Multi-biometric authentication or identification method
By using generative models on terminals and servers to generate synthetic data from different biometric traits for identity authentication, the privacy and security issues caused by centralized database storage are solved, achieving efficient and secure biometric authentication, suitable for authentication systems with a large number of users.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- WORLDLINE SA(FR)
- Filing Date
- 2024-09-16
- Publication Date
- 2026-06-16
Smart Images
Figure CN122228495A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of biometric identification, particularly voice and facial recognition. More specifically, it relates to a method for authenticating or identifying an individual based on first candidate biometric data and second candidate biometric data obtained from two different biometric traits of the individual. Background Technology
[0002] Biometric authentication / identification is the automatic identification of people using specific traits, which are automatically measurable, robust, and specific physical (biological) or personal behavioral characteristics that can be used to verify (in the case of authentication) or determine (in the case of identification) an individual's identity. Biometric technology improves security and ease of use.
[0003] Several biometric modalities have already been used, such as fingerprints, face, iris, and voice.
[0004] The problem is that managing biometric databases is complex and expensive. In practice, traditional technologies require storing so-called reference biometric data (of authorized individuals, called registered individuals) in a database so that it can be compared with entirely new biometric data (called candidate data) obtained on-site from individuals seeking to identify / authenticate themselves. All of this biometric data remains personal data, and its processing involves confidentiality and privacy considerations and is subject to strict regulation, particularly under the GDPR.
[0005] A technique called "zero biometric database" has therefore been proposed, in which reference biometric data (specifically, an individual's facial data) is encoded and stored only on the individual's terminal (smartphone) for representation via a dynamic QR code. To authenticate themselves, the individual displays their QR code on their terminal and presents it to the terminal device, which decodes the QR code to reconstruct a reference biometric template that can be compared with candidate biometric data (facial images) acquired by the terminal device.
[0006] This approach avoids any need for centralized biometric databases, but introduces other problems:
[0007] Mobile devices are essential (but they can be lost, run out of battery, have damaged screens, etc.), so this is not a “pure” biometric method.
[0008] In practice, QR codes store a limited amount of information compared to a complete facial image, thus offering a lower level of security.
[0009] This invention improves upon this situation. Summary of the Invention
[0010] The present invention relates in a first aspect to a method for authenticating or identifying an individual, characterized in that the method is implemented via a data processing device of a terminal and / or a first server, comprising the following steps:
[0011] (a) Obtain first candidate biometric data from an individual’s first biometric trait and second candidate biometric data from a second biometric trait of the individual that is different from the first biometric trait;
[0012] (b) By applying the generative model to the first candidate biometric data and at least one reference unique identifier, synthetic biometric data corresponding to the second biometric trait is constructed; the reference unique identifier is not biometric data and is stored in the data storage device of the terminal and / or the first server.
[0013] (c) Authenticate or identify the individual based on a comparison between the second candidate biometric data and the synthetic biometric data.
[0014] Based on favorable and non-restrictive features:
[0015] There are multiple reference unique identifiers, each associated with a reference individual; in step (b), synthetic biometric data is generated for each reference unique identifier; and in step (c), second candidate biometric data is compared with each synthetic biometric data.
[0016] The reference unique identifier is the encoding of the second reference biometric data, which also corresponds to the second biometric trait associated with the reference unique identifier; the second reference biometric data is not stored by the data storage device of the terminal and / or the first server.
[0017] The method includes pre-registration of data for authentication or identification of reference individuals, the registration being performed by a data processing device of a first server, and includes the following steps:
[0018] (A) Obtain second reference biometric data derived from the second biometric traits of a reference individual;
[0019] (B) Generate a reference unique identifier by applying a coding model to second reference biometric data;
[0020] (C) The reference unique identifier is stored in the data storage device (22a) of the first server (2a), but the second reference biometric data is not stored.
[0021] The generative model is a learned model, and the method includes a preliminary step (a0): using the data processing device of a second server to learn at least the parameters of the generative model from the databases of first and second learned biometric data.
[0022] Step (a0) for each of a plurality of pairs of first and second learned biometric data from the database includes:
[0023] By applying the encoding model to the second learned biometric data, a learned unique identifier (associated with the pair) is generated.
[0024] An attempt was made to reconstruct the second learned biometric data by applying the generative model to the first learned biometric data and the learned unique identifier.
[0025] Step (a0) for each of the multiple combinations of first and second pairs of first and second learned biometric data in the database includes:
[0026] generate:
[0027] The first learned unique identifier is obtained by applying an encoding model to the second learned biometric data of the first pair, and
[0028] The second learned unique identifier is obtained by applying the encoding model to the second pair of second learned biometric data.
[0029] Attempt to build:
[0030] The first synthetic data is obtained by applying a generative model to a first pair of first-learned biometric data and a second-learned unique identifier, and
[0031] The second synthetic data is obtained by applying a generative model to the first learned biometric data and the first learned unique identifier in a second pair.
[0032] So that neither the first synthetic biomarker nor the second synthetic data is consistent with the second biomarker data of the first pair and / or the second biomarker data of the second pair.
[0033] The generative model is a CNN-type neural network, specifically an autoencoder.
[0034] Step (a) includes acquiring first candidate biometric data from an individual's first biometric trait using a first biometric acquisition device of the terminal, and acquiring second candidate biometric data from an individual's second biometric trait using a second biometric acquisition device of the terminal, and transmitting the first candidate biometric data and the second candidate biometric data to a first server; steps (b) and (c) are implemented by the data processing device of the first server.
[0035] The first biometric trait is the voice, and the second biometric trait is the face.
[0036] According to a second aspect, the invention relates to an apparatus for authenticating or identifying an individual, comprising a data processing device and a data storage device storing at least one reference unique identifier, characterized in that the data processing device is configured to:
[0037] Obtain first candidate biometric data from an individual's first biometric trait and second candidate biometric data from an individual's second biometric trait that is different from the first biometric trait;
[0038] By applying the generative model to the first candidate biometric data and the at least one reference unique identifier, synthetic biometric data corresponding to the second biometric trait is constructed.
[0039] The individual is authenticated or identified based on a comparison of second candidate biometric data and synthetic biometric data. This is based on both advantageous and non-limiting characteristics:
[0040] The device is also configured for data registration to authenticate or identify reference individuals.
[0041] The data processing device is also configured to:
[0042] Obtain second reference biometric data derived from the second biometric trait of the reference individual;
[0043] A unique reference identifier is generated by applying the coding model to the second reference biometric data;
[0044] The reference unique identifier is stored in the data storage device, but the second reference biometric data is not stored.
[0045] According to the third aspect, the invention proposes a set of a terminal, a first server according to the second aspect, and an optional second server for learning the generative model.
[0046] According to the fourth and fifth aspects, the invention relates to a computer program product including code instructions for performing a method for authenticating or identifying an individual according to the first aspect; and a computing device readable storage medium having the computer program product recorded thereon, the computer program product including code instructions for performing a method for authenticating or identifying an individual according to the first aspect. Attached Figure Description
[0047] Other features and advantages of the invention will become apparent from the following description of preferred embodiments. This description will be given with reference to the accompanying drawings, in which:
[0048] [ Figure 1 ] Figure 1 This is a schematic diagram of a system implementing the method according to the present invention.
[0049] [ Figure 2 ] Figure 2 This is a flowchart illustrating the steps of an embodiment of the method according to the present invention.
[0050] [ Figure 3a ] Figure 3a The arrangement of the neural network used in operation in a preferred embodiment of the method according to the invention is shown;
[0051] [ Figure 3b ] Figure 3b The arrangement of the neural network used during learning is shown in a preferred embodiment of the method according to the invention. Detailed Implementation
[0052] Architecture
[0053] This invention relates to a method for authenticating and identifying individuals, that is, for use in situations such as... Figure 1 The system shown determines or verifies the identity of an individual presenting themselves in front of terminal 1, for example, to conduct transactions by said individual, but is also used for access control, guaranteed geolocation, etc.
[0054] The individual whose identity is to be verified is considered a "candidate" in contrast to a "reference" individual whose identity is known.
[0055] Terminal 1 includes a data processing device 11, such as a processor, and optionally a data storage device 12 (memory) and an interface device 13 (screen). This is typically a mobile terminal such as a smartphone, but alternatively, terminal 1 may be a tablet computer, a personal computer, or even a fixed device such as an access control terminal, or any other device held and controlled by an entity with which authentication / identification must be performed.
[0056] Furthermore, as will be seen, terminal 1 advantageously includes first and second biometric acquisition devices 14, 15, namely, devices for acquiring first biometric data from a first biometric trait of an individual and acquiring second biometric data from a second biometric trait of an individual that is different from the first biometric trait.
[0057] Biometric traits can be, for example, the shape of an individual's face, their voice, one or more fingerprints, the shape of their ears, the shape of their hands (or any other part of their body), or one or more irises. This can also refer to the structure of one or more fingers or the network of veins in one or both hands. The term "modality" is also used to specify various biometric traits.
[0058] According to the preferred embodiment described below, the first biometric trait is the individual's voice (and the first biometric data is then the individual's speech signal), and the second biometric trait is the shape of the individual's face (and the second biometric data is then a photograph of that face). However, interchangeability or any other combination of two different biometric traits may also be implemented.
[0059] In the above embodiments, the first and second biometric acquisition devices 14 and 15 are specifically a microphone 14 and a camera 15, but in the case of other biometrics, they can be a fingerprint scanner, an eye tracker, etc. It should be noted that these devices 14 and 15 are typically integrated into the terminal 1, but they can also be peripheral devices connected to the rest of the terminal 1, such as a webcam or headset connected (wired or wirelessly) to a smartphone-type terminal 1. The whole assembly constitutes the terminal 1.
[0060] In all cases, the biometric data acquisition devices 14, 15 advantageously include data processing capabilities (which may be implemented by the processing device 11) adapted to extract biometric data from the "raw" data representing the biometric traits, based on the nature of the biometric traits and possible processing of the image or signal representing the biometric traits. Various sound and image processing techniques for extracting biometric data are known to those skilled in the art. As a non-limiting example, the extraction of biometric data may include segmenting an audio signal and separating segments corresponding to speech. Alternatively, it may involve extracting specific points (fingerprint details) or, in the case of an image of an individual's face, the shape / appearance of the face, etc. Alternatively or additionally, convolutional neural networks may be used. It is also possible that no such processing is performed, and the biometric data directly corresponds to the raw data (e.g., an iris image).
[0061] This aspect will not be described in detail here, and it will be assumed that for an individual to be identified / authenticated (i.e., a candidate individual), for any desired pair of biometric traits, it is possible to obtain first candidate biometric data from the individual's first biometric trait and second candidate biometric data from the individual's second biometric trait, which is different from the first biometric trait. In other words, each individual is associated with a pair including first and second biometric data corresponding to different biometric traits.
[0062] It should be noted that, similar to the first and second candidate biometric data, first and second reference data can also be introduced; that is, “expected” biometric data corresponding to the same first and second biometric traits but associated with a known and authorized reference individual. If the candidate individual is indeed the reference individual or one of the reference individuals (in the case of authentication and identification, respectively), then the first candidate biometric data must be consistent with the first reference biometric data associated with the reference individual, and the second candidate biometric data must be consistent with the second reference biometric data associated with the reference individual.
[0063] However, such reference data is unnecessary for implementing this method, and in particular, it does not need to be stored by any of the involved devices. As we will see, it may only be necessary to briefly access an individual's second reference biometric data to perform their registration, but this data can be processed on the fly and is never stored in practice. The first reference data does not even need to be acquired.
[0064] Therefore, with no personal data stored, there is no risk of data breach (and thus no potential privacy issues), and regulatory constraints (such as those under the GDPR) no longer apply.
[0065] This method is implemented by terminal 1 and / or a first server 2a, which may be the same as terminal 1 or remote and connected via network 10 such as the Internet. Advantageously, a second server 2b (which, as we will see, is a learning device) is provided, which is typically remote (i.e., within network 10); however, it may also be combined with the first server 2a. The first server 2a may, for example, be an authentication server for a banking entity.
[0066] Each server 2a, 2b also includes data processing units 21a, 21b (typically processors) and data storage units 22a, 22b (memory, such as hard disk drives). As will be seen, the data processing unit 21b of the second server 2b (but may also be the data processing unit of the first server 2a) may store at least one learning database, which is actually a pair of first and second learning biometric data belonging to the same individual. This data is preferably public data (rather than reference biometric data) in order to continue to avoid any privacy-related issues.
[0067] method
[0068] This method relates to a biometric authentication or identification method based on two biometric traits, which can verify the identity of an individual without storing reference biometric data for either of the two traits.
[0069] refer to Figure 2The method begins with step (a) of obtaining first candidate biometric data corresponding to a first biometric trait of an individual and second candidate biometric data corresponding to a second biometric trait of the individual that is different from the first biometric trait. More specifically, step (a) includes directly acquiring the two candidate biometric data by acquisition devices 14 and 15 of terminal 1, or receiving the data from terminal 1 by server 2a (where applicable, the data is encrypted).
[0070] When the primary biometric trait is voice, as explained, the first candidate biometric data is typically the individual's speech audio signal. The term "the individual's speech audio signal" refers to the audio recording of that individual as a "speaker," that is, the speech spoken when reading a sentence (pre-planned or unpre-planned). It should be noted that this document does not limit specific speaker identification technologies; therefore, the speech can be, for example:
[0071] -A specific pre-defined sentence;
[0072] - The expected sentence, that is, for example, the sentence to be read aloud is displayed in challenge / response mode on terminal 1;
[0073] - Any sentence;
[0074] -etc.
[0075] When the second biometric trait is the shape of the face, the second candidate biometric data is an image of the face or data derived from it (e.g., the location of facial feature points).
[0076] In step (b), data processing device 11 or 21a constructs synthetic biometric data that also corresponds to a second biometric trait of the individual by applying a generative model to the first candidate biometric data and at least one reference unique identifier. The unique identifier or code is not biometric data and is typically an alphanumeric value. It could be, for example, a QR code, as in the prior art. As will be seen, this specifically refers to the encoding of the second biometric data that also corresponds to the second reference biometric trait.
[0077] It should be noted that in the case of multiple reference unique identifiers, step (b) may include constructing several synthetic data, i.e., constructing one for each reference unique identifier (in the case of identification).
[0078] In other words, each reference individual is associated with a reference unique identifier (which thus represents the expected identity), which in practice represents the second reference biometric data but does not disclose any information about the reference individual.
[0079] As explained, preferably only the reference unique identifier is stored in the data storage devices 12, 22a of terminal 1 and / or the first server 2a (and / or possibly the second server 2b), rather than the reference biometric data.
[0080] This method cleverly leverages work demonstrating that different biometric traits of an individual are "correlated" and can predict second biometric data corresponding to a second biometric trait of the same individual from first biometric data corresponding to a first biometric trait of the individual. In other words, synthetic biometric data (because it is artificial) corresponding to the second biometric trait (e.g., a facial image if the second biometric trait is the shape of the face) can be generated. This synthetic biometric data result is consistent with the second biometric data, i.e., sufficiently similar, and, for example, represents the same face in a recognizable manner.
[0081] For example, the paper “Speech2Face: Learning the Face Behind a Voice” (Tae Hyun Oh, Tali Dekel, Changil Kim, Inbar Mosseri, William T. Freeman, Michael Rubinstein, and Wojciech Matusik at the 2019 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)) describes an autoencoder model that includes a first encoder block and a second decoder block.
[0082] Each block is a CNN-type neural network (convolutional neural network). The encoder block (e.g., a feature extraction block of a CNN for speech recognition) generates code representing first biometric data (voice), and the decoder block (e.g., a GAN, generative adversarial network) generates second biometric data (face) from said code. In fact, it should be observed that the code also represents the second biometric data in practice.
[0083] As explained, the code is not biometric data, but rather a “feature map” that can be further compressed in the form of the aforementioned unique identifier.
[0084] like Figure 3a As shown, the current method proposes a model for generating synthetic data that takes not only the first biometric data as input but also a unique identifier. This introduces additional entropy into the generative model and ensures that if the first biometric data and the input unique identifier are associated with the same individual—that is, when the unique identifier is indeed the encoding of the second biometric data of the same individual—the model can construct synthetic biometric data consistent with the individual's second biometric data.
[0085] In other words, as explained, since biometrics are correlated, the unique identifier "complements" the first biometric data, and the predictive quality of the second biometric data is significantly improved if the two input data correspond to the same individual: for example, in the case of faces, not only is the individual identifiable, but the generated images are also highly similar. Conversely, although it is theoretically possible to reconstruct the second biometric data solely from the identifier (as proposed in the prior art "zero biometric database"), this is ineffective in practice. Therefore, this invention proposes not to reconstruct the second biometric data based on the identifier, but rather to use the identifier only to enhance the reconstruction based on the first biometric data, which is fundamentally different. Furthermore, the identifier can be further compressed, making it completely impossible to reconstruct biometric data solely from the identifier. In other words, the reference unique identifier is preferably obtained by encoding the second reference biometric data, in which way it is impossible to reconstruct the second reference biometric data from the unique identifier, which further enhances the protection of user privacy, as the unique identifier is unique data stored in the database, thus avoiding any regulatory constraints as explained.
[0086] Furthermore, this method further enhances security: conversely, if the two input data of the generating model (the first biometric data and the unique identifier) are associated with different individuals (which constitutes a fraudulent attempt), the model is perturbed, and the synthesized data is no longer similar to the second biometric data at all, thus representing, for example, the face of another individual that does not exist (a "mixture" between the individual associated with the first biometric data and the individual associated with the second biometric data and the unique identifier), thereby making any authentication / identification impossible.
[0087] Finally, it is worth noting that this method is particularly suitable for a large number of users: as long as the database is significantly reduced and only alphanumeric data of the smallest size compared to biometric data is stored, computation and access times will be greatly reduced.
[0088] Finally, in step (c), the data processing device 11 or 21a authenticates / identifies the individual based on a comparison between the second candidate biometric data and the synthetic biometric data.
[0089] It should be understood that the property of the generative model is that the synthetic biometric data is consistent with the second reference biometric data associated with the reference unique identifier (i.e., also belonging to the same reference individual) if and only if the first candidate biometric data is consistent with the hypothetical first reference biometric data associated with the reference unique identifier (i.e., belonging to the same reference individual—it should be understood that in practice, the first reference biometric data is theoretical, not stored, or even unnecessary to exist).
[0090] Therefore, this technology enables the use of generative models to simulate second reference biometric data without storing it.
[0091] In authentication, there is a single unique identifier associated with the expected reference individual (i.e., the identity that the candidate individual claims to possess), and the comparison result is binary.
[0092] In the identification process, multiple reference unique identifiers exist, each associated with a reference individual. In step (b), synthetic biometric data is generated for each reference unique identifier, and in step (c), second candidate biometric data is compared with each synthetic biometric data. The candidate individual is then identified as a reference individual associated with a reference unique identifier, based on which synthetic biometric data consistent with the second candidate biometric data can be constructed (if a match exists—otherwise, the candidate individual is rejected).
[0093] In all cases, security is indeed multifactorial because:
[0094] 1- If the voice of the candidate individual is not the voice of the reference individual, the generated synthetic data is inconsistent with any second biometric data and systematically represents an individual face that does not exist and is therefore neither the candidate individual nor any reference individual, which makes any positive result in step (c) impossible, even if the candidate individual happens to be similar to one of the reference individuals.
[0095] 2- Even if an individual attempts to impersonate a reference individual using a deepfake of their voice, the model will correctly reconstruct the second reference biometric data of that reference individual, but this second reference biometric data will not be consistent with the second candidate biometric data.
[0096] For the implementation of step (c), any known method can be used, particularly by calculating a so-called similarity score, for example, from distance (between the synthetic biometric data and the second candidate biometric data), and then comparing it to a threshold. It should be understood that two biometric data sets, even those from the same individual, will never be completely identical, but if the score exceeds the threshold (i.e., the distance is less than the minimum acceptable distance), they are considered consistent (i.e., the data are identical). It should be noted that the threshold can be predetermined or dynamic, depending on the context.
[0097] Alternatively, particularly in identification, classification models can be used to attempt to associate second candidate biometric data with each synthetic biometric data.
[0098] For generative models, CNN-type neural networks can be used, particularly autoencoders, and especially the autoencoder proposed in the previously mentioned Speech2Face paper, which has a feature extraction block followed by a GAN. If other biometric features besides speech / face are needed, simply change the feature extraction block, specifically by selecting one from the networks adapted to the first biometric feature: for example, VGG would be adapted to facial biometric features, while a speech transcription network would be suitable for facial biometric features.
[0099] study
[0100] As explained, the generated model is typically learned from a database of first and second pairs of learned biometric data associated with the same individual (who is not a reference individual). Each pair may also be associated with a corresponding unique identifier, but it should be reiterated that this identifier can be generated on the fly from the second biometric data, particularly by using an coded model. It should be noted that the learning database is preferably a public database (i.e., each pair is associated with a "public" individual, such as a celebrity or a volunteer individual) to avoid any storage of reference biometric data.
[0101] Preferably, the encoding model is a model jointly learned with the generative model, or a pre-learned model that may be readily available, such as a feature extraction block of a neural network adapted to the first biometric trait (e.g., the proposed VGG). It should be noted that the encoding model may include a final block that represents the output (typically a matrix) of the feature extraction block using alphanumeric identifiers via any known technique. For example, this could be a fully connected layer connecting the values of the boxes in the matrix.
[0102] The possible classification models in step (c) can also be learned, and always on a publicly available basis.
[0103] In this respect, the method advantageously includes a preparatory step (a0): learning at least the parameters for generating a model from the learning database (first and second learning biometric data pairs) via the data processing device 21b of the second server 2b.
[0104] For a pair (preferably multiple, or even all) of first and second learned biometric data in the database, a generative model is trained to reconstruct the second learned biometric data from the first learned data and a unique identifier obtained by applying the encoded model to the second learned biometric data.
[0105] Therefore, refer to Figure 3b In a preferred embodiment, step (a0) for each pair of multiple (or even all) pairs of first and second learned biometric data in the database includes:
[0106] By applying the encoding model to the second learned biometric data, a learned unique identifier (associated with the pair) is generated.
[0107] A second set of learned biometric data is reconstructed by applying a generative model to the first set of learned biometric data and the learned unique identifier. This learning aims to minimize the reconstruction error (the distance between the synthetic image actually constructed by the generative model and the second set of learned biometric data) by adjusting the parameters of the generative model. It should be noted that this type of learning (“input reconstruction”) is standard in autoencoders. It is called an “attempt” of reconstruction because, initially, the synthetic biometric data and the second set of learned biometric data will not be perfectly identical, but this distance will decrease as the parameter values evolve and as new pairs are tested. The encoding model can also be fine-tuned.
[0108] It should be noted that, in order to increase discriminativeness and prevent the generative model from relying on a single input, it can be provided with erroneous samples (two different pairs of first and second biometric data interleaved), and trained to prevent it from reconstructing the second data.
[0109] In this embodiment, step (a0) for each of a plurality of combinations of first and second pairs of first and second learned biometric data from the database (i.e., there exists first learned biometric data of the first pair, second learned biometric data of the first pair, first learned biometric data of the second pair different from the first pair (i.e., a pair associated with two different individuals), and second reference biometric data of the second pair), includes:
[0110] -generate:
[0111] The first learned unique identifier (associated with the first pair), obtained by applying the encoding model to the second learned biometric data of the first pair, and
[0112] The second learned unique identifier (as associated with the second pair) is obtained by applying the encoding model to the second learned biometric data of the second pair;
[0113] - Attempt to build:
[0114] The first synthetic data is obtained by applying a generative model to a first pair of first learned biometric data and a second learned unique identifier, and / or
[0115] The second synthetic data is obtained by applying a generative model to the first learned biometric data of the second pair and to the first learned unique identifier.
[0116] So that neither the first synthetic biomarker nor the second synthetic data is consistent with the second biomarker data of the first pair and / or the second biomarker data of the second pair.
[0117] This learning aims to maximize the reconstruction error (the distance between the synthetic image actually constructed by the generative model and two second-learned biometric data sets) by adjusting the parameters of the generative model. It should be noted that this type of learning (discrimination of fake samples) is typical for GANs. Again, this is called an attempt at reconstruction because there may initially be a certain number of false positives (the synthetic biometric data matching one of the second-learned biometric data sets), but this distance will increase as the parameter values evolve and as new pairs are tested. The encoded model can also be fine-tuned again.
[0118] register
[0119] The method advantageously includes pre-registration of data for authentication or identification of reference individuals, which is typically implemented by the data processing apparatus 21a of the first server 2a in such a way that the registration can be controlled. As for the authentication or identification method, it can be implemented entirely or partially on each of the first and second servers 2a and 2b. According to a preferred mode, there is a true separation where the first server 2a has both authentication / identification and registration functions, while the second server 2b only has learning functions. It should be noted that the registration is independent of the learning (step (a0)) and can be implemented before or after the learning.
[0120] Registration can be initiated by an individual on their terminal 1, assuming they can authenticate themselves independently (e.g., through other biometric factors and / or in the presence of an authority figure), but not on terminal 1 for security reasons.
[0121] According to the preferred embodiment shown in Figure 3, the method includes the following steps:
[0122] (A) Obtain second reference biometric data from the second biometric traits of the reference individual (this is equivalent to step (a) of the method according to the first aspect; it should be understood that first reference biometric data is not required).
[0123] (B) Generate a reference unique identifier by applying the encoding pattern to the second reference biometric data;
[0124] (C) The reference unique identifier is stored on the data storage device 22a of the first server 2a, but the second reference biometric data is not stored. It should be understood that the second reference biometric data may be stored briefly for the time required to perform step (B), but will be deleted in step (C) if necessary. Preferably, the unique identifier is generated on the fly.
[0125] server
[0126] According to a second aspect, the invention relates to apparatus for implementing the method according to the invention. Specifically, the first server 2a and / or terminal 1 serve as authentication / identification devices, and the first server 2b and / or the second server 2b serve as registration devices. However, the second server 2b is the only server responsible for learning.
[0127] The authentication / identification device includes data processing devices 11 and 21a and data storage devices 12 and 22a. Terminal 1 has an interface 13, and most importantly, it has biometric data acquisition devices 14 and 15 (e.g., a microphone and a camera).
[0128] Devices 11 and 21a are configured as follows:
[0129] Obtain first candidate biometric data from an individual's first biometric trait and second candidate biometric data from an individual's second biometric trait that is different from the first biometric trait;
[0130] By applying the generative model to the first candidate biometric data and the at least one reference unique identifier, synthetic biometric data corresponding to the second biometric trait is constructed.
[0131] The individual is authenticated or identified based on a comparison between second candidate biometric data and synthetic biometric data.
[0132] Device 21a may be further configured as follows:
[0133] Obtain second reference biometric data derived from the second biometric traits of a reference individual;
[0134] A unique reference identifier is generated by applying an encoding model to second reference biometric data;
[0135] The reference unique identifier is stored on the data storage device 22a, but the second reference biometric data is not stored.
[0136] According to the third aspect, a set of terminal 1, first server 2a, and optionally second server 2b is proposed. All these components 1, 2a, and 2b can be connected via network 10.
[0137] Computer program products
[0138] According to the fourth and fifth aspects, the invention relates to a computer program product including code instructions for executing (particularly on data processing devices 11, 21a, 21b) a method for authenticating or identifying an individual according to the first aspect of the invention, and a computing device readable storage device (memory 12, 22a, 22b) on which the computer program product is located.
Claims
1. A method for authenticating or identifying an individual, characterized in that, The method is implemented by the data processing apparatus (11, 21a) of the terminal (1) and / or the first server (2a), and includes the following steps: (a) Obtaining first candidate biometric data from a first biometric trait of the individual and second candidate biometric data from a second biometric trait of the individual that is different from the first biometric trait; (b) By applying the generative model to the first candidate biometric data and at least one reference unique identifier, synthetic biometric data corresponding to the second biometric trait is constructed; the reference unique identifier is not biometric data and is stored on the data storage device (12, 22a) of the terminal (1) and / or the first server (2a); (c) Authenticate or identify the individual based on a comparison between the second candidate biometric data and the synthetic biometric data.
2. The method according to claim 1, wherein, There are multiple reference unique identifiers, each of which is associated with a reference individual. In step (b), synthetic biometric data is generated for each reference unique identifier, and in step (c), the second candidate biometric data is compared with each synthetic biometric data.
3. The method according to any one of claims 1 and 2, wherein, The reference unique identifier is the encoding of the second reference biometric data, which also corresponds to the second biometric trait associated with the reference unique identifier; the second reference biometric data is not stored by the data storage device (12, 22a) of the terminal (1) and / or the first server (2a).
4. The method according to claim 3, comprising pre-registration of data for authentication or identification of a reference individual, said registration being performed by the data processing apparatus (21a) of the first server (2a), comprising the following steps: (A) Obtain second reference biometric data from the second biometric traits of the reference individual; (B) A reference unique identifier is generated by applying the encoding pattern to the second reference biometric data; (C) The reference unique identifier is stored on the data storage device (22a) of the first server (2a), but the second reference biometric data is not stored.
5. The method according to any one of claims 1 to 4, wherein, The generative model is a learning model, and the method includes a preliminary step (a0): the data processing device (21b) of the second server (2b) learns at least the parameters of the generative model from a database of the first and second pairs of learning biometric data.
6. The method according to any one of claims 3 and 4 and a combination of claim 5, wherein, Step (a0) for each of a plurality of pairs of first and second learned biometric data in the database includes: By applying the encoding model to the second learned biometric data, a learned unique identifier (associated with the pair) is generated. The second learned biometric data is reconstructed by applying the generative model to the first learned biometric data and the learned unique identifier.
7. The method according to claim 6, wherein, Step (a0) for each of a plurality of combinations of first and second pairs of first and second learned biometric data from the database includes: generate: The first learned unique identifier is obtained by applying the encoding model to the second learned biometric data of the first pair, and The second learned unique identifier is obtained by applying the encoding model to the second pair of second learned biometric data; Attempt to build: The first synthetic data is obtained by applying the generative model to the first learned biometric data and the second learned unique identifier of the first pair, and The second synthetic data is obtained by applying the generative model to the first learned biometric data and the first learned unique identifier of the second pair. This ensures that neither the first synthetic biomarker nor the second synthetic data is consistent with the second biomarker data of the first pair and / or the second biomarker data of the second pair.
8. The method according to any one of claims 1 to 7, wherein, The generative model is a CNN-type neural network, specifically an autoencoder.
9. The method according to any one of claims 1 to 8, wherein, Step (a) includes acquiring the first candidate biometric data from the first biometric trait of the individual through the first biometric acquisition device (14) of the terminal (1), and acquiring the second candidate biometric data from the second biometric trait of the individual through the second biometric acquisition device (15) of the terminal (1), and transmitting the first candidate biometric data and the second candidate biometric data to the first server (2a); steps (b) and (c) are implemented by the data processing device (21a) of the first server (2a).
10. The method according to any one of claims 1 to 9, wherein, The first biometric trait is voice, and the second biometric trait is face.
11. A device (11, 2a) for authenticating or identifying an individual, comprising a data processing device (11, 21a) and a data storage device (12, 22a) storing at least one reference unique identifier, characterized in that, Its included data processing apparatus (11, 21a) is configured as follows: Obtain first candidate biometric data from a first biometric trait of the individual and second candidate biometric data from a second biometric trait of the individual that is different from the first biometric trait; By applying the generative model to the first candidate biometric data and the at least one reference unique identifier, synthetic biometric data corresponding to the second biometric trait is constructed. The individual is authenticated or identified based on a comparison between the second candidate biometric data and the synthetic biometric data.
12. The device according to claim 11, further defined as a data registration for authentication or identification of a reference individual, characterized in that, The data processing device (21a) is further configured to: Obtain second reference biometric data derived from the second biometric trait of the reference individual; A unique reference identifier is generated by applying an encoding pattern to the second reference biometric data; The reference unique identifier is stored on the data storage device (22a), but the second reference biometric data is not stored.
13. A computer program product comprising code instructions that, when executed on a computer, are used to perform a method for authenticating or identifying an individual according to any one of claims 1 to 10.
14. A computing device readable storage device having a computer program product recorded thereon, the computer program product including code instructions for performing a method for authenticating or identifying an individual according to any one of claims 1 to 10.