Method for classifying signals, associated device and computer program

Abstract

Method for classifying signals, associated device and computer program This method for classifying digital signals by implementing a classification model trained by machine learning on a set of training signals, the classification model providing, for an input digital signal, a classification vector of size equal to a predefined number of classes, each component of the classification vector having a value representative of a probability of the input digital signal belonging to a corresponding class, comprises for an input signal to be classified, steps of: -calculating (42) a score of the input digital signal to be classified, said score being representative of a probability of the input digital signal belonging to a statistical distribution of the training signals, -comparing (44) the score to a predetermined threshold, and in the event of said threshold being exceeded,and - emission (46) of an alarm in order to indicate the presence of a supposedly corrupted digital input signal. Figure for the abstract: Figure 3,

Description

Title of the invention: Method for classifying signals, associated device and computer program

[0001] The present invention relates to a method for classifying signals by implementing a classification model trained by machine learning on a set of training signals.

[0002] The invention also relates to an associated digital signal classification device and an associated computer program.

[0003] The invention lies in the field of signal classification, and more particularly of signals representative of physical quantities such as digital image signals, by using machine learning. The invention also applies to audio signals.

[0004] One of the objectives of the invention is to provide a classification method and device which are robust to adversarial attacks on the signals to be classified, which are also known as adversarial example attacks.

[0005] Digital image classification, implemented by classification models trained by machine learning, for example classification models based on deep learning neural networks, has experienced great growth in recent years.

[0006] Such classification models are used in various fields of application, in particular in the field of autonomous or semi-autonomous vehicles, for example motor vehicles or unmanned aircraft. Indeed, to carry out movement without a pilot on board, the vehicles are equipped with sensors, for example digital image sensors, as well as communication interfaces making it possible to acquire signals, for example digital images, from systems external to the vehicle. The digital images received are then analyzed, and in particular classified, in particular to detect obstacles to be avoided, and therefore reduce the risk of an accident.

[0007] Other fields of application include the field of industrial production supervision, quality inspection, as well as the field of medical imaging among others.

[0008] In all these fields of application, the accuracy of the classification result is very important, even critical to achieve the task to be accomplished. For example, in the field of autonomous or semi-autonomous vehicles, an erroneous classification result can lead to an erroneous action and potentially a collision of the vehicle. with another vehicle, a pedestrian etc.

[0009] The possibility of attacking digital image classification models trained by machine learning has been demonstrated. Indeed, such models are trained on a training database, e.g. a training digital image database, which contains images of objects belonging to a plurality of predefined classes. The term "object" is used here in the broad sense, for example in the case of autonomous motor vehicles the object classes include for example: car, motorcycle, bicycle, pedestrian, road sign, pole etc. The classification model associates a class label with an input digital image or a vector containing probability values ​​of belonging to each of the predefined classes, which makes it possible to deduce the most probable class.We know of attacks called "adversarial example attacks", consisting of a modification, imperceptible to the human eye, of a digital image, leading to a bad classification of the digital image by the classification model.

[0010] The aim of the invention is then to propose a method for classifying digital signals making it possible to improve the accuracy of the classification, including from signals which are corrupted by an attack of the type mentioned above.

[0011] To this end, the subject of the invention is a method for classifying digital signals by implementing a classification model trained by machine learning on a set of training signals, the classification model providing, for an input digital signal, a classification vector of size equal to a predefined number of classes, each component of the classification vector having a value representative of a probability of the input digital signal belonging to a corresponding class, the method being implemented by a calculation processor of a programmable electronic device. This method comprises, for an input signal to be classified, steps of:

[0012] -calculation of a score of the input digital signal to be classified, said score being representative of a probability of the input digital signal belonging to a statistical distribution of the learning signals,

[0013] -comparison of the score to a predetermined threshold, and in the event of said threshold being exceeded, and

[0014] - emission of an alarm to indicate the presence of a digital input signal supposedly corrupt.

[0015] Advantageously, the proposed classification method comprises a score calculation making it possible to detect whether an input signal to be classified is part of a statistical distribution of all the learning signals, and depending on this score, if the input signal to be classified is assumed to be corrupted, an alarm is emitted.

[0016] The signal classification method according to the invention can also present

[0017]

[0018]

[0019]

[0020]

[0021]

[0022]

[0023]

[0024]

[0025]

[0026]

[0027]

[0028] one or more of the characteristics below, taken independently or in all technically conceivable combinations. The classification model implements a neural network transforming a digital input signal into an intermediate vector representative of the signal digital input, and a prediction module implementing a function of prediction associating a classification vector with each intermediate vector. The calculation of the score implements a gradient of the logarithm of a pro density reliability of the input digital signal. The score is calculated by the following formula: VJogfi^exp^X)[y]) ) Where Vx is the gradient operator, f ( ) is the parameterized function implemented by the neural network, being the index component y of the inter vector mediate when the input digital signal is signal X. The method further comprises, in the event that said threshold is exceeded, steps of: - transformation of the input signal into a transformed signal, the transformation comprising an application of an iterative method of statistical sampling of said probability density of the set of training signals, and providing the transformed signal as input to the classification model. Advantageously, the input signal to be classified is modified by transformation into a transformed signal that belongs to the statistical distribution of the set of training signals. Consequently, the classification performance is advantageously improved, including for input signals potentially corrupted by an adversary attack. The iterative statistical sampling method is a Langevin sampling method. The transformation step further comprises, after application of the iterative statistical sampling method to obtain a first transformed signal, a projection of the first transformed signal to obtain a second transformed signal, the first transformed signal being represented in the form of a first vector, the second transformed signal being represented in the form of a second vector, the projection comprising a calculation of the second vector closest to the first vector according to a predetermined metric and belonging to a set of vectors at a predetermined distance, according to said metric, from an input vector representative of the digital input signal. The metric is for example the infinite norm. The predetermined threshold is obtained in advance based on at least part of the signals from the training set and simulated corrupted signals

[0029] According to another aspect, the invention relates to a device for classifying digital signals by implementing a classification model trained by machine learning on a set of training signals, the classification model providing, for an input digital signal, a classification vector of size equal to a predefined number of classes, each component of the classification vector having a value representative of a probability of the input digital signal belonging to a corresponding class. This device comprises a calculation processor configured to implement, for an input signal to be classified:

[0030] - a module for calculating a score of the digital input signal to be classified, said score being representative of a probability of the digital input signal belonging to a statistical distribution of the learning signals,

[0031] - a module for comparing the score to a predetermined threshold, and in the event that said threshold is exceeded,

[0032] - a module for transmitting an alarm in order to indicate the presence of a signal input digital assumed to be corrupted.

[0033] According to a variant, the device is configured to implement, in addition:

[0034] - a module for transforming the input digital signal into a transformed signal, the transformation comprising an application of an iterative method of statistical sampling of said probability density of the training data set,

[0035] - a module for providing the transformed signal as input to the classification model.

[0036] According to one characteristic, the module for transforming the input digital signal into a transformed signal is further configured to carry out, after application of the iterative statistical sampling method to obtain a first transformed signal, a projection of the first transformed signal to obtain a second transformed signal, the first transformed signal being represented in the form of a first vector, the second transformed signal being represented in the form of a second vector, the projection comprising a calculation of the second vector closest to the first vector according to a predetermined metric and belonging to a set of vectors at a predetermined distance, according to said metric, from an input vector.

[0037] According to another aspect, the invention relates to an information recording medium, on which are stored software instructions for the execution of a signal classification method as briefly described above, when these instructions are executed by a programmable electronic device.

[0038] According to another aspect, the invention relates to a computer program comprising software instructions which, when implemented by an electronic device programmable electronics, implement a signal classification method as briefly described above.

[0039] Other characteristics and advantages of the invention will emerge from the description given below, for information purposes only and in no way limiting, with reference to the appended figures, among which:

[0040] [Fig-1] [Fig.l] illustrates a system comprising a classification device according to an embodiment;

[0041] [Fig.2] [Fig.2] schematically represents a neural network classification model and a classification vector obtained as output;

[0042] [Fig.3] [Fig.3] is a block diagram of the main steps of a method for classifying digital signals according to one embodiment;

[0043] [Fig.4] [Fig.4] is a block diagram of the steps of a statistical sampling method implemented by the digital signal classification method according to one embodiment.

[0044] [Fig.l] schematically represents a system 2 implementing a device 4 for classifying digital signals, called digital input signals, by implementing a classification model trained by machine learning.

[0045] For example, in one embodiment, the input digital signals are input digital images, the method then being applicable for the automatic classification of digital images.

[0046] The system 2 is for example a system for guiding autonomous or semi-autonomous vehicles, and also comprises a device 6 for guiding a vehicle (not shown), which is in communication via a communication link 5, wired or wireless, with the device 4 for classifying input signals, more simply called a classification device.

[0047] The classification device 4 is a programmable electronic device, e.g. a computer, and comprises, in one embodiment, one or more processors 8, an electronic memory unit 10, an input / output interface 12 and a communication interface 14, these elements being configured to communicate with each other via a communication bus 15 internal to the device 4.

[0048] The electronic memory unit 10 is configured to store a classification model 20 previously trained by machine learning on a set of learning signals, for example on a database of digital learning images.

[0049] According to a variant, the previously trained classification model is stored and executed on a remote server (not shown), or on a set of interconnected servers of the server cloud type (or “cloud” in English) accessible via a communication link.

[0050] In one embodiment, the classification model implements a neural network, for example a convolutional neural network known by the acronym CNN (for “convolutional neural network”).

[0051] In a known manner, a neural network comprises an ordered succession of layers of neurons, each of which takes its inputs from the outputs of the previous layer.

[0052] More precisely, each layer comprises neurons taking their inputs from the outputs of the neurons of the previous layer, or from the input variables for the first layer.

[0053] Alternatively, more complex neural network structures can be envisaged with a layer that can be connected to a layer further away than the immediately preceding layer.

[0054] Each neuron is also associated with an operation, i.e. a type of processing, to be carried out by said neuron within the corresponding processing layer.

[0055] Each layer is connected to the other layers by a plurality of synapses. A synaptic weight is associated with each synapse, and each synapse forms a connection between two neurons. It is often a real number, which takes both positive and negative values. In some cases, the synaptic weight is a complex number.

[0056] Each neuron is capable of performing a weighted sum of the value(s) received from the neurons of the previous layer, each value then being multiplied by the respective synaptic weight of each synapse, or link, between said neuron and the neurons of the previous layer, then applying an activation function, typically a non-linear function, to said weighted sum, and delivering at the output of said neuron, in particular to the neurons of the following layer which are connected to it, the value resulting from the application of the activation function. The activation function makes it possible to introduce a non-linearity into the processing carried out by each neuron. The sigmoid function, the hyperbolic tangent function, the Heaviside function are examples of activation functions.

[0057] As an optional addition, each neuron is also capable of applying, in addition, an additive factor, also called bias, before the application of the activation function, and the value delivered at the output of said neuron is the image by the activation function of the sum of the bias value and the weighted sum previously carried out.

[0058] In a convolutional neural network, each neuron in the same layer has exactly the same connection pattern as its neighboring neurons, but at different input positions. The connection pattern is called a convolution kernel or, more often, a "kernel" in reference to the corresponding English term.

[0059] A fully connected layer of neurons is a layer in which the neurons of said layer are each connected to all the neurons of the previous layer.

[0060] Such a type of layer is more often referred to as “fully connected” and sometimes referred to as “dense layer”.

[0061] The classification device 4 is configured to perform a robust classification of input digital signals, which are typically digital images 22, a digital image being for example represented in the form of a matrix of pixels. The digital images are for example captured by digital image capture devices external to the classification device.

[0062] For example, in the case of application in a system 2 for guiding an autonomous vehicle, the digital images are provided by one or more capture devices arranged on the vehicle, or external to the vehicle.

[0063] The digital images 22 are temporarily stored in the electronic memory unit 10.

[0064] The processor 8 of the signal classification device 4 is configured to execute:

[0065] - a module 24 calculating a score of the input digital signal to be classified, the score being representative of a probability of the input digital signal belonging to a statistical distribution of the learning signals;

[0066] - a module 26 for comparing the score to a predetermined threshold,

[0067] - a module 28 for transmitting an alarm, for example to the device 6, in order to indicate, in the event of the threshold being exceeded, the presence of a digital input signal presumed to be corrupted.

[0068] The processor 8 is also configured to execute, in one embodiment:

[0069] - a module 30 for transforming the input digital signal into a transformed signal, the transformation comprising an application of an iterative method of statistical sampling of said probability density of the learning data set, and optionally a projection,

[0070] - a module 32 for supplying the transformed signal as input to the classification model.

[0071] In one embodiment, the modules 24, 26, 28, 30, 32 are implemented in the form of software instructions forming a computer program, which, when executed by a programmable electronic device, implements a method for classifying digital input signals as described.

[0072] In a variant not shown, the modules 24, 26, 28, 30, 32 are each produced in the form of programmable logic components, such as FPGAs (Field Programmable Gate Arrays), microprocessors, GPU components (GPUs) General-purpose processing on graphics processing), or even dedicated integrated circuits, such as ASICs (from the English Application Specific Integrated Circuit).

[0073] The computer program comprising software instructions is further capable of being recorded on a non-transitory, computer-readable information recording medium. This computer-readable medium is, for example, a medium capable of storing electronic instructions and of being coupled to a bus of a computer system. For example, this medium is an optical disk, a magneto-optical disk, a ROM memory, a RAM memory, any type of non-volatile memory (for example EPROM, EEPROM, FLASH, NVRAM), a magnetic card or an optical card.

[0074] An embodiment of a method for classifying digital signals will be described below with reference to Figures 2 and 3.

[0075] The method implements a classification model (or classifier) ​​of digital signals, e.g. digital images, previously trained by machine learning on a set of training signals (or training data set) to classify each input signal according to a predefined number K of classes, K being an integer greater than or equal to 2.

[0076] At the output of the classification model, a classification vector is obtained, of size K equal to the number of classes, each component of the classification vector having a value representative of a probability of belonging of the input digital signal to the corresponding class.

[0077] In one embodiment, as schematically illustrated in [Fig.2], the classification model 20 comprises for example a convolutional neural network (CNN) 34, parameterized by a set of parameters whose values ​​are calculated during the preliminary training phase, the neural network 34 providing as output an intermediate vector, Vint, of size K, representative of the input digital signal. The classification model 20 also comprises a prediction module 36 implementing a prediction function associating the corresponding classification vector with each intermediate vector.

[0078] Considering that the neural network 34 is characterized by a parameterized function noted f X being a digital input signal, the intermediate vector Vint(X) obtained has components noted VinJ [ y ] = f ( X ) [ y ], 1S y < K, y being a class index among the K classes.

[0079] In one embodiment, the prediction module 36 applies the normalized exponential function, also called the Softmax function, which converts a vector of K real numbers into a probability distribution over K choices. Thus, the corresponding classification vector Vclass obtained has the following components:

[0080] [Math.l]

[0081] In other words, each component of the classification vector Vclass is a conditional probability associated with the index class y knowing the input signal X (or probability of the label y for the input data X), which is also written according to the mathematical formula expressed above as a function of the components of the intermediate vector Vint:

[0082] [Math.2]

[0083]

[0084]

[0085]

[0086] [Fig.3] is a block diagram of the main steps of a method for classifying digital signals according to one embodiment. A digital input signal Eo is provided as input to the process. The method then comprises a step 40 of classifying the input digital signal Eo, by applying a previously trained digital signal classification model (or classifier), as described above. The method then comprises a step 42 of calculating a score relating to the classification of the digital input signal Eo, the calculated score is representative of a probability of the digital input signal belonging to a statistical distribution of the learning signals of the parameterized function y Q.

[0087] In other words, the score is calculated based on a probability density of all the learning signals.

[0088] Preferably, the calculation of the score is performed by applying a metric known as approximate mass, as described in the article “Your classifier is secretly an energy based model and you should treat it like one” by W. Grathwohl et al, published in ICLR 2020.

[0089] The formula for calculating the score is:

[0090] [Math.3] ^(^o) = V.Jog^Eo) ) = V jexp^ [y]))

[0091] Where VA is the gradient operator, and Q j is the logarithm of the probability density "A, also called log likelihood, of the digital signal Eo, also called example Eo.

[0092] Indeed, as shown in the article “Your classifier is secretly an energy based model and you should treat it like one” by W. Grathwohl et al, published in ICLR 2020, the following formula applies, considering [MATH 2]:

[0093] [Math.4] P g (x ) v) = Z((f)

[0094] Where p is the probability of signal X belonging to a statistical distribution of the learning signals of the parameterized function y Q, also called the true semblance of X, and Z(0) is a constant.

[0095] Given the difficulty of calculating the constant Z(0), the derivative of the log likelihood is used, as shown in the formula [MATH 3]: the constant Z(0) is then eliminated.

[0096] The score S(E0) thus calculated, or the approximate mass of Eo, is representative of the membership of the signal Eo to a statistical distribution of the learning signals of the parameterized function f A, and more generally of a probability that the signal Eo J fh) or so-called healthy data, not corrupted by an adverse attack.

[0097] According to the score calculation formula applied, the higher the score S(E0), the higher the probability that the signal Eo is uncorrupted.

[0098] The method then comprises a comparison 44 of the score calculated in step 42 with a previously determined threshold r.

[0099] For example, the threshold r is calculated on a set of validation signals distinct from the set of training signals, so as to maximize the classification accuracy.

[0100] According to a variant, the threshold r is calculated on a set of signals comprising at least part of the signals of the training set and the simulated corrupted signals.

[0101] If the calculated score S(E0) is greater than the previously determined threshold r, an alarm transmission 46 is implemented. For example, alarm information is transmitted to an external device, for example to the device 6, in order to indicate the presence of an input signal presumed to be corrupted.

[0102] If the calculated score S(Eo) is lower than the threshold r, the comparison step 44 is followed by the classification of a new input signal.

[0103] If the calculated score S(Eo) is greater than the threshold r, the method further comprises, according to one embodiment, a transformation 48 of the digital input signal Eo into a transformed signal Er, so as to transform the input signal Eo into a “healthy” signal Er, the classification of which by the previously trained classification model provides a correct result.

[0104] Preferably, the transformation 48 comprises a step 50 of iterative sampling by the Langevin sampling method, which is a statistical sampling according to said probability density n A.

[0105] In an embodiment illustrated in [Fig.4], the iterative sampling 50 comprises, starting from an initial signal Xo (also called initial sample) the following steps: initialize 52 an iteration counter t at t=0, then calculate 54 the sample at t+1 by the following formula, expressed on a signal Xt+i as a function of the signal XJa, the formula applying to each component of the respective signals:

[0106] [Math.5] =X, + fv. ( log(p e (X,) ) + ^5Z,

[0107] Which is also written:

[0108] [Math.6] X l+l ^X t + ^S(X t )+ylôZ !

[0109] Where ô is a parameter representing the step of the gradient rise, and Zt is a reduced centered Gaussian (i.e. with zero mean and variance equal to 1).

[0110] For example, the value of the parameter ô is between 1 / 256 and 8 / 256 for a digital image in which each pixel is coded on 8 bits.

[0111] It is then checked in verification step 56 whether a stopping condition is validated. For example, the stopping condition relates to the number of iterations, and the iteration counter t is then compared to a maximum number Nmax of iterations. For example, Nmax is greater than or equal to 1000, and for example equal to 100000 depending on the complexity of the signals processed. For example, for digital images, the larger the images, the higher Nmax. For example, MNIST images (black and white, 28x28 pixels) will not require as many iterations as a set of digital images such as cifarlO (RGB color images, 32x32 pixels).

[0112] If the maximum number Nmax of iterations is reached, then the method continues with step 60 described below.

[0113] If the maximum number Nmax of iterations is not reached, then the method continues with an update 58 of the iteration counter (i.e. the value of t is increased by one), and a return to step 54 of calculating a next sample.

[0114] Thus, at the end of the iterative sampling step 50, the input signal is transformed into a first transformed signal.

[0115] Returning to [Fig.3], the transformation step 48 further preferably comprises a step 60 of projecting the first transformed signal to obtain a second transformed signal.

[0116] According to one embodiment, the projection step 60 implements a mathematical projection according to the infinite norm on a ball of radius e around the input signal Eo, represented in the form of an input vector Vo. For example, the value of the parameter e is between 1 / 256 (0.0039) and 8 / 256 (0.031) for a digital image in which each pixel is coded on 8 bits.

[0117]

[0118]

[0119]

[0120]

[0121]

[0122]

[0123]

[0124]

[0125]

[0126]

[0127]

[0128] The first transformed signal being represented in the form of a first vector Vi and the second transformed signal also being represented in the form of a second vector V2, the projection is written: [Math.7] V^projiV^ The following relationship is verified: [Math. 8] said^V^ vQ < e Where dist() is the distance according to the chosen metric. For the so-called infinite norm metric, each of the vectors having components {^¢, 1 £ q £ TV) and {V2, q , 1SÇSJV} , the relation [MATH 8] is written: [Math.9] max| V2, q - q\ < e Alternatively, norms other than the infinite norm can be used, for example the Euclidean norm (or L2 norm) or the L1 norm (sum of the absolute values ​​of the components of the vector). The second transformed signal is then again provided as input to the classifier at classification step 40, and a correct classification is then performed. The projection step 60 advantageously makes it possible to obtain a second transformed signal which is closer, in the sense of a chosen metric, to the input signal. In particular, for image signals, this results in perceptual proximity. Indeed, statistical sampling provides, in many cases, a first transformed signal which is perceptually distant from the input signal. However, adversarial attacks consist of generating corrupted signals perceptually close to the initial signals, precisely so as to conceal the attack for a human observer. Thus, a transformed signal perceptually distant from the input signal is not a good candidate as a corresponding healthy signal. The projection step makes it possible to remedy this drawback of statistical sampling and to provide a second transformed signal, obtained from the first transformed signal and perceptually closer to the input signal than the first transformed signal. The invention has been described above in the case of a classification model implementing implements a convolutional neural network and a prediction module, but it is not limited to this application case. The invention also applies with a so-called hybrid classification model, consisting of a generative model and a classifier, when the training function of the hybrid model is modified as proposed in the article “Improving Normalizing Flows With the Approximate Mass for Out-of-Distribution Detection” by Samy Chali et al, published at the CVPR Workshop 2023.

[0129] Advantageously, for several types of classification models, the invention makes it possible to detect corrupted input signals, while the corruption applied by a malicious third party is not perceptible to a human, and also makes it possible to transform such corrupted input signals into healthy signals which are then correctly classifiable by the classification model.

[0130] Thus, advantageously, the applied opposing attack is countered.

Claims

Claims

1. Method for classifying digital signals by implementing a classification model trained by machine learning on a set of training signals, the classification model providing, for an input digital signal, a classification vector of size equal to a predefined number of classes, each component of the classification vector having a value representative of a probability of the input digital signal belonging to a corresponding class, the method being implemented by a calculation processor of a programmable electronic device and being characterized in that it comprises, for an input signal to be classified, steps of: -calculating (42) a score of the input digital signal to be classified, said score being representative of a probability of the input digital signal belonging to a statistical distribution of the training signals, -comparing (44) the score with a predetermined threshold,and in the event of said threshold being exceeded, and - emission (46) of an alarm in order to indicate the presence of a digital input signal assumed to be corrupted.,

2. Method according to claim 1, in which the classification model implements a neural network (34) transforming an input digital signal into an intermediate vector representative of the input digital signal, and a prediction module (36) implementing a prediction function associating with each intermediate vector, a classification vector.

3. The method of claim 2, wherein the calculation (42) of the score implements a gradient of the logarithm of a probability density of the input digital signal.

4. Method according to claim 3, wherein said score is calculated by the following formula: Vxlog^^exp^tX)[y]) ) Where V x is the gradient operator, f ( ) is the parameterized function implemented by the neural network, f being the index component y of the intermediate vector when the input digital signal is the signal X.

5. A method according to any one of claims 1 to 4, comprising in furthermore, in the event of exceeding said threshold, steps of: - transformation (48) of the digital input signal into a transformed signal, the transformation comprising an application of an iterative statistical sampling method according to said statistical distribution of the set of learning signals, - supplying the transformed signal as input to the classification model.

6. The method of claim 5, wherein the iterative statistical sampling method is a Langevin sampling method.

7. A method according to any one of claims 5 or 6, wherein the transforming step (48) further comprises, after applying the iterative statistical sampling method to obtain a first transformed signal, a projection (60) of the first transformed signal to obtain a second transformed signal, the first transformed signal being represented in the form of a first vector, the second transformed signal being represented in the form of a second vector, the projection (60) comprising a calculation of the second vector closest to the first vector according to a predetermined metric and belonging to a set of vectors at a predetermined distance, according to said metric, from an input vector.

8. The method of claim 7 wherein said metric is the infinite norm.

9. Method according to one of claims 1 to 8, wherein said predetermined threshold is obtained beforehand as a function of at least a part of the signals of the training set and of the simulated corrupted signals.

10. A computer program comprising software instructions which, when executed by a programmable electronic device, implement a signal classification method according to claims 1 to 9.

11. Device for classifying digital signals by implementing a classification model trained by machine learning on a set of training signals, the classification model providing, for an input digital signal, a classification vector of size equal to a predefined number of classes, each component of the classification vector having a value representative of a probability of the input digital signal belonging to a corresponding class, the device being characterized in that it comprises a calculation processor configured to implement, for an input signal to be classified: - a calculation module (24) for calculating a score of the digital input signal to be classified, said score being representative of a probability of the digital input signal belonging to a statistical distribution of the learning signals, - a comparison module (26) of the score with a predetermined threshold, and in the event of said threshold being exceeded, - a module (28) for transmitting an alarm in order to indicate the presence of a supposedly corrupted digital input signal.

12. Device according to claim 11, configured to implement, in addition: - a module (30) for transforming the input digital signal into a transformed signal, the transformation comprising an application of an iterative method of statistical sampling of said probability density of the training data set, - a module (32) for providing the transformed signal as input to the classification model.

13. Device according to claim 12, in which the module for transforming the input digital signal into a transformed signal is further configured to carry out, after application of the iterative statistical sampling method to obtain a first transformed signal, a projection of the first transformed signal to obtain a second transformed signal, the first transformed signal being represented in the form of a first vector, the second transformed signal being represented in the form of a second vector, the projection comprising a calculation of the second vector closest to the first vector according to a predetermined metric and belonging to a set of vectors at a predetermined distance, according to said metric, from an input vector.

Images