Secure authentication

Abstract

A method for authenticating an authorized user is disclosed, the method comprising: receiving a request to access a resource; in response to receiving the request to access the resource, triggering irradiation of the object with light, and triggering generation of an image of the object while the object is irradiated with light; receiving a template image of an authorized user; providing the image and the template image to a data-driven model trained with a plurality of training images obtained by generating a plurality of synthetic representations of the person generated from representations of the person to determine whether an object associated with the image corresponds to an authorized user, and enhancing the plurality of synthetic representations of the person to generate a plurality of training images; the authorized user is allowed to access the resource in response to determining that the object corresponds to the authorized user.

Description

Technical Field

[0001] This disclosure relates to methods for authenticating authorized users, methods for generating multiple training images suitable for training a data-driven model to determine whether an object associated with an image corresponds to an authorized user, methods for training a data-driven model to determine whether an object associated with an image corresponds to an authorized user, a non-transitory computer-readable storage medium, uses of multiple training images, uses of a data-driven model trained based on multiple training images, and devices and / or systems for authenticating authorized users. Background Technology

[0002] Secure authentication of authorized users requires a large amount of training data. Collecting such a large amount of training data requires significant resources, but it improves the reliability of the authentication algorithm. Summary of the Invention

[0003] Any disclosures, embodiments, and examples described herein relate to the methods, systems, apparatuses, chemical products, and computer elements listed above and below. Advantageously, the benefits provided by any embodiments and examples also apply to all other embodiments and examples.

[0004] In one aspect, this disclosure relates to a method for authenticating an authorized user, the method comprising: receiving a request to access a resource; in response to receiving the request to access the resource, triggering the illumination of an object with light, and triggering the generation of an image of the object when the object is illuminated; receiving a template image of an authorized user; providing the image and the template image to a data-driven model to determine whether an object associated with the image corresponds to an authorized user, wherein the data-driven model is trained based on a plurality of training images obtained by: generating a plurality of synthetic representations of a person based on a person's representation, and enhancing the plurality of synthetic representations of the person to generate a plurality of training images; and allowing the authorized user to access the resource based on the determination that the object corresponds to an authorized user.

[0005] On the other hand, this disclosure relates to a device and / or system for authenticating an authorized user, the device and / or system comprising: a processor; and a memory storing instructions that, when executed by the processor, configure the device and / or system to perform any of the methods described herein.

[0006] On the other hand, this disclosure relates to the use of data-driven models trained on multiple training images obtained through any of the methods described herein.

[0007] On the other hand, this disclosure relates to the use of training images obtained through any of the methods described herein for training data-driven models for certified authorized users.

[0008] On the other hand, this disclosure relates to an authentication system that is trained using training images obtained by any of the methods described herein and / or trained by any of the methods described herein.

[0009] On the other hand, this disclosure relates to a method for authenticating an authorized user, the method comprising: receiving a request to access a resource; in response to receiving the request to access the resource, triggering the generation of an image of an object based on light received from the direction of the object; receiving a template image of an authorized user; determining whether an object associated with the image corresponds to the authorized user by providing the image and the template image to a data-driven model, wherein the data-driven model is trained using multiple training images obtained by: generating multiple synthetic representations of a person generated from a representation of a person, and enhancing the multiple synthetic representations of the person to generate multiple training images; and allowing the authorized user to access the resource based on the determination that the object corresponds to the authorized user.

[0010] On the other hand, this disclosure relates to a method for obtaining multiple training images for training a data-driven model for determining whether an object associated with an image corresponds to an authorized user, the method comprising: providing a representation of a person, generating multiple synthetic representations of the person based on the representation of the person, generating multiple training images by enhancing the multiple synthetic representations, and providing the multiple training images.

[0011] On the other hand, this disclosure relates to a method for authenticating an authorized user, the method comprising: receiving an image of an object; providing the image to a data-driven model to determine whether an object associated with the image corresponds to an authorized user, wherein the data-driven model is trained based on a plurality of training images generated by: generating a plurality of synthetic representations of a person based on a person's representation, and enhancing the plurality of synthetic representations of a person to generate a plurality of training images; and allowing the authorized user to access resources based on the determination that the object corresponds to an authorized user.

[0012] On the other hand, this disclosure relates to a method for training a data-driven model for determining whether an object associated with an image corresponds to an authorized user, the method comprising: providing a plurality of training images generated by any of the methods described herein, training the data-driven model based on the plurality of training images, and optionally, providing the trained data-driven model.

[0013] On the other hand, this disclosure relates to a device and / or system for authenticating an authorized user, the device and / or system comprising: a processor; and a memory storing instructions that, when executed by the processor, configure the device and / or system to perform any of the methods described herein.

[0014] On the other hand, this disclosure relates to a method for training a data-driven model for determining whether an object associated with an image corresponds to an authorized user, the method comprising: receiving a request for training the data-driven model; providing a plurality of training images obtained by: generating a plurality of synthetic representations of a person generated from a representation of a person, and augmenting the plurality of synthetic representations of the person to generate a plurality of training images; training the data-driven model based on the plurality of training images; and optionally providing the data-driven model.

[0015] On the other hand, this disclosure relates to an apparatus, particularly a communication apparatus for authenticating an authorized user, the apparatus comprising: an illumination source for illuminating an object with light; a camera for generating an image of the object when it is illuminated; and a processor for receiving a template image of an authorized user, providing the image and the template image to a data-driven model to determine whether an object associated with the image corresponds to an authorized user, wherein the data-driven model is trained using multiple training images obtained by: generating multiple synthetic representations of a person generated from a representation of a person, and enhancing the multiple synthetic representations of the person to generate multiple training images; and allowing the authorized user to access resources based on the determination that the object corresponds to an authorized user. Example

[0016] In the following sections, the terminology and / or technical fields used herein and / or the scope of this disclosure will be outlined by way of definition and / or examples. Where examples are given, it should be understood that this disclosure is not limited to those examples.

[0017] Secure authentication requires reliable user identification. To achieve this, data-driven models are trained using thousands to hundreds of thousands of different images. Generating so many different images for training these models requires significant time and resources. These images need to represent the vast majority of perspectives a user can present to the authentication device and / or system. However, common authentication algorithms may fail to distinguish between people who look alike, such as siblings, especially twins. Furthermore, adapting current authentication algorithms to new scenarios is time-consuming. Therefore, it is desirable to improve the reliability of authentication algorithms while simultaneously enhancing their adaptability.

[0018] This can be achieved by training a data-driven model for authenticating authorized users based on multiple training images generated as described in this paper. These training images can be generated based on readily available and easily collected representations of people. Based on these representations, multiple synthetic representations of a person can be created. This increases the number of different training images to be generated and thus improves the reliability of the data-driven model trained on these images. Multiple synthetic representations of a person can be associated with different perspectives of the person, for example, by changing the orientation of the person's head between two synthetic representations of the person. This makes it possible to effectively distinguish between people with similar appearances. Therefore, this feature aligns with the goal of improving the reliability of the authentication algorithm. Furthermore, enhancing multiple synthetic representations to generate multiple training images allows for the customization of the conditions of the training images, such as background or lighting, based on already available training data. Therefore, the already available training data can be further utilized, while the training images enrich the training data, thereby significantly improving the reliability of the authentication algorithm. Finally, training the data-driven model based on the training images described in this paper allows for the implementation of customized and effective authentication algorithms.

[0019] These and other objectives are addressed by the subject matter of the independent claims, and will become apparent upon reading the following description. The dependent claims relate to embodiments of the content of this disclosure.

[0020] In this embodiment, authentication may refer to facial authentication. The training images may show at least a portion of a person's face. The representation of a person may be a representation of a person's face.

[0021] In this embodiment, the light may be infrared light. Infrared light may include near-infrared light, mid-infrared light, and / or far-infrared light. Near-infrared light may be in the range of 780 nm to 3000 nm, excluding the value of 3000 nm. Mid-infrared light may be in the range of 3 µm to 15 µm, excluding the value of 15 µm. Far-infrared light may be in the range of 15 µm to 1000 µm. Infrared light may be associated with wavelengths between 700 nm and 800 nm and / or between 1000 nm and 1200 nm.

[0022] In embodiments, the data-driven model can be an augmented data-driven model and / or a liveness data-driven model. A data-driven model can refer to a model suitable for describing one or more non-linear relationships between input data and output data. Input data can refer to data to be provided to the data-driven model and / or data received by the data-driven model. Output data can be data to be received from the data-driven model and / or data to be provided by the data-driven model. Therefore, the data-driven model can determine output data based on transforming input data via one or more non-linear relationships. In this context, input data can be, for example, one or more representations of an image and / or a person. Output data can be, for example, whether the object associated with the image can be an instruction from an authorized user, and / or one or more training images.

[0023] A data-driven model can be trained on multiple training images to determine similarity scores associated with at least two of the training images. If the similarity score falls within a predefined range, it indicates that the two training images depict the same person. The predefined range can be a numerical range specified by at least one threshold. If the similarity score exceeds the threshold, it indicates that the two training images depict the same person.

[0024] A data-driven model used to determine whether an object associated with an image corresponds to an authorized user may include one or more embedding layers. These embedding layers may be configured to reduce the dimensionality of the image. The embedding layers may transform the image and / or template image into a dimensional representation, such as a tensor, particularly a two-dimensional or one-dimensional tensor. The data-driven model used to determine whether an object associated with an image corresponds to an authorized user may be trained and / or parameterized to reduce the dimensionality of the image and / or template image. The data-driven model may be trained and / or parameterized to generate the dimensionality-reduced representation associated with the image and / or template image. Alternatively, receiving the template image may include receiving a tensor associated with the template image. The tensor associated with the template image may be obtained by providing the template image to one or more embedding layers of, for example, a second data-driven model. The second data-driven model may include one or more encoders. Additionally or alternatively, the data-driven model used to determine whether an object associated with an image corresponds to an authorized user may include one or more classification layers. The one or more classification layers may be configured to receive tensors associated with an image and tensors associated with a template image, and / or to classify the tensors associated with the image and tensors associated with the template image based on whether the image and the template image can be associated with an authorized user, preferably the same authorized user. Additionally or alternatively, the data-driven model for determining whether an object associated with an image corresponds to an authorized user may include one or more mathematical relationships, particularly Euclidean distance, and / or cosine similarity, for determining the distance between the tensors associated with the image and the tensors associated with the template image.

[0025] In an embodiment, the liveness data-driven model can be a liveness classification data-driven model. The liveness classification model can be adapted to classify pattern images. By doing so, the liveness classification model can classify whether an object associated with a pattern image is a living person. The liveness classification model can include an encoder and / or one or more classification layers. The liveness classification model can receive a pattern image, preferably a partial image generated from the pattern image, at the encoder. The encoder can transform the pattern image, preferably a partial image generated from the pattern image, into a dimensionality-reduced representation associated with the pattern image. The dimensionality-reduced representation associated with the pattern image can be a tensor associated with the pattern image, particularly a one-dimensional tensor or a two-dimensional tensor. For this purpose, the encoder can include one or more convolutional layers. One or more classification layers of the liveness classification data-driven model can receive the dimensionality-reduced representation associated with the pattern image and can map the dimensionality-reduced representation associated with the pattern image to an indication of whether an object is likely a living organism. The indication of whether an object is a living organism can be a numerical value. This numerical value can indicate the level of confidence associated with determining whether an object is a living organism.

[0026] Therefore, the method described herein can further include generating a partial image from the pattern image by cropping the pattern image. Providing a pattern image can refer to providing a partial image generated from the pattern image.

[0027] An encoder can be adapted to reduce the dimensionality of an image. Therefore, an encoder can be adapted to generate a dimensionality-reduced representation of an image and / or a template image. In an embodiment, the encoder may include one or more convolutional layers. Therefore, the data-driven model may include a convolutional neural network.

[0028] In an embodiment, training images may refer to images suitable for training a data-driven model. Training images may depict people. Preferably, two or more training images may depict the same person, and one or more training images may depict different people. Thus, multiple training images may be associated with two or more different people. In an embodiment, the person associated with a training image may be unrelated to existing people, and / or may be a synthetically generated representation of a person and / or a synthetic image of a person.

[0029] In embodiments, a representation may refer to a visual representation of an object. The object may be a living organism, such as a person. Preferably, the representation of a person may be a representation of at least a portion of a person, particularly at least a portion of a person's face. The visual representation may be adapted to identify the object and / or indicate the object. The visual representation may include features associated with the object. Features associated with the object may characterize the object. The visual representation of a first object may differ from the visual representation of a second object. Examples of representations may include sketches, textual descriptions of the object, images, at least a portion of the object's outline, or combinations thereof. In embodiments, the person associated with the representation of a person may be unrelated to existing people and / or may be a synthetically generated representation of a person. The representation of a person may be a synthetic representation of a person.

[0030] In this embodiment, the user can be a user of the device. The method can be a computer-implemented method. An authorized user can be a user authorized to access resources.

[0031] In an embodiment, allowing an authorized user to access a resource may include allowing the authorized user to perform at least one operation using a device, computing device, and / or system. The resource may be a device, system, computing device, functionality of a computing device, functionality of a device, functionality of a system, and / or entity. Additionally and / or alternatively, allowing an authorized user to access a resource may include allowing the authorized user to access an entity. An entity may be a physical entity and / or a virtual entity. A virtual entity may be, for example, a database. A physical entity may be an access-restricted area. An access-restricted area may be one of the following: a secure area, a room, an apartment, a vehicle, a portion of the examples mentioned above, etc. The device and / or system may be locked. The device and / or system may be unlocked by an authorized user. The device and / or system may be unlocked by determining that an object associated with an image corresponds to an authorized user and optionally by determining that an object associated with a pattern image may be a living organism.

[0032] The device and / or system may be adapted to compare an image associated with the object initiating the authentication process and a template image associated with an authorized user. The template image associated with the authorized user may be generated during the registration process. A user who has completed the registration process can be referred to as a registered user. The template image associated with the authorized user generated during the registration process may be stored in the memory of the device and / or system. Storing the tensor associated with the template image may refer to storing a dimensionality-reduced representation associated with the template image. The dimensionality-reduced representation associated with the template image may be stored in memory. The device and / or system may be adapted to perform at least one action. The dimensionality-reduced representation may include at least one tensor. The tensor may indicate features associated with the image.

[0033] In an embodiment, providing an image to a data-driven model to determine whether an object associated with the image corresponds to an authorized user may include a dimensionality-reduced representation of the image that generated the object, particularly through the data-driven model. The dimensionality-reduced representation of the image may include fewer data points than the image itself. Preferably, the dimensionality-reduced representation of the image may be a feature vector. The feature vector may include multiple numerical values. The feature vector may indicate one or more features associated with the object. The one or more features may be associated with an image of the object. The one or more features may characterize the object and / or be suitable for distinguishing the object from a second object. The data-driven model may determine whether an object associated with an image corresponds to an authorized user based on the dimensionality-reduced representation of the image and / or a dimensionality-reduced representation of a template image. For this purpose, the data-driven model may receive the dimensionality-reduced representation of the image and / or the dimensionality-reduced representation of the template image at one or more classification layers of the data-driven model, and the dimensionality-reduced representations of the image and the template image may be mapped to an indication of whether the object corresponds to an authorized user. Allowing an object to access resources based on determining that the object corresponds to an authorized user may include receiving an indication that the object corresponds to an authorized user, particularly from one or more classification layers of the data-driven model.

[0034] In this embodiment, receiving a template image from an authorized user may refer to receiving a dimensionality-reduced representation of the template image. Providing the template image to a data-driven model may refer to providing the dimensionality-reduced representation of the template image to the data-driven model. The template image may be generated during the authorized user's registration process.

[0035] The dimensionality reduction representation of the template image may include fewer data points than the template image itself. The dimensionality reduction representation of the template image may include multiple numerical values. The dimensionality reduction representation of the template image may indicate one or more features associated with an authorized user. These one or more features may be associated with the template image of the authorized user. These one or more features may characterize the authorized user and / or may be suitable for distinguishing the authorized user from an unauthorized user (such as a second object).

[0036] Using dimensionality-reduced representations of images and / or template images allows for resource-efficient authentication of authorized users while maintaining reliability, as the size of the input data used for authentication is reduced.

[0037] In embodiments, these methods may further include receiving a template image, and in particular a dimensionality-reduced representation of the template image. The template image may represent an authorized user. The template image, and in particular the dimensionality-reduced representation of the template image, may be further provided to a data-driven model for determining whether an object associated with the image corresponds to an authorized user.

[0038] In this embodiment, the user can be an authorized user.

[0039] In this embodiment, the device may be a smartphone, smartwatch, computer, etc.

[0040] In an embodiment, the method further includes: triggering patterned light to illuminate an object, and generating a patterned image when the object can be illuminated by patterned light; providing the patterned image to a liveness data-driven model to determine whether the object associated with the image corresponds to a living person, wherein the liveness data-driven model can be trained based on historical patterned images and corresponding indications of whether objects associated with these historical patterned images correspond to living people; and further allowing an authorized user to access resources based on the determination that the object corresponds to a living person. The liveness data-driven model can be trained and / or parameterized to receive patterned images and map the patterned images to indications of whether the object associated with the patterned image is a living organism.

[0041] In one embodiment, the patterned coherent infrared light may include fewer than 4,000 beams. In another embodiment, projecting the patterned coherent infrared light onto an object may result in fewer than 4,000 projected light spots. In yet another embodiment, the patterned coherent infrared light may include fewer than 3,000 beams, preferably fewer than 2,000 beams, and most preferably fewer than 1,000 beams. In yet another embodiment, projecting the patterned coherent infrared light onto an object may result in fewer than 3,000 projected light spots, preferably fewer than 2,000 light spots, and most preferably fewer than 1,000 light spots.

[0042] When an object is illuminated by patterned light, a patterned image can be projected onto the object. The patterned light can comprise two or more beams. Projecting patterned light onto an object causes a pattern to be projected onto the object. Determining whether the object is a living person is advantageous because it enhances the security of authentication. For example, liveness detection can determine whether an authorized user is being presented during the authentication process or whether a deceptive object is presenting an authorized user. Therefore, this feature contributes to reliable authentication and enables the detection of deceptive objects. The patterned light can be patterned infrared light, particularly patterned infrared light associated with wavelengths of infrared light as described herein. Using infrared light is advantageous because it may be invisible to humans. Therefore, detection of deceptive masks can be performed without the object, and especially the imposter, being aware of it.

[0043] In this embodiment, the infrared light can be coherent infrared light or coherently patterned infrared light. Projecting coherent light can cause an infrared light pattern to be projected onto an object, wherein the pattern includes two or more light spots, and wherein the light spots include multiple speckles. Therefore, irradiating a user with coherently patterned infrared light can result in the formation of multiple speckles. The formation of speckles may be related to the material associated with the object. The speckles can indicate whether the object is a living person. Therefore, using coherently patterned infrared light improves the reliability of the authentication process.

[0044] In this embodiment, human representation can be received via a user interface. By doing so, the training of the data-driven model can be controlled, thereby enabling a customized and reliable authentication process.

[0045] In this embodiment, the representation of a person may include an image of a person, a sketch of a person, a text description of a person, at least a portion of a person's outline, or a combination thereof. By doing so, training images can be generated using readily available data. Therefore, resources and time spent capturing training images can be saved, while training images can be tailored to the application domain of the data-driven model for authentication.

[0046] In an embodiment, generating multiple synthetic representations of a person based on the person's representation may include generating a 3D model of the person based on the person's representation. Generating the 3D model of the person may include identifying one or more features of the person associated with the person's representation and mapping those features to one or more features of a topological map of the person, particularly the user's face. The one or more features of the person may be key points of the user's face. Additionally or alternatively, generating the 3D model of the person may include providing the person's representation to a 3D data-driven model. The 3D data-driven model may be parameterized and / or trained to receive the person's representation and generate a 3D model based on it. The 3D data-driven model may be parameterized and / or trained to generate depth information associated with the person's representation. The 3D data-driven model may include an encoder configured to reduce the dimensionality of the person's representation and / or transform the person's representation into a machine-processable representation of the person, such as a tensor, particularly a two-dimensional or one-dimensional tensor.

[0047] Multiple synthetic representations can be generated based on a 3D human model by selecting one or more orientations of the 3D human model and / or selecting one or more viewpoints of the 3D human model.

[0048] Furthermore, generating multiple synthetic representations of a person based on their representation can include generating a first representation of the person, altering the orientation of the person's 3D representation (especially after the first representation has already been generated), and generating a second representation of the person. Multiple synthetic representations of a person can include both the first and second representations. Altering the orientation of the person's 3D representation can refer to rotating and / or translating it. By doing so, training images associated with the person at different viewpoints and / or orientations can be obtained. This enables reliable authentication of authorized users.

[0049] In an embodiment, multiple composite representations of a person may differ in terms of the person's perspective, the position of at least a portion of the person in the representation, or combinations thereof.

[0050] Different perspectives on a person can refer to different orientations of a person relative to a viewpoint, which is related to the person's representation, the person's synthetic representation, training images, images, etc.

[0051] In an embodiment, multiple training images may depict one or more people. At least a portion of the multiple training images associated with a person may differ in the following ways: the viewpoint of that person in the representation, the position of at least a portion of that person in the representation, the brightness of the training images, the background within the training images, or a combination thereof. By doing so, the generated training images can be customized to meet the needs of the authentication process. Therefore, these features facilitate effective and customized authentication of authorized users.

[0052] In this embodiment, the background may include scenery and / or one or more objects unrelated to the objects associated with the image. The background can be defined based on the objects. The background may include at least a portion of the environment surrounding the objects.

[0053] In an embodiment, generating multiple training images may include providing multiple synthetic representations of a person to an augmented data-driven model and receiving the multiple training images from the augmented data-driven model. The augmented data-driven model may be trained based on historical representations of a person and corresponding training images. The augmented data-driven model may be trained to generate multiple training images associated with changes in the person's visual perception and / or changes in the person's position within the image and / or relative to the person's surrounding environment (e.g., background). Changes in position may include, for example, altering the viewpoint of the person, the size of the person (particularly relative to the person's surrounding environment (e.g., background), and altering brightness, contrast, hue, saturation values, etc., associated with the person's synthetic representation. Changes in the person's visual perception may include altering facial expressions, altering clothing associated with the person, and altering features of the person (including, for example, hair, nails, lips, etc.). Preferably, the augmented data-driven model may be trained to alter the viewpoint of the person in the person's representation, the position of at least a portion of the person in the person's representation, the brightness of the training images, the person's surrounding environment (e.g., the background in the training images), or combinations thereof. Augmented data-driven models can be trained to generate multiple training images containing a greater number of human-associated features than are included in the human representation. Therefore, augmented data-driven models can be trained to increase the number of features associated with the object. Augmented data-driven models can be trained to add one or more features associated with the person and / or background to generate multiple training images. Augmented data-driven models can be configured to map multiple synthetic representations of a person to multiple training images.

[0054] In embodiments, the method may further include preprocessing the multiple synthetic representations of a person before providing them to an augmented data-driven model. Preprocessing the multiple synthetic representations may include applying at least one image enhancement technique, particularly applying at least one of the multiple synthetic representations of a person. Additionally or alternatively, preprocessing the multiple synthetic representations may include generating a representation of a contour associated with the person. Further, providing the person's representation to the augmented data-driven model may include providing the representation of the contour associated with the person to the augmented data-driven model. The augmented data-driven model may be trained and / or parameterized to receive representations of contours associated with multiple people and provide multiple training images generated from the representations of contours associated with multiple people. In embodiments, the method may further include generating a representation of the person's contour, and wherein providing the person's representation to the augmented data-driven model includes providing the representation of the person's contour to the augmented data-driven model.

[0055] By doing so, training images can be customized to meet the needs of the authentication process, enabling customized and reliable authentication for authorized users.

[0056] In an embodiment, enhancing multiple synthetic representations of a person to generate multiple training images can refer to enhancing multiple synthetic representations of a person by adding one or more features associated with the person to one or more of the multiple synthetic representations of the person used to generate the multiple training images. The number of distinct pixel values ​​associated with a training image can be greater than the number of distinct pixel values ​​associated with the person's representation. Therefore, adding one or more features associated with the person can increase the number of distinct pixel values. Furthermore, enhancing multiple synthetic representations of a person to generate multiple training images can refer to enhancing multiple synthetic representations of a person by adding one or more features associated with the background to generate multiple training images.

[0057] In embodiments, the human-associated feature may be a part of a person, particularly the face. The human-associated feature may be adapted to characterize and / or identify at least a part of a person. One or more of the multiple training images may include one or more human-associated features. Examples of features may include eyes, nose, cheeks, eyebrows, lips, chin, wrinkles, moles, age spots, beauty marks, shadows cast by various parts of the face, one or more hairs, at least a portion of skeletal structure associated with a person (particularly the face), etc. The background-associated feature may be a part of the background scene, preferably an item in the background (e.g., an object associated with the background).

[0058] In embodiments, image enhancement techniques may include at least one of the following, or combinations thereof: scaling, cropping, rotating, blurring, distorting, shearing, resizing, folding, changing contrast, changing brightness, adding noise, multiplying by at least a portion of pixel values, filtering, adjusting color, applying convolution, imprinting, sharpening, flipping, averaging pixel values.

[0059] In an embodiment, the augmented data-driven model can be trained to add noise to a human representation (particularly through one or more encoder blocks) to obtain a noisy human representation, and to remove noise from the noisy human representation (particularly through one or more decoder blocks) to obtain multiple training images. The augmented data-driven model may include multiple encoder blocks and / or decoder blocks. Encoder blocks may be adapted to reduce the dimensionality of the human representation. Decoder blocks may be adapted to increase the dimensionality of the human representation. The method may further include passing the human representation through the encoder blocks of the augmented data-driven model, and / or passing the human representation, particularly the noisy human representation, through the decoder blocks of the augmented data-driven model. Passing the human representation through encoder blocks may be referred to as backdiffusion. Passing the human representation through decoder blocks may be referred to as forwarddiffusion.

[0060] In this embodiment, the synthetic representation of a person can be a synthetically generated representation of a person. Therefore, the synthetic representation of a person can be a representation generated by digitally processing a person's representation. The synthetic representation of a person can be generated independently of an image generation unit (such as a camera). The synthetic representation can include image data suitable for visually representing a person.

[0061] In an embodiment, training a data-driven model can refer to retraining the data-driven model. A request can be a request to retrain the data-driven model. The data-driven model can be trained based on multiple historical training images. Receiving a request to retrain the data-driven model is triggered and / or initiated by a registration process associated with a user. The registration process can be a process for generating template images for authorized users. Additionally or alternatively, receiving a request to retrain the data-driven model is triggered and / or initiated by providing the data-driven model with a test image when associated with the authorized user to determine if the object associated with the test image is a deceptive object. By doing so, the trained model can be improved. This saves time and resources compared to training the model from scratch. Furthermore, the trained model can be customized to meet the needs of specific use cases, and the reliability of the data-driven model is improved. Attached Figure Description

[0062] The disclosure will be further described below with reference to the accompanying drawings. In the drawings and the disclosure, the same reference numerals are intended to refer to the same or similar elements, components and / or parts.

[0063] Figure 1A An embodiment of device 102 for authenticating users is shown.

[0064] Figure 1B An example of a system for authenticating users is shown.

[0065] Figure 2 An example of a method for authenticating users is shown.

[0066] Figure 3 An example of obtaining multiple training images is shown.

[0067] Figure 4 An example of generating training images by enhancing human representations is shown.

[0068] Figure 5 An example of generating a 3D model 514 of a person based on a representation 516 is shown.

[0069] Figure 6 An example of obtaining multiple training images is shown. Detailed Implementation

[0070] The following embodiments are merely examples for implementing the methods, systems, or application devices disclosed herein and should not be considered limiting.

[0071] Figure 1A An embodiment of device 102 for authenticating users is shown.

[0072] Device 102 may include an illumination source 110, a camera 112 including a sensor 114, a processor 104, and / or a memory 116. The illumination source 110 may emit light, preferably infrared light, toward the object 106. Infrared light may be unidentifiable to the object 106. The sensor 114 of the camera 112 may be sensitive to the light emitted by the illumination source 110. Therefore, the sensor 114 may be adapted to generate an image of the object 106 when a user may be illuminated by light emitted from the illumination source 110. The processor 104 may receive the image of the object 106. The processor 104 may process the image of the object 106. By doing so, the processor 104 may determine whether the object 106 corresponds to an authorized user. This may include determining whether the image of the object corresponds to a representation of visual features associated with an authorized user. The representation of visual features associated with an authorized user may be a template image and / or a dimensionality-reduced representation of a template image. For this purpose, the processor may execute instructions stored in the memory 116. Figure 2The document describes an example of authenticating users.

[0073] Alternatively, the illumination source 110 and / or camera 112 may be included in a second and / or third device. In this example, the processor 104 may be communicatively coupled to the second and / or third device to trigger the illumination source 110 to emit light and trigger the sensor 114 to generate an image of the object 106 when it is illuminated.

[0074] The processor can receive requests to access resources, such as unlocking a device. For this purpose, the device may include a user interface. Object 106 can request access to resources through the user interface. For example, the device could be a telephone. Object 106 may wish to control the telephone. For this purpose, object 106 may need to authenticate. Object 106 can request access to resources and / or authenticate by using the device's touchscreen display. This can trigger a signal to the device's processor 104. Based on the received signal, processor 104 triggers illumination source 110 to emit light.

[0075] Figure 1B An example of a system for authenticating users is shown.

[0076] The system may include a first device 120 and a second device 118. In an embodiment, the first device 120 may include a processor 104 and a memory 116. The first device 120 may be communicatively connected to the second device 118, which includes an illumination source 110 and a camera 112 (including a sensor 114). For example, the first device 120 may be connected to the second device 118 by means of a cloud service. In particular, the processor 104 and / or the memory 116 may be part of a cloud service. The second device 118 may be configured to provide an image generated by the sensor 114 to the first device 120. Receiving a request to access a resource may refer to receiving a signal that triggers the illumination source 110 to emit light at the illumination source 110. In response to receiving this signal, the illumination source 110 may be triggered to illuminate the object 106 with light. Further, the camera 112 may receive a signal that triggers the generation of an image of the object 106. Receiving a request to access a resource may further include receiving a signal at the camera 112.

[0077] Processor 104 can determine whether an object associated with an image corresponds to an authorized user by providing the image to a data-driven model and / or running the data-driven model as input data to determine whether an object associated with the image corresponds to an instruction from an authorized user. Based on whether an object associated with an image corresponds to an instruction from an authorized user, the processor can provide a signal to grant access to the authorized user. Providing a signal to grant access to the authorized user can be referred to as allowing the authorized user to access the resource.

[0078] Figure 2 An example of a method for authenticating users is shown.

[0079] An image 204 of the object can be provided. Providing the image of the object may include receiving a request to access a resource, triggering the illumination of the object with light in response to receiving the request to access the resource, and triggering the generation of an image of the object when the object is illuminated by light. The image may be generated by sensor 114 when the object can be illuminated by light emitted from illumination source 110, such as... Figure 1A and Figure 1B As described in the context, images can be fed to a data-driven model to determine whether an object associated with the image corresponds to an authorized user. The data-driven model can be trained based on multiple training images. Training images can be as follows: Figure 3 and Figure 4 The data is generated as described in the context. A data-driven model can be trained to determine whether an object associated with an image corresponds to an authorized user. An authorized user can be authorized to access resources. Therefore, a data-driven model can be trained to determine whether an object associated with an image can be authorized to access resources.

[0080] For this purpose, a data-driven model may, for example, receive an image at its input layer. The data-driven model may include an encoder. The encoder may be adapted to reduce the dimensionality of the image. Thus, the encoder may be adapted to generate a dimensionality-reduced representation of the image. In an example, the dimensionality-reduced representation may be a feature vector. The data-driven model may be trained to determine whether an object associated with an image corresponds to an authorized user by determining a similarity score associated with the image (specifically, its dimensionality-reduced representation) and a template image (specifically, its dimensionality-reduced representation). Therefore, the data-driven model may receive a template image, specifically its dimensionality-reduced representation. The similarity score may be the distance between the dimensionality-reduced representation of the image and the dimensionality-reduced representation of the template image, specifically the distance in the feature space. In an example, the data-driven model may determine the distance between a feature vector and a template vector. If the distance between the feature vector and the template vector can be within a predefined range, the data-driven model can provide an indication that the object may be an authorized user. If the distance between the feature vector and the template vector can be outside a predefined range, the data-driven model can provide an indication that the object may be an unauthorized user. Providing an indication that the object may be an authorized user may result in allowing authorized users access to resources. Therefore, allowing an authorized user to access a resource based on determining that an object corresponds to an authorized user may include receiving an instruction from a data-driven model that the object may be an authorized user and allowing the authorized user to access the resource.

[0081] Figure 3 An example of obtaining multiple training images 312, 314, and 316 is shown.

[0082] To obtain multiple training images 312, 314, and 316, a human representation 304 can be provided and / or received. The human representation 304 can be, for example, a sketch and / or image of a person generated by a camera. Based on the human representation 304, multiple synthetic representations of the person 306, 308, and 310 can be generated. These multiple synthetic representations 306, 308, and 310 can be associated with different viewpoints of the person, different positions of the person in the representation, different backgrounds, etc. Furthermore, the multiple training images 312, 314, and 316 can be associated with different viewpoints of the person, different positions of the person in the representation, different backgrounds, etc.

[0083] Generating multiple synthetic representations of a person 306, 308, 310 may include generating a 3D representation of a person based on the person's representation. Unless otherwise specified, the person's representation may be 2D. The 3D representation of a person may indicate the topological structure associated with the person. Generating a 3D representation of a person 302 may include adding depth information to the person's representation. The 3D representation of a person 302 can be generated by projecting the person's representation onto a 3D model of the person's topology. The 3D model of the person's topology may indicate depth information associated with the person based on human anatomy. The 3D model of the person's topology may specify one or more features associated with multiple people. Generating a 3D representation of a person 302 may include projecting the person's representation onto the 3D model of the person's topology. For this purpose, features associated with the person's representation 304 may be detected. The person's representation 304 can be projected onto the 3D model of the person's topology by matching the features of the person associated with the person's representation with the features of the 3D model of the person's topology.

[0084] Additionally or alternatively, generating a 3D representation of a person 302 may include generating depth information based on the person's representation 304. Generating depth information based on the person's representation 304 may include determining the distances between two or more features of the person associated with the person's representation 304, and determining depth information based on those distances. For example, from a viewpoint associated with the person's representation 304, the distance between the eyes and nose, and the distance between the eyes, may be correlated with the distance between the nose and eyes. Therefore, depth information can be obtained based on the correlation between the person's representation 304 and depth information. This correlation may be obtained based on the person's historical representations 304 and corresponding depth information. For example, a 3D model of the person's topology may be obtained based on the distances between two or more features of the person.

[0085] Based on the 3D representation of a person 302, multiple composite representations of a person 306, 308, and 310 can be generated. As mentioned above, the multiple composite representations of a person 306, 308, and 310 may differ in the orientation of the person in the representation 304. Therefore, the multiple composite representations of a person 306, 308, and 310 can be generated by rotating and / or mirroring the 3D representation of a person 302.

[0086] Based on multiple synthetic representations of humans, 306, 308, and 310, it is possible to... Figure 4 Multiple training images are generated as described in the context.

[0087] Figure 4 An example of generating training images by enhancing human representation 408 is shown.

[0088] The human representation 408 can be one of several composite representations of the human, such as 306, 308, and 310. It can be... Figure 3 The representation 408 of the person being obtained is described in the context of the given information (as one of several composite representations 306, 308, and 310).

[0089] Representation 408 can be preprocessed 422. Preprocessing of the human representation 422 may include generating a representation of the human's contours associated with the human representation. The human representation may include information about colors associated with a portion of the human. The representation of the human contours may include the contour lines of at least a portion of the human. The representation of the human contours may indicate and / or may be adapted to define the contours of the human associated with training images, particularly training images generated based on the human representation.

[0090] Human representations can be provided to augmented data-driven models. Augmented data-driven models can be adapted to augment input data to generate output data. In this context, an augmented data-driven model can generate training images by augmenting human representations. Therefore, human representations can be provided to an augmented data-driven model. Human representations can be processed by the augmented data-driven model, specifically by passing human representations through one or more layers of the model. Human representations can be provided to an augmented data-driven model to modify at least a portion of the human representation. Modifying at least a portion of the human representation can yield training images.

[0091] In an embodiment, providing a human representation to an augmented data-driven model may include preprocessing the human representation to generate a representation of the human contour, and providing the representation of the human contour to the augmented data-driven model. The augmented data-driven model can process the human contour representation into a training image, particularly by augmenting the human contour representation. For this purpose, algorithms such as the Canny edge detection algorithm can be applied.

[0092] In an embodiment, providing a human representation to an augmented data-driven model to generate training images may include passing the human representation through multiple encoder blocks, including a first encoder block 412 of dimension x and an nth encoder block 414 of dimension xn. Passing the human representation through the multiple encoder blocks 412, 414 may add noise to the human representation and / or result in a noisy representation of the human. This may be referred to as backdiffusion. Preferably, the noise may be Gaussian noise. The dimension x associated with the first encoder block 412 may be greater than the dimension xn of the nth encoder block 414. Therefore, the multiple encoder blocks may be adapted to reduce the dimensionality of the human representation and / or add noise to the human representation.

[0093] Additionally or alternatively, providing human representations to an augmented data-driven model to generate training images may include passing the human representations through multiple decoder blocks, including a first decoder block 416 of dimension xn and an m-th decoder block 418 of dimension x. The dimension of the m-th decoder block 416 may correspond to the dimension of the first encoder block 412. Further, the dimension of the n-th encoder block 414 may correspond to the dimension of the first decoder block 416. Passing the human representations through the multiple decoder blocks 416, 418 can remove noise from the human representations (particularly the noisy representations of the human received from the encoder blocks, preferably from the last encoder block of the augmented data-driven model), and / or can yield training images. This may be referred to as forward diffusion. Preferably, the noise may be Gaussian noise. The dimension x associated with the m-th decoder block 418 may be greater than the dimension xn of the first block 416. Therefore, multiple decoder blocks may be adapted to increase the dimension of the human representations (particularly the noisy representations of the human), and / or remove noise from the human representations (particularly the noisy representations of the human).

[0094] By doing so, training images can include more human-associated features than a human representation. Training images can correspond to human-generated images. Therefore, training images may be more realistic than human representations. Human-associated features can be various parts of a person, particularly the human face, such as eyes, nose, cheeks, eyebrows, lips, chin, wrinkles, and shadows cast on the face.

[0095] Additionally or alternatively, enhancing a human representation may include applying at least one image enhancement technique to the human representation. This image enhancement technique may be adapted to increase the number of features associated with the human. The image enhancement technique may add one or more features associated with the human to the human representation, and / or may sharpen the human representation, particularly for generating training images.

[0096] Examples of available augmented data-driven models can include stable diffusion models, preferably ControlNet. ControlNet can be a machine learning architecture derived from a stable diffusion model.

[0097] Figure 5 An example of generating a 3D model 514 of a person based on a representation 516 is shown.

[0098] The 3D data-driven model may include an encoder 502 and a decoder 512. The 3D data-driven model may be a generative model. The 3D data-driven model may be trained, for example, in an adversarial manner. Therefore, the generative model may be and / or may include at least a portion of a generative adversarial network. This may include training the 3D data-driven model to generate a 3D model of a person, while simultaneously training a detector model to distinguish the 3D model of a person generated by the 3D data-driven model from a real or realistic 3D model of a person.

[0099] The human representation 516 can be received at the encoder 502 of the 3D data-driven model in image format, text format, etc. To process the human representation 516 in image format, the encoder 502 is configured to receive image data, particularly 2D image data, and reduce the dimension of the image data to a tensor associated with the human representation 510. To process the human representation 516 in text format, the encoder 502 is configured to receive text data and transform the dimension of the text data to a tensor associated with the human representation 510. This allows the 3D data-driven model to efficiently process the machine-processable format associated with the human representation 516. In other words, the encoder 502 can transform the human representation 516 into a latent space. For this purpose, the encoder 502 may include one or more convolutional layers configured to reduce the dimension of the human representation 516. The tensor associated with the human representation 510 can be received by the decoder 512 of the 3D data-driven model. The decoder 512 may include one or more deconvolutional layers. Deconvolutional layers can be configured to increase the dimension of the tensor associated with the human representation 510, preferably to obtain a 3D human model 514. The 3D human model 514 may include multiple voxels.

[0100] Figure 6 An example of obtaining multiple training images is shown.

[0101] The 3D model 514 of the human can specify depth information associated with the human. Multiple composite representations 624 can specify 2D views of the human from predefined viewpoints. Therefore, multiple composite representations 624 can include views of the 3D model 602 of the human from one or more predefined viewpoints, particularly multiple different viewpoints.

[0102] This disclosure has also been described in conjunction with various preferred embodiments and examples. However, by studying the accompanying drawings, this disclosure, and the claims, those skilled in the art, as well as those practicing the claimed subject matter, will understand and implement other variations. It is particularly noteworthy that any steps presented can be performed in any order; that is, this disclosure is not limited to a specific order of these steps. Furthermore, it is not required that different steps be performed at a specific location or node in a distributed system; that is, each step can be performed on different nodes using different devices / data processing.

[0103] As used herein, "determine" also includes "initiating or causing determination," "generate" also includes "initiating and / or causing generation," and "provide" also includes "initiating or causing determination, generation, selection, sending, and / or receiving." "Initiating or causing an action" includes any processing signal that triggers a computing node or device to perform a corresponding action.

[0104] In the claims and specification, the word "comprising" or "including" or similar wording does not exclude other elements or steps and should not be construed as limiting oneself to the listed elements or steps. The indefinite article "a" or "an" does not exclude multiple. A single element or other unit may perform the function of several entities or items recited in the claims. The fact that certain measures are recited only in mutually different dependent claims does not indicate that a combination of these measures cannot be used in advantageous implementations or that additional elements may be included.

[0105] Within the scope of this disclosure, provision may include any interface configured to provide data. This may include application programming interfaces, human-machine interfaces (such as displays), and / or software module interfaces. Provision may include transmitting or submitting data to the interface, particularly displaying data to a user or having data used by a receiving entity.

[0106] Any disclosures and embodiments described herein relate to the methods, systems, devices, and computer program elements listed above, and vice versa. Advantageously, the benefits provided by any embodiment and example also apply to all other embodiments and examples, and vice versa.

Claims

1. A method for authenticating an authorized user, the method comprising: Receive requests to access resources. In response to receiving the request to access the resource, trigger the illumination of the object with light, and Trigger the generation of an image of the object when it is illuminated by light. Receive the template image from the authorized user. The image and the template image are provided to a data-driven model to determine whether the object associated with the image corresponds to the authorized user. This data-driven model is trained using multiple training images obtained by generating multiple synthetic representations of the person from the person's representation, and augmenting these synthetic representations to generate multiple training images. In response to determining that the object corresponds to the authorized user, the authorized user is allowed to access the resource.

2. The method as described in claim 1, wherein, Receiving the template image from the authorized user means receiving a dimensionality-reduced representation of the template image, and providing the template image to the data-driven model means providing the dimensionality-reduced representation of the template image to the data-driven model.

3. The method of claim 1 or 2, further comprising triggering patterned light illumination of the object, and generating a patterned image while the object is illuminated by patterned light; and providing the patterned image to a liveness data-driven model to determine whether an object associated with the image corresponds to a living human, wherein, The liveness data-driven model is trained using historical pattern images and indications of whether objects associated with these historical pattern images correspond to live people; and further, in response to determining that an object corresponds to a live person, the authorized user is allowed access to the resource.

4. A method for obtaining a plurality of training images for training a data-driven model for determining whether an object associated with an image corresponds to an authorized user, the method comprising: The provider's statement, Generate multiple composite representations of the person based on the person's representation. Multiple training images are generated by enhancing these multiple synthetic representations. Receive the multiple training images.

5. A method for training a data-driven model to determine whether an object associated with an image corresponds to an authorized user, the method comprising: Receive a request to train the data-driven model. Multiple training images are provided, which are obtained by generating multiple synthetic representations of the person from the person's representation, and augmenting the multiple synthetic representations of the person to generate multiple training images. The data-driven model is trained using these multiple training images. This data-driven model can be provided optionally.

6. The method of claim 5, wherein, The representation of a person includes an image of the person, a sketch of the person, a textual description of the person, at least a portion of the person's outline, or a combination thereof.

7. The method of claim 5 or 6, wherein, Training the data-driven model means retraining the data-driven model, and wherein the request is a request to retrain the data-driven model, and wherein the data-driven model is trained using multiple historical training images, and wherein training the data-driven model means retraining the data-driven model, and wherein receiving the request for retraining the data-driven model is triggered and / or initiated by a registration process associated with the user, and / or wherein receiving the request for retraining the data-driven model is triggered and / or initiated by providing the test image to the data-driven model when the test image is associated with the authorized user, in order to determine that the object associated with the test image is a deceptive object.

8. The method according to any one of claims 1 to 7, wherein, Generate multiple composite representations of a person based on that person's representation, including generating a 3D representation of that person based on that person's representation.

9. The method according to any one of claims 1 to 8, wherein, The plurality of training images depict one or more people, and wherein at least a portion of the plurality of training images associated with a person differs in the following ways: the viewpoint of the person in the representation of the person, the position of at least a portion of the person in the representation of the person, the brightness of the training images, the background within the training images, or a combination thereof.

10. The method according to any one of claims 1 to 9, wherein, Generating the plurality of training images includes providing a plurality of synthetic representations of the person to an augmented data-driven model, and receiving the plurality of training images from the augmented data-driven model, wherein the augmented data-driven model is trained using the person’s historical representations and corresponding training images.

11. The method of any one of claims 1 to 10, further comprising generating a representation of the person's contour, wherein, Providing the person's representation to the augmented data-driven model includes providing the person's silhouette representation to the augmented data-driven model.

12. The method according to any one of claims 1 to 5 or 7 to 11, wherein, The person's representation is the registration image generated during the registration process of that person, and in particular the authorized user.

13. Use of training images obtained by any one of claims 4 or 6 to 11 for training a data-driven model for certified authorized users.

14. A device and / or system for authenticating authorized users, the device and / or system comprising: processor; as well as A memory storing instructions that, when executed by the processor, configure the device and / or system to perform the method as described in any one of claims 1 to 11.

15. An authentication system trained using training images obtained by the method of any one of claims 4 or 7 to 12 and / or trained by the method of any one of claims 6 to 12.

Images