as genuine) or “no result available” (=not detectable as genuine, which can mean either genuine but not detectable from the smartphone imaging, or not detectable due to a fake bullion) (https://www.pamp.com/veriscan/).

There is therefore a need for improved methods and systems to predict the detectability of counterfeited objects and to reliably classify objects as counterfeited objects when using any digital authentication algorithm of the prior art.

It is an object of the invention to provide improved methods and systems to predict the detectability of counterfeited objects. Summary

The present invention is based on the finding that the use of a predictive machine learning model to predict a detectability value of the genuineness of an object in combination with a genuineness detection algorithm (also referred herein as an authentication algorithm) allows to identify genuine and counterfeited objects. Further, 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 sets of digital signal representations of genuine objects with their associated detectability value of the genuineness of an object. The predictive machine learning model allows to predict a detectability value of the genuineness of an object to be identified and the genuineness detection algorithm allows to determine from the predicted detectability value of the genuineness of an object if the genuineness detection algorithm can or cannot detect the object as a genuine object. Therefore, the predicted detectability value identifies whether a digital signal representation of the object to be identified is sufficient for the genuineness detection algorithm to identify this object as genuine. The predicted detectability value is a result of the environment in which the digital signal representation of an object has been obtained (captured) with a sensor and can be also defined as the detectability value of the genuineness measurement of an object. In next step, if the objects to be identified can be detected as a genuine object than if the genuineness detection algorithm does not identify this object as genuine, it can be determined that this is counterfeited.

In one embodiment provided is computer-implemented method for training a predictive machine learning model to predict a detectability value of the genuineness of an object, the method comprising the steps to: a) obtain a genuineness detection algorithm, which produces a detectability value of the genuineness of a type of object from a digital signal representation of this object, wherein the detectability value determines whether a digital signal representation of this object is sufficient to identify this object as genuine; b) obtain one or more genuine objects of one type; c) obtain, with a sensor, a set of digital signal representations of each of the genuine objects, wherein each digital signal representation of each of the genuine objects is obtained from the sensor operating under a different capture condition; d) input each of the digital signal representation of each of the genuine objects (obtained in step c)) to the genuineness detection algorithm (obtained in step a)) and output a detectability value of the genuineness of the object, so that each of the digital signal representation of each of the genuine objects has the associated detectability value of the genuineness of an object; e) train a predictive machine learning model to predict the detectability value of the genuineness of an object using the sets of digital signal representations of each of the genuine objects with the associated detectability values of the genuineness of an object (obtained in step d).

In another embodiment provided is a computer-implemented method for predicting a detectability value of the genuineness of an object, the method comprising the steps to: a) obtain an object to be detected; b) obtain, with a sensor, a digital signal representation of the object to be detected; c) obtain a predictive machine learning model to predict a detectability value of the genuineness of an object, wherein the predictive machine learning model is trained according to the method of the invention; d) input the digital signal representation of the object to be detected (obtained in step b)) to the predictive machine learning model (obtained in step c)) and output a predicted detectability value of the genuineness of an object for the digital signal representation of the object to be detected.

In another embodiment provided is a computer-implemented method for identifying if an object is a genuine or a counterfeited object, the method comprising the steps to: a) obtain a.1) an object to be identified; a.2) with a sensor, a digital signal representation of the object to be identified; a.3) a genuineness detection algorithm, which returns as an output a genuineness decision from the digital signal representation of the object to be identified, a.4) a predictive machine learning model to predict a detectability value of the genuineness of the object to be identified with the genuineness detection algorithm (obtained in step a.3)), wherein the predictive machine learning model is trained according to the method of the invention; b) input the digital signal representation of the object to be identified (obtained in step a.2)) to the genuineness detection algorithm (obtained in step a.3)) and output

- that the object to be identified is identified as genuine, or

- that the object to be identified cannot be identified as genuine, and if the object to be identified cannot be identified as genuine then c) input the digital signal representation of the object to be identified (obtained in step a.2)) to the predictive machine learning model (obtained in step a.4)) and output a predicted detectability value of the genuineness of an object for the digital signal representation of the object to be identified; d) determine from the predicted detectability value of the genuineness of an object (obtained in step c)) that the genuineness detection algorithm (obtained in step a.3))

- can detect the object as a genuine object and determine the object to be identified as a counterfeited object, or

- that the object to be identified cannot be identified as genuine, and optionally determine the object to be identified as not-detectable.

Description of the figures

Figure 1 shows an example of a processing workflow of a prior art authentication algorithm.

Figure 2 shows an example of a processing workflow of a method for predicting a detectability value of the genuineness of an object with the use of the predictive machine learning model (noted on figure as a “ML model”).

Figure 3 panel a) shows a robot arm (330) manipulating an imaging device (310) to capture multiple pictures of a genuine product element (320); panel b) shows a robot arm (330) manipulating an imaging device (310) to capture multiple pictures as digital representations of a genuine banknote physical object (340).

Figure 4 panel a) illustrates the side view of a system comprising a fixed imaging device and two robot arms (410, 420), one manipulating a genuine physical object (101) and one manipulating a lighting device (430); panel b) illustrates the side view of a system comprising a conveyor (450) for carrying a genuine physical object (101) and two robot arms (410, 420), one manipulating an imaging device and one manipulating a lighting device (430).

Figure 5 shows an example of a processing workflow of a method for identifying if an object is a genuine or a counterfeited object according to one embodiment (embodiment 1).

Figure 6 shows another example of a processing workflow of a method for identifying if an object is a genuine or a counterfeited object according to another embodiment (embodiment 2).

Figure 7 shows 13 images of a genuine object as described in Example 1. Based on two images (panel 7A & 7B) the object can be recognized as genuine and eleven images based on which the object cannot be recognized as genuine (panels 7C to 7M).

Detailed description The terms “object” or “item” or “object item” (used interchangeably) refer to something material that may be perceived by the senses. It can be a manufactured object or an artisanal object. Examples of objects include, but are not limited to a security document, a precious metal, a banknote, a watch, a leather product such as a bag, a part of an object such as a component, a label, a package, a printed surface, an embossed surface, a metallized surface, and the like. The objects that have the same characteristics belong to the same / one “type of objects” or “class of objects”.

The terms “genuine object” or “real object” or “authentic object” (used interchangeably) refer to an object that is exactly what it appears to be, and is not false or an imitation. In other words, a genuine object is an original, a real, an authentic, not fake or counterfeit. The genuine object may have added or integrated security features that can be detected by an authentication algorithm.

The terms “counterfeit object” or “fake object” or “not authentic object” (used interchangeably) refer to an object or item that is made to be an imitation of something genuine and meant to be taken as genuine, and is false or an imitation. In other words, a counterfeit object is a forgery, copy, imitation, or fake.

The “genuine/counterfeit object” can be characterised as having “a detectability property”, which allows to identify if an object is detectable as genuine/counterfeit or not detectable as genuine/counterfeit. The detectability property can be identified or recognised based on a detectability value of an object. In the context of the present invention “a detectability value of the genuineness of an object” or “a detectability value of a genuineness” is used. The “detectability value of the genuineness of an object” or “a detectability value” refers to a value identified from a digital signal representation of this object, wherein the detectability value identifies whether a digital signal representation of this obj ect is sufficient to identify this obj ect as genuine. In the context of the invention, the predicted detectability value identifies whether a digital signal representation of an object is sufficient for the genuineness detection algorithm to identify this object as genuine. “Sufficient” means that a digital signal representation of an object (and not an object) has an adequate signal (or quality of signal) for the genuineness detection algorithm to identify an object as genuine. Illustration of whether a digital signal representation of an object is sufficient to identify this object as genuine are shown in Example 1. Thus, “detectability value of the genuineness of an object” does not relate to the object, but to the digital signal representation of this object. In other words, it is a “detectability value of the genuineness measurement of an object”. Therefore, it is not a function of an object, but rather function of the environment in which the digital signal representation of this object has been obtained (captured) with a sensor. Examples of detectability value of the genuineness of an object include but are not limited to a binary label such as detectable or not detectable, a ternary label such as detectable or not detectable or unknown, or a scalar value such as a signal to noise ratio (SNR) measurement, a difference measurement, a distance metrics, or the like known in the art. In a possible embodiment the label may be 0 for non-detectable and 1 for detectable (see Example 2). In an alternate possible embodiment, the label may be 1 for non- detectable and 0 for detectable. In a possible embodiment the detectability value may be a scalar value. In a possible embodiment, the detectability value may be a signal processing metrics. A signal processing metrics may be for instance the signal-to-noise ratio (SNR) of a cross-correlation of the captured digital signal representation with a template digital signal representation reference for the object to be authenticated. In an alternate possible embodiment, the detectability metrics may also be a simple distance measurement (for instance, a difference) between one extracted feature from the captured digital signal representation and the matching reference feature from a digital signal representation template. In an alternate possible embodiment, the detectability metrics may be a composite distance measurement between a set of features from the captured digital signal representation and the matching set of reference features from a digital signal representation template. Examples of composite distance measurement include the L0, LI, L2 norms and other ways of measuring distances between sets of values in statistic modelling.

The terms “a digital signal representation” or “a digital representation” of an object refer to a representation of on object in the form of digital data. Examples of a digital signal representations of an object include but are not limited to a binary image, a digital sound record, chemical composition, spectral representation of a wave acquired by a spectrometer hardware, being electromagnetic as in a case of images, or mechanical/pressure as in the case of sound or the like, or a combination of the above in the case of multi-modal capture. The digital signal representation of an object may be obtained from acquiring a signal captured with a sensor. In a one embodiment, a digital signal representation of an object is a binary image.

In a one embodiment, a digital signal representation of an object is obtained (i.e., captured), with a sensor. Each of digital signal representations of an object may be obtained from the sensor operating under a different capture condition.

The term “prediction” refers to inferring, with a statistical analysis model or a predictive machine learning model, a detectability value from a digital signal representation of an object. Prediction may be defined as a mean to output a value of potentially multiple dimensions, from a potentially multi-dimensional input value never seen before, by using a model. The model can come from a set of acquired observations or it can be an analytical/a priori model defined from a set of known relationships.

A “machine learning model” refers to a data model or a data classifier which has been trained using a supervised, semi-supervised or unsupervised learning technique as known in the data science art, as opposed to an explicit statistical model. The data input may be represented as a ID signal (vector), a 2D signal (matrix), or more generally a multidimensional array signal (for instance a tensor, or a RGB color image represented as 3 *2D signals of its Red, Green and Blue color decomposition planes - 3 matrices), and/or a combination thereof. A multidimensional array is mathematically defined by a data structure arranged along at least two dimensions, each dimension recording more than 1 value.

In the case of a deep learning classifier, the data input is further processed through a series of data processing layers to implicitly capture the hidden data structures, the data signatures and underlying patterns. Thanks to the use of multiple data processing layers, deep learning facilitates the generalization of automated data processing to a diversity of complex pattern detection and data analysis tasks. The machine learning model may be trained within a supervised, semi-supervised or unsupervised learning framework. Within a supervised learning framework, a model learns a function to map an output result from an input data set, based on example pairs of inputs and matching outputs. Examples of machine learning models used for supervised learning include Support Vector Machines (SVM), regression analysis, linear regression, logistic regression, naive Bayes, linear discriminant analysis, decision trees, k- nearest neighbor algorithms, random forest, artificial neural networks (ANN) such as convolutional neural networks (CNN), recurrent neural networks (RNN), fully-connected neural networks, long short-term memory (LSTM) models, and others; and/or a combination thereof. A model trained within an unsupervised learning framework infers a function that identifies the hidden structure of a data set, without requiring prior knowledge on 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 a combination thereof. Semi-supervised learning (SSL) is a machine learning framework within which one can train a model using both labeled and unlabeled data. Data augmentation methods can be optionally used to produce artificial data samples out of a scarce set of real data samples and increase the number and diversity of data used for model training. Unlabeled data, when used in conjunction with a small amount of labeled data, can produce considerable improvement in learning accuracy compared to other frameworks. This approach is particularly interesting when only part of the available data is labeled.

A “convolutional neural network” or “CNN” refers to a machine learning model which uses multiple data processing layers, known as convolutional layers, to represent the input data in a way which is best suited to solve a classification or regression task. During training, weight parameters are optimized for each CNN layer using optimization algorithms known in the art such as the backpropagation algorithm to perform a stochastic gradient descent. At runtime, the resulting trained CNN may then process very efficiently the input data, for instance to classify it into the right data output labels with as little false positives and false negatives as possible in the case of a learnt classification task. Convolutional neural networks may also be combined with recurrent neural networks to produce a deep learning classifier.

The term “genuineness detection algorithm” in the context of this invention refers to an algorithm that produces a value of the genuineness of a type of object from a digital signal representation of this object and/or also returns as an output a genuineness decision. Thus, in some embodiments a value of genuineness is referred to as a detectability value of the genuineness. The genuineness decision can for example be decided from a pre-defined threshold value of the detectability value, or an aggregate of multiple detectability values computed from different subsampling of the digital representation, like in the case of multiple crops on the same acquired input image.

The genuineness detection algorithm may also be referred to as an “authentication algorithm”. In the context of the present invention and also in prior art (figure 1), the “genuineness detection algorithm” or “authentication algorithm” may be any authentication algorithm which takes as input a digital signal representation of the object to be identified/authenticated. As will be apparent to those skilled in the art of authentication, prior art authentication algorithms usually return as output either a genuineness decision as a binary label “authenticated genuine”/ “non authenticated/detected genuine” or a ternary label “authenticated genuine”/ “non authenticated genuine”/ “unknown”. However, and as seen in figure 1 a label of “non authenticated/detected genuine” is not necessarily equal to a definitive label “fake” (dashed line in figure 1). Similarly, a label of “not detected” is not equal to a definitive label “unknown”. Prior art authentication algorithms may provide label ‘fake’ only if acquisition conditions are totally controlled, like it is the case for a flatbed scanner, so that one shot acquisition is sure to provide a detectable digital representation of the object, or if they were trained with the use of fake items or synthetically generated representation of fake items. Prior art authentication algorithms may also comprise internal signal processing algorithms to calculate a measurement of how the digital signal representation differs from a reference template digital signal representation. When such a measurement is available, it enables to quantify the detectability value of the genuineness of an object as a scalar value (for instance ranging from 0 not detectable at all as a genuine object to 100 perfectly detectable). The resulting scalar value may then be further used by the authentication algorithm to classify the genuineness of the object to be authenticated, for instance using a predefined threshold to discriminate between measurement values corresponding respectively to “non authenticated genuine” (lower range below the threshold) and “authenticated genuine” (higher range above the threshold) decisions.

Figure 1 shows an example of a processing workflow of a prior art authentication algorithm (or genuineness detection algorithm) (100) in line with the authentication methods described for instance in WO0225599, W004028140, cloud of micro-holes W006087351, or US10332247. Such an authentication algorithm (100) takes as input one or more digital signal representations of an object as may be captured with a sensor such as for instance an image sensor. The genuineness detection algorithm (100) may optionally pre-process the captured digital signal representations, for instance by using geometrical transforms (e.g. scaling, rotating, translating, downsampling, upsampling, cropping, etc.), frequency domain transforms (e.g. Fourier transform, Discrete Cosine Transform DCT, etc), filters (for instance, low-pass filters, high-pass filters, equalizers, etc.) to produce a set of pre-processed digital signal representations of the object that are suitable for comparative analysis against reference templates (for instance, using cross-correlation of cropped areas of an input image capture versus matching cropped areas of genuine reference images stored in a template database). Out of the comparative analysis, the genuineness detection algorithm (100) may output a measured scalar value of the genuineness (for instance, a Signal to Noise Ratio SNR scalar value out of the cross-correlation calculation). The genuineness detection further comprises a decision module to determine from the latter value (for instance by comparing it to a pre-determined threshold) if the object can be detected as a genuine one with high confidence, or if it cannot be detected. A non-detection event occurs either because the object is actually a fake (but one cannot identify it as such), or because the digital signal representation of the object does not enable the genuineness detection algorithm (100) to detect the genuineness of the object with high enough confidence.

The term a “predictive machine learning model” or “predictive model” or “a predictive machine learning model to predict a detectability value of the genuineness of an object” in the context of this invention refers to a machine learning model that can be trained to predict a detectability value of the genuineness of an object from at least one digital signal representation of an object. In one embodiment, a predictive machine learning model is trained according to the methods of the invention.

The term “pre-processing” refers to a set of digital operations leading to the transformation of raw data or a signal captured with a sensor or a raw digital signal representation to a pre- processed digital signal representation that can be used for example by the predictive machine learning algorithm or genuineness detection algorithm (authentication algorithm). Examples of the known method of pre-processing include but are not limited to geometrical transforms (e.g., scaling, rotating, translating, downsampling, upsampling, cropping, etc.), frequency domain transforms or other domain transform (e.g. Fourier transform, Discrete Cosine Transform DCT, Wavelet transforms, etc), filters (for instance, low-pass filters, high-pass filters, equalizers, etc.) and the like. Background suppression algorithms or image registration algorithms can be applied as pre-processing steps. Fully convolutional neural networks, UNet networks and spatial transformers network can even be trained to optimize correlation between reference images. High dynamic range pre-processing combining multiple digital representations.

Methods for training a predictive machine learning model and uses of a predictive machine learning model

In one embodiment is provided a computer-implemented method for training a predictive machine learning (ML) model to predict a detectability value of the genuineness of an object based on the at least one digital signal representation of at least one genuine object.

A training with the at least one digital signal representation of counterfeit objects might not be of interest since there is no definitive number of such objects and new counterfeit objects can be produced resulting in the algorithms that are made redundant and require continuous updates. Therefore, and according to the methods of the invention, performing the training on genuine objects allows for the reliable further classification of objects. In other words, the methods of the invention are based on positive detection.

In an alternative embodiment, the training may be performed with use of genuine objects in conjunction with a small amount of genuine objects without embedded security/ authentication features, which therefore act as potential representations of counterfeited objects. These representations are positioned in the exact same point of views as a similar genuine object, and are mapped to the corresponding genuine object observed detectability value of genuineness. This may have an effect in an increased learning efficiency. Therefore, a computer-implemented method for training a predictive machine learning model according to the invention allow to obtain a predictive machine learning model to predict a detectability value of the genuineness of an object.

A computer-implemented method for training a predictive machine learning model according to the invention includes a step of obtaining a training data set comprising a set of digital signal representations of each of the used genuine objects, wherein each digital signal representation has the associated detectability value of the genuineness of an object. It is understood that a set of digital signal representations may be obtained by capturing digital signal representations with a use of a sensor. Thus, captured digital signal representations of an object may be obtained.

In one embodiment is provided a computer-implemented method for training a predictive machine learning model to predict a detectability value of the genuineness of an object, the method comprising the steps to: a) obtain a genuineness detection algorithm, which produces a detectability value of the genuineness of a type of object from a digital signal representation of this object; b) obtain one or more genuine objects of one type; c) obtain a set of digital signal representations of each of the genuine objects; d) input each of the digital signal representation of each of the genuine objects (obtained in step c)) to the genuineness detection algorithm (obtained in step a)) and output a detectability value of the genuineness of the object, so that each of the digital signal representation of each of the genuine objects has the associated detectability value of the genuineness of an object; e) train a predictive machine learning model to predict the detectability value of the genuineness of an object using the sets of digital signal representations of each of the genuine objects with the associated detectability values of the genuineness of an object (obtained in step d).

In one embodiment, the detectability value determines (or identifies) whether a digital signal representation of on object is sufficient to identify (or characterise) this object as genuine. However, the detectability value solely does not enable to determine (or identify) whether a digital signal representation of an object represents a fake object.

In one embodiment is provided a computer-implemented method for training a predictive machine learning model, wherein the genuineness detection algorithm is selected from detectors of surface fingerprints and detectors of product markings. In one embodiment, a set of digital signal representations of each of the genuine objects is obtained with a sensor. In further embodiment, a set of digital signal representations of each of the genuine objects is obtained with a sensor, wherein each digital signal representation of each of the genuine objects is obtained from the sensor operating under a different capture condition. In one embodiment, a predictive machine learning model that is trained according to the methods of the invention is for use in a computer-implemented method to identify if an object is a genuine or counterfeited object.

In one embodiment is provided a computer-implemented method for training a predictive machine learning model, wherein the selected genuineness detection algorithm and a predictive machine learning model to be trained are chosen to be able to process the digital signal representations of a specific type of genuine object. The selected genuineness detection algorithm can be any suitable algorithm known in the art, such as selected from detectors of surface fingerprints and detectors of product markings. In particular an AlpVision fingerprint detector, an AlpVision cryptoglyph detector, a taggant detector, a Scantrust secure graphic detector, a SICPA security ink detector and the like.

In one embodiment, a method for training a predictive machine learning model according to the invention uses a genuineness detection algorithm (e.g., obtained in step a)) that produces the detectability value of the genuineness of an object, wherein the value may be a scalar value, a label, a hash, a vector, a multidimensional vector/a tensor, an image, a matrix, a distribution curve and the like, preferably a scalar value or a label.

In one embodiment, a method for training a predictive machine learning model according to the invention uses a genuineness detection algorithm (e.g., obtained in step a)) that produces the detectability value of the genuineness of an object, wherein the value is a scalar value, such as selected from a signal to noise ratio (SNR) measurement, a difference measurement, and a distance metrics. In a preferred embodiment, the detectability value of the genuineness of an object is a scalar value, preferably SNR, which quantifies the strength of the security feature signal. Using a continuous variable as target of the predictive machine learning model has the advantage of removing an hyperparameter compared to the use of a categorical variable. When using a categorical variable, one needs to explicitly select the threshold representing boundary between detectable and not detectable predictions before training.

In an alternative embodiment, a method for training a predictive machine learning model according to the invention uses a genuineness detection algorithm (e.g., obtained in step a)) that produces the detectability value of the genuineness of an object, wherein the value is a label, such as selected from a binary label (such as detectable or not detectable; or such as pass or no result available), and a ternary label (such as detectable or not detectable or unknown) or any categorization of a continuous variable, such as one hot encoding of the value for each integer steps.

In one embodiment, a method for training a predictive machine learning model according to the invention is based on selected one type of object (e.g., in step b). Therefore, different types of objects may require different predictive machine learning model that are trained separately considering differences in properties of types of objects.

In one embodiment, a method for training a predictive machine learning model according to the 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 objects (e.g., in step b), preferably at least 100, more preferably at least 10. The effect of using more than one genuine object is increased size of a training set, which in turn allows for increased predictive power of obtained predictive machine learning model of the invention. All the objects used in the training methods belong to one class or type of objects. In one embodiment, a method for training a predictive machine learning model according to the invention uses a set of digital signal representations of each of the genuine objects (e.g., in step c)), wherein the set of digital signal representations of each of the genuine objects may consists of or comprise one, at least 1, at least 10, at least 100, at least 200, at least 1000, at least 10000, at least 20000 or at least 40000 digital signal representations, preferably at least 10 or at least 100, more preferably at least 200 digital signal representations. The effect of using increased number of digital signal representations (e.g., at least 200 for each genuine object) is increased size of a training set, which in turn allows for increased predictive power of obtained predictive machine learning model of the invention. The effect of using reduced number of digital signal representations (e.g., at least 10 for each genuine object) is obtaining a sufficient size of a training set, which in turn allows for a good predictive power of obtained predictive machine learning model of the invention. The skilled in the art would use common methods to determine the appropriate training set size so that to achieve best results of the training.

It is understood that a training thus comprises sets of digital signal representations of each of the genuine objects, such as at least 200 digital signal representations for each of 10 genuine objects amounting to at least 2000 digital signal representations in one training set, or such as at least 200 digital signal representations for each of 100 genuine objects amounting to at least 20000 digital signal representations in one training set. In one preferred embodiment, a method for training a predictive machine learning model according to the invention uses a set of at least 200 digital signal representations of each of at least 50 genuine objects. In one embodiment, a method for training a predictive machine learning model according to the invention uses the training sets of digital signal representations of each of the genuine objects with the associated detectability values of the object genuineness wherein all detectability values are labelled as detectable, or wherein all detectability values are labelled as non-detectable, or wherein part of detectability values are labelled as detectable and remaining detectability values are labelled as non-detectable. Preferably, the proportions of the number of digital signal representations in the set with detectable values relative to the number of digital signal representations in the set with non-detectable values are selected from 50:50, 70:30, 80:20, 90: 10 of detectable to non-detectable. The effect of using a training set with a mix of associated detectability values (e.g., detectable and non-detectable) is in general to provide an increased predictive power of obtained predictive machine learning model of the invention.

In one embodiment provided is a computer-implemented method for training a predictive machine learning model according to the invention, wherein each of digital signal representation of each of the genuine object is obtained from acquiring a signal captured with a sensor, such as 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, a 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 one embodiment, a sensor is a camera.

In one embodiment, a signal captured with a sensor may be optionally further subjected to a step of signal pre-processing with a signal pre-processing algorithm, so that to obtain a digital signal representation.

In one embodiment provided is a computer-implemented method for training a predictive machine learning model according to the invention, wherein, obtaining a set of digital signal representations of each of the genuine objects further comprises transforming, with a signal pre-processing method, the acquired digital signal representation into a digital signal representation suitable to be input to the genuineness detection algorithm.

In one embodiment, a pre-processing of a signal captured with a sensor may include a step of pre-processing by a genuineness detection algorithm (authentication algorithm) according to known methods with the use of known systems as described herein. Examples of a preprocessing methods include but are not limited to geometrical transforms such as scaling, rotating, translating, down-sampling, up-sampling, cropping, and the like; frequency domain transforms such as Fourier transform, Discrete Cosine Transform DCT, and the like; and filters such as low-pass filters, high-pass filters, equalizers, and the like.

In another embodiment, a computer-implemented method for training a predictive machine learning model according to the invention may use two sets of digital signal representations of each of the genuine objects, wherein one set is an input to the genuineness detection algorithm, and another set is an input to the predictive machine learning model. This different sets may be obtained based on different pre-processing steps.

In one exemplary embodiment, a computer-implemented method for training a predictive machine learning model according to the invention may use a AlpVision Cryptoglyph detector as selected genuineness detection algorithm, wherein the signal captured with a sensor is cropped to a larger field of view than the one used for the Cryptoglyph detector and downsampled so that to obtain a digital signal representation suitable for further processing.

In an alternative exemplary embodiment, a computer-implemented method for training a predictive machine learning model according to the invention may use a surface fingerprint detector as selected genuineness detection algorithm, wherein the surface fingerprint detector does not perform downsampling of a signal captured with a sensor and obtain a digital signal representation suitable for further processing. Since a microstructure with comparable distribution is also present on counterfeit objects, this will not prevent the model from identifying a digital representation of a counterfeit as detectable, while improving the rejection of unrelated objects.

In one embodiment provided is a computer-implemented method for training a predictive machine learning model according to the invention, wherein each of digital signal representation of each of the genuine object is acquired with the sensor at different sensor position and/or orientation. The different sensor position and/or orientation is in relation to the genuine object. In one embodiment, the different sensor position and/or orientation is selected from a range of possible sensor’s positions and/or orientation. In one embodiment, the different sensor positions and/or orientations are pre-determined. In one embodiment, the different sensor positions and/or orientations are the same for each genuine object.

In one embodiment provided is a computer-implemented method for training a predictive machine learning model according to the invention, wherein the sensor and the genuine object relative position and/or orientation differ to produce a different sensor capture condition.

In one embodiment, the different sensor position and/or orientation is controlled (or provided) by a robot arm. In an embodiment where two sensors are used, at least one or at least two robot arms may be used. Examples of use of a robot arm can be seen on figures 3 and 4. In another embodiment, the different sensor position and/or orientation is controlled (provided) by a human operator that positions the sensor. Human operator may position this sensor manually or with a suitable device.

In one embodiment provided is a computer-implemented method for training a predictive machine learning model according to the invention, wherein each of digital signal representation of each of the genuine object is acquired at different genuine object position and/or orientation. The different genuine object position and/or orientation is in relation to the sensor. In one embodiment, the different genuine object position and/or orientation is selected from a range of possible genuine object’s positions and/or orientation. In one embodiment, the different genuine object position positions and/or orientations are pre-determined. In one embodiment, the different genuine object positions and/or orientations are the same for each genuine object.

In one embodiment, the different genuine object position and/or orientation is controlled (provided) by a robot arm that positions the genuine object. Examples of use of a robot arm can be seen figures 3 and 4.

In one embodiment, the different genuine object position and/or orientation is controlled (provided) by a conveyor that positions the genuine object. An example of use of a conveyor is seen figure 4b).

In one embodiment, the different genuine object position and/or orientation is controlled (provided) by a human operator that positions the genuine obj ect. Human operator may position this genuine object manually or with a suitable device.

In one embodiment provided is a computer-implemented method for training a predictive machine learning model according to the invention, wherein each of digital signal representation of each of the genuine object is acquired under at least one predetermined physical environment parameter value, such that the value of the physical environment parameter changes the digital signal representation of each of the genuine object at the predetermined object position and orientation and/or at the 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.

Therefore, it is understood that wherein a set of digital signal representations of each of the genuine objects is obtained with a sensor, each digital signal representation of each of the genuine objects is obtained from the sensor operating under a different capture condition. This different capture conditions are understood as different physical environment parameter values or as characterized by at least one physical environment parameter around the genuine object. In one embodiment provided is a computer-implemented method for training a predictive machine learning model according to the invention, wherein each of the digital signal representation of each of the genuine object is obtained at a capture condition, that is characterized by at least one physical environment parameter around the genuine object.

In one embodiment provided is a computer-implemented method for training a predictive machine learning model according to the invention, wherein each of the digital signal representation of each of the genuine object is obtained at a capture condition, that is characterized by physical environment parameters around the genuine object. It one embodiment the at least one physical environment parameter (value) is changing around the object. This may be due to changing relative position and/or orientation of the sensor and the genuine object. This results in that each digital signal representation of each of the genuine objects is obtained from the sensor operating under a different capture condition. In other words, capture conditions differ from each other in that at least one physical environment parameter value differs between captures with a sensor of each of digital signal representations of one of the genuine objects.

In the most general case, there are three different variables in the physical environment parameters which may impact the digital representation signal of a physical item as captured with at least one sensor:

- The position and orientation of the object in the 3 -dimensional space, which can be modelled with up to 6 variables (“6DOF”) corresponding to the translational and angular degrees of freedom for a rigid object body; for flexible objects (like banknotes, textiles, etc...) the position and orientation may be further modelled from the position and/or orientations of vertices over a warping grid.

- The position and orientation of the sensor in the 3 -dimensional space, which can be modelled with up to 6 variables (“6DOF”) corresponding to the translational and angular degrees of freedom for a rigid object body, either as an absolute value in space or as a difference value relative to the object to be sensed; this includes for instance tilting, scaling, rotation, translation of a camera sensor.

- Certain variations in the physical environment which may directly impact the sensor capture, such as for instance the lighting conditions (position, intensity, direction, spectrum of each light, ambient lighting, presence of smoke, etc) for image sensors or the ambient noise conditions for audio sensors.

In one embodiment, the capture condition, in the method of training as well as in a method of identification, is further characterised by the parametrization or state of a sensor. Examples of a sensor parametrization, wherein a sensor is a camera include camera resolution, autofocus, aperture, focal length, field of view, white balance, and the like. Examples of a sensor state, wherein a sensor is a camera include camera temperature (possibly causing CMOS thermal noise), aging of electronic components, or presence of dirt, fingerprints, and the like on the camera. In one specific example, the capture conditions are characterized by presence of dirt and/or fingerprints on a camera lenses.

In one embodiment, the capture condition, in the method of training as well as in a method of identification, is further characterised by the quality or state of an object. The quality or state of an object may depend on how the object has aged or how it has been handled over its lifetime, e.g., dirt surfaces, scratches, colour loss due to UV exposure, deformation due to shocks or wet environment, oxidation, and the like. Examples of such objects include an object with dirt on surface, scratches, colour loss due to UV exposure, deformation due to shocks or wet environment, oxidation, and the like.

Depending on the actual applications, various system setups may be used to control, with an object positioner, at least one variable position and/or orientation parameter for the training item at capture time. In a possible embodiment, a mechanical setup with an automation control may be used, such as a robot arm or a conveyor with their software controllers, to precisely manipulate the training item and control the training item variable position and/or orientation parameter. In another possible embodiment, the training item may be placed in a fixed position and orientation and at least one other variable (such as the sensor position or parametrization) in the physical environment around the training item may be varied.

Depending on the actual applications, various system setups may be used to control, with a sensor positioner, at least one variable position and/or orientation parameter for the sensor at capture time. In a possible embodiment, a mechanical setup with an automation control may be used, such as a robot arm. In another possible embodiment, a set of several sensors may be placed at different fixed positions and orientations around the training item, and each sensor may be sequentially controlled to take a different capture of the training item, each capture corresponding to a different position and orientation of the training item relative to the fixed training item. Depending on the actual applications, at least one physical environment parameter may be automatically setup by a physical environment parameter controller. For instance, when using a smartphone for the sensor capture, a dedicated smartphone app may be developed which controls the smartphone lighting towards the object, for instance using the smartphone flashlight in torch mode. More generally, the physical environment around the items to be captured for training may be adapted with at least one physical environment control device such as a lamp, a speaker, and/or a mechanical actuator.

Examples of a physical environment lamp variable characteristics include, but are not limited to: color temperature, polarization, emission spectrum, intensity, beam shape, impulsion shape, lamp orientation, lamp distance towards the object, etc. Examples of an actuator variable characteristics include, but are not limited to: the volume of water or air projected towards the object; the force applied to brushing it, knocking on it, shearing it, bending it; the temperature of heating or cooling it; the distance of approaching a magnet towards it; the variable placement of movable light reflectors, blockers, diffusers or filters, such as for example Wratten color filters, or interference filters; etc.

In a possible embodiment, a series of digital representations of a training item may be captured with an imaging device as a series of image acquisitions over time under an orientable neon lamp with variable intensity, then under an orientable LED lamp with variable intensity, then under direct sunlight at different times of the day. With this scheme, the set of digital representations will inherently represent several light sources with different spectra and variable intensities and variable positions as input to the machine learning classifier production engine.

In a possible embodiment, a series of digital representations of a training item may be captured with an imaging device as a series of image acquisitions over time using different parametrizations of the sensor.

In another possible embodiment, a series of digital representations of a training item may be captured with the Apple iPhone with a version equal to the iPhone 7 or more recent, by independently controlling the two torch LEDs, each of which has a different, adjustable, variable color temperature. As will be apparent to those skilled in the art of imaging, the resulting set of digital representations will inherently represent variable spectral reflectance, transmittance and radiance environments as input to the machine learning classifier production engine.

In one embodiment provided is a computer-implemented method for training a machine learning model according to the invention, wherein each of the digital signal representation of each of the genuine object is a digital image representation of the object, wherein the at least one physical environment parameter around the genuine object is illumination of the object by a lighting device, and wherein the illumination of the object is controlled by a robot arm, a conveyor or a human operator that positions the lighting device relative to the genuine object. It one embodiment the at least one physical environment parameter is changing around the object.

In one further embodiment provided is a computer-implemented method for training a machine learning model according to the invention, wherein the sensor is a smartphone camera and the lighting device is the smartphone flash.

The above-described embodiments may also be combined to provide as broad an input digital representation space of the training items as possible. In the most general case, each of the three different variables in the physical environment which may impact the digital representation signal of a physical item as captured with at least one sensor may be varied step-by-step in different possible ranges to produce a diversity of digital representations of each input item, so that the machine learning can better anticipate the diversity of end user physical environments that its produced classifier solution will encounter at the time of detection:

- The position and orientation of the object in the 3 -dimensional space can be varied step by step along any of the position or orientation axes, for instance by increments of 1 mm in translation and 1° in Euler angles;

- The position and orientation of the sensor in the 3 -dimensional space can be varied step by step along any of the position or orientation axes, for instance by increments of 1 mm in translation and 1° in Euler angles;

- Various physical environment parameters may be simply varied as on/off, or as step by step increments in any variable associated with the underlying physical component, for instance by increments of 10 lux in light intensity or 1 dB in sound intensity or 1 Hz in frequency intensity, etc;

- Various sensor parametrization settings may be simply varied as enabled by the sensor controller API, for instance by varying the resolution, the focal length, the aperture, the field of view, the white balance or other optical parameters of a camera sensor on a smartphone.

In one embodiment, the trained predictive machine learning model may be a machine learning classifier. In an alternative embodiment, the trained predictive machine learning model may be a machine learning regressor.

In one embodiment, the trained predictive machine learning model may be an artificial neural network tool (ANN) like a deep learning model, and a convolution 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 others may 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, wherein the training set comprises or consists of the digital signal representations of each of the genuine objects, wherein each of the digital signal representation has the associated detectability value of the genuineness of an object.

In one embodiment is provided a computer-implemented method for predicting a detectability value of the genuineness of an object, the method comprising the steps to: a) obtain an object to be detected; b) obtain a digital signal representation of the object to be detected; c) obtain a predictive machine learning model to predict a detectability value of the genuineness of an object; d) input the digital signal representation of the object to be detected (obtained in step b)) to the predictive machine learning model (obtained in step c)) and output a predicted detectability value of the genuineness of an object for the digital signal representation of the object to be detected.

In one embodiment is provided a computer-implemented method for predicting a detectability value of the genuineness of an object, the method comprising the steps to: a) obtain an object to be detected; b) obtain a digital signal representation of the object to be detected; c) obtain a predictive machine learning model to predict a detectability value of the genuineness of an object, wherein the predictive machine learning model is trained according to the methods of the invention; d) input the digital signal representation of the object to be detected (obtained in step b)) to the predictive machine learning model (obtained in step c)) and output a predicted detectability value of the genuineness of an object for the digital signal representation of the object to be detected.

In one embodiment is provided a computer-implemented method for predicting a detectability value of the genuineness of an object, wherein the detectability value may be further used to determine (or identify) whether a digital signal representation of on object is sufficient to identify (or characterise) this object as genuine.

In one embodiment is provided a computer-implemented method for predicting a detectability value of the genuineness of an object, wherein a digital signal representation of the object to be detected is obtained (or captured) with a sensor. In further embodiment, a digital signal representation of the object to be detected is obtained with a sensor, wherein this digital signal representation of the object is obtained from the sensor operating under a given capture condition. This given (or one) capture condition is understood as given physical environment parameter values, and/or as characterized by at least one physical environment parameter around the object, and/or as characterized by the parametrization or state of the sensor, and/or as characterized by the quality or state of the object.

In one embodiment is provided a computer-implemented method for predicting a detectability value of the genuineness of an object with the use of a predictive machine learning model, wherein the method may comprise an algorithm referred herein as detectability prediction algorithm (200) (figure 2).

In one embodiment is provided a computer-implemented method for predicting a detectability value of the genuineness of an object based on the at least one digital signal representation of such an object to be detected. In one embodiment, a computer-implemented method for predicting a detectability value of the genuineness of an object uses more than one digital signal representations of the object to be detected, wherein these digital signal representations are obtained and processed sequentially or in parallel. This enhances the confidence of the prediction. For example, if the predictions are considered to be independent for different digital signal representations, then if N prediction above a given detectable prediction threshold are needed to consider the object as detectable, the false detectable prediction rate is effectively divided by N. In one embodiment is provided a computer-implemented method for predicting a detectability value of the genuineness of an object based on the at least two, at least 10, at least 25, at least 50, at least 100 or at least 250 digital signal representations of such an object to be detected, preferably at least 50.

In one embodiment, a computer-implemented method for predicting a detectability value of the genuineness of an object is for use in a computer-implemented method to identify if an object is a genuine or counterfeited object.

In one embodiment is provided a computer-implemented method for predicting a detectability value of the genuineness of an object, wherein the selected predictive machine learning model is chosen to be able to process the digital signal representations of a specific type of an object to be detected. It is understood that a predictive machine learning model is trained based on the same type of object as an object to be identified in the method for predicting a detectability value of the genuineness of an object.

In one embodiment, a computer-implemented method for predicting a detectability value of the genuineness of an object outputs a predicted detectability value of the genuineness of an object to be detected, wherein the value may be a continuous value, a categorical value, a hash, a vector, a tensor, an image, a matrix and the like. In one embodiment, a computer-implemented method for predicting a detectability value of the genuineness of an object output a predicted detectability value of the genuineness of an object to be detected, wherein the value is a scalar value, such as selected from a signal to noise ratio (SNR) measurement, a difference measurement, and a distance metrics. In a preferred embodiment, the detectability value of the genuineness of an object is a continuous scalar value such as SNR. Using a continuous value as target of the predictive machine learning model has the advantage of removing an hyperparameter compared to the use of a categorical variable. When using a categorical variable, one needs to explicitly select the threshold representing boundary between detectable and not detectable predictions before training.

In another embodiment, a computer-implemented method for predicting a detectability value of the genuineness of an object output a predicted detectability value of the genuineness of an object to be detected, wherein the value is a label, such as selected from a binary label (such as detectable or not detectable), and a ternary label (such as detectable or not detectable or unknown).

In one embodiment, a computer-implemented method for predicting a detectability value of the genuineness of an object is suitable for one type of object (e.g., in step a). Therefore, different types of objects may require different methods for predicting a detectability value that may use different predictive machine learning models that are trained separately considering differences in properties of types of objects.

In one embodiment provided is a computer-implemented method for predicting a detectability value of the genuineness of an object, wherein the digital signal representation of the object to be detected is obtained from acquiring a signal captured with a sensor, such as 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, a 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 one embodiment, a signal captured with a sensor may be optionally further subjected to a step of signal pre-processing with a signal pre-processing algorithm, so that to obtain a digital signal representation. This pre-processing step may be included in a detectability prediction algorithm (200).

In one embodiment provided is a computer-implemented method for predicting a detectability value of the genuineness of an object, wherein obtaining a digital signal representation of the object to be identified further comprises transforming, with a signal pre-processing method, 1 the acquired digital signal representation into a digital signal representation suitable to be input to a predictive machine learning model. Similarly, this step may be included in a detectability prediction algorithm (200).

In one embodiment, a pre-processing of a signal captured with a sensor may include a step of pre-processing by a predictive machine learning model according to known methods with the use of known systems as described herein.

Figure 2 shows an example of a processing workflow of a method for predicting a detectability value of the genuineness of an object with the use of the predictive machine learning model (noted on figure as a “ML model”), wherein this method includes an algorithm that may be referred herein as detectability prediction algorithm (200). Figure 2 shows an exemplary detectability prediction algorithm 200 according to certain embodiments of the present disclosure. Such a detectability prediction algorithm (200) takes as input one or more digital signal representations of an object as may be captured with a sensor such as for instance an image sensor. The detectability prediction algorithm (200) may optionally pre-process the captured digital signal representations, for instance by using geometrical transforms (e.g. scaling, rotating, translating, downsampling, upsampling, cropping, etc.), frequency domain transforms (e.g. Fourier transform, Discrete Cosine Transform DCT, etc), filters (for instance, low-pass filters, high-pass filters, equalizers, etc) to produce a set of digital signal representations of the object that are suitable as input to a machine learning model which has been trained to predict a value of the detectability of the genuineness for objects of the same type when using a genuineness detection algorithm (100) suitable for determining the genuineness of objects of this same type from its captured digital signal signal representations. From the machine learning model prediction, the detectability prediction algorithm (200) may then output a predicted value of the detectability of the genuineness of the object according to its captured digital signal representations.

Methods for identifying if an object is a genuine or a counterfeited object and uses thereof In one embodiment is provided a computer-implemented method for identifying if an object is a genuine or a counterfeited object, the method comprising the steps to: a) obtain a.1) an object to be identified; a.2) a digital signal representation of the object to be identified; a.3) a genuineness detection algorithm, which returns as an output a genuineness decision from the digital signal representation of the object to be identified, a.4) a predictive machine learning model to predict a detectability value of the genuineness of the object to be identified with the genuineness detection algorithm (obtained in step a.3)), wherein the predictive machine learning model is trained according to the method of the invention; b) input the digital signal representation of the object to be identified (obtained in step a.2)) to b. l.) the predictive machine learning model (obtained in step a.4)) and output a predicted detectability value of the genuineness of an object for the digital signal representation of the object to be identified; and determine from the predicted detectability value of the genuineness of an object that the genuineness detection algorithm (obtained in step a.3)) can or cannot detect the object as a genuine object; b.2) input the digital signal representation of the object to be identified (obtained in step a.2)) to the genuineness detection algorithm (obtained in step a.3)) and output that the object to be identified is identified as genuine or that the object to be identified cannot be identified as genuine, c) provide a decision based on the output from the predictive machine learning model (obtained in step b.l)) and the genuineness detection algorithm (obtained in step b.2)), wherein c. l) optionally the object can be detected as a genuine object (step b.l)) and the object is identified as a genuine object (step b.2)) than provide output of genuine object; c.2) the object cannot be detected as a genuine object (step b.1)) and optionally the object is identified as a genuine object (step b.2)) than provide output of not- detectable object, and optionally repeat step steps a.2), b) and c); c.3) the object can be detected as a genuine object the object (step b.1)) and the object is not identified as a genuine object (step b.2)) than provide output of counterfeit object; c.4) the object cannot be detected as a genuine object (step b.1)) and optionally the object is not identified as a genuine object (step b.2)) than provide output of not- detectable object, and optionally repeat step steps a.2), b) and c).

In one embodiment is provided a computer-implemented method for identifying if an object is a genuine or a counterfeited object, wherein a digital signal representation of the object to be identified is obtained (or captured) with a sensor. In further embodiment, a digital signal representation of the object to be identified is obtained with a sensor, wherein this digital signal representation of the object is obtained from the sensor operating under a given capture condition. This given (or one) capture condition is understood as given physical environment parameter values, and/or as characterized by at least one physical environment parameter around the object, and/or as characterized by the parametrization or state of the sensor, and/or as characterized by the quality or state of the object. Physical environment parameters are as defined above.

In one embodiment is provided a computer-implemented method for identifying if an object is a genuine or a counterfeited object, wherein the predicted detectability value (obtained e.g., in step b.l) of the method) identifies whether a digital signal representation of this object is sufficient to identify this object as genuine. It is understood that meant is whether a digital signal representation of this object is sufficient for the genuineness detection algorithm (obtained e.g., at step a.3) of the method) to identify this object as genuine. However, the predicted detectability value solely does not enable to determine (or identify) whether a digital signal representation of an object represents a fake object.

In a further embodiment, is provided a computer-implemented method for identifying if an object is a genuine or a counterfeited object according to the invention, wherein to identify an object to be a genuine object

- the object to be identified is detected as a genuine object (step b .2)) by the genuineness detection algorithm; or

- the object to be identified is detected as a genuine object (step b .2)) by the genuineness detection algorithm and it is determined that the object can be detected as a genuine object from the predicted detectability value of the genuineness of an object (step b.1)).

In a further embodiment, is provided a computer-implemented method for identifying if an object is a counterfeited object according to the invention, wherein to identify an object to be a counterfeited object

- the object can be detected as a genuine object from the predicted detectability value of the genuineness of an object (step b.1)) and the object is not identified as a genuine object by the genuineness detection algorithm (step b.2)).

It is understood that the security feature is present in the object to be identified in order to determine from the predicted detectability value of the genuineness of an object that the genuineness detection algorithm can or cannot detect the object as a genuine object. In the event that the presence or absence of security feature is not known then the methods of the invention allow to determine from the predicted detectability value of the genuineness of an object that the genuineness detection algorithm could or could not detect the object as a genuine object. It is understood that a computer-implemented method for identifying if an object is a genuine or a counterfeited object according to the invention uses a combination of output from the predictive machine learning model and the genuineness detection algorithm, wherein the two algorithms can be connected in series, in parallel, or may also be combined according to any logical operation such as Addition, Subtraction, Multiplier, Divider, Convolution, And, Or, Xor, Not, Comparison above, Comparison below, Comparison equal and Comparison differs. Exemplary embodiments of the algorithms connected in series, i.e., the result of the upstream one is passed to the downstream one, are presented herein as embodiment 1 (figure 5) or embodiment 2 (figure 6), wherein these are not limiting examples.

In one embodiment is provided a method for identifying if an object is a genuine or a counterfeited object according to the invention that with uses of a predictive machine learning model (such as within a genuineness detection algorithm (100)) and a detectability prediction algorithm (200), wherein the combination of the two algorithms provides that the method comprises an algorithm referred herein as a “fake authentication algorithm” (500) on figure 5 or (600) on figure 6.

It is understood that the methods and systems of the invention allow to detect of counterfeited objects due to the combination of use of a genuineness detection algorithm and a predictive machine learning model trained according to the methods of the invention.

In one embodiment is provided a method for identifying if an object is a genuine or a counterfeited object, wherein the selected genuineness detection algorithm and a predictive machine learning model are chosen to be able to process the digital signal representations of a specific type of object. The selected genuineness detection algorithm can be any suitable algorithm known in the art, such as selected from detectors of surface fingerprints and detectors of product markings. In particular an AlpVision fingerprint detector, an AlpVision cryptoglyph detector, a taggant detector, a Scantrust secure graphic detector, a SICPA security ink detector and the like.

It is understood that any authentication algorithm outputs a label ‘genuine object’ that may include true positive or false positive cases, or a label ‘counterfeit object’ that may include true negative or false negative cases.

In a further embodiment, the effect in decreasing the number of false negative cases within a ‘counterfeit object’ label is achieved by using further statistical methods and parameters within or in combination with a genuineness detection algorithm. For example, several predictions of detectability for different acquisitions of digital representations of the same object can be made at runtime in order to form a distribution and select meaningful quantities such as maximum or average prediction, or confidence interval, and the like known in the art.

In one embodiment is provided a computer-implemented method for identifying if an object is a genuine or a counterfeited object based on the at least one digital signal representation of such an object to be detected. In one embodiment, a computer-implemented method for identifying if an object is a genuine or a counterfeited object uses more than one digital signal representations of the object to be detected, wherein these digital signal representations are obtained and processed sequentially or in parallel. This enhances the confidence of the prediction. For example, if the predictions are considered to be independent for different digital signal representations, then if N predictions above a given detectable prediction threshold are needed to consider the object as detectable, the false detectable prediction rate is effectively divided by N.

In another embodiment is provided a computer-implemented method for identifying if an object is a genuine or a counterfeited object based on the at least two, at least 10, at least 25, at least 50, at least 100 or at least 250 digital signal representations of such an object to be detected, preferably at least 50. Using a plurality of digital signal representations of an object to be detected allows to provide for example statistical distribution of these representations. This may have the effect in decreasing the number of false negative cases within a ‘counterfeit object’ label and increasing confidence level in true negative cases within a ‘counterfeit object’ label.

The at least two digital signal representations of an object to be detected may be obtained for example by obtaining a camera sensor capture for about 5-15s such as about 10s, minimum about 25-150 frames such as about 50 frames.

It is understood that using at least two digital signal representations has further effect in decreasing the number of false negative cases within a ‘counterfeit object’ label.

It is understood that a predictive machine learning model is trained based on the same type of object as an object to be identified in methods for identifying if an object is a genuine or a counterfeited object according to the invention. Illustration is provided in Example 2.

In one embodiment, in a method for identifying if an object is a genuine or a counterfeited object, in one step this methods output a predicted detectability value of the genuineness of an object to be detected, wherein the value may be a scalar value, a label, a ground truth, a regressor, a hash, a vector, a multidimensional vector, an image, a matrix, or any categorization of a continuous variable, such as one hot encoding of the value for each integer steps and the like, preferably a scalar value or a label. In one embodiment, in a method for identifying if an object is a genuine or a counterfeited object, in one step this method outputs a predicted detectability value of the genuineness of an object to be detected, wherein the value is a scalar value, such as selected from a signal to noise ratio (SNR) measurement, a difference measurement, and a distance metrics. In a preferred embodiment, the detectability value of the genuineness of an object is a scalar value, such as SNR. Using a continuous variable as target of the predictive machine learning model has the advantage of removing an hyperparameter compared to the use of a categorical variable. When using a categorical variable, one needs to explicitly select the threshold representing boundary between detectable and not detectable predictions before training.

In one embodiment, in a method for identifying if an object is a genuine or a counterfeited object, in one step this method outputs a predicted detectability value of the genuineness of an object to be detected, wherein the value is a label, such as selected from a binary label (such as detectable or not detectable), and a ternary label (such as detectable or not detectable or unknown).

In one embodiment, a method for identifying if an object is a genuine or a counterfeited object according to the invention, is suitable for one type of object (e.g., in step a). Therefore, different types of objects may require different methods for identifying if an object is a genuine or a counterfeited object that may use different predictive machine learning models that are trained separately considering differences in properties of types of objects.

In one embodiment provided is a method for identifying if an object is a genuine or a counterfeited object according to the invention, wherein the digital signal representation of the object to be detected is obtained from acquiring a signal captured with a sensor, such as 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, a 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 one further embodiment provided is a method for identifying if an object is a genuine or a counterfeited object according to the invention, wherein the sensor is a smartphone camera and the lighting device is the smartphone flash.

In one embodiment, a signal captured with a sensor may be optionally further subjected to a step of signal pre-processing with a signal pre-processing algorithm, so that to obtain a digital signal representation. This pre-processing step may be included in a genuineness detection algorithm (100) and/or a detectability prediction algorithm (200). In one embodiment provided is a method for identifying if an object is a genuine or a counterfeited object according to the invention, wherein obtaining a digital signal representation of the object to be identified further comprises transforming, with a signal preprocessing method, the acquired digital signal representation into a digital signal representation suitable to be input to a predictive machine learning model and/or a genuineness detection algorithm. Similarly, this step may be included in a genuineness detection algorithm (100) and/or a detectability prediction algorithm (200). In one embodiment, a pre-processing of a signal captured with a sensor may include a step of pre-processing by a predictive machine learning model according to known methods with the use of known systems as described herein. In one embodiment, a pre-processing of a signal captured with a sensor may include a step of pre-processing by a genuineness detection algorithm according to known methods with the use of known systems as described herein.

In another embodiment, a method for identifying if an object is a genuine or a counterfeited object according to the invention may use two sets of digital signal representations of each of the genuine objects, wherein one set is an input to the genuineness detection algorithm, and another set is an input to the predictive machine learning model. This different sets may be obtained based on different pre-processing steps.

In one exemplary embodiment, a method for identifying if an object is a genuine or a counterfeited object according to the invention may use a Cryptoglyph detector as selected genuineness detection algorithm, wherein the signal captured with a sensor is cropped to a larger field of view than the one used for the Cryptoglyph detector.

In an alternative exemplary embodiment, a method for identifying if an object is a genuine or a counterfeited object according to the invention may use a surface fingerprint detector as selected genuineness detection algorithm, wherein the surface fingerprint detector does not perform downsampling of a signal captured with a sensor and obtain a digital signal representation suitable for further processing. Since a microstructure with comparable distribution is also present on counterfeit objects, this will not prevent the model from identifying a digital representation of a counterfeit as detectable, while improving the rejection of unrelated objects.

In one embodiment (herein exemplary embodiment 1) is provided a computer-implemented method for identifying if an object is a genuine or a counterfeited object, the method comprising the steps to: a) obtain a.1) an object to be identified; a.2) with a sensor, a digital signal representation of the object to be identified; a.3) a genuineness detection algorithm, which returns as an output a genuineness decision from the digital signal representation of the object to be identified, a.4) a predictive machine learning model to predict a detectability value of the genuineness of the object to be identified with the genuineness detection algorithm (obtained in step a.3)), wherein the predictive machine learning model is trained according to the method of the invention; b) input the digital signal representation of the object to be identified (obtained in step a.2)) to the genuineness detection algorithm (obtained in step a.3)) and output

- that the object to be identified is identified as genuine, or

In one embodiment (herein exemplary embodiment 1) is provided a computer-implemented method for identifying a counterfeited object, the method comprising the steps to: a) obtain a.1) an object to be identified; a.2) with a sensor, a digital signal representation of the object to be identified; a.3) a genuineness detection algorithm, which returns as an output a genuineness decision from the digital signal representation of the object to be identified, a.4) a predictive machine learning model to predict a detectability value of the genuineness of the object to be identified with the genuineness detection algorithm (obtained in step a.3)), wherein the predictive machine learning model is trained according to the method of the invention; b) input the digital signal representation of the object to be identified (obtained in step a.2)) to the genuineness detection algorithm (obtained in step a.3)) and output

- optionally that the object to be identified is identified as genuine, or

- can detect the object as a genuine object and determine the object to be identified as a counterfeited object.

- that the object to be identified is identified as genuine, or

- that the object to be identified cannot be identified as genuine, and then c) input the digital signal representation of the object to be identified (obtained in step a.2)) to the predictive machine learning model (obtained in step a.4)) and output a predicted detectability value of the genuineness of an object for the digital signal representation of the object to be identified; d) determine from the predicted detectability value of the genuineness of an object (obtained in step c)) that the genuineness detection algorithm (obtained in step a.3))

- that the object to be identified cannot be identified as genuine, and determine the object to be identified as not-detectable, if the object is determined as not detectable (in step d)) then e) obtain, with a sensor, another digital signal representation of the object to be identified; f) input another digital signal representation of the object to be identified (obtained in step e)) to the genuineness detection algorithm (obtained in step a.3)) and output

- that the object to be identified is identified as genuine, or

- that the object to be identified cannot be identified as genuine, and if the object to be identified cannot be identified as genuine then g) input the digital signal representation of the object to be identified (obtained in step e)) to the predictive machine learning model (obtained in step a.4)) and output another predicted detectability value of the genuineness of an object for the digital signal representation of the object to be identified; h) determine from the predicted detectability value of the genuineness of an object (obtained in step g)) that the genuineness detection algorithm (obtained in step a.3))

- can detect the object as a genuine object and determine the object to be identified as a counterfeited object or

- that the object to be identified cannot be identified as genuine, and optionally determine the object to be identified as not-detectable and at least once repeat steps e) to h).

Figure 5 shows a processing workflow of a method for identifying if an object is a genuine or a counterfeited object according to exemplary embodiment 1.

In one embodiment (herein exemplary embodiment 2) is provided a computer-implemented method for identifying if an object is a genuine or a counterfeited object, the method comprising the steps to: a) obtain a.1) an object to be identified; a.2) with a sensor, a digital signal representation of the object to be identified; a.3) a genuineness detection algorithm, which returns as an output a genuineness decision from the digital signal representation of the object to be identified, a.4) a predictive machine learning model to predict a detectability value of the genuineness of the object to be identified with the genuineness detection algorithm (obtained in step a.3)), wherein the predictive machine learning model is trained according to the method of the invention; b) input the digital signal representation of the object to be identified (obtained in step a.2)) to the predictive machine learning model (obtained in step a.4)) and output a predicted detectability value of the genuineness of an object for the digital signal representation of the object to be identified; c) determine from the predicted detectability value of the genuineness of an object (obtained in step b)) that the genuineness detection algorithm (obtained in step a.3)) can detect the object as a genuine object or optionally that the object to be identified cannot be identified as genuine, and further optionally determine the object to be identified as not-detectable, d) input the digital signal representation of the object to be identified (obtained in step a.2)) to the genuineness detection algorithm (obtained in step a.3)) and output

- that the object to be identified is identified as genuine, or

- that the object to be identified cannot be identified as genuine, and determine the object to be identified as a counterfeited object.

In one embodiment (herein exemplary embodiment 2) is provided a computer-implemented method for identifying a counterfeited object, the method comprising the steps to: a) obtain a.1) an object to be identified; a.2) with a sensor, a digital signal representation of the object to be identified; a.3) a genuineness detection algorithm, which returns as an output a genuineness decision from the digital signal representation of the object to be identified, a.4) a predictive machine learning model to predict a detectability value of the genuineness of the object to be identified with the genuineness detection algorithm (obtained in step a.3)), wherein the predictive machine learning model is trained according to the method of the invention; b) input the digital signal representation of the object to be identified (obtained in step a.2)) to the predictive machine learning model (obtained in step a.4)) and output a predicted detectability value of the genuineness of an object for the digital signal representation of the object to be identified; c) determine from the predicted detectability value of the genuineness of an object (obtained in step b)) that the genuineness detection algorithm (obtained in step a.3)) can detect the object as a genuine object or that the object to be identified cannot be identified as genuine, and optionally determine the object to be identified as not-detectable, if the object to be identified can be identified as genuine, then d) input the digital signal representation of the object to be identified (obtained in step a.2)) to the genuineness detection algorithm (obtained in step a.3)) and output

- that optionally the object to be identified is identified as genuine, or

In one embodiment (herein exemplary embodiment 2) computer-implemented method for identifying if an object is a genuine or a counterfeited object, the method comprising the steps to: a) obtain a.1) an object to be identified; a.2) with a sensor, a digital signal representation of the object to be identified; a.3) a genuineness detection algorithm, which returns as an output a genuineness decision from the digital signal representation of the object to be identified, a.4) a predictive machine learning model to predict a detectability value of the genuineness of the object to be identified with the genuineness detection algorithm (obtained in step a.3)), wherein the predictive machine learning model is trained according to the method of the invention; b) input the digital signal representation of the object to be identified (obtained in step a.2)) to the predictive machine learning model (obtained in step a.4)) and output a predicted detectability value of the genuineness of an object for the digital signal representation of the object to be identified; c) determine from the predicted detectability value of the genuineness of an object (obtained in step b)) that the genuineness detection algorithm (obtained in step a.3))

- can detect the object as a genuine object or cannot detect the object as a genuine object, and determine the object to be identified as not- detectable, if the object is identified as not-detectable (in step c), then e) obtain, with a sensor, another digital signal representation of the object to be identified; f) input another digital signal representation of the object to be identified (obtained in step e)) to the predictive machine learning model (obtained in step a.4)) and output another predicted detectability value of the genuineness of an object for the digital signal representation of the object to be identified; g) determine from the predicted detectability value of the genuineness of an object (obtained in step f)) that the genuineness detection algorithm (obtained in step a.3)) can detect the object as a genuine object and then perform step h) or

- that the object to be identified cannot be identified as genuine, and optionally determine the object to be identified as not-detectable and at least once repeat steps e) to g); h) input another digital signal representation of the object to be identified (obtained in step e)) to the genuineness detection algorithm (obtained in step a.3)) and output

- that the object to be identified is identified as genuine, or

Figure 6 shows a processing workflow of a method for identifying if an object is a genuine or a counterfeited object according to exemplary embodiment 2.

It is understood that provided are exemplary embodiments of computer-implemented methods for identifying if an object is a genuine or a counterfeited object, wherein the predicted detectability value, obtained for one or another digital signal representation of the object to be identified, identifies whether a digital signal representation of this obj ect is sufficient to identify this object as genuine. It is understood that meant is whether a digital signal representation of this object is sufficient for the genuineness detection algorithm (obtained e.g., at step a.3) of the method) to identify this object as genuine. In the exemplary embodiments 1 (illustrated by figure 5) and the exemplary embodiments 2 (illustrated by figure 6), when a digital signal representation of this object is sufficient for the genuineness detection algorithm (obtained e.g., at step a.3) of the method) to identify this object as genuine, the genuine detection algorithm returns as an output a genuineness decision from the digital signal representation of the object to be identified.

It will be apparent to those skilled in the art of genuineness detection that the system architecture choice among the various exemplary embodiments 1 or 2 may depend on application-dependent requirements, such as the respective computational performance of the genuine detection algorithm (100) and the detectability prediction (200), and/or the ratio of genuine versus counterfeited objects to be processed by the system, and/or the exposure to reverse engineering of the genuine detection algorithm (100). (thus preferably using embodiment 2 - figure 6).

In a possible embodiment, the detectability prediction model (200) and the genuineness detection algorithm (100) may run on the same device (for instance as a smartphone app, or on a remote server in connection with the local device operating the sensor and the digital signal representation captures). In another possible embodiment, the detectability prediction model (200) and the genuineness detection algorithm (100) may run on different devices. For instance, in one embodiment, the detectability prediction model may run first on a local device as the front-end application while the detection algorithm runs second on a remote server as a backend application which is only called when the frontend predicted detectability value of the genuineness of an object indicates that the genuineness detection algorithm can detect the object as a genuine object (i.e., following the method steps of figure 6). This method is particularly advantageous when the genuineness detection algorithm requires complex computations which may not be handled fast enough on the user device, or when the brand owner requires a server-side control of the authentication application. It is also advantageous when there is a need to isolate the genuine detection algorithm from hacker attacks by only giving access to it when the predicted detectability value of the genuineness of an object indicates that the genuineness detection algorithm can detect the object as a genuine object. Conversely, in an alternative embodiment, the genuineness detection algorithm model may run first on a local device as the front-end application while the detectability prediction model runs second on a remote server as a backend application which is only called when the frontend genuineness detection algorithm fails to detect the object as a genuine object (i.e., following the method steps of figure 5).

In one embodiment, wherein in computer-implemented methods for identifying if an object is a genuine or a counterfeited object, the object to be identified is labelled as not-detectable, at least 2, at least 5, at least 10 or at least 50 further (or another) digital signal representations of the object to be identified may be obtained, preferably at least 50, and processed again by the methods (as seen on figures 5 and 6). The methods for identifying if an object is a genuine or a counterfeited object repeat processing of further digital signal representations of the object to be identified until counterfeited object is identified or until maximum allowed number of loops have been reached (time-out). In one embodiment, a maximum allowed number of loops is reached when in at least 20% of trials the object is predicted detectable.

In one embodiment, the methods for identifying if an object is a genuine or a counterfeited object repeat processing of further digital signal representations of the object to be identified until genuine object is identified or until a maximum allowed number of loops is reached. The maximum allowed number of loops is reached when in at least 20% of trials the object is predicted detectable.

In one embodiment, is provided a data processing apparatus comprising means for carrying out the methods of the invention as described herein.

In one embodiment, is provided a data processing apparatus comprising instructions which, when the program is executed by the apparatus, cause the apparatus to carry out the methods of the invention as described herein.

In one embodiment, is provided a computer-readable medium comprising instructions which, when executed by a computer, cause the computer to carry out the methods of the invention as described herein.

It is understood that the methods of the invention are computer-implemented or computer- based methods.

Examples

Example 1 - different signal representations of the same object under different capture conditions

Exemplified is a preparation of a training set for training a predictive machine learning model according to the methods of the invention, wherein one object is available for training and testing the algorithms.

Thirteen images (i.e., digital signal representation) of a genuine object (here electronic circuit) have been obtained (i.e., captured) by a sensor (here a smartphone camera). These images were input to the genuineness detection algorithm (authentication algorithm).

In this example, the output detectability value of the genuineness of an object was a binary label such as detectable or not detectable as genuine.

Figure 7 shows these 13 images of a genuine object, that is a single set, i.e., one object only with 13 digital signal representations thereof. In two images (image reference 7 A & 7B) the object can be recognized as genuine (i.e., the associated detectability value of the genuineness of the object is detectable). In the other eleven images the object cannot be recognized as genuine (i.e., the associated detectability value of the genuineness of the object is non- detectable) (image references 7C to 7M). Examples of images in which the object cannot be recognized as genuine include the image where the object is on non-compatible with the image background (Figure 7C), the image where the object is blurred (Figure 7D), the image where the object is blurred and the object on the image is cropped (Figure 7E), the image where color distortions were introduced so the object is not well visible (Figure 7F), the image where the object on the image is cropped (Figure 7G), the image where disturbing light is present so the object is not well visible (Figure 7H), the image where the object is overexposed (Figure 71), the image where the object on the image is small (Figure 7J), the image where the object on the image is in a tilted position (Figure 7K), the image where the object is not sufficiently exposed (Figure 7L), the image where only a part of the object on the image is present and an zoomed portion of the object is seen (Figure 7M).

Example 2 - A method of training and a method of testing (identifying) with the use of different signal representations of objects under different capture conditions

Exemplified is a preparation of a training set for use in a computer-implemented method for training a predictive machine learning model (Example 2A) and a preparation of a testing set for use in testing said trained predictive machine learning model (from 2A) in a framework of a method for identifying if an object is a genuine or a counterfeited object (Example 2B). Multiple objects of one type are available for training and testing the algorithms.

2 A) - training set for use in a method for training a predictive machine learning model

Four genuine objects of the same type were obtained (labelled 1, 2, 3 and 4). Four images (i.e., digital signal representation) of a genuine object #1 have been obtained (i.e., captured) by a sensor (labelled (object.capture) 1.1, 1.2, 1.3, 1.4; Table 1, col. 1). Three images of a genuine object #2 have been obtained by a sensor (labelled 2.1, 2.2, 2.3). Four images of a genuine object #3 have been obtained by a sensor (labelled 3.1, 3.2, 3.3, 3.4). Four images of a genuine object #4 have been obtained by a sensor (labelled 4.1, 4.2, 4.3, 4.4). Therefore, four sets of digital signal representations of each of the genuine objects (all of the same type) were obtained, providing a training set of 15 digital signal representations (4+3+4+4=15; Table 1, col. 2). It is understood that each of the image has been captured at a given capture conditions. These images were input to a genuineness detection algorithm (authentication algorithm) with returns an authentication decision as “Pass” or “No result available” (i.e., output a detectability value of the object genuineness for each of the digital signal representations of each of the genuine objects. The output was a detectability value of the genuineness of an object in the form of a binary label which can be either detectable as a genuine, meaning authentication success (“Pass”) or not detectable as genuine (“No result available”). As a reminder “detectable” is different from “detected”. This way, 15 digital signal representations from 4 sets (one set for each of the four genuine objects used for training) with the associated detectability values of the object genuineness were produced (Table 1, col. 4). These sets of digital signal representations of each of the genuine objects with the associated detectability values of the object genuineness (i.e., 4 sets of images; 1 set for each of the 4 genuine objects; summing to 15 images in a training set) were used for training of a predictive machine learning model to predict the detectability value of an object genuineness from a digital signal representation of an object.

Table 1

2B) - testing set for use in a method for identifying if an object is a genuine or a counterfeited Exemplified is a computer-implemented method for identifying if an object is a genuine or a counterfeited obj ect according to the invention, wherein the predictive machine learning model to predict a detectability value of the genuineness of the object to be identified, was trained according to the Example 2A, and wherein the obj ect to be identified is the same type of obj ect as used for training purposes in Example 2A.

Five objects of known property i.e., three genuine objects labelled #5, #7, and #8; and two fake objects labelled #6 and #9 (Table 2, col. 2) were used in a test as objects to be authenticated/identified. One or more images (i.e., digital signal representations) of these objects have been obtained by a sensor (labelled (pbject.capture) 5.1, 5.2, etc.) (Table 2, col. 1). It is understood that each of the image has been captured at a given capture condition.

In line with embodiment 1 illustrated by Figure 5, each digital signal representation of the object to be identified was input to the same genuineness detection algorithm as was used for training. Output was that the object to be identified is identified as genuine (“Pass”) or that the object to be identified cannot be identified as genuine (“No result available”) (Table 2, col. 3). If the object could not be identified as genuine, each digital signal representation of the object to be tested was input to the predictive machine learning model trained according to the methods of Example 2A. Output was a predicted detectability value of the genuineness of an object for the digital signal representation of the object to be identified (Table 2, col. 4). This allows to determine from the predicted detectability value of the genuineness of an object that the genuineness detection algorithm (used for authentication) can or cannot detect the object as a genuine object. Note that if the object could be identified as genuine in the first step by the genuineness detection algorithm, there is no need to apply the detectability prediction as there is no need to separate fakes from non detectable objects.

Finally, provided was object authentication final result (Table 2, col. 5) which was a decision based on the output from the predictive machine learning model (Table 2, col. 4) and on the output from the genuineness detection algorithm (Table 2, col. 3).

As summarised in Table 2:

- digital signal representations 5.1, 7.1, 7.2, 7.3, 7.4, 8.2, 8.3, 8.4: if the object is detected as a genuine object by the genuineness detection algorithm (100) (col. 3), it can be derived that the object is detectable as a genuine object (detectability value prediction (200) not needed, col. 4) and then provide output of genuine object (col. 5);

- digital signal representations 6.1, 9.1 : if the object is not identified as a genuine object by the genuineness detection algorithm (col. 3) while the predictive machine learning model indicates that the object is detectable as a genuine object (col. 4), then provide output of counterfeit (fake) object (col. 5); - digital signal representations 5.2, 8.1, 9.2: if the object is not identified as a genuine object by the genuineness detection algorithm (col. 3) and the predictive machine learning model indicates that the object is not detectable as a genuine object (col. 4), then provide output of “unable to authenticate” object (col. 5). Table 2

Further Applications

The advantages of the proposed invention may be further exemplified when experimenting the limitations of authentication of Euro banknotes with the ValiCash application as available in the Apple AppStore as of July 2023. This application is an example of a genuineness detection algorithm that produces a detectability value of the genuineness of a Euro banknote from a digital image representation of the Euro banknote artwork, using a smartphone camera capture. The application is sensitive to the capture conditions; with an overlay on the screen aligned to the banknote artwork, it forces the end user to manually position of the smartphone relative to the banknote with the right geometry (for instance ensuring that the banknote lays flat). However, even with the proper geometrical positioning, it is possible, in particular under certain capture conditions (light reflections, dirty camera) that the algorithm cannot detect the genuineness of a Euro banknote even if it is a genuine one. The app then suggests steps of manual checks to the end user (visual verification of the raised print, manual number, watermark). But the end user will still have a doubt whether the banknote is a fake one or a genuine one and will most likely assume it is a fake. In contrast, the proposed invention enables to primarily detect if the capture conditions are suitable for the underlying genuineness detection algorithm, thus enabling to re-orient the user to manually check the capture conditions rather than manually checking the banknote features. For instance, the application may suggest to clean or dry the sensor lens in a foggy or dusty environment, and/or to avoid to put the object under a transparent plastic sheet or piece of glass to have it lay flat, as this may cause flash spotlight reflections on the plastic or glass surface which may disturb the digital image representation signal processing for genuineness detection.

Other examples of applications which may be improved thanks to the proposed invention include:

- the gold bar security features employing a high resolution image processing algorithm to verify the authenticity of a gold bar according to the London Bullion Market Association independent authority standards (https://www.lbma.org.uk/good-delivery/gold-bar-security- features#-); - the covert anticounterfeiting technologies employing a digital image processing algorithm to detect counterfeited goods such as tobacco industry manufactured products; pharmaceuticals in accordance with the European False Medicines Directive (FMD) regulation and/or with the US Drug Quality and Security Act (DQSA) anti-counterfeiting protocol; and in general any goods or security documents for which digital anticounterfeiting technologies have been developed based on digital signal processing algorithms.

Claims

Claims:

1. A computer-implemented method for training a predictive machine learning model to predict a detectability value of the genuineness of an object, the method comprising: a) obtain a genuineness detection algorithm, which produces a detectability value of the genuineness of a type of object from a digital signal representation of an object, wherein the detectability value determines whether a digital signal representation of this object is sufficient to identify this object as genuine; b) obtain one or more genuine objects of the same type; c) obtain, with a sensor, a set of digital signal representations of each of the genuine objects, wherein each digital signal representation of each of the genuine objects is obtained from the sensor operating under a different capture condition; d) input each of the digital signal representations of each of the genuine objects (obtained in step c)) to the genuineness detection algorithm (obtained in step a)) and output a detectability value of the object genuineness for each of the digital signal representations of each of the genuine objects; e) train a predictive machine learning model to predict the detectability value of an object genuineness from a digital signal representation of this object using the sets of digital signal representations of each of the genuine objects with the associated detectability values of the object genuineness (obtained in step d).

2. The computer-implemented method for training a predictive machine learning model according to claim 1, wherein the genuineness detection algorithm is selected from detectors of surface fingerprints and detectors of product markings.

3. The computer-implemented method for training a predictive machine learning model according to any of the claims 1-2, wherein the trained predictive machine learning model is for use in a computer-implemented method to identify if an object is a genuine or counterfeited object.

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

5. The computer-implemented method for training a predictive machine learning model according to any of the claims 1-4, wherein each of the digital signal representations of each of the genuine objects is obtained from acquiring a signal captured with 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, a 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 of the claims 1-5, wherein the sensor and the genuine object relative position and/or orientation differ to produce a different sensor capture condition.

7. The computer-implemented method for training a machine learning model according to claim 6, wherein the different sensor position and/or orientation is controlled by a robot arm or a human operator that positions the sensor relative to the genuine object.

8. The computer-implemented method for training a machine learning model according to claim 6, wherein the different genuine object position and/or orientation is controlled by a robot arm, a conveyor or a human operator that positions the genuine object relative to the sensor.

9. The computer-implemented method for training a machine learning model according to any of the claims 1-8, wherein each of the digital signal representation of each of the genuine object is obtained at a capture condition, that is characterized by physical environment parameters around the genuine obj ect.

10. The computer-implemented method for training a machine learning model according to claim 9, wherein each of the digital signal representation of each of the genuine object is a digital image representation of the object, wherein at least one physical environment parameter around the genuine object is changing between different capture conditions, wherein this parameter is illumination of the object by a lighting device, and wherein the illumination of the object is controlled by a robot arm, a conveyor or a human operator that positions the lighting device relative to the genuine object.

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 the smartphone flash.

12. A computer-implemented method for predicting a detectability value of the genuineness of an object, the method comprising: a) obtain an object to be detected; b) obtain, with a sensor, a digital signal representation of the object to be detected; c) obtain a predictive machine learning model to predict a detectability value of the genuineness of an object, wherein the predictive machine learning model is trained according to the method of any of the claims 1-11; d) input the digital signal representation of the object to be detected (obtained in step b)) to the predictive machine learning model (obtained in step c)) and output a predicted detectability value of the genuineness of an object for the digital signal representation of the object to be detected.

13. A computer-implemented method for identifying if an object is a genuine or a counterfeited object, the method comprising: a) obtain a.1) an object to be identified; a.2) with a sensor, a digital signal representation of the object to be identified; a.3) a genuineness detection algorithm, which returns as an output a genuineness decision from the digital signal representation of the object to be identified a.4) a predictive machine learning model to predict a detectability value of the genuineness of the object to be identified with the genuineness detection algorithm (obtained in step a.3)), wherein the predictive machine learning model is trained according to the method of any of the claims 1-11; b) input the digital signal representation of the object to be identified (obtained in step a.2)) to the genuineness detection algorithm (obtained in step a.3)) and output

- that the object to be identified is identified as genuine, or

14. The computer-implemented method for identifying if an object is a genuine or a counterfeited object according to claim 13, wherein the object to be identified cannot be identified as genuine in step d), and the object is determined as not-detectable, wherein the method comprises further steps: e) obtain, with a sensor, another digital signal representation of the object to be identified; f) input another digital signal representation of the object to be identified (obtained in step e)) to the genuineness detection algorithm (obtained in step a.3)) and output

- that the object to be identified is identified as genuine, or

15. A computer-implemented method for identifying if an object is a genuine or a counterfeited object, the method comprising: a) obtain a.1) an object to be identified; a.2) with a sensor, a digital signal representation of the object to be identified; a.3) a genuineness detection algorithm, which returns as an output a genuineness decision from the digital signal representation of the object to be identified, a.4) a predictive machine learning model to predict a detectability value of the genuineness of the object to be identified with the genuineness detection algorithm (obtained in step a.3)), wherein the predictive machine learning model is trained according to the method of any of the claims 1-11; b) input the digital signal representation of the object to be identified (obtained in step a.2)) to the predictive machine learning model (obtained in step a.4)) and output a predicted detectability value of the genuineness of an obj ect for the digital signal representation of the object to be identified; c) determine from the predicted detectability value of the genuineness of an object (obtained in step b)) that the genuineness detection algorithm (obtained in step a.3)) can detect the object as a genuine object; d) input the digital signal representation of the object to be identified (obtained in step a.2)) to the genuineness detection algorithm (obtained in step a.3)) and output

- that the object to be identified is identified as genuine, or

16. The computer-implemented method for identifying if an object is a genuine or a counterfeited object according to claim 15, wherein the object to be identified cannot be identified as genuine in step c), and the object is determined as not-detectable, wherein the method comprises further steps: e) obtain, with a sensor, another digital signal representation of the object to be identified; f) input another digital signal representation of the object to be identified (obtained in step e)) to the predictive machine learning model (obtained in step a.4)) and output another predicted detectability value of the genuineness of an object for the digital signal representation of the object to be identified; g) determine from the predicted detectability value of the genuineness of an object (obtained in step f)) that the genuineness detection algorithm (obtained in step a.3)) can detect the object as a genuine object, and then perform step h) or

- that the object to be identified is identified as genuine, or

17. A computer-implemented method for identifying if an object is a genuine or a counterfeited object according to any of the claims 13-16, wherein the genuineness detection algorithm is selected from detectors of surface fingerprints and detectors of product markings.