FR3137475A1 - Method and device for estimating the authenticity of audio or video content and associated computer program - Google Patents

Abstract

The present invention relates to a method for estimating the authenticity of audio or video content, the audio or video content being represented by a set of input values, the method comprising steps of: - determination, by means of 'a processing system (5) and on the basis of said set of input values, an intermediate vector (w) in accordance with a learned distribution, the processing system (5) being configured, beforehand, to produce, at output, output vectors distributed according to the learned distribution and designed so that a generator network produces content of the same type as the audio or video content when said output vectors are applied as input to this generator neural network, and - estimation of a level of authenticity (m) of the audio or video content by application of classification means (20) to said intermediate vector (w). The invention also relates to a device for estimating the authenticity of audio or video content and an associated computer program. Figure for abstract: Fig. 1

Description

Method and device for estimating the authenticity of audio or video content and associated computer program Technical field of the invention

The present invention relates to the technical field of determining the authenticity of data, and in particular the authenticity of audio or video content.

It relates in particular to a method and a device for estimating the authenticity of audio or video content and an associated computer program.

State of the art

Hyperfaking (or “ deepfake ” according to the term of Anglo-Saxon origin commonly used) is a technique which consists of modifying existing data, for example video data, to manipulate it and associate with this data a message different from that which was initially associated.

This technique is particularly known for reproducing a person's voice and making them say invented things. The detection of manipulated data therefore constitutes a major challenge.

For example, it is known to use artificial neural network structures to detect possible falsified data.

The use of these artificial neural network structures makes this detection very effective. However, these artificial neural network structures are often trained on data manipulated in a certain way (for example, with the use of a particular compression method for video data). Thus, as soon as the structural elements of the analyzed data (for example the shape or resolution of the data) change, the performance of the artificial network structures used is greatly reduced.

The article “ ID- Reveal : Identity- aware Deepfake Video Detection ” by D. Cozzolino, A. Rössler, J. Thies, M. Nießner and L. Verdoliva, 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021 , pp. 15088-15097, doi: 10.1109/ICCV48922.2021.01483 describes an implementation of artificial neural network structures making it possible to overcome these changes in the structural elements of the analyzed data.

In this document, the neural network structures are not trained to detect traces of data manipulation but to analyze facial expressions and movements present in the images considered. Thus, by analyzing facial expressions and movements and comparing them to expressions known to be authentic, the relevant neural network structure detects a possible falsification of the initial image (or video).

However, the implementation of such a neural network structure is very cumbersome and requires the use of large data formats. Furthermore, knowledge of authentic data for the analyzed faces is necessary to successfully identify possible falsifications of the expressions and movements of these faces.

Presentation of the invention

In this context, the present invention proposes to improve the determination of the authenticity of audio or video content.

More particularly, according to the invention, we propose a method for estimating the authenticity of audio or video content, the audio or video content being represented by a set of input values, the method comprising steps of:

- determination, by means of a processing system and on the basis of said set of input values, of an intermediate vector in accordance with a learned distribution, the processing system being configured, beforehand, to produce, as output, output vectors distributed according to the learned distribution and designed so that a generator network produces content of the same type as the audio or video content when said output vectors are applied to the input of this generator network, and

- estimation of a level of authenticity of the audio or video content by application of classification means to said intermediate vector.

Thus, the use of an intermediate vector, which can have a reduced dimension compared to the dimension of the initial audio or video content, makes it possible to improve simplicity and analysis time. Furthermore, the use of this intermediate vector does not reduce the robustness of the level of authenticity obtained because this intermediate vector, thanks to the use of the learned distribution, only includes the relevant parameters necessary for determining this level of authenticity. 'authenticity.

Other non-limiting and advantageous characteristics of the process according to the invention, taken individually or in all technically possible combinations, are as follows:

- it is also provided, prior to the estimation step, a preliminary process for learning the distribution comprising steps of:

- supply, at the input of a redistribution network, of data distributed according to a random distribution, so as to obtain, at output, vectors distributed according to a distribution (initial or current), said vectors being supplied at the input of the network generating so as to provide, as output, a set of learning values, and

- training of the redistribution network and the generator network so as to update said distribution (with a view to optimizing distribution);

- it is also planned, during the determination stage, sub-stages of:

- supply, at the input of a generator network, of said intermediate vector so as to obtain, at the output of said generator network, a generated set of values, and

- updating said intermediate vector so as to optimize a cost function representing a distance between the set of input values and said generated set of values,

said updated intermediate vector being used as an intermediate vector during the estimation step;

- the updating step is implemented by a gradient descent method;

- it is planned, prior to the estimation stage, stages of:

- supply, at the input of the generator network, of said updated intermediate vector so as to obtain, at the output of said generator network, an updated generated set of values, and

- new updating of said intermediate vector so as to optimize a cost function representing a distance between the set of input values and said updated generated set of values;

- the determination step comprises supplying, at the input of the processing system, said set of input values so as to obtain, at the output of the processing system and on the basis of the learned distribution, said intermediate vector;

- the set of input values comprises a first number of values and the intermediate vector comprises a second number of values, the second number of values being strictly less than the first number of values;

- the second number of values is between one hundredth and one thirtieth of the first number of values;

- the second number of values is less than or equal to 512;

- there is provided, upstream of the estimation step, a step of training the classification means from a plurality of intermediate vectors each associated with authentic or falsified audio or video content in such a way that the estimation of the level of authenticity for an intermediate vector associated with authentic audio or video content gives a first result and that the estimation of the level of authenticity for an intermediate vector associated with falsified audio or video content gives a second result distinct from the first result;

- the set of input values is obtained by extracting part of the audio or video content; And

- the set of input values is made up of values associated with pixels of an image.

The present invention also relates to a device for estimating the authenticity of audio or video content, the audio or video content being represented by a set of input values, the device comprising:

- a processing system configured to determine, on the basis of said set of input values, an intermediate vector in accordance with a learned distribution, the processing system being configured to produce, as output, output vectors distributed according to the learned distribution and designed so that a generator network produces content of the same type as the audio or video content when said output vectors are applied to the input of this generator network, and

- a module for estimating a level of authenticity of the audio or video content by applying classification means to said intermediate vector.

The present invention also relates to a computer program comprising instructions executable by a processor and designed to implement an estimation method as introduced previously when these instructions are executed by the processor.

Of course, the different characteristics, variants and embodiments of the invention can be associated with each other in various combinations as long as they are not incompatible or exclusive of each other.

Detailed description of the invention

In addition, various other characteristics of the invention emerge from the appended description made with reference to the drawings which illustrate non-limiting forms of embodiment of the invention and where:

represents a set of elements brought into play as part of the method for estimating a level of authenticity in accordance with the present invention,

represents a set of artificial neural networks brought into play as part of a preliminary learning method in accordance with the present invention,

represents, in functional form, a device for estimating a level of authenticity configured to implement a method for estimating a level of authenticity of audio or video content in accordance with the invention,

represents, in flowchart form, an example of a method for estimating a level of authenticity of audio or video content in accordance with the present invention,

represents, in flowchart form, an example of a preliminary learning method usable in the context of the invention, and

represents, in flowchart form, an example of a preliminary training method usable in the context of the invention.

There represents an example of a set 1 of elements involved in the present invention. This set 1 here comprises a processing system 5 and classification means 20.

This set 1 of elements is configured to process input data. This input data is for example in the form of images. As will be seen later, these images come for example from audio or video content.

Below, we note that each image is represented, using a plurality of pixels, by a plurality of components for each pixel. Generally, each image is represented by three color components for each pixel (a red component R, a green component G and a blue component B). In other words, each image is represented by a set of values associated with the different pixels forming it.

The processing system 5 here comprises an optimization block 10, a transmission block 12 and a generator network 34.

The processing system 5 has the characteristics of an encoder, that is to say it is configured to condense the information provided to it as input.

In practice, from input data, here from an image Im provided as input to the processing system 5 ( ) that is to say a set of input values associated with it, this processing system 5 is configured to provide, at output, an intermediate vector w of reduced size compared to the set of values entry.

In other words, considering that the set of input values includes a first number of values and that the intermediate vector includes a second number of values, the second number of values is here strictly less than the first number of values. For example here, the second number of values is between one hundredth and one thirtieth of the first number of values. For example, the second number of values is less than 512.

Thus, the processing system 5 provides, as output, an intermediate vector w represented in a particular space called “latent space”. The intermediate vector w is also called a “latent vector”. This latent vector therefore corresponds to a new representation of the image provided as input to the processing system 5. This new representation retains the most relevant information of the image Im provided as input, that is to say those which characterize the best this image Im, with a much lower dimension than that of the image Im.

Generally speaking, the processing system 5 implements an inversion function associated with a generation function.

This inversion function is for example implemented via the processing system 5 using the optimization block 10, according to optimization steps described below. The inversion function is for example of the type presented in the article “ Image2StyleGAN+ +: How to Edit the Embedded Images? » by Abdal R., Qin Y. and Wonka P., 10.48550/ARXIV.1911.11544, 2019.

The generation function is implemented by the generator network 34 described below.

As shown on the , the set 1 of elements also includes the classification means 20. These classification means 20 are based for example here on the use of an algorithm commonly called "decision tree forest" (or " Random Forest " according to the commonly used Anglo-Saxon designation of origin). More details on decision tree forests can be found in the article " Random forests ", by Breiman L., in Machine learning , 45(1), 5-3, 2001.

So, as shown in , from the intermediate vector w received as input, the classification means 20 estimate a level m of authenticity of the set of input values associated with the image Im. This level m of authenticity takes for example a first result m ₁ when the set of input values is associated with authentic audio or video content. For example, it takes a second result m ₂ when the set of input values is associated with falsified audio or video content. The second result m ₂ is distinct from the first result m ₁ .

Alternatively, the classification means may comprise an artificial neural network configured to receive, as input, the intermediate vector w and provide, at output, an estimate of the level of authenticity of the audio or video content associated with the intermediate vector w (provided entrance).

Advantageously according to the invention, the processing system 5 is configured to determine the intermediate vector w in accordance with a distribution p _w learned beforehand. This distribution p _w is learned before implementing the estimation method according to the invention, according to a preliminary learning method described below.

This preliminary learning process (which will be described later with reference to the ) is implemented via a structure 35 of artificial neural networks. This structure 35 of artificial neural networks is represented on the .

As shown in the , the structure 35 of artificial neural networks comprises a so-called redistribution network 32 (or “mapping network” according to the Anglo-Saxon designation of origin) and the generator network 34 (included in the processing system 5 introduced previously).

The redistribution network 32 and the generator network 34 are for example of the type used in StyleGAN generative adversarial networks. More details on StyleGAN networks can be found in the article “ Analyzing and Improving the Image Quality of StyleGAN ” by Karras T., Laine S., Aittala M., Hellsten J., Lehtinen J. and Aila T. , 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 8107-8116, 2020.

In practice, the redistribution network 32 comprises, for example, eight layers here. The redistribution network 32 is for example of the multilayer perceptron type (here with eight layers).

As shown in , the redistribution network 32 receives, as input, data z distributed according to a random distribution. This redistribution network 32 is then configured to provide, at output, a latent vector u represented in a predetermined latent space.

The generator network 34 is an artificial neural network which here comprises a plurality of layers. Each of these layers is for example a convolution layer combined with a normalization function and a non-linearity function.

In practice, the structure 35 of artificial neural networks is structured via nodes forming the different layers of the redistribution network 32 and the generator network 34. These nodes are characterized by a plurality of weighting coefficients which can be adjusted, prior to the implementation of the process for estimating the level of authenticity itself, thanks to a training phase as described subsequently in the preliminary learning process.

As shown on the , the generator network 34 is configured to receive, as input, the latent vector u obtained at the output of the redistribution network 32 and to provide as output a set of values x (associated with an image).

The present invention relates more particularly to the estimation of the level m of authenticity of audio or video content. For this, the invention also relates to a device 50 for estimating the authenticity of audio or video content. There represents, in functional form, such an estimation device 50 configured to implement the invention.

This estimation device 50 comprises a control unit 52 provided with a processor 54 and a memory 56.

This estimation device 50 is designed to implement a set of modules and functional blocks. For example, it includes an estimation module, the optimization block 10 and the transmission block 12. Each of the different modules and blocks is for example produced by means of computer program instructions stored by the memory 56 of the control unit 52 and designed to implement the module concerned when these instructions are executed by the processor 54 of the control unit 52.

There is a flowchart representing an example of a method for estimating the authenticity of audio or video content implemented in the context described above.

As indicated previously, prior to implementing the estimation process, a preliminary learning process is implemented in order to learn the distribution p _w (which then allows the determination of the intermediate vector w).

There is a flowchart representing an example of a preliminary learning process. In practice, this preliminary learning method makes it possible to determine the weighting coefficients associated with the plurality of nodes of the structure 35 of artificial neural networks.

As shown in this figure, this preliminary method comprises a first step (step E50) of initializing the weighting coefficients of the structure 35 of artificial neural networks to initial values. These initial values are for example random values.

The preliminary learning process then continues in step E52 during which data z distributed according to a random distribution are provided as input to the structure 35 of artificial neural networks. The vector u is obtained at the output of the redistribution network 32. A set of learning values x is then obtained at the output of the neural network structure 35.

The process then continues in step E54 forming a step of learning the redistribution network 32 and the generator network 34. Generally speaking, this step aims to update the weighting coefficients of the structure 35 of artificial neural networks ( more particularly the respective weighting coefficients of the redistribution network 32 and the generator network 34) so that the content obtained at the output of the structure 35 of artificial neural networks comes as close as possible to a reference content (i.e. here reference content which is known to correspond to authentic content). In practice this reference content may include images (verified as authentic) from a database.

This learning step is based for example on the optimization of a so-called progressive adversarial cost function as described in the article “ Progressive Growing of GANs for Improved Quality , Stability , and Variation ” by Karras T., Aila T. , Laine S. and Lehtinen J., doi: 10.48550/ARXIV.1710.10196, 2017.

Finally, the weighting coefficients are modified in such a way as to minimize the cost function, for example over several iterations, until the optimal weighting coefficients are obtained which effectively minimizes the cost function. These optimal weighting coefficients then make it possible to obtain vectors u distributed according to a certain determined distribution p _w (called “learned distribution p _w ”).

Alternatively, optimization of the cost function can be implemented via a predetermined number of iterations. In this case, the weighting coefficients obtained are not necessarily optimal but form a correct approximation also making it possible to obtain vectors u distributed according to a suitable distribution p _w (therefore forming a good approximation of the learned distribution previously described). Such a variant makes it possible in particular to obtain this distribution more quickly.

This learned distribution p _w is then used in the method of estimating the authenticity of audio or video content described below.

The process for estimating the authenticity of audio or video content is therefore implemented following the preliminary learning process.

As shown in the , the estimation process begins in step E2 during which the control unit 52 receives audio or video content. This is audio or video content that we want to determine if it is authentic or if it has been falsified (totally or in part).

To do this, in step E4, the processor 54 determines a set of input values associated with the audio or video content. This set of input values is obtained by extracting part of the audio or video content.

More precisely, in the case of video content (that is to say that this content comprises at least one image, possibly a sequence of images), the latter is broken down into the form of a plurality of images according to well-known methods (by breaking down, for example, the video content into a sequence of a plurality of still shots, at regular time intervals, each then corresponding to an image).

In the case of audio content, the sound components can be transcribed in the form of components of an image as is, for example, described in the article “ Learning and controlling the source- filter representation of speech with a variational autoencoder ” , de Sadok, S., Leglaive, S., Girin, L., Alameda-Pineda, form of spectrograms which are graphs representing frequency versus time. On these spectrograms, a color scale reflects the intensity of each sound component. The intensity of each sound component represented by a color can therefore be used as the pixel value of an image (then corresponding to the image used below).

Thus, whether the initial content received in step E2 is audio or video content, the method of estimating its authenticity is based on input data formed from an image formatted for example in the manner described below. -After.

On each image, the processor 54 first detects the presence of the face of at least one person. This is done, for example, by identifying certain particularities allowing a face to be defined.

Then, once the face is detected on the image, the processor 54 extracts the image around the detected face. It “crops” the image so as to keep only the part of the image including the detected face.

Then, the processor 54 adapts the part of the image comprising the face so as to carry out facial alignment. This then consists of locating characteristic points of the face in order to identify the geometric structure of the face concerned.

More precisely, the processor 54 is based on the elements of the face containing the most semantic information (such as the eyes, the nose or the mouth) in order to determine the geometry of the components of the face. This then makes it possible to reliably represent the “non-rigid” elements of the face. The processor 54 then determines a modified image on which the face is “aligned”.

Finally, at the end of this step E4, the processor 54 breaks down the modified image into a set of input values associated with the different pixels forming it. As defined previously, the set of input values is formed from the values associated with the different pixels of the image (here the modified image).

As shown in the , the estimation process continues in step E6. During this step, the processor 54 provides, as input to the processing system 5, and more particularly as input to the optimization block 10, the set of input values obtained in step E4. This processing system 5 then provides, as output, an initial vector w ₀ . This initial vector w ₀ is for example determined randomly.

Alternatively, in order to accelerate the optimization process described below, the initial vector can be determined in the manner described in the article “ In-Domain GAN Inversion for Real Image Editing ” by Zhu, Jiapeng and Shen, Yujun and Zhao, Deli and Zhou, Bolei, 10.48550/ARXIV.2004.00049, 2020.

In order to obtain the best possible estimate of the level of authenticity of the audio or video content received, the estimation method includes a process of optimizing the intermediate vector described below by steps E8 to E20. The method of optimizing the intermediate vector w is for example here an iterative optimization method by gradient descent.

For this, the estimation method comprises a step E8 of initialization to the value 0 of an index i. This index i designates the current round of this optimization method. During this step, the processor 54 also initializes the intermediate vector. The initial vector w ₀ determined in step E6 is used as initialization value.

In step E10, as is also illustrated in the , for the current round, the processor 54 provides, as input to the generator network 34, the current value of the intermediate vector w _i (for the first round, it is therefore the initial vector w ₀ ). The generator network 34 then provides, as output, a set of generated values corresponding to this current intermediate vector w _i . In other words, the generator network 34 provides, as output, the image Y _i corresponding to the current intermediate vector w _i .

The two sets of values, the set of input values and the set of generated values are then compared, by the optimization block 10, to evaluate whether the current intermediate vector w _i forms a relevant representation of the image initial Im from the audio or video content received in step E2.

More precisely, in step E12, the processor 54 determines a cost function L. This cost function L depends on all of the input values (step E4) and on all of the values generated (step E10). . This cost function makes it possible to quantify the difference between the image received as input (in step E2) and that generated from the current intermediate vector w _i .

In practice, during this step E12, the cost function L is here represented by a distance between the set of input values and the set of generated values.

The cost function L is expressed in the following form:

with Im the image received as input (in step E2), Y _i the image generated from the intermediate vector wi, λ a variable parameter depending on the application, N the total number of pixels in each image Im or Y _i , the notation ||Y _i , Im|| ₂ representing the Euclidean distance between the two elements Im and Y _i and Lpercept a function defined by the following expression:

with the notation vgg ₁₆ corresponding to the application of an artificial neural network of the vgg16 type to the image Im or the image Y _i , the result obtained, for each image being for example an associated vector of 4096 values.

As shown in the , the process continues in step E14 during which the processor 54 evaluates whether the set of values generated makes it possible to optimize the cost function L. For example, the processor 54 evaluates here whether the distance between the set of input values and the set of values generated is sufficiently small. This distance evaluates for example the average squared difference between all the input values and all the generated values. To do this, the processor 54 compares this distance to a predetermined value ε.

If, in step E16, the cost function (that is to say here the distance between the set of input values and the set of values generated) is greater than this predetermined value ε, this means that the generated set of values, and therefore the current intermediate vector w _i concerned, do not form a relevant representation of the image resulting from the audio or video content. The process then continues in step E18. During this step, the processor 54 updates the current intermediate vector w _i on the basis of the determined cost function. More precisely, the updated intermediate vector w _act is given by the following expression:

with w _act , the updated intermediate vector, w _i , the current intermediate vector, η a constant, L(w _i , x) the cost function depending on the set of input values and the generated set of values And the gradient operator.

Then, in step E20, the index i is incremented. The updated intermediate vector w _act becomes the current value of the intermediate vector. A new iteration is then implemented and the process resumes at step E12.

On the other hand, if, in step E16, the cost function L is less than the predetermined value ε, the current intermediate vector w _i can be considered as corresponding to a relevant representation of the input image. The current intermediate vector w _i is therefore used for the rest of the estimation process. The current intermediate vector w _i optimized (and used for the rest of the estimation process) is denoted “intermediate vector w”.

This intermediate vector w is determined in accordance with the learned distribution p _w during the preliminary learning process described above. Thanks to the preliminary learning process described above, the intermediate vector obtained is such that, when this intermediate vector is supplied as input to the generator network 34, the latter provides, at output, content of the same type (i.e. say for example an image representing a person with the same identity, pose and expression as in the reference image).

In other words, the generator network 34 makes it possible to obtain, from an intermediate vector w following the learned distribution p _w , plausible content (compared to reference content which is known to be authentic).

Thus, the processing system 5 as it is designed can be seen as an encoder which makes it possible to produce intermediate vectors distributed according to a certain previously learned distribution p _w . Advantageously, when the generator network 34 is applied to the intermediate vector w, this then makes it possible to find, substantially, the input data.

The transmission block 12 transmits, to the classification means 20, then, in step E22, this intermediate vector w (here under the control of the optimization block 20 when the optimization stopping condition is met) , for the implementation of step E24.

As shown on the , the process continues in step E24 of estimating the level of authenticity of the audio or video content received in step E2.

For this, the processor 54 provides the intermediate vector w, transmitted in step E22, to the classification means 20. The classification means 20 then estimate the level m of authenticity of the set of input values associated with the 'image from the audio or video content received in step E2. The classification means 20 then provide, as output, a first result m ₁ or a second result m ₂ . The first result m ₁ and the second result m ₂ are distinct from each other. Here, the first result m ₁ is interpreted for example as an authentic nature of the audio or video content analyzed while the second result m ₂ means that the audio or video content received in step E2 has been falsified.

The classification means 20 are trained prior to the implementation of step E24 according to a preliminary training method. There is a flowchart representing an example of a preliminary training process. This preliminary training process is for example implemented prior to the implementation of the estimation process (therefore before step E2).

This preliminary training process is implemented for audio or video content whose “authentic” or “falsified” status is known. These contents are, for example, those from the FaceForensics++ database.

As shown in the , this method begins with a step E70 of initializing the node values of the classification means 20 to initial values. These initial values are for example random values.

The preliminary training process then continues in step E72. During this step, the processor 54 receives reference audio or video content whose status (authentic or falsified) is known.

Then, similarly to step E4 described previously, processor 54 breaks down the audio or video content into a plurality of images and determines, for each image, a set of reference input values (step E74).

In step E76, the processor 54 provides, as input to the processing system 5, the set of reference input values. Similar to step E6 described, an intermediate reference vector is determined at the end of this step E76.

This intermediate reference vector is then optimized during steps E78 to E90 in the same way as in steps E8 to E20 described previously.

In step E92, the processor 54 provides, as input to the classification means 20, the (updated) intermediate reference vector.

Then, in step E94, the processor 54 adjusts the node values of the classification means so that the estimation of the level of authenticity for an intermediate reference vector associated with authentic audio or video content gives the first result m ₁ and that the estimation of the level of authenticity for an intermediate reference vector associated with falsified audio or video content gives the second result m ₂ distinct from the first result m ₁ .

Advantageously according to the invention, the use of the intermediate vector, having a considerably reduced dimension compared to the dimension of the image part analyzed, makes it possible to improve the simplicity and analysis time. Furthermore, this reduction in dimension (therefore apparently of information available to carry out the estimation of authenticity) does not reduce the robustness of the level of authenticity obtained because ultimately the intermediate vector only includes the relevant parameters necessary for the determination of this level. For example, the method according to the invention does not take into account information concerning lighting or backgrounds, information which is not useful for determining the authenticity of the content from the faces present in the analyzed image. .

In addition, the neural networks and classification means involved in the invention do not require a costly training phase (in time, in available memory).

The present invention presents a particularly advantageous application for the detection of audio or video content which may have been falsified but which would be broadcast, intentionally, as authentic content. This is particularly advantageous in order to identify content disseminating false information (or “fake news”), to fight against cyberharassment or to fight against certain scams linked to possible blackmail based on falsified content.

Claims (14)

Method for estimating the authenticity of audio or video content, the audio or video content being represented by a set of input values, the method comprising steps of:
- determination (E6), by means of a processing system (5) and on the basis of said set of input values, of an intermediate vector (w) in accordance with a learned distribution (p _w ), the system of processing (5) being configured, beforehand, to produce, as output, output vectors distributed according to the learned distribution (p _w ) and designed so that a generator network (34) produces content of the same type as the content audio or video when said output vectors are applied to the input of this generator network (34), and
- estimation (E22) of a level of authenticity (m) of the audio or video content by application of classification means (20) to said intermediate vector (w). Method according to claim 1, in which there is also provided, prior to the estimation step, a preliminary method of learning the distribution (p _w ) comprising steps of:
- supply (E52), at the input of a redistribution network (32), of data (z) distributed according to a random distribution, so as to obtain, at the output, vectors (u) distributed according to a distribution (p _w ) , said vectors (u) being supplied as input to the generator network (34) so as to provide, at output, a set of learning values (x), and
- drive (E54) of the redistribution network (32) and the generator network (34) so as to update said distribution (p _w ). Method according to claim 1 or 2, in which it is provided, during the determination step (E6), sub-steps of:
- supply (E10), at the input of a generator network (34), of said intermediate vector (w) so as to obtain, at the output of said generator network (34), a generated set of values, and
- updating (E18) of said intermediate vector (w) so as to optimize a cost function representing a distance between the set of input values and said generated set of values,
said updated intermediate vector being used as intermediate vector (w) during the estimation step. Method according to claim 3, in which the updating step (E18) is implemented by a gradient descent method. Method according to claim 3 or 4, in which steps of:
- supply, at the input of the generator network (34), of said updated intermediate vector so as to obtain, at the output of said generator network (34), an updated generated set of values, and
- new updating of said intermediate vector (w) so as to optimize a cost function representing a distance between the set of input values and said updated generated set of values. Method according to any one of claims 1 to 5, in which the determination step comprises supplying, at the input of the processing system (5), said set of input values so as to obtain, at the output of said processing system processing (5) and on the basis of the learned distribution (p _w ), said intermediate vector (w). Method according to any one of claims 1 to 6, in which the set of input values comprises a first number of values and the intermediate vector (w) comprises a second number of values, the second number of values being strictly lower to the first number of values. Method according to claim 7, wherein the second number of values is between one hundredth and one thirtieth of the first number of values. Method according to claim 7 or 8, wherein the second number of values is less than or equal to 512. Method according to any one of claims 1 to 9, in which there is provided, upstream of the estimation step (E22), a step of training the classification means (20) from a plurality of intermediate vectors each associated with authentic or falsified audio or video content in such a way that the estimation of the level of authenticity (m) for an intermediate vector (w) associated with authentic audio or video content gives a first result and that the The estimation of the level of authenticity (m) for an intermediate vector (w) associated with falsified audio or video content gives a second result distinct from the first result. Method according to any one of claims 1 to 10, wherein the set of input values is obtained by extracting part of the audio or video content. Method according to any one of claims 1 to 11, in which the set of input values is formed of values associated with pixels of an image. Device (50) for estimating the authenticity of audio or video content, the audio or video content being represented by a set of input values, the device (50) comprising:
- a processing system (5) configured to determine, on the basis of said set of input values, an intermediate vector (w) in accordance with a learned distribution (p _w ), the processing system (5) being configured to produce , at output, output vectors distributed according to the learned distribution (p _w ) and designed so that a generator network (34) produces content of the same type as the audio or video content when said output vectors are applied as input of this generator network (34), and
- a module for estimating a level of authenticity (m) of the audio or video content by application of classification means (20) to said intermediate vector (w). Computer program comprising instructions executable by a processor (54) and designed to implement a method according to one of claims 1 to 12 when these instructions are executed by the processor (54).