Secure classification method for input data using a neural network with convolution
Patent Information
- Application Number
- DE602020041544
- Authority / Receiving Office
- DE · DE
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2019-05-07
- Filing Date
- 2020-05-07
- Publication Date
- 2024-11-20
- Estimated Expiration
- 2040-05-07
AI Technical Summary
Current biometric classification methods using convolutional neural networks (CNNs) face challenges in secure data sharing and collaboration between entities, as existing solutions either leak national biometric characteristics or require central manipulation, leading to potential security breaches.
A method involving a main server and secondary servers to learn and aggregate CNN parameters from confidential databases while maintaining data confidentiality, using a 'draw and discard' mechanism with noise introduction to create a common representation space for classification, ensuring secure data processing and aggregation.
This approach enables secure collaboration among entities by maintaining data confidentiality and preventing leaks, while improving recognition effectiveness without exposing sensitive information.
Description
GENERAL TECHNICAL FIELD
[0001] The present invention relates to the field of biometrics, and in particular to a method for secure classification of input data using a convolutional neural network, for authentication / identification. STATE OF THE ART
[0002] Neural networks are widely used for data classification.
[0003] After a phase of machine learning (generally supervised, i.e. on a base of already classified reference data), a neural network “learns” and becomes capable of applying the same classification to unknown data on its own.
[0004] Convolutional Neural Networks (CNNs) are a type of neural network in which the connection pattern between neurons is inspired by the visual cortex of animals. They are therefore particularly suited to a particular type of classification, which is image analysis. They effectively enable the recognition of objects or people in images or videos, particularly in security applications (automatic surveillance, threat detection, etc.).
[0005] CNNs are particularly well known for their use in the police and anti-terrorism fields. More specifically, police forces have databases of photographs, for example, of the faces of individuals involved in cases. It is then possible to train CNNs to recognize faces in video surveillance data, in particular to detect wanted individuals. Similarly, we can imagine that governments have biometric databases, for example, of passport fingerprints. It is then possible to train CNNs to recognize the fingerprints of specific individuals. YU HUAPENG ET AL: "Transfer Learning for Face Identification with Deep Face Model", 2016 7TH INTERNATIONAL CONFERENCE ON CLOUD COMPUTING AND BIG DATA (CCBD), IEEE ,DOI: 10.1109 / CCBD.2016.014, describes transfer learning for facial recognition.
[0006] Today, a problem that arises is that these databases are confidential and restricted (especially national). However, it would be desirable, for example, for the police forces of several states to cooperate and improve the overall effectiveness of reconnaissance, without it being possible to trace confidential data.
[0007] However, this would in any case imply that one entity (for example the police force of a state) would have its CNNs learn from the databases of another entity (the police force of another state), i.e. transmitting to each other in clear the databases of photographs or other biometric traits, which is not possible today.
[0008] In application FR1852351, a clever solution was proposed which, despite the non-existence of a common representation space, allows "teachers" to learn about each of the confidential data sets, so as to be able to jointly train a "student" who will ultimately have knowledge of the various confidential data sets.
[0009] More specifically, teachers will generate a learning database for the student by classifying raw public data using a teacher "voting" system, possibly noisy to avoid being able to trace back to individual sets of confidential data used for the teachers' learning.
[0010] This technique is entirely satisfactory, but has the disadvantage that learning, for each base, takes place in isolation. It is then possible to imagine that although the images used for learning are individually protected, national biometric characteristics could still leak. Another problem comes from the fact that teachers must be manipulated at a central level to teach the student and that it is at this level that there can be security breaches.
[0011] So the situation could be further improved. PRESENTATION OF THE INVENTION
[0012] According to a first aspect, the present invention relates to a method for learning parameters of a convolutional neural network, CNN, for classification of personal data, in order to recognize an individual by means of personal data of said individual, the method comprising the implementation of steps of:(a1) Obtaining by data processing means of a main server, the parameters of a set of at least one first CNN; (a2) For a first CNN of said set: Learning by the data processing means of said main server or the data processing means of a secondary server having a confidential training database already classified, from a public training database already classified, the parameters of a final representation block of a second CNN corresponding to the first CNN to which said representation block has been added; Relearning by data processing means of said secondary server, from its confidential training database already classified, the parameters of the second CNN; Transmission to the main server of the parameters of a third CNN corresponding to the second CNN without a final representation block;(a3) Replacement by the data processing means of the main server of a first CNN of said set of first CNNs by the third CNN; (a4) Aggregation of said set of at least one first CNN into a fourth CNN.;
[0013] According to other advantageous and non-limiting characteristics: said data are, in particular biometric data, in particular facial images; steps (a2) then (a3) are implemented for a plurality of secondary servers each having a confidential training database already classified; steps (a2) then (a3) are repeated a given number of times so as to update said set of first CNNs; the retraining in step (a2) of the parameters of the second transmitted CNN comprises the introduction of noise; said set obtained in step (a1) comprises a plurality of first CNNs, step (a2) being implemented for a first CNN randomly selected from said set, and the first CNN replaced by the third CNN in step (a3) being randomly selected from said set; the first CNNs randomly selected in steps (a2) and (a3) are independently selected;step (a4) comprises processing data from a second public database still unlabeled, independently by each of the first CNNs, and aggregating the results of this processing; then learning, from the second public training database now labeled with the aggregated results, the parameters of said fourth CNN; the results of the processing of a data item from a second public database by the first CNNs are feature vectors, the aggregation of the results of this processing consisting of averaging the feature vectors obtained; said representation block of the second CNN establishes a mapping from a representation space of the first CNN to a classification space; said representation block consists of a fully connected layer;either the learning of the parameters of the final representation block of the second CNN is implemented by the data processing means of said main server, and step (a2) comprises the subsequent transmission to said secondary server of the parameters of said second CNN, or the learning of the parameters of the final representation block of the second CNN is implemented by the data processing means of said secondary server, and step (a2) comprises the prior transmission to said secondary server of the parameters of said first CNN. ;
[0014] According to a second aspect, a method for classifying input data is proposed, characterized in that it comprises the implementation of steps of: (a) Training a fourth CNN in accordance with the method according to the first aspect; (b) Classification by the data processing means of the terminal of said input data, by means of the fourth CNN.
[0015] According to a third and a fourth aspect, the invention relates to a computer program product comprising code instructions for executing a method according to the first aspect of learning parameters of a convolutional neural network, CNN, or according to the second aspect of classifying an input data; and a storage means readable by a computer equipment on which a computer program product comprises code instructions for executing a method according to the first aspect of learning parameters of a convolutional neural network, CNN, or according to the second aspect of classifying an input data. PRESENTATION OF FIGURES
[0016] Other characteristics and advantages of the present invention will appear on reading the following description of a preferred embodiment. This description will be given with reference to the appended drawings in which: There figure 1is a diagram of an architecture for implementing the methods according to the invention; The figure 2 schematically represents the steps of the learning method according to the invention; The figure 3 schematically represents how the first, second and third CNNs used in the methods according to the invention are defined in relation to each other. DETAILED DESCRIPTION Architecture
[0017] According to two complementary aspects of the invention, the following are proposed: a method for training parameters of a convolutional neural network (CNN) for data classification; and a method for classifying input data using a CNN learned using the first method.
[0018] These two types of processes are implemented within an architecture as represented by the figure 1, thanks to one or more servers 1a, 1b, 1c and a terminal 2. The server(s) 1a, 1b, 1c are the learning equipment (implementing the first method) and the terminal 2 is classification equipment (implementing the second method), for example video surveillance data processing equipment.
[0019] In all cases, each equipment 1a, 1b, 1c, 2 is typically a remote computer equipment connected to a wide area network 10 such as the internet network for the exchange of data. Each comprises data processing means 11a, 11b, 11c, 21 of the processor type, and data storage means 12a, 12b, 12c, 22 such as a computer memory, for example a disk.
[0020] A server 1c is a so-called main server which does not have a training database (or at least not a confidential database, but we will see that it has access to at least a first public training database). The role of this server 1c can quite easily be accomplished by one or other of the servers 1a, 1b, but preferably it is a separate server (i.e. partitioned) to avoid any risk of disclosure of the confidential databases of the servers 1a, 1b. Similarly, the terminal 2 could be confused with the main server 1c. Indeed, at least one server 1a, 1b called secondary (or "client") stores a confidential training database, i.e. a set of already classified data (as opposed to the so-called input data that we are trying to classify).Preferably, there are at least two confidential learning databases already classified, stored on two separate secondary servers (1a and 1b in the . figure 1 ), without interactions: server 1a does not have access to the database of server 1b and vice versa.
[0021] For example, these are databases of national police forces from two states.
[0022] As such, the said input or learning data are preferably personal data, i.e. personal data of an individual (for which confidentiality is therefore necessary), and in particular either biometric data (which are by definition personal to their owner), or alphanumeric personal data (typically name, first name, date and place of birth, address, etc.). Note that we will not be limited to these two types of personal data, and that we can, for example, take a signature, a seal, etc.
[0023] Generally speaking, the classification of personal data is typically the recognition of the individual concerned, and as such the present method of classifying input personal data may be a method of identifying / authenticating an individual by means of personal data of said individual.
[0024] Note that according to an unclaimed aspect, the present method is not limited to personal data, and the input or training data may be representative of images, sounds, etc. (said classification being recognition of an object, animal, etc.).
[0025] In the preferred embodiment of biometric data, the input or training data are typically again representative of images or even directly of images of biometric traits (faces, fingerprints, irises, etc.), or directly of preprocessed data from the biometric traits (for example the position of minutiae in the case of fingerprints). In the remainder of this description, the example of facial images will be taken, but it will be understood that this embodiment is not limited.
[0026] Generally speaking, personal data, and in particular biometric data, are "non-quantifiable" data (in the sense that a face or a fingerprint has no numerical value), for which there is no common space of representation. CNN
[0027] A CNN generally contains four types of layers that successively process the information: the convolution layer which processes blocks of the input one after the other; the non-linear layer which allows non-linearity to be added to the network and therefore to have much more complex decision functions; the pooling layer which allows several neurons to be grouped into a single neuron; the fully connected layer which connects all the neurons of a layer to all the neurons of the previous layer.
[0028] The nonlinear layer activation function NL is typically the function ReLU (Rectified Linear Unit, i.e. Linear Rectification Unit) which is equal to f(x) = max(0, x) and the pooling layer (denoted POOL ) the most used is the function MaxPool2 × 2 which corresponds to a maximum between four values of a square (we put four values together into one).
[0029] The convolution layer, denoted CONV,and the fully connected layer, denoted FC , generally correspond to a scalar product between the neurons of the previous layer and the weights of the CNN.
[0030] Typical CNN architectures stack a few pairs of layers CONV → NL then add a layer POOL and repeat this pattern [(CONV → NL) p< → POOL ] until a sufficiently small output vector is obtained, then end with two fully connected layers FC .
[0031] Here is a typical CNN architecture: INPUT → [[CONV → NL] p< → POOL] n< → FC → FC
[0032] In the context of this process, we can notably use CNNs as proposed in the documents Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman: Deep Face Recognition. BMVC 2015 ,and DeepVisage: Making face recognition simple yet with powerful generalization skills. Abul Hasnat, Julien Bohne, Jonathan Milgram, Stephane Gentric, Liming Chen, ICCV 2017 , these networks being particularly effective for face recognition. Principle
[0033] We know of so-called federated learning algorithms that allow data confidentiality to be maintained. For example, the document H. Brendan McMahan, Daniel Ramage, Kunal Talwar, Li Zhang Learning Differentially Private Recurrent Language Models. https: / / arxiv.org / pdf / 1710.06963v3.pdf, allows learning LSTM (Long Short-Term Memory) type models used in language recognition, by successively submitting a shared model to several users (secondary servers).
[0034] More precisely, for a selected group of secondary servers, each generates an update of a model received on the basis of all or part of its confidential database in the form of a vector, which is "clipped" (i.e. averaged if necessary) so as to limit its norm, and noisy.
[0035] More recently, a new learning algorithm was proposed in the paper Vasyl Pihur, Aleksandra Korolova, Frederick Liu, Subhash Sankuratripati, Moti Yung, Dachuan Huang, Ruogu Zeng. Differentially-Private "Draw and Discard" Machine Learning. https: / / arxiv.org / pdf / 1807.04369v2.pdf .
[0036] This "Draw and Discard" algorithm allows for learning Generalized Linear Model (GLM) models. GLM generalizes linear regression by allowing the linear model to be related to the response variable via a link function and by allowing the magnitude of the variance of each measurement to be a function of its predicted value.
[0037] The algorithm proposes to initialize a set of k models, and repeatedly update a randomly chosen model on the confidential database (by adding noise to this learning) and then replace a randomly chosen model from the set (not necessarily the same) with the updated model. The selection of the model to be updated is the "draw" and the replacement of a model is the "discard". After a certain number of iterations, the models can be aggregated.
[0038] These techniques are satisfactory, but are not usable in the case of identification / authentication of individuals, due to the non-existence of a common representation space. More precisely, the different confidential databases relate to different people: it is not possible to simply transfer a model from one secondary server to another for updating since the individuals in other databases are mutually unknown to them.
[0039] This learning method proposes a clever solution to the problem of the non-existence of a common representation space. Learning process
[0040] According to a first aspect, is proposed with reference to the figure 2 the learning method, implemented by the data processing means 11a, 11b, 11c of at least one secondary server 1a, 1b and the main server 1c.
[0041] The learning method begins with a “preparatory” step (a1) of obtaining, by the data processing means 11c of a main server 1c, the parameters of at least one first CNN, preferably a plurality of first CNNs, forming a set of k first CNNs, with k typically between 10 and 30, advantageously around 20.
[0042] We understand that this is an initialization step implemented in any possible way, the first CNN(s) being able to be taken as is (with parameter values taken arbitrarily and / or chosen randomly) or pre-learned for example on a public learning database (see below).
[0043] All the first CNNs advantageously have the same architecture, and they are preferably CNNs for extracting "features" (i.e. characteristics, i.e. differentiating traits), in particular biometric, and not necessarily classification CNNs. More precisely, the first CNNs are not necessarily capable of returning a classification result for a piece of data, but at least of producing a representation of the "features" of this data in a representation space (called "feature space"). To reformulate further, the first CNNs have as output not a classification result for an input piece of data, but a vector of characteristics representing (in an abstract manner) the input piece of data in said representation space.
[0044] In a step (a2) we will implement the "draw". Step (a2) is thus implemented for a first CNN of said set selected randomly if necessary. Naturally, if the set includes a single first CNN, this is automatically selected. It will be understood in this respect that step (a2) (with step (a3) which will be described later) can be repeated a large number of times, in particular several hundred or even several thousand times.
[0045] Normally, according to the draw mechanism, the parameters of the first CNN should be transmitted to a secondary server 1a, 1b having a confidential training database already classified, for updating.
[0046] However, each first CNN potentially has its own representation space, so this update is impossible. Thus, we will cleverly transform the first CNN selected into a second CNN, so that whatever the representation space of the first CNN, all the second CNNs will have a common representation space.
[0047] To achieve this, from a base, this time public, of already classified learning data (which will be called the first public database to distinguish it from a possible public learning database, this time again unclassified, which will be used in a particular embodiment), the main server 1c or the secondary server 1a, 1b learns parameters of a final representation block B of a second CNN corresponding to the first selected CNN to which said representation block B has been added.
[0048] In the case of multiple implementations of step (a2), the sub-step of transforming each selected first CNN into a second CNN is implemented using the same first public learning database already classified so as to learn the parameters of a plurality of second CNNs, all the second CNNs advantageously again having the same architecture.
[0049] In reference to the figure 3, we understand that the second CNNs correspond to first CNNs each completed with the final block B called representation (or IP layer, for "Inner Product"). More precisely, we give the first CNNs a common classification capacity (in said first public database) from their feature extraction capacity. We thus obtain a common space for representing the classifications. We understand that the representation block B can take many forms, in particular at least one fully connected layer FC, and preferably exactly one final FC layer. Generally speaking, the representation block B "maps" (i.e. establishes a correspondence) the representation space of the first CNN to the classification space.
[0050] This sub-step must be understood as a step of learning only the missing parameters of the second CNNs, i.e. we keep the parameters already learned for the first CNNs and we “re-train” these networks so as to learn the representation block B, and this on the same data, which forces the appearance of the common space. Advantageously, this learning is not noisy, because the data is public so it is not necessary.
[0051] The skilled person will be able to use any already classified public database for the implementation of this sub-step (and if necessary during the initialization of the first CNNs), knowing that by "public" we mean that the data are not confidential and can be shared freely. The first public database is different from each of the confidential databases. We can either specifically create this first public database, or use an available database. In the case where the data are photographs of faces, we will cite for example the IMDB (Internet Movie DataBase) database, from which it is possible to extract images of N=2622 individuals, with up to 1000 images per individual, see Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman: Deep Face Recognition. BMVC 2015 .
[0052] Note that the learned classification can, rather than a simple binary response, graduate the responses to obtain nuances that can be exploited by also knowing which are the closest classes, see the document Geoffrey Hinton, Oriol Vinyals, Jeff Dean: Distilling the Knowledge in a Neural Network. NIPS 2014 . This simply means that where a classic classification determines THE class that it considers to be that of an input data (for example the name of the person in the photo), the present classification can indicate one or more potential classes, possibly with a percentage of confidence.
[0053] If it is the central server 1c which has implemented the learning of the parameters of the final representation block B of the second CNN, the parameters of the second CNN obtained (for the first selected CNN, i.e. the parameters of the first CNN initially available plus the parameters of the newly learned representation block B) are transmitted to said secondary server 1a, 1b having a confidential learning database already classified. Alternatively, the parameters of the first selected CNN are transmitted upstream to said secondary server 1a, 1b, and it is the data processing means 11a, 11b of the latter which implement the learning of the parameters of the final representation block B of the second CNN.
[0054] Note that in the case of a plurality of secondary servers 1a, 1b, one can either also randomly select the one to which the transmission (of the parameters of the first selected CNN or of the second learned CNN) is made, or proceed in turn.
[0055] Then, step (a2) comprises the relearning by the data processing means 11a, 11b of said secondary server 1a, 1b, from its already classified confidential learning database, of the parameters of the second transmitted CNN. In other words, the secondary server 1a, 1b having received or learned the second CNN updates it on the basis of its confidential data. It is noted that all or part of its base is used, in particular a batch of chosen size. All or part of the parameters of the second CNN can be locked during this relearning (in particular those of the very first layers) to increase efficiency, but preferably the entire network is relearning, this is what is called fine-tuning.
[0056] Preferably, the relearning comprises the introduction of noise, so as to avoid being able to go back to the confidential databases of the secondary servers 1a, 1b. Those skilled in the art may refer to the noise techniques proposed in the document H. Brendan McMahan, Daniel Ramage, Kunal Talwar, Li Zhang Learning Differentially Private Recurrent Language Models or the document Vasyl Pihur, Aleksandra Korolova, Frederick Liu, Subhash Sankuratripati, Moti Yung, Dachuan Huang, Ruogu Zeng. Differentially-Private "Draw and Discard" Machine Learning (and generally implement any other characteristic of the learning technique described in this document) or any other known "differential privacy" technique.
[0057] After retraining, only the parameters of a third CNN corresponding to the second CNN without the final representation block B are transmitted back to the main server 1a, 1b. In other words, only the parameters of the second CNN corresponding to the layers from the beginning to the representation block are sent back.
[0058] We thus understand that the third CNN obtained has exactly the structure of the first CNN selected and corresponds to this first updated CNN. Its representation space has evolved and is now "a little more common" to the extent that new data has been taken into consideration.
[0059] It can therefore be replaced within the said set of first CNNs, in accordance with the “discard” mechanism.
[0060] Thus, in step (a3), the data processing means 11c of the main server 1c replace in said set of first CNNs a first CNN selected, if necessary, randomly by the third CNN. As explained previously, this random drawing of a first CNN is preferably independent of the random drawing of a first CNN in step (a2). Thus, there is a probability 1 / k of replacing the first CNN already selected in step (a2), i.e. of updating it (it is recalled that the third CNN corresponds to the first selected updated CNN), but there is a probability (k-1) / k of replacing another first CNN. Naturally, if said set only comprises a single first CNN, this is the replaced CNN.
[0061] The third CNN is then considered for the remainder of the method as a first CNN in its own right (i.e. the set of first CNNs is updated), and steps (a2) then (a3) can be repeated as explained on the new set of first CNNs. It is entirely possible that the third CNN is in other occurrences of steps (a2) and (a3) selected for new update or replaced by another third CNN. Preferably, the secondary servers 1a, 1b implement step (a2) in parallel and independently on their confidential database: first CNNs are randomly selected for each of these secondary servers 1a, 1b, and third CNNs are received in response for each of them.
[0062] In a step (a4) said set of first CNNs is aggregated. It is understood that at this stage said set of first CNNs has been updated a very large number of times, so that the probability that there remains an original first CNN is tiny. All the CNNs present have undergone multiple retrainings in accordance with step (a2), so that it can be assumed that their representation spaces have converged (even if in practice they will never be identical).
[0063] Aggregation means obtaining from said set a single CNN, called the fourth CNN, for use in classification.
[0064] A first simple embodiment of step (a4) may consist of simply taking as the fourth CNN any first CNN of said set. If the set contains only one CNN, this is taken as the fourth CNN. Alternatively, if there are a plurality of first CNNs and if we want to benefit from the consolidation of all these networks, another approach will be initially to learn the transformation matrices of the representation spaces of the different first CNNs to ensure that the first k CNNs work in the same space. This will make it possible to have a consolidated feature vector by averaging the outputs of the k networks.
[0065] We can therefore consider a “teacher / student” approach, in order to train a fourth CNN which, on an input data, must reproduce the average of the characteristic vectors output by each of the first k CNNs.
[0066] To do this, we can advantageously proceed in a similar way to what is proposed in the document Nicolas Papernot, Martín Abadi, Úlfar Erlingsson, Ian Goodfellow, Kunal Talwar: Semi-supervised Knowledge Transfer for Deep Learning from Private Training Data. ICLR 2017 , or in the application FR1852351 which uses this teacher / student approach.
[0067] More precisely, the first CNNs constitute k teachers who will be able to train a student.
[0068] Step (a4) thus comprises the processing of data from a second public database that is still unlabeled, by means of said first CNNs (the result being a classification of the input data or simply a vector of characteristics), then the learning, from the second public learning database now associated with the result of said processing, i.e. labeled (i.e. classified or associated with a vector of characteristics), of the parameters of the fourth CNN. In a classification case, the fourth CNN again comprises a final representation block B (in particular an FC layer), and thus preferably has an architecture similar to that of the second CNNs.
[0069] In a known manner, step (a4) preferably comprises, for each data item of said second public database, the processing of this data item independently by each of the first CNNs, and the aggregation of the results of this processing, i.e. the vote of the teachers. In the case where the results are characteristic vectors, the vote of the teachers advantageously consists of taking the average of the k representation vectors obtained.
[0070] Again, this step may include the introduction of noise into the classification results before aggregation, but this is not mandatory since it is already impossible to go back to the confidential data in the databases of the secondary servers 1a, 1b. Indeed, the entirety of step (a4) is implemented at the level of the main server 1c which has all the first CNNs, and the secondary servers 1a, 1b are no longer used.
[0071] Finally, when the fourth CNN, i.e. the student, is obtained, its eventual final representation block B can be deleted to return to a feature extraction CNN. Classification process
[0072] According to a second aspect, the method for classifying an input data item is proposed, implemented by the data processing means 21 of the terminal 2. The classification method comprises two steps: in a first step (a) the learning of a fourth CNN as defined previously is implemented, and in a second step (b) the data processing means 21 of the terminal 2 classify said input data item, by means of the fourth CNN.
[0073] This process is implemented in a classic way, we just understand as explained that knowing only the fourth CNN, there is no possibility of going back to the confidential databases which allowed the learning of the first CNNs. Computer program product
[0074] According to a third and a fourth aspect, the invention relates to a computer program product comprising code instructions for the execution (in particular on the data processing means 11a, 11b, 11c, 21 of one or more servers 1a, 1b, 1c or of the terminal 2) of a method according to the first aspect of the invention for learning parameters of a CNN or a method according to the second aspect of the invention for classifying an input data item, as well as storage means readable by computer equipment (a memory 12a, 12b, 12c, 22 of one or more servers 1a, 1b, 1c or of the terminal 2) on which this computer program product is found.
Claims
1. Method for learning parameters of a convolutional neural network, CNN, for classifying personal data, in order to recognize an individual by way of a personal datum of said individual, the method comprising implementing steps of: (a1) data processing means (11c) of a main server (1c) obtaining parameters of a set of at least one first CNN; (a2) for a first CNN of said set: - the data processing means (11c) of said main server (1c) or data processing means (11a, 11b) of a secondary server (1a, 1b) possessing a database of previously classified confidential learning data learning the parameters of a final representation block (B) of a second CNN corresponding to the first CNN plus said representation block (B), based on a database of previously classified public learning data; - data processing means (11a, 11b) of said secondary server (1a, 1b) relearning the parameters of the second CNN, based on its database of previously classified confidential learning data; - transmitting the parameters of a third CNN corresponding to the second CNN without a final representation block (B) to the main server (1c); (a3) the data processing means (11c) of the main server (1c) replacing a first CNN of said set of first CNNs with the third CNN; (a4) aggregating said set of at least one first CNN into a fourth CNN.
2. Method according to Claim 1, wherein said personal data are biometric data, in particular images of faces.
3. Method according to either of Claims 1 and 2, wherein steps (a2) and then (a3) are implemented for a plurality of secondary servers (1a, 1b) each possessing a database of previously classified confidential learning data.
4. Method according to one of Claims 1 to 3, wherein steps (a2) and then (a3) are repeated a given number of times so as to update said set of first CNNs.
5. Method according to one of Claims 1 to 4, wherein the relearning of the transmitted parameters of the second CNN in step (a2) comprises introducing noise.
6. Method according to one of Claims 1 to 5, wherein said set obtained in step (a1) comprises a plurality of first CNNs, step (a2) being implemented for a first CNN selected randomly from said set, and the first CNN replaced by the third CNN in step (a3) being selected randomly from said set.
7. Method according to Claim 6, wherein the first CNNs selected randomly in steps (a2) and (a3) are randomly selected independently.
8. Method according to one of Claims 1 to 7, wherein step (a4) comprises each of the first CNNs independently processing data in a second database of public data that have not yet been labelled, and aggregating the results of this processing; and then learning the parameters of said fourth CNN based on the second database of public learning data that have now been labelled with the aggregated results.
9. Method according to Claim 8, wherein the results of the processing of a datum in a second database of public data carried out by the first CNNs are feature vectors, aggregating the results of this processing consisting in averaging the obtained feature vectors.
10. Method according to one of Claims 1 to 9, wherein said representation block (B) of the second CNN establishes a mapping of a representation space of the first CNN to a classification space.
11. Method according to one of Claims 1 to 10, wherein said representation block (B) consists of a fully connected layer.
12. Method according to one of Claims 1 to 11, wherein either the learning of the parameters of the final representation block (B) of the second CNN is implemented by the data processing means (11c) of said main server (1c), and step (a2) comprises subsequently transmitting the parameters of said second CNN to said secondary server (1a, 1b), or the learning of the parameters of the final representation block (B) of the second CNN is implemented by the data processing means (11a, 11b) of said secondary server (1a, 1b), and step (a2) comprises transmitting the parameters of said first CNN to said secondary server (1a, 1b) beforehand.
13. Method for classifying an input datum, characterized in that it comprises implementing steps of: (a) training a fourth CNN in accordance with the method according to one of Claims 1 to 12; (b) the data processing means (21) of the terminal (2) classifying said input datum by way of the fourth CNN.
14. Computer program product comprising code instructions for executing a method according to one of Claims 1 to 12 for learning parameters of a convolutional neural network, CNN, or for classifying an input datum according to Claim 13 when said program is executed by a computer.
15. Storage medium able to be read by a computer equipment and on which a computer program product comprises code instructions for executing a method according to one of Claims 1 to 12 for learning parameters of a convolutional neural network, CNN, or for classifying an input datum according to Claim 13 when the code instructions are executed by a computer.
