Method for classifying images of fingerprints using a hybrid quantum-classical machine learning system

Abstract

Disclosed is a method for providing a trained machine learning system for classifying images of fingerprints, the method comprising: providing a training data set, comprising sets of images of fingerprints, wherein each image in the training data set further comprises a binary label encoding a class designation; using the training data set to train a machine learning system, wherein the machine learning system comprises a classical machine learning model implemented on a classical processor and a quantum machine learning model implemented on a quantum processing unit (QPU); encoding the image data of each image in the training data set into a feature vector using the classical machine learning model, obtaining a family of feature vectors; using the family of feature vectors obtained from the images in the training data set to train the classical machine learning model and the quantum machine learning model, for providing the trained machine learning system, comprising a trained classical machine learning model and a trained quantum machine learning model.

Description

METHOD FOR CLASSIFYING IMAGES OF FINGERPRINTS USING A HYBRID QUANTUM-CLASSICALMACHINE LEARNING SYSTEMFIELD OF THE INVENTION

[0001] The present invention relates to the field of image processing using trained machine learning models, and particularly to the classification of images of fingerprints.BACKGROUND

[0002] High-security systems are applied for the identification of individuals across a variety of domains, including law enforcement, criminal investigations, forensics, the restriction of access of individuals to secure facilities and the restriction of access of individuals to secure computer systems, wherein the computer system may store sensitive data. A method for the reliable identification of individuals is the analysis of biometric data, including the matching of fingerprints.

[0003] An individual may choose to provide a manipulated fingerprint in order to circumvent the high-security system using the analysis of fingerprints for identification. An individual may also possess a fingerprint having a form and shape similar to the fingerprint of another person. Given the consequences of an individual wrongfully being granted access to a restricted secure facility or computer system, or the consequences of an individual wrongfully being matched to the fingerprint of another person during a legal proceeding or forensic analysis, there arises a need for accurate methods for the processing and analysis of images of fingerprints.

[0004] The publication entitled "Fingerprint classification and identification algorithms for criminal investigation: A survey" by Win et al., published in "Future Generation Computer Systems, 110, 758-771" in September 2020 describes fingerprint classification algorithms and fingerprint enhancement techniques.

[0005] The publication entitled "Quantum Support Vector Machine for Big Data Classification" by Rebentrost et al., published in Physical Review Letters 113, 130503 on the 25.09.2014describes a quantum machine learning algorithm called a quantum support vector machine (QSVM).

[0006] The publication entitled "On Neural Quantum Support Vector Machines" by Simon and Radons, published on arXiv on the 24.11.2023 describes Neural Quantum Support Vector Machines (NQSVMs).SUMMARY OF THE INVENTION

[0007] It is an objective to provide for a system and method for the classification of images of fingerprints using a machine learning system, wherein the machine learning system is a hybrid quantum-classical machine learning system. The objectives underlying the invention are solved by the features of the independent claims.

[0008] Images of fingerprints are provided in digital format. The images of fingerprints possess characteristic patterns, said characteristic patterns comprising loops, whorls and archs. The locations and arrangements of said characteristic patterns allows for fingerprints to be associated with specific individuals. Each individual possesses a unique arrangement of said characteristic patterns, making the fingerprint of an individual unique to said individual.

[0009] Images of fingerprints can be manipulated. Examples of manipulation of images of fingerprints comprises blurring, distortion, rescaling, replacement and / or rearrangement of characteristic patterns of the images of fingerprints. Said manipulation can be performed with the goal of circumventing a security system relying on the analysis of fingerprints for identification of an individual, wrongfully granted access to a secure location to an unauthorized individual based on a manipulated fingerprint.

[0010] The analysis of images of fingerprints can be formulated as a classification problem, wherein said classification problem is solvable using a machine learning model. A preferred embodiment of said machine learning model is a binary classifier, wherein said binary classifier assigns an unlabeled input data point one of two labels, wherein each label encodes a class designation. A preferred embodiment of the binary classifier is a support vector machine (SVM). The support vector machine is a supervised machine learning model trained on a training dataset of labeled images of fingerprints, wherein each image in the training dataset further comprises a label according to one of the two classes of the binary classifier.

[0011] The machine learning model is trained on the training dataset, providing a trained binary classifier. Unlabeled images are provided to the trained machine learning model, wherein said unlabeled images are assigned a label according to those images in the training dataset most closely resembling the unlabeled image.

[0012] For the analysis of images of fingerprints, said machine learning model can be used to distinguish between manipulated and authentic fingerprints. Said machine learning model can further be used to distinguish between fingerprints of different individuals.

[0013] Advantageously, the information in an image of a fingerprint can be encoded into feature vectors before being provided as input to the machine learning model, wherein said feature vector comprise a lower-dimensional representation of the image, comprising characteristic features most representative of the encoded fingerprint. Said encoding into feature vectors reduces the complexity of the machine learning model, as the classification is performed in a lower-dimensional space. Alternatively, said encoding allows for the representation of more information given a constant dimension of the input of the machine learning model.

[0014] The machine learning model can comprise a classical machine learning model, wherein the classical machine learning model is executed on a classical computer system using a classical processor. The machine learning model can further comprise a quantum machine learning model, wherein the quantum machine learning model is executed on a quantum computer using a quantum processing unit (QPU), wherein the quantum processing unit refers to a component that may be used to perform quantum computations by operating on one or more interconnected qubits.

[0015] A use of quantum computers for execution of the quantum machine learning model allows for quantum effects to be used for the classification problem, said quantum effects comprising superposition and entanglement.

[0016] Advantageously, exploitation of quantum effects including superposition and entanglement in quantum algorithms allows for the encoding of problems into higherdimensional space than would be possible using purely classical algorithms. Said encoding into higher-dimensional spaces is enabled by the entanglement of qubits, wherein a set of N qubits allows for the encoding of features into an exponentially larger state space of dimension 2W. Advantageously, entanglement can be used to encode a larger number of features into said exponentially larger state space than would be possible using purely classical algorithms.

[0017] As a further advantage, the use of quantum effects allows for quantum algorithms to be applied in quantum machine learning models for the classification problem, wherein said quantum algorithms provide computational advantages compared to purely classical models.

[0018] Embodiments provide a method for providing a trained machine learning system for classifying images of fingerprints, the method comprising: providing a training data set, comprising sets of images of fingerprints, wherein each image in the training data set further comprises a binary label encoding a class designation; using the training data set to train a machine learning system, wherein the machine learning system comprises a classical machine learning model implemented on a classical processor and a quantum machine learning model implemented on a quantum processing unit (QPU); encoding the image data of each image in the training data set into a feature vector using the classical machine learning model, obtaining a family of feature vectors; using the family of feature vectors obtained from the images in the training data set to train the classical machine learning model and the quantum machine learning model, for providing the trained machine learning system, comprising a trained classical machine learning model and a trained quantum machine learning model.

[0019] Advantageously, the use of hybrid quantum-classical machine learning systems combines the benefits of classical and quantum computing. In a preferred embodiment, the machine learning system comprises the classical machine learning model and the quantum machine learning model, wherein the classical machine learning model processes the provided input image data in a first step, generating feature vectors comprising a lower-dimensional representation of the input image data. The quantum machine learning model processes said feature vectors obtained as output from the classical machine learning model in a second step. The feature vectors are encoded into quantum states of a quantum processing unit. The quantum machine learning model applies a quantum algorithm to said quantum states, providing measurement outcomes of the quantum processing unit as output. Said measurement outcomes are used to compute the binary labels corresponding to the class designations of the input images.

[0020] Embodiments provide a use of the trained machine learning system, the use comprising the following steps: receiving a test data set, comprising sets of images of fingerprints, from a user; using the trained classical machine learning model to encode the image data of each image in the test data set into a feature vector, obtaining a family of feature vectors; sending the family of feature vectors to a quantum cloud service provider, including instructions for encoding thefamily of feature vectors into a quantum state of a quantum processing unit for the trained quantum machine learning model; using the trained quantum machine learning model to obtain measurement outcomes from a quantum processing unit via the quantum cloud service provider; processing the obtained measurement outcomes to obtain binary labels for the classification of the images in the test data set; sending the set of binary labels of the images in the test data set to the user.

[0021] In a preferred embodiment of the use of the trained machine learning system, the binary labels encoding the class designations of the images of fingerprints may encode the class designations "authentic fingerprint" and "manipulated fingerprint". In an alternative embodiment, said binary labels may encode the class designations "fingerprint of person A" and "fingerprint of person B".

[0022] Exemplary embodiments of the classical machine learning model include a neural network, in particular a convolutional neural network (CNN), a support vector machine (SVM) or a combination of a neural network and / or a support vector machine, in particular a neural support vector machine (NSVM). In a preferred embodiment, the classical machine learning model further comprises the application of a dimensionality reduction technique on the feature vectors. In a preferred embodiment, said dimensionality reduction technique comprises an autoencoder neural network. Alternative embodiments of the dimensionality reduction technique comprise a principal component analysis (PCA), a linear discriminant analysis (LDA) or a t-stochastic neighbor embedding (t-SNE) method.

[0023] The input image data comprises images possessing a width in pixels and a height in pixels. The dimension of the input image data is the product of the width in pixels and the height in pixels. Said dimension of the input image data is too large to be processed directly by a quantum machine learning model on available and near-term quantum processing units. In a preferred embodiment, the classical machine learning model comprises a convolutional neural network. Convolutional neural networks can advantageously be applied to raw input image data to reduce the dimension of the input image data.

[0024] In a preferred embodiment, the dimensionality reduction technique comprises an autoencoder neural network. Said autoencoder neural network advantageously provides a compressed representation of the information contained in the input provided to saidautoencoder neural network, wherein said input comprises the output of the convolutional neural network.

[0025] In a preferred embodiment, the classical machine learning model is a pretrained model, wherein a transfer learning technique is used to transfer results obtained from the pretrained model to another application. For example, the transfer learning may be used to apply knowledge gained during the training of the pretrained model forthe extraction of characteristic features from images besides images of fingerprints to the extraction of characteristic features from images of fingerprints.

[0026] The transfer learning technique comprises a pretrained model and advantageously reduces the total training time of the machine learning system to the training time of the quantum machine learning model. In an embodiment using the transfer learning technique, the total time required for the execution of the method of the present subject matter is advantageously reduced to the time required forthe use of the pretrained model forthe generation of feature vectors, the training of the quantum machine learning model, the use of the quantum machine learning model using said feature vectors, and the processing of the measurement outcomes obtained by the quantum machine learning model to compute the binary labels of the input image data.

[0027] An advantage of quantum machine learning models is an improved performance given a low number of training samples in the training dataset compared to classical machine learning models. In an embodiment comprising the transfer learning technique, the method of the present subject matter can be performed using a smaller number of training samples in the training set compared to an embodiment in which the classical machine learning model comprises an untrained classical machine learning model and a training step for said classical machine learning model.

[0028] Exemplary embodiments of the quantum machine learning model include a quantum convolutional neural network (QCNN) and a quantum support vector machine (QSVM). Preferred embodiments of the machine learning system comprise a neural quantum support vector machine (NQSVM), wherein the classical machine learning model of the neural quantum support vector machine (NQSVM) comprises a neural support vector machine (NSVM) comprising at least one neural network and a dimensionality reduction technique. The quantum machine learning model of the neural quantum support vector machine (NQSVM) comprises a quantum supportvector machine (QSVM). The use of a quantum support vector machine provides an exponential advantage in computational complexity compared to a classical support vector machine. Advantages of the use of NSVMs and / or NQSVMs are the ability to incorporate and exploit domain knowledge in the model architecture by adapting the neural network to a given data type and the ability of quantum computers to access kernel functions in high-dimensional spaces for the construction of quantum kernels.

[0029] In a preferred embodiment, the same training data set of images of fingerprints is used to train the classical machine learning model and the quantum machine learning model. In an alternative embodiment, separate training data sets of images of fingerprints are used to train the classical machine learning model and the quantum machine learning model.

[0030] In a preferred embodiment, the feature vectors are mapped to quantum states on the quantum processing unit using angle encoding. In angle encoding, the features given by the feature vector of an image are mapped to quantum states by the use of single-qubit rotation gates parameterized by a rotation angle, wherein the rotation angle is dependent on the feature vector being encoded. Advantageously, angle encoding allows for a total of N values to be encoded on a set of N qubits, wherein the encoding requires only a single layer of quantum gates in a quantum circuits, therefore limiting the complexity of the resulting quantum circuit. In an alternative embodiment, the feature vectors are mapped to quantum states on the quantum processing unit using amplitude encoding. Advantageously, amplitude encoding allows for an exponentially larger number of values of feature vectors to be encoded in the same set of qubits, allowing for a total of 2Wvalues to be encoded in a set of N qubits.

[0031] Advantageously, the present subject matter provides an improved method for solving the problem of accurately classifying images of fingerprints, wherein a combination of a classical machine learning model and a quantum machine learning model in a hybrid quantum-classical machine learning system allows for the benefit of generating a set of feature vectors using the classical machine learning model, wherein said feature vectors encode the characteristic features of a fingerprint most relevant for a classification task, and wherein said feature vectors are given as input to a quantum machine learning model to benefit from the computational capacities of quantum computers, including the exploitation of quantum effects such as superposition and entanglement.

[0032] As a further advantage of the present subject matter, the use of the quantum machine learning model allows for computations to be performed in a high-dimensional state space of a set of entangled qubits, wherein a set of N entangled qubits are capable of encoding states in a 2w-dimensional space. Said encoding in a high-dimensional space allows the hybrid quantum- classical machine learning system to detect more complex patterns in the encoded data, yielding an improved generalization of the provided results, compared to a purely classical machine learning model. The reduction of image data to feature vectors encoding the characteristic features of the fingerprint is further advantageous in allowing the subsequent quantum machine learning model to use quantum circuits of reduced complexity compared to a direct encoding of image data into quantum states, allowing the method provided in the present subject matter to be performed on currently available and / or near-term quantum processing units.

[0033] As a further advantage of the present subject matter, quantum machine learning algorithms have been found to show improved performance compared to classical machine learning algorithms in particular for the case of noisy data and / or cases where the amount of available data for training a machine learning model is limited.

[0034] Embodiments provide a software package comprising a web interface and a trained machine learning system, wherein the software package allows a user to perform the use of the trained machine learning system. Said software package comprises a computer program installed on a computer system, wherein the software package handles the interactions between the user, the input data provided by the user, the computer system hosting the trained classical machine learning model and the quantum cloud service provider providing access to the quantum processing unit (QPU). The execution of the computer program included in the software package causes the computer system to perform the method for classifying images of fingerprints.

[0035] In a preferred embodiment of the method or software package provided in the present subject matter, the images of fingerprints are preprocessed according to the following steps: standardizing (203) images to a common format comprising a fixed number of pixels in the directions of width and height, obtaining a set of formatted images; cropping (205) the images to the fingerprint, obtaining a set of cropped images; statistical normalization (207) in accordance with the normalization method used by the classical machine learning model.

[0036] In a preferred embodiment of the method or software package, the provided images of fingerprints are gray scale or RGB (red, green, blue) images. In a preferred embodiment, the formatted images are represented by a tensor of data fields, comprising a width in pixels, aheight in pixels, and a single color value if the image is a gray scale image or three color values if the image is an RGB (red, green, blue) image.

[0037] It is understood that one or more of the aforementioned embodiments may be combined as long as the combined examples are not mutually exclusive.BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In the following, examples are described in greater detail making reference to the drawings in which:

[0039] Figure 1 is a block diagram of the interactions between the software package executing the method of claim 1, the user, the quantum cloud service provider and the quantum processing unit.

[0040] Figure 2 is a flowchart of the preprocessing steps applied to input image data.

[0041] Figure 3 is a flowchart of the training of the hybrid quantum-classical machine learning system.

[0042] Figure 4 is a flowchart of the information flow of the method of claim 1 as applied to input image data.

[0043] Figure s is a block diagram of an exemplary computer system for implementing at least part of the present method in accordance with an example of the present subject matter. While the invention has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative or exemplary and not restrictive; the invention is not limited to the disclosed examples.DETAILED DESCRIPTION

[0044] In the following, similar elements are denoted by the same reference numerals.

[0045] The block diagram in Fig. 1 is intended to be read in conjunction with Fig. 5.In Fig. 1, the software package 100 is installed on a server 108. The server is a computer system 602, comprising a processor 602, memory unit 605, a bus 607, a network adapter 609, a storage system 611, external devices 613 and an I / O interface 619. The software package 100 comprises the machine learning system 102. The machine learning system 102 may be the untrainedmachine learning system or the trained machine learning system, wherein the trained machine learning system further comprises a set of learnable parameters of the classical machine learning model 104 and the quantum machine learning model 106. The server 108 further comprises a web interface 101. The server 108 further comprises the machine learning system 102, wherein the machine learning system 102 includes the classical machine learning model 104 and the quantum machine learning model 106.

[0046] The computer device 110 is a computer system 602 comprising a processor 602, memory unit 605, a bus 607, a network adapter 609, a storage system 611, external devices 613 and an I / O interface 619. The user 112 may interact with the computer device 110, for example for providing a set of unlabeled images of fingerprints. Exemplary embodiments of the computer device 110 comprise a desktop computer, a laptop, a tablet and a mobile device.

[0047] The computer device 110 of the user 112 may interact with the web interface 101, wherein the web interface is hosted on the server 108. Embodiments of the web interface may comprise a graphical user interface, an upload function for receiving images of fingerprints, an interface to the software package 100 and the machine learning system 102 included therein, and a display function for displaying the results of an execution of the method for classifying fingerprints to the user 112 via the computer device 110.

[0048] The server 108 is further capable of interacting with a quantum cloud service provider 114. In a preferred embodiment, the interaction between the server 108 and the quantum cloud service provider 114 is handled by the software package 100. The quantum cloud service provider 114 receives instructions from the server 108 via the software package 100, wherein said instructions comprise instructions for implementing a quantum circuit for executing the quantum machine learning model 106 on the quantum processing unit 116. Said instructions comprise a choice of encoding to use for mapping feature vectors to quantum states, wherein the choice comprises an angle encoding and an amplitude encoding. Said instructions further comprise a number for how many shots of the quantum circuit are to be performed by the quantum processing unit, wherein a shot indicates a single execution of the quantum circuit implementing the quantum machine learning model, including a measurement and the obtaining of a measurement value. Said instructions may further comprise error mitigation and / or error correction techniques to be applied to the quantum circuit implementing the quantum machine learning model. If the quantum machine learning model 106 is a trained quantum machine learning model, said instructions may further comprise the trained learnable parameters of the quantum machine learning model.

[0049] The quantum cloud service provider 114 executes the quantum circuit implementing the quantum machine learning model as provided by the server 108 on the quantum processing unit 116. The quantum cloud service provider 114 then receives measurement outcomes 308 from the quantum processing unit 116 and sends these measurement outcomes 308 to the server 108.

[0050] In a preferred embodiment, the measurement outcomes 308 are provided to the server 108 as a dictionary, wherein the keys of the dictionary are bitstrings denote the measured state of the qubits on the quantum processing unit and wherein the values associated with said keys denote the number of times said measured state was obtained. The total number of all values in the dictionary sums up to the number of shots performed on the quantum processing unit, wherein a shot denotes a single initialization, execution and measurement of a quantum circuit.

[0051] The server 108 receives the measurement outcomes 308 of the quantum processing unit 116 from the quantum cloud service provider 114. Using said measurement outcomes 308, the server 108 executes postprocessing steps on a classical processor 603 to compute the binary labels 300 corresponding to the class designations of the unlabeled images of fingerprints from the measurement outcomes 308. The computed binary labels 300 are returned as an output to the computer device 110 of the user 112 via the web interface 101.

[0052] The flowchart in Fig. 2 shows a preprocessing method 200 applied to the input image data before processing by the machine learning system 102. A set images of fingerprints is provided by the user 112 as raw image data in digital form. Said images possess a width measured in pixels and a height measured in pixels. Said images may be colored RGB images or gray scale images. The digital form of the images provides RGB values per pixel for RGB images or gray scale values per pixel for gray scale images. In step 201, the raw image data is received as input image data by the software package 100.

[0053] In step 203, the input image data is standardized to conform to the requirements of a homogeneous format. Embodiments of said homogeneous format comprises fixed values for the width measured in pixels, the height measured in pixels and / or the relation between the width and the height. In a preferred embodiment, the ratio between the width and the height is set to 3:1. In a preferred embodiment, the requirements of the homogeneous format are set by the software package 100. In a preferred embodiment, the homogeneous format is a tensor with data fields for width in pixels, height in pixels and one data value per pixel for gray scale images or three data values per pixel for RGB images.

[0054] In step 205, the images are cropped to the fingerprint, obtaining cropped homogeneously formatted images. The cropping removes parts of the input image data not belonging to the fingerprint, advantageously reducing the information provided to the machine learning system 102.

[0055] In step 207, statistical normalization of the cropped homogeneously formatted images is performed, obtaining preprocessed image data. The statistical normalization is performed in accordance with a normalization method specified by the classical machine learning model 104.

[0056] If the input image data comprises gray scale images, the steps 203, 205 and 207 are performed for the gray scale values of the input image. If the input image data comprises RGB images, the steps 203, 205 and 207 are performed for all RGB values of the input image.

[0057] In step 209, the preprocessed image data is provided by the software package 100. The preprocessed image data possesses the required format for use in the machine learning system 102.

[0058] The flowchart in Fig. 3 shows the method steps of the present subject matter for the training of the hybrid quantum-classical machine learning system and the application of said machine learning system to images of fingerprints.

[0059] In step 301, the server 108 receives a dataset of images of fingerprints comprising a training dataset 118 containing labeled images of fingerprints and a test dataset 120 containing unlabeled images of fingerprints, wherein the dataset of images of fingerprints has been preprocessed according to the preprocessing method 200. In the following steps 303, 305 and 307, an untrained hybrid quantum-classical machine learning system 102, comprising an untrained classical machine learning model 104 and an untrained quantum machine learning model 106, is trained.

[0060] In step 303, the training dataset 118 is used to train an untrained classical machine learning model 104. In a preferred embodiment, the classical machine learning model 104 comprises a pretrained model.

[0061] In a preferred embodiment, the machine learning system 102 comprises a neural quantum support vector machine. A neural quantum support vector machine is a neural support vector machine, wherein the kernel of the neural support vector machine is a quantum kernel. A quantum kernel is a matrix, wherein the entries of said matrix quantify the overlap betweenquantum states corresponding to feature vectors of images in the training dataset 118. The overlap of quantum states is a measure of the similarity of said states. In said embodiment, the classical machine learning model 104 comprises a neural network, a classical support vector machine and a dimensionality reduction technique 306. In said embodiment the quantum machine learning model comprises a quantum support vector machine, wherein the quantum support vector machine is used to compute the kernel of the classical support vector machine as a quantum kernel.

[0062] In said embodiment, the neural network of the classical machine learning model 104 is a pretrained model. In a preferred embodiment, said pretrained model is a pretrained convolutional neural network. In a first step of the classical machine learning model 104, using the transfer learning technique, said pretrained model is used to generate feature vectors in step 305 by providing the labeled images in the training dataset 118, wherein each image in the training dataset 118 generates a separate feature vector. In a second step of said preferred embodiment of the classical machine learning model 104, said feature vectors are provided as input for a dimensionality reduction technique 306, wherein said dimensionality reduction technique 306 comprises an autoencoder neural network.

[0063] The autoencoder neural network is trained to copy its input to its output, wherein a lower-dimensional encoding is used to represent the input. Said lower-dimensional encoding minimizes the reconstruction error of the output of the autoencoder neural network while reducing the dimension of the feature vectors. Advantageously, the autoencoder neural network is an unsupervised learning technique and does not require input from the user 112 or the server 108 for its training.

[0064] The feature vectors obtained as output from the second step of the classical machine learning model 104 are vectors of a fixed length k. In a preferred embodiment, the fixed length k of the feature vectors generated in step 305 matches the length k of the inputs used for training the quantum support vector machine on the feature vectors. In an exemplary embodiment, the length k of the feature vector that represents the image of a fingerprint may be defined based on the number of qubits in the quantum processing unit 116.

[0065] In a preferred embodiment, the neural network of the classical machine learning model 104 and the quantum support vector machine of the quantum machine learning model 106 are trained using the same training dataset 118. Advantageously, said training using the same training dataset 118 enables the machine learning system 102 to extract the most relevantcharacteristic features in the images of fingerprints, improving classification accuracy. In an alternative embodiment, the training dataset 118 used for training the classical machine learning model 104 differs from the training dataset 118 used for training the quantum machine learning model.

[0066] In another embodiment, the quantum machine learning model 106 comprises a quantum convolutional neural network. In said embodiment, the feature vectors provided by the classical machine learning model 104 are mapped to quantum states using an encoding scheme. Said quantum states are then processed by a quantum circuit consisting of repeated convolutional layers and pooling layers. Said convolutional layers and pooling layers are followed by a measurement, wherein the measurement outcome is used to compute a binary label 300 of the input image data.

[0067] In step 307, an untrained quantum machine learning model is trained using the feature vectors generated in step 305 from the classical machine learning model 104. Said feature vectors are mapped to quantum states using a quantum feature map. In a preferred embodiment, the mapping of a feature vector to a quantum state is performed using angle encoding, wherein the angle encoding is implemented in a quantum circuit by a layer of singlequbit rotation gates applied to said set of qubits, wherein the angles of ration of the rotation gates are parametrized based on the feature vector being encoded.

[0068] The quantum feature map performs encoding of classical data into quantum state space, feature vectors are mapped to the entangled quantum state of a set of qubits, wherein a set of N qubits can encode a feature vector of length 2W. A preferred embodiment of the quantum feature map is a parametrized quantum circuit < >(%). In a preferred embodiment, the quantum feature map < >(%) comprises a parametrized quantum circuit comprising a layer of Hadamard gates applied to a set of qubits, followed by a layer of single-qubit rotation gates applied to said set of qubits, wherein the angles of ration of the rotation gates are parametrized based on the classical feature vector being encoded. The layer of rotation gates is followed by a set of operations comprising multi-qubit entangling gates and single-qubit gates. Said parametrized quantum circuit is repeatedly measured, wherein said measurement corresponds to an evaluation of the quantum kernel.

[0069] In a preferred embodiment, the quantum kernel is a fidelity quantum kernel, wherein the kernel function K(x, y) is defined as the overlap of two quantum states <|)(x)| and|4>(y)) defined by a parametrized quantum circuit as K(x,y) = | 4)(%)|< )(y)}|2, wherein thedata point x is mapped to quantum state by the quantum feature map (j)(x) via the parametrized quantum circuit.

[0070] An alternative embodiment of the quantum feature map comprises a quantum circuit implementing a second-order Pauli-Z expansion for the encoding of the classical feature vectors.

[0071] In step 309, a trained quantum machine learning model 106 is obtained. In a preferred embodiment, the trained quantum machine learning model is a trained quantum support vector machine, comprising the following: a feature map for mapping input feature vectors from the classical machine learning model 104 to quantum states on the quantum processing unit 116 and a quantum kernel containing values quantifying the similarity between feature vectors of images in the training dataset 118. The trained quantum machine learning model 106 is stored on the server 108 and used to classify unlabeled images of fingerprints provided by the user 112.

[0072] A preferred embodiment of the quantum processing unit 116 is a trapped-ion quantum computer. Alternative embodiments of the quantum processing unit 116 comprise quantum computers based on superconducting Josephson junctions, quantum computers based on NV centers in diamond, quantum computers based on neutral atoms and optical quantum computers.

[0073] The flowchart in Fig. 4 shows the steps of the data processing pipeline applied to the input image data. In step 401, the input image data comprising the images of fingerprints is provided to the server 108. The input image data may comprise unlabeled images provided by the user 112 via the web interface 101 of the software package 100, wherein said unlabeled images are used as the test dataset 120 for the machine learning system 102. The input image data may further comprise labeled images of the training dataset 118.

[0074] In step 200, the image data received by the server is preprocessed and normalized as described by the preprocessing method in Fig. 2, obtaining preprocessed input image data.

[0075] In step 403, the preprocessed input image data is used to generate feature vectors using the trained classical machine learning model of the machine learning system as described in step 305.

[0076] Step 405, information is sent to a quantum cloud service provider 114, said information comprising: the feature vectors generated in step 403, a quantum kernel matrix for the trained quantum machine learning model, a choice of encoding scheme, wherein the encoding schemecomprises angle encoding or amplitude encoding, a choice of quantum machine learning model 106, wherein the choice of quantum machine learning model 106 comprises a quantum support vector machine, neural quantum support vector machine or quantum neural network, and execution instructions. Embodiments of said execution instructions comprise a number of shots to be executed on the quantum processing unit 116, instructions for the mapping of physical qubits on the quantum processing unit to logical qubits of the quantum machine learning model 106, instructions for error mitigation, instructions for error correction and / or instructions for an optimization of the quantum circuit specified by the quantum machine learning model 106 to the quantum processing unit 116. The quantum circuit specified by the quantum machine learning model 106 is then executed on the quantum processing unit 116 by the quantum cloud service provider 114 in accordance with the sent information and execution instructions, obtaining measurement outcomes 308.

[0077] In step 407, said measurement outcomes 308 are received by the server 108 from the quantum cloud service provider 114.

[0078] In step 409, postprocessing steps are performed on the server 108, wherein the measurement outcomes 308 are used to compute the binary labels 300 corresponding to the class designations of the images of the input image data.

[0079] In step 411, said class designations of the images of the input image data are provided to the user 112 via the web interface 101 of the software package 100.

[0080] Fig. 5 shows a block diagram of an exemplary computer system for implementing the present method in accordance with an example of the present subject matter. The computer system may include the computer device 301 of the user 302, wherein the computer device 110 may refer to a desktop computer, laptop, tablet or mobile device. The computer system may alternatively refer to the server 108.

[0081] In a preferred embodiment, the computer system 602 includes the server 108, the computer device 110 of the user 112 or the joint system of the server 108 and the computer device 110 of the user 112. In operation, the computer system 602 may be configured to execute the interactions between the software package 100, the server 108, the computer device 110, the web interface 101 and the quantum cloud service provider 114. Said computer system 602 may be configured to execute the method for classifying images of fingerprints using the trained hybrid quantum-classical machine learning system as described in the preceding figures Fig. 1 to Fig. 4.

[0082] The components of the computer system 602 may include, but are not limited to, one or more processors or processing units 603, a storage system 611, a memory unit 605, and a bus 607 that couples various system components including memory unit 605 to processor 603. The storage system 611 may include for example a hard disk drive (HDD). The memory unit 605 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory.

[0083] The computer system 602 may also communicate with one or more external devices such as a keyboard, a pointing device, a display 613, etc.; one or more devices that enable a user to interact with computer system 602; and / or any devices (e.g., network card, modem, etc.) that enable the computer system 602 to communicate with one or more other computing devices. Such communication can occur via I / O interface(s) 619. Still yet, the computer system 602 can communicate with one or more networks such as a local area network (LAN), a general wide area network (WAN), and / or a public network (e.g., the Internet) via a network adapter 609. As depicted, the network adapter 609 communicates with the other components of the client system 602 via bus 607.

[0084] The memory unit 605 is configured to store applications that are executable on the processor 603. For example, the memory unit 605 may comprise an operating system as well as one or more application programs. The application programs comprise instructions that when executed enable to perform the method described with reference to Fig. 2.

[0085] As will be appreciated by one skilled in the art, aspects of the present invention may be embodied as an apparatus, method, computer program or computer program product. Accordingly, aspects of the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment (including firmware, resident software, microcode, etc.) or an embodiment combining software and hardware aspects that may all generally be referred to herein as a "circuit," "module" or "system." Furthermore, aspects of the present invention may take the form of a computer program product embodied in one or more computer readable medium(s) having computer executable code embodied thereon. A computer program comprises the computer executable code or "program instructions".

[0086] The term "computer system" refers to data processing hardware and encompasses all kinds of apparatus, devices, and machines for processing data, including by way of example, a programmable processor, a computer, or multiple processors or computers. The apparatus can also be or further include special purpose logic circuitry, e.g., a central processing unit (CPU), aFPGA (field programmable gate array), or an ASIC (application specific integrated circuit). In some implementations, the data processing apparatus and / or special purpose logic circuitry may be hardware-based and / or software-based. The apparatus can optionally include code that creates an execution environment for computer programs, e.g., code that constitutes processor firmware, a protocol stack, a database management system, an operating system, or a combination of one or more of them. The present disclosure contemplates the use of data processing apparatuses with or without conventional operating systems, for example LINUX, UNIX, WINDOWS, MAC OS, ANDROID, IOS or any other suitable conventional operating system.

[0087] Any combination of one or more computer readable medium(s) may be utilized. The computer readable medium may be a computer readable storage medium. A 'computer-readable storage medium' as used herein encompasses any tangible storage medium which may store instructions which are executable by a processor of a computing device. The computer-readable storage medium may be referred to as a computer-readable non-transitory storage medium. The computer-readable storage medium may also be referred to as a tangible computer readable medium. In some embodiments, a computer-readable storage medium may also be able to store data which is able to be accessed by the processor of the computing device.

[0088] 'Computer memory' or 'memory' is an example of a computer-readable storage medium. Computer memory is any memory which is directly accessible to a processor. 'Computer storage' or 'storage' is a further example of a computer-readable storage medium. Computer storage is any non-volatile computer-readable storage medium. In some embodiments computer storage may also be computer memory or vice versa.

[0089] A 'processor' as used herein encompasses an electronic component which is able to execute a program or machine executable instruction or computer executable code. References to the computing device comprising "a processor" should be interpreted as possibly containing more than one processor or processing core. The processor may for instance be a multi-core processor. A processor may also refer to a collection of processors within a single computer system or distributed amongst multiple computer systems. The term computing device should also be interpreted to possibly refer to a collection or network of computing devices each comprising a processor or processors. The computer executable code may be executed by multiple processors that may be within the same computing device or which may even be distributed across multiple computing devices.

[0090] Computer executable code may comprise machine executable instructions or a program which causes a processor to perform an aspect of the present invention. Computer executable code for carrying out operations for aspects of the present invention may be written in any combination of one or more programming languages, including an object oriented programming language such as Java, Smalltalk, C++ or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages and compiled into machine executable instructions. In some instances the computer executable code may be in the form of a high level language or in a pre-compiled form and be used in conjunction with an interpreter which generates the machine executable instructions on the fly.

[0091] Generally, the program instructions can be executed on one processor or on several processors. In the case of multiple processors, they can be distributed over several different entities. Each processor could execute a portion of the instructions intended for that entity. Thus, when referring to a system or process involving multiple entities, the computer program or program instructions are understood to be adapted to be executed by a processor associated or related to the respective entity.

[0092] While the invention has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative or exemplary and not restrictive; the invention is not limited to the disclosed examples.REFERENCE SIGNS LIST100 Software package101 Web interface102 Machine learning system104 Classical machine learning model106 Quantum machine learning model107 Classical processor108 Server110 Computer device112 User114 Quantum cloud service provider (QCSP)116 Quantum processing unit (QPU)118 Training dataset120 Test dataset200 Preprocessing method201 Receiving raw image data as input203 Standardizing raw image data205 Cropping image data207 Statistical normalization209 Providing preprocessed image data300 Binary labels301 Receiving dataset303 Training classical machine learning model305 Feature vectors306 Dimensionality reduction technique307 Training quantum machine learning model308 Measurement outcomes309 Obtaining trained quantum machine learning model401 Providing input image data403 Generating feature vectors405 Sending information to quantum cloud service provider407 Receiving measurement outcomes409 Postprocessing411 Providing class designations602 Computer system603 Processors605 Memory unit607 Bus609 Network adapter611 Storage system613 External devices619 I / O interface

Claims

CLAIMS1. A method for providing a trained machine learning system (102) for classifying images of fingerprints, the method comprising:Providing a training data set (118), comprising sets of images of fingerprints, wherein each image in the training data set further comprises a binary label (300) encoding a class designation,Using the training data set (118) to train a machine learning system (102), wherein the machine learning system (102) comprises a classical machine learning model (104) implemented on a classical processor (107) and a quantum machine learning model (106) implemented on a quantum processing unit (QPU) (116),Encoding the image data of each image in the training data set (118) into a feature vector (305) using the classical machine learning model (104), obtaining a family of feature vectors (305),Using the family of feature vectors obtained from the images in the training data set (118) to train the classical machine learning model (104) and the quantum machine learning model (106), for providing the trained machine learning system, comprising a trained classical machine learning model (104) and a trained quantum machine learning model (106).

2. The method of claim 1, wherein the binary labels (300) of the images in the training data set (118) encode the class designations "authentic fingerprint" and "manipulated fingerprint".

3. The method of claim 1, wherein the binary labels (300) of the images in the training data set (118) encode the class designations "fingerprint of person A" and "fingerprint of person B".

4. The method according to any of the preceding claims, wherein the classical machine learning model (104) comprises a neural network, in particular a convolutional neural network (CNN), a support vector machine (SVM) or a combination of a neural network and a support vector machine.

5. The method according to any of the preceding claims, wherein the classical machine learning model (104) comprises a pretrained model and a transfer learning technique.

6. The method according to any of the preceding claims, wherein the classical machine learning model (104) further comprises the application of a dimensionality reduction technique (306) on the family of feature vectors.

7. The method according to any of the preceding claims, wherein the dimensionality reduction technique (306) comprises an autoencoder neural network.

8. The method according to any of the preceding claims, wherein the quantum machine learning model (106) comprises a quantum convolutional neural network (QCNN), quantum support vector machine (QSVM) and / or a neural quantum support vector machine (NQSVM).

9. The method according to any of the preceding claims, wherein the feature vectors (305) obtained from the classical machine learning model are used to train the quantum machine learning model (106).

10. The method according to any of the preceding claims, wherein separate training data sets (118) are used to train the classical machine learning model (104) and the quantum machine learning model (106).

11. The method according to any of the preceding claims, wherein the feature vectors (305) are mapped to quantum states on the quantum processing unit (116) using angle encoding, encoding a total of N values in a set of N qubits.

12. The method according to any of the preceding claims, wherein the feature vectors (305) are mapped to quantum states on the quantum processing unit using amplitude encoding, encoding a total of 2Wvalues in a set of N qubits.

13. A use of the trained machine learning system (102) as obtained using the method of claim 1, the use comprising the following steps:Receiving a test data set (120), comprising sets of images of fingerprints, from a user,Using the trained classical machine learning (104) model to encode the image data of each image in the test data set (120) into a feature vector (305), obtaining a family of feature vectors,Sending the family of feature vectors (305) to a quantum cloud service provider (114), including instructions for encoding the family of feature vectors (305) into a quantum state of a quantum processing unit (116) for the trained quantum machine learning model (106),Using the trained quantum machine learning model (106) to obtain measurement outcomes (308) from a quantum processing unit (116) via the quantum cloud service provider (114),Postprocessing the obtained measurement outcomes (308) to obtain binary labels (300) for the classification of the images in the test data set,Sending the set of binary labels of the images in the test data set to the user.

14. A software package (100) comprising a web interface (101) and a trained machine learning system (102) as obtained using the method of claim 1, wherein the software package (100) allows a user to perform the use of the trained machine learning system (102) of claim 13.

15. The method or software package (100) according to any of the preceding claims, wherein the images of fingerprints are preprocessed according to the following steps:Standardizing (203) images to a common format comprising a fixed number of pixels in the directions of width and height, obtaining a set of formatted images, Cropping (205) the images to the fingerprint, obtaining a set of cropped images, Statistical normalization (207) in accordance with the normalization method used by the classical machine learning model.

16. The method or software package (100) according to any of the preceding claims, wherein the images of fingerprints are gray scale or RGB images.

17. The method or software package according to any of the preceding claims, wherein the formatted images are represented by a tensor of data fields, comprising a width in pixels, a height in pixels, and a single color value if the image is a gray scale image or three color values if the image is an RGB image.

Images