Method for automatically detecting digital forgery

Abstract

The invention relates to a method (100) for automatically detecting digital forgery for the purpose of verifying the conformity of a digital copy (DD) of an official document (OD). Such a method (100) primarily comprises a step (130) of processing one or more images (DF1, …, DFj, …DFm) extracted from the digital copy (DD) in order to produce one or more images (DF1', …, DFj', …DFm') expressing a residual noise, and a step (140) of analysing this residual noise in order to produce a conformity result (SCj, ASC) in relation to the digital copy (DD). Such a method (100) is designed to be implemented by the processing unit of an electronic entity (20) for verifying such conformity, possibly in response to a request from a third-party electronic entity (40) operated by a state or service provider.

Description

[0001] Automatic method for detecting digital forgery

[0002] The invention relates to the field of detecting digital forgery or more generally any attempt to fabricate an image or video containing one or more forged parts, i.e. added before broadcast or injected in real time, so as to give the impression that said image or video is authentic.

[0003] The invention finds its application particularly in the context of remote verification of an official document (identity card, certificate, etc.) and fraud detection (also known by the acronym DVFD, from the English expression "Document Verification and Fraud Detection"), that is, in the absence of a local agent physically present to judge the authenticity of said official document. To this end, a user is asked to submit a digital document resulting from a capture of the official document to a remote electronic verification entity responsible for confirming or rejecting the conformity of the digital copy with the official document.Such a presentation is made using a personal electronic device, for example, a smartphone, connected via a wired (such as the internet) or wireless (such as the mobile phone network) connection to the electronic verification entity. When the latter confirms the conformity of the presented digital copy, the user of the personal electronic device is considered to have proven their identity or demonstrated their status and may claim to benefit from a service or access to a physical, electronic, or digital resource.

[0004] Such a remote verification procedure therefore requires the submission of a copy or digital twin of a physical document in the form of an image or video captured using a matrix camera on the personal electronic device or any third-party electronic device capable of such capture. This image or video can be created before the digital copy is presented to an electronic verification entity. It can also be created in real time, in response to a capture request from the electronic verification entity, delivered by a resident application on the personal electronic device or by a remote application hosted on an application server.Real-time development, i.e. during the presentation process, is sometimes required to reduce the risk of falsification whereby a fraudster would have ample opportunity to fabricate a fraudulent digital document from scratch prior to a submission for compliance verification.

[0005] In all cases, a remote electronic verification entity must perform an authenticity test of the presented digital copy (image or video) since the source of said digital copy, in this case a personal electronic device, cannot be considered a reliable source.

[0006] Falsification can be classified as exogenous when it involves integrating one or more foreign elements into the image of a physical document to alter that image by adding or replacing content. It can be classified as endogenous when it involves taking one or more elements already present in an image of a physical document to modify its content. More difficult to detect, such endogenous falsification can consist of taking letters, numbers, or shapes already present to alter the integrity of the original document by modifying, on the fraudulent digital copy, an address, a nationality, a date of birth, accentuating or reducing hair or any other distinctive feature of a portrait, or the structural characteristics or appearance of a graphic object present on the official document.

[0007] To complicate matters for fraudsters, an official document attesting to an identity, origin, or status may contain one or more holograms, similar to the physical holograms found on banknotes, passports, or certificates of guarantee. Such a physical hologram is designed to partially cover the content of an official document. It can be transparent or opaque and may appear or change color depending on the lighting or viewing angle. By exploiting such holograms, it becomes difficult to alter the content of an official document through internal or external forgery without damaging the hologram and revealing the forgery.

[0008] A fraudster might also be tempted to create a completely fake official document and then scan it for submission to an electronic verification body. This way, the fraudster avoids the risk of altering the appearance of a physical hologram after scanning the fraudulent official document. To do this, they use a standard, generic "base card" without any physical hologram, scan it using video capture, and then overlay a photograph and / or other text or graphic content to mimic that of a legitimate official document. Acquiring equipment capable of applying a physical hologram is expensive and complex. Digital forgery, on the other hand, involves injecting a digital hologram using appropriate image processing software.Indeed, fraudulent digital holograms exist that are designed to simulate the optical behavior of a physical hologram in response to a specific capture angle or lighting conditions. Such digital holograms are so realistic that forgeries can fool existing remote electronic verification entities, particularly when these entities use photographs or videos created prior to a verification submission.

[0009] Technological advancements have made it possible to falsify documents or digital copies in real time during the verification process. The fraudster's electronic device includes a suitable application and / or is configured to inject such a digital hologram in real time during the capture of an image or video in response to a request from a remote electronic verification entity. This includes situations where the entity invites the fraudster to scan the counterfeit document while simultaneously altering its orientation and / or position relative to the matrix camera. Because this injected digital hologram dynamically and realistically simulates the optical behavior of a physical hologram, the falsification can defeat even the most sophisticated electronic verification entities.

[0010] Some work has been conducted or used to detect attempts at falsification. For example, the paper by Aliev MA et al., titled "Algorithm for choosing the best frame in a video stream in the task of identity document recognition," dated February 1, 2021, XP093283539, discloses a technique for extracting a relevant frame from a video stream. The paper "Fake Education Documentation Detection using Image Processing and Deep Learning" by Chandra Praba Mrs et al., dated March 27, 2021, XP093283205, discloses a solution involving training a neural network on denoised images to determine areas of interest in digital copies of documents.Gaël Mahfoudi's doctoral thesis (UTT Doctoral Thesis, Authentication of Digital Images and Videos, January 1, 2021, XP093042178) focuses on specific work to detect attempts to falsify a digital portrait using morphing techniques. This involves exploiting noise levels within two specific regions of interest: one selected within the face and the other at its periphery. However, there is currently no effective technique for detecting, by analyzing a digital copy of an official document, any attempt at falsification, particularly through the embedding of fraudulent elements within such a digital copy.

[0011] It therefore becomes necessary to design a process for verifying the conformity of a digital copy implemented by an electronic verification entity, designed to detect such digital falsifications.

[0012] The invention addresses the previously mentioned drawbacks and risks and promotes the deployment of remote verification of the conformity of a digital document.

[0013] Among the many advantages provided by the implementation of the invention, we can mention that the process offers a method arranged for:

[0014] - detect endogenous and / or exogenous falsification in the form of one or more images of a fraudulent document created respectively prior to a verification request or in real time as part of such a verification;

[0015] - produce an automatic or human decision aid aimed at recognizing or rejecting the conformity of a digital copy in the form of a sequence of images taken from a video;

[0016] - optimize the implementation time of such a verification and / or the material resources required for such a verification by selecting one or more particularly relevant images to implement an efficient verification;

[0017] - to iterate a request to capture a document whose conformity one seeks to verify when the digital copy of such a document does not meet the criteria required to carry out a relevant verification;

[0018] - exploit residual noise induced by the image sensor that produced a digital copy of a document as a source of detection of an attempt at falsification by embedding a fraudulent element.

[0019] To this end, the invention provides a method for detecting digital forgery, said method being designed to be implemented by a processing unit of an electronic entity for verifying the conformity of a digital copy of an official document, said method comprising a step of analyzing at least one image taken from said digital copy and producing a conformity result for the digital copy. To detect an attempt to fraudulently embed an element within said digital copy, said method:

[0020] - includes, prior to the implementation of said analysis and production of a conformity result step, a processing step of said at least one image taken from said digital copy to produce an image expressing the residual noise of each of said at least one image taken from said digital copy, said processing step consisting of the implementation of: o a denoising substep of said at least one image taken from said digital copy and production of a denoised image of each of said at least one image taken from said digital copy; o a substep of production of the image expressing the residual noise of each of said at least one image taken from said digital copy as being the result of an absolute difference between the denoised image and the image taken from said digital copy from which said denoised image originates;

[0021] - said step of analysis and production of a conformity result is arranged to produce said conformity result from the image expressing the residual noise of each of said at least one image taken from said digital copy by the implementation of a first model previously trained from images expressing respectively the residual noise of lawful images or obtained by digital falsification.

[0022] According to an advantageous embodiment, the denoising substep of said at least one image taken from said digital copy can be carried out by wavelets.

[0023] As an example of a preferred implementation, the step of analyzing and producing a conformity result could consist of implementing a first model previously trained on images representing, respectively, the residual noise of legitimate images or images obtained through digital manipulation. Such an initial model implementation could then consist of the processing unit of the electronic verification entity using a convolutional neural network.

[0024] When a digital copy of a document consists of a video, a method according to the invention may include a preliminary step of extracting at least one image from said digital copy. Such a step may then be arranged to extract a first set of images from said digital copy, said first set comprising at least two distinct images, the processing and analysis steps of at least one image from said digital copy being implemented iteratively to process and analyze at least two images belonging to said first set of images from said digital copy.

[0025] When a digital copy of an official document consists of a video, in order to optimize the performance of implementing such a process according to the invention, said process may further include a step of producing a subset of said first set of images to produce a second set of images satisfying a selection criterion, said subset production step being implemented upstream of the iterative implementation of the processing and analysis steps of an image taken from said digital copy, so that said iteratively implemented steps only relate to images belonging to the second set of images.

[0026] When a verification of the conformity of a digital copy requires a video showing a variety of orientations of said official document during shooting by a camera, in order to reveal different faces of interest of said document with regard to varying capture angles for example to detect the lawful behavior of a physical hologram conveyed by said official document, a step in producing a second set of images satisfying a selection criterion may advantageously consist of associating a pose attribute with each image belonging to the second set of images, said pose attribute belonging to a set of determined pose attribute values.One such embodiment may, more specifically, consist of implementing a second model previously trained on images describing documents analogous to the official document. These analogous documents exhibit positions and orientations relative to the image-producing sensors, and pose characteristics are associated with the pose attribute values ​​of the determined set of pose attribute values. In an advantageous embodiment, the implementation of such a second model may consist of using a multilayer neural network by the processing unit of the electronic verification entity.In order to reject a digital copy unsuitable for such a conformity check, a method according to the invention may include a step of rejecting the digital copy when said second set of images contains a number of images associated with one of the values ​​of said set of determined pose attribute values, less than a predetermined threshold.

[0027] Regardless of the embodiment chosen of a method for detecting digital forgery according to the invention, the conformity result produced by the step of analyzing and producing such a conformity result from the image expressing the residual noise of each of said at least one image taken from said digital copy may consist of at least one value associated with said at least one image taken from said digital copy, expressing a probability of forgery of said analyzed image.

[0028] Alternatively or in addition, such a process may include a step of producing an aggregate conformity result of the digital copy consisting of a calculation based on respective probability values ​​of falsification of at least two images analyzed and previously taken from said digital copy.

[0029] According to a second object, the invention relates to a computer program product comprising program instructions which, when executed by a computer processing unit, cause the implementation of a method for detecting digital forgery according to the invention.

[0030] According to a third object, the invention relates to an electronic entity for verifying the conformity of a digital copy of an official document comprising a program memory and a processing unit, respectively adapted to record and execute the instructions of such a computer program product.

[0031] According to a fourth object, the invention also relates to a system comprising a first electronic entity adapted according to the invention to verify the conformity of a digital copy of an official document, a second electronic entity operated by a state or a service provider, said second electronic entity being arranged to:

[0032] - transmit a digital copy of an official document whose conformity must be analyzed by the said first electronic entity;

[0033] - to utilize the compliance result produced by said first electronic entity.

[0034] Other features and advantages will become clearer upon reading the following description and examining the accompanying figures, including:

[0035] - Figure 1 illustrates a system for verifying the conformity of an official document by means of an electronic verification entity in communication with a personal electronic device of the holder of said official document and an electronic entity operated by a state or a service provider;

[0036] - Figure 2 illustrates the implementation of a verification process implemented by such an electronic verification entity adapted according to the invention;

[0037] - Figure 3 illustrates an implementation of such a verification process applied to the detection of exogenous falsification by the injection or forging of a digital hologram.

[0038] Figure 1 presents the functional architecture of a system 1 for verifying the conformity of a digital copy DD describing an official document OD. This system 1 includes an electronic verification entity 20 designed to detect endogenous or exogenous falsification during the production of said digital copy DD. Such falsification detection can be part of the broader context of an enrollment request with a government authority or a trusted third party, or even within the context of a request for access to a service (to conduct a banking transaction, view multimedia digital content) or access to a physical or electronic resource, etc.System 1, as illustrated in Figure 1, includes a second electronic entity 40, operated by a state or service provider, which can rely on the electronic verification entity 20 to verify the conformity of the digital copy DD. According to this advantageous embodiment, the electronic entity 40 sends the digital copy DD and awaits the result of the conformity verification performed by the electronic verification entity 20 in the form of a score or a unit probability (SCM) when the digital copy DD consists of an image, or possibly in the form of an aggregate score or probability (ASC) when the digital copy DD consists of a video or a plurality of images. Alternatively, the electronic entities 20 and 40 could constitute a single physical entity.

[0039] Such a system 1 further comprises a personal electronic device 10 associated with an individual 2 wishing to prove his identity, a quality or the holding of a certificate to the electronic entity 40, the latter relying on the electronic verification entity 20.

[0040] The electronic verification unit 20 comprises a processing unit 21 in the form of one or more microprocessors. It is connected to the outside world via a wired (Internet or intranet) or wireless communication network 30, using suitable communication means 24. This electronic verification unit 20 may include a human-machine interface 25 for output and / or input (e.g., a screen, a keyboard, a touchscreen, etc.) enabling the reading and / or input of information to allow for the implementation of maintenance operations, for example.

[0041] To adapt its operation and transform a simple computer into an electronic verification unit 20 within the meaning of the invention, the computer may include a program memory 23, either integrated with or separate from a data memory 22, for storing one or more computer programs P whose program instructions automatically trigger the implementation of appropriate functional processes. The term "memory" 22 or 23 refers to any computer memory, whether volatile or non-volatile. Non-volatile memory is computer memory whose technology retains its data even in the absence of an electrical power supply. It may contain data resulting from input, calculations, measurements, and / or program instructions.The main non-volatile memories currently available are electrically writable, such as EPROM (Erasable Programmable Read-Only Memory), or electrically writable and erasable, such as EEPROM (Electrically Erasable Programmable Read-Only Memory), flash, SSD (Solid-State Drive), etc. Non-volatile memories are distinguished from so-called "volatile" memories, whose data is lost when power is removed.The main volatile memories currently available utilize RAM technologies ("Random Access Memory" according to Anglo-Saxon terminology or also called "vm" memory >>), DRAM (dynamic vm memory, requiring regular updating), SRAM (static vm memory requiring such updating during an electrical under-powering), DPRAM or VRAM (particularly suited to video), etc.

[0042] An electronic entity 20 may also include means 26 supplying it with the electrical energy necessary for its operation (battery(ies), mains).

[0043] Like an electronic verification entity 20, a personal electronic device 10 includes a processing unit, an input / output interface, and data and / or program memory. Such a device also includes a raster image sensor, such as a camera. As mentioned previously, the principle of remotely verifying the authenticity of an official document is well-established. It consists of scanning, upon invitation, an official document OD using the raster image sensor of the personal electronic device 10 to produce a digital copy (image(s) or video) DD of said official document OD, and then presenting, directly or indirectly, said digital copy DD to an electronic verification entity 20 responsible for confirming or refuting the conformity of the digital copy DD with the official document OD through an analysis of said digital copy.According to the SCM, ASC result of said verification, the application or sovereign electronic entity or service provider 40 gives a favorable or unfavorable response (i.e. grants or rejects a request for access to a service or enrollment) to a request from the user of said personal electronic device who has presented a digital copy DD deemed compliant or non-compliant.

[0044] As previously explained, falsification techniques are legion.

[0045] Figure 3 illustrates, respectively, a legitimate digital copy DFj-A of an official document OD (situation A) and a fraudulent copy DFj-B (situation B) of such an official document OD obtained through exogenous falsification. Visually, the digital copies DFj-A and DFj-B may initially appear legitimate. Unfortunately, this is also true for many current electronic verification entities.20

[0046] Figure 2 illustrates a method 100 for the automatic detection of a falsification of a digital copy DD of an official document OD. Such a method is designed to be translated into instructions for a computer program product that can be loaded into a program memory 23 of an electronic verification unit 20 of a system 1 such as that previously described in connection with Figure 1. Such a method 100 is also designed to be implemented by a processing unit 21 of such an electronic unit 20, thereby adapting the operation of the latter so that said electronic unit 20 automatically and particularly effectively highlights a falsification of a digital copy DD of an official document OD.

[0047] The functional example of a method 100 illustrated by Figure 2 falls within a use case in which the digital copy DD of an official document OD is presented to said electronic verification entity 20 in the form of a video produced by a user using a personal electronic device, such as the device 10 in Figure 1, in response to an invitation that said user must film their official document OD, for example a national identity card or a passport, dynamically changing the orientation and viewing angles during the capture of said video, so that the latter relates not only the front and back of said official document OD but also presents it from various viewing angles, for example by rotating said official document at vertical or horizontal angles, revealing the presence of characteristic elements such as holograms HA,as shown in image DFj-A, which could have been extracted from such a DD video in Figure 3. By changing the orientation and / or viewing angles, such physical holograms embedded in a legitimate official document become more or less visible, disappear, fade, or are colored differently. Obviously, as indicated in image DFj-B in Figure 3, a fraudulent DD video can also reveal a fraudulent digital hologram HB, arranged to adopt optical behavior analogous to that of a physical hologram HA when said HB hologram is embedded in said fraudulent DD video. It is therefore relevant to extract several images from the DD video in step 110 to test the conformity of said images. Obviously, when a presented digital DD copy consists of only a single image,Such a step 110 is not implemented. The single image is directly analyzed in steps 130 and 140, which we will detail later. Step 110 is therefore optional for the implementation of a process 100 according to the invention.

[0048] In the context of a DD copy in the form of a video, step 110 can thus extract a plurality of n images DF1, ..., DFi, ..., DFn, according to a given sampling or extraction frequency (for example, five, ten, or twenty-five images per second), or extract a determined number of images (for example, between five and thirty images) drawn randomly or deterministically from the DD video. The invention is not limited to these extraction examples alone, in order to meet various verification needs or requirements.Thus, when the said electronic verification entity is requested by a third-party entity 40 of a service provider, a request addressed by the latter to the said electronic verification entity 20 may include the digital copy DD as well as one or more configuration parameters for the implementation of the process 100, including the aforementioned sampling frequency or number of images. Alternatively, such a request may convey a level of requirement for conformity verification that the electronic verification entity 20 can translate into such configuration parameters for the implementation of the process 100.

[0049] Since the digital copy DD in the form of a video is made by a non-professional user, the first set DFN of extracted images may contain some almost identical and therefore redundant, or even irrelevant, images, particularly if the framing of the official document OD is approximate, said official document OD going totally or partially outside the field of capture of the camera of the device 10. The invention therefore provides that, when the copy DD consists of a video, a step 120 subsequent to the extraction 110 can be implemented (optionally but advantageously) to reduce said first set DFN and produce a second set DFM comprising images, possibly fewer in number but more relevant, for efficient and subsequent verification of the conformity of the digital copy DD.

[0050] Such a step 120 of producing a subset of said first DFN set of images to produce a second DFM set of relevant images may include a first substep 121 consisting of retaining only those images from said first DFN set that satisfy a selection criterion XC. A first selection criterion XC may consist of detecting within the image a region of interest describing a view of an official document that is not truncated. Alternatively or in addition, such a criterion may also consist of verifying a necessary and sufficient coverage rate of said view of the official document OD with respect to the rest of the image (for example, on the order of eighty percent), thus attesting that said document was not captured from too far away, the digital copy DD losing precision and / or sharpness for the remainder of the conformity analysis. The invention is not limited to these choices of selection criteria alone.When the official document (OD) has several faces of interest, for example, a front and a back, it is relevant to ensure that the image corpus constituted by the DFN or DFM set includes images depicting both the front and back. The same would apply to other faces of interest, or to verify that the thickness of the captured official document is accurate. Images depicting certain profile views of the official document (OD) would be relevant for analysis. The selection step 120 for producing a second DFM set of images can therefore include, in addition to a substep 121 for satisfying a selection criterion XC, a substep 122 consisting of associating a pose attribute with each image DF1, ..., DFj, ...DFm of the DFM set.Such a pose attribute can take predetermined values ​​within a set of predetermined values ​​respectively reflecting the fact that the image presents said document predominantly from the front (recto), back (verso), side, etc. In order to be able to associate such a pose attribute or in other words to constitute such a categorization of images, the invention provides that step 120 of production of a second DFM set of images can consist of an implementation of a model M2, possibly previously trained from a corpus of images LS2 describing respectively documents analogous to the official document OD, said analogous documents describing on the images of said corpus LS2 various positions and orientations with respect to the image production sensors, pose characteristics respectively associated with the pose attribute values ​​of the FA set of pose attribute values ​​determined previously mentioned.

[0051] For example, such a step 120 could consist of implementing a "deep learning" type model in the form of a multilayer neural network, such as the YOLO model, Single Shot Detector (SSD), or CenterNet. Such M2 models from this non-exhaustive list, or other equivalent models, will be chosen to implement the invention because of their ability to detect reference objects (in this case, an official document OD) in a single step or in very few steps, while achieving a very high level of accuracy and high speed of detection and categorization (or assignment of pose attributes as mentioned previously).

[0052] The optional and advantageous implementation of step 120, particularly substep 122, thus makes it possible to validate a sufficient DFM set of images to produce a conformity check of the high-quality DD digital copy, i.e., one that is not partial or incomplete. Therefore, a method 100 according to the invention may include a step 160 for rejecting a DD digital copy when said DFM set of images contains a number of images associated with one of the values ​​of said FA set of pose attribute values ​​that is less than a predetermined threshold, for example, equal to one or a greater number (for example, between two and twenty) of distinct images showing the OD document in the same pose. Thus, such a method 100 may include a test 123 to ensure that a sufficient number of images associated respectively with the different predetermined pose attribute values ​​is present.If not (situation illustrated by link 23-n in figure 2) process 100 is completed and the person is asked to produce a new digital copy DD describing sufficient content to conduct a conformity analysis in accordance with the invention.

[0053] When the DFM corpus of m images DF1, ..., DFj, ..., DFm is deemed satisfactory by process 100, the latter comprises respective preprocessing steps 130 and analysis steps 140 of each of these preprocessed images to produce a score or, more generally, a plural conformity result. Thus, an SCM set of m scores SC1, ..., SCj, ..., SCm can be produced for an analysis 140 of m images DF1, ..., DFj, ..., DFm taken from the digital copy DD describing the official document OD. As mentioned previously, in the specific case where said digital copy DD is limited to a single image DF1, steps 110 and 120 are not implemented, and said image DF1 is directly preprocessed to be ultimately analyzed and produce a single score SC1.

[0054] In accordance with the invention, to detect many types of attempted falsification, the pretreatment step 130 can be adapted to maximize the chances of detecting falsification during the implementation of the subsequent step 140. Thus, depending on whether the attempted falsification is endogenous or exogenous, said step 130 can be adapted accordingly. Figure 3 illustrates the implementation of steps 130 and 140 on two DFj images, depending on whether said DFj image is taken from a legitimate DD video (situation A, said DFj image is referenced DFj-A on the left of said Figure 3) or from a fraudulent video, in which a digital HB hologram has been superimposed or forged during the digitization of the official document, in this case a counterfeit document (situation B, said DFj image is referenced DFj-B on the right of said Figure 3). Image DFj-A presents a background image of a national identity card containing a physical HA hologram.When the matrix image sensor (or camera) of the personal device 10 scans the official document, the sensor, not being perfect, results in the image DFj-A containing residual noise directly from the sensor. Therefore, every sensor can be considered to possess a unique "signature" directly conveyed by the image it produces—in this case, the residual noise of the image. Such noise is generally homogeneous, deterministic, and relatively independent of the captured subject. When an image is forged or falsified, the embedded fraudulent element, however realistic, is not captured by the sensor. The resulting image, in this case DFj-B, thus exhibits residual noise that breaks with the matrix sensor's "signature," and this noise is generally correlated with the embedded fraudulent element—in this case, the digital hologram HB.

[0055] According to a preferred embodiment, particularly suitable for highlighting an overlay of a fraudulent digital hologram, step 130 can consist of calculating and producing an image DFj', from the image DFj, which expresses the residual noise of the latter.

[0056] In conjunction with Figure 2, Figure 3 presents such a step 130 in the presence (situation A) of a DFj-A image taken from a legitimate DD video and a DFj-B image taken from a fraudulent DD video (situation B). Thus, a preprocessing step 130 can consist of implementing a first sub-step 131 of wavelet denoising of said DFj-A and DFj-B images to obtain respectively denoised images DFj-A” and DFj-B” of the DFj-A and DFj-B images. Step 130 then includes a second substep 132, applied to each denoised image DFj-A” and DFj-B” of production of the image expressing the residual noise DFj-A' with respect to image DFj-A and DFj-B' with respect to image DFj-B. Such an image DFj-A' or DFj-B' is the result of an absolute difference between the denoised image DFj-A” or DFj-B” and the image DFj-A or DFj-B from which said denoised image DFj-A” or DFj-B originates.According to a preferred embodiment of step 140 described later, since DFj-A or DFj-B images are generally expressed in color, the invention provides that the pixel values ​​of the raster images expressing residual noise DFj-A' or DFj-B' can be normalized. Thus, these pixels describe light intensities ranging from 0 to 255 or from -1 to 1.

[0057] In the example illustrated by Figure 3, it appears that the residual noise expressed by the image DFj-A' is more homogeneous and less dependent on the content of the image DFj-A, whether or not it is covered by the physical hologram HA. In contrast, the image DFj-B' exhibits more heterogeneous residual noise, particularly in areas corresponding to the areas in the image DFj-B affected by the overlay of the fraudulent digital hologram HB. In this example, the residual noise described by image DFj-B, deliberately exaggerated for pedagogical reasons, indicates digital falsification. Related to Figure 2, for which a DFM set of images DF1, ..., DFj, ..., DFm is extracted in step 110 from a digital copy DD of an official document OD and then selected in step 120 for analysis, such a step 130 consists of generating a DFM set' of images representing the respective residual noise of said images DF1, ..., DFj, ..., DFm obtained by the absolute differences between the denoised images DF1”, ..., DFj”, ..., DFm” and the images DF1, ..., DFj, ..., DFm from which said denoised images DF1”, ..., DFj”, ..., DFm are respectively derived. Analyzing an image exhibiting residual noise to determine whether the image from which said image exhibiting residual noise is derived is fraudulent or legitimate cannot be performed manually. A method 100 according to the invention therefore includes a step 140 to perform such an analysis automatically and objectively. Thus, in connection with Figure 2, said step 140 analyzes the images of the set DFM' (that is, the set of images DF1', ... DFj', ... DFm' expressing residual noise) and produces a conformity result in the form of an SCM set of conformity results SC1, ... SCj, ..., SCm respectively specific to the images DF1, ..., DFj, ...DFm taken from said digital copy DD.Each result or score SCj for an image DFj can express a probability of falsification of said analyzed image DFj, or be normalized to express two predetermined Boolean values, for example, "False" or '0' to express a legitimate image and "True" or '1' to express an image that has been manifestly falsified. To produce such a result SCj, the invention provides that step 140 can consist of the processing unit implementing said method 100 using a model M1 previously trained from an image corpus LS1 expressing, respectively, the residual noise of legitimate images or images obtained by digital falsification. Step 140 may advantageously consist of implementing such an M1 model in the advantageous form of a convolutional neural network, for example the EfficientNet-bO convolutional neural network (https: / / arxiv.org / abs / 1905-1 1946#) pre-trained on an LS1 image corpus, in this case the ImageNet corpus.Such a network can be tuned, through supervised learning, using a proprietary training image bank consisting of images or videos representing official documents (national identity cards, passports, etc.) displaying genuine physical holograms (images or videos, for example, labeled "True" or "0") or fraudulent digital holograms (images or videos, for example, labeled "Fake" or "1"). The invention is not limited to these examples of image or video labeling. Other label values ​​could be used instead of the aforementioned examples to characterize training content containing genuine and fake holograms, respectively. During such training, the volume of training data can be artificially increased by using linear transformations, luminance increases or decreases, etc.

[0058] When a DFM set of images extracted from a digital copy DD is plural (i.e., it contains a number m of images DF1 to DFm greater than or equal to two), the invention provides that a method 100 may include a step 150 for producing an aggregated conformity result ASC of the digital copy DD. Such a "consolidated" ASC result may consist of a calculation based on falsification probability values ​​from the SCM set respectively associated with the images of the analyzed DFM set, reflecting the residual noise of the DFm images previously extracted (step 110) from said digital copy DD and selected (step 120) for preprocessing (step 130). Such a calculation may consist of the product of the expressed SCM probabilities, producing a resulting probability that the digital copy DD is falsified.Such a resulting probability can be normalized to express a first value of "True" or '0' for a digital copy conforming to a legitimate official document, or a second value of "False" or '1' for a clearly fraudulent digital copy. The SCM or ASC output can be used by a third-party electronic entity, such as entity 40 operated by a state or service provider, as illustrated in Figure 1.

[0059] The execution of a process 100 by an electronic entity, such as the electronic verification entity 20 illustrated by the example in Figure 1, can be triggered automatically by adapting said electronic entity 20 by loading instructions from a suitable computer program product into a program memory 23 of the latter.

[0060] A method for detecting digital forgery 100 has been described through examples, in connection with Figures 1 to 3, illustrating a digital copy DD consisting of a video DD during which a user of a personal electronic device 10 orients an official document with respect to different axes relative to the field of capture of the camera of said device 10. As already stated, steps 110, 120 and 150 are optional because they cannot be implemented for a digital copy DD which consists of a single image from a single capture of such a document.

[0061] Furthermore, the previous examples described one or more images of an official document conveying a physical hologram HA or a forged hologram HB as part of a falsification.

[0062] A method for detecting digital forgery 100 according to the invention can be adapted to detect endogenous forgeries, that is, forgeries resulting from the displacement of elements of a lawfully produced image that is then forged to replace other original elements of said image, for example, to change a date of birth. The principle of steps 110 to 150 of method 100 remains. Step 140 can be adapted to use a model M1 trained by supervised learning from a training image corpus LS1 composed of images or videos of official documents that do not exhibit any endogenous forgery (images or videos labeled "True" or "0") and images or videos of official documents whose digital copy exhibits endogenous forgery, for example, via a technique known as "copy and move" (images or videos labeled "False" or "1").Similar to the M1 model trained to detect forgeries by injection or forging of digital holograms, the volume of training data for detecting endogenous forgeries can be artificially increased by the use of linear transformations, increases or decreases in luminance, etc.

Claims

22 DEMANDS 1. A method (100) for detecting digital forgery, said method (100) being designed to be implemented by a processing unit (21) of an electronic entity for verifying the conformity of a digital copy (DC) of an official document (OD), said method (100) comprising a step (140) of analyzing at least one image taken from said digital copy (DC) and producing a conformity result (SCj, ASC) of the digital copy (DC); said method (100) being characterized in that: - the process includes, prior to the implementation of said step (140) of analysis and production of a conformity result, a processing step (130) of said at least one image (DF1, ..., DFj, ...DFm) taken from said digital copy (DD) to produce an image (DFT, ..., DFj', ...DFm') expressing the residual noise of each of said at least one image (DF1, ..., DFj, ...DFm) taken from said digital copy (DD), said processing step (130) consisting of the implementation of: o a substep (131) of denoising said at least one image (DF1, ..., DFj, ...DFm) taken from said digital copy (DD) and of producing a denoised image (DF1, ..., DFj, ...DFm") of each of said at least one image (DF1, ..., DFj, ...DFm) taken from said digital copy (DD); o of a substep (132) of image production expressing the residual noise (DF1', ..., DFj', ...DFm') of each of said at least one image (DF1, ..., DFj, ...DFm) taken from said digital copy (DD) as being the result of an absolute difference between the denoised image (DF1 ”, ..., DFj”, ...DFm”) and the image (DF1 , ..., DFj, ...DFm) taken from said digital copy (DD) from which said denoised image is derived;. - said step (140) of analysis and production of a conformity result is arranged to produce said conformity result (SCj, ASC) from the image (DF1 ', DFj', ...DFm') expressing the residual noise of each of said at least one image (DF1 , DFj, ...DFm) taken from said digital copy (DD) by the implementation of a first model (M1 ) previously trained from images (LS1 ) expressing respectively the residual noise of lawful images or obtained by digital falsification.

2. Method (100) according to claim 1, wherein the substep (131) of denoising said at least one image (DF1, ..., DFj, ...DFm) taken from said digital copy (DD) is carried out by wavelets.

3. Method (100) according to claim 1 or 2, wherein the implementation of the first model (M1) consists of the operation by the processing unit (21) of the electronic verification entity (20) of a convolutional neural network.

4. Method (100) according to any one of claims 1 to 3, comprising a preliminary step (1 10) of extracting at least one image (DF1 , ..., DFi, ...DFn) from said digital copy (DD) when the latter consists of a video.

5. A method (100) according to claim 4, wherein said extraction step (110) is arranged to extract a first set (DFN) of images from said digital copy (DD), said first set (DFN) comprising at least two distinct images (DF1, ..., DFi, ...DFn), the processing (130) and analysis (140) steps of at least one image taken from said digital copy (DD) being implemented iteratively to process and analyze at least two images (DF1, ..., DFi, ...DFn) belonging to said first set (DFN) of images taken from said digital copy (DD).

6. A method (100) according to claim 5, comprising a step (120) of producing a subset of said first set (DFN) of images to produce a second set (DFM) of images satisfying (121) a selection criterion (XC), said step (120) of producing a subset being carried out upstream of the iterative implementation of the processing steps (130) and analysis (140) of an image taken from said digital copy (DD), so that said iteratively carried out steps only relate to images (DF1, DFj, ...DFm) belonging to the second set of images (DFM).

7. Method (100) according to claim 6, wherein the step (120) of producing a second set (DFM) of images satisfying (121) a selection criterion, consists of associating (122) a pose attribute with each image (DF1, ..., DFj, ...DFm) belonging to the second set of images (DFM), said pose attribute belonging to a set (FA) of determined pose attribute values.

8. A method (100) according to claim 7, wherein the step (120) of producing a second set (DFM) of images satisfying (121) a selection criterion (XC) consisting of associating (122) a pose attribute (FA) with each image (DF1, ..., DFj, ...DFm) belonging to the second set of images (DFM), consists of implementing a second model (M2) previously trained from images (LS2) describing respectively documents analogous to the official document (OD), said analogous documents having positions and orientations with respect to the image production sensors, characteristic of poses 25 respectively associated with the pose attribute values ​​of the set (FA) of determined pose attribute values.

9. Method (100) according to claim 8, wherein the implementation of the second model (M2) consists of the processing unit (21) of the electronic verification entity (20) operating a multilayer neural network.

10. Method (100) according to any one of claims 7 to 9, comprising a step (160) of rejecting the digital copy (DD) when said second set (DFM) of images comprises a number of images associated with one of the values ​​of said set (FA) of determined pose attribute values, less than a predetermined threshold.

11. A method (100) according to any one of claims 1 to 10, wherein the conformity result (SCM, ASM) produced by the step (140) of analyzing and producing such a conformity result from the image (DFM') expressing the residual noise of each of said at least one image (DFM) taken from said digital copy (DD) consists of at least one value (SCM) associated with said at least one image (DFM) taken from said digital copy (DD), expressing a probability of falsification of said analyzed image.

12. Method according to claim 1 1, comprising a step (150) of producing an aggregate conformity result (ASC) of the digital copy (DD) consisting of a calculation on respective probability of forgery (SCM) values ​​of at least two images (DFM') analyzed and previously taken from said digital copy (DD). 26 13. Product computer program (P) comprising program instructions which, when executed by a processing unit of a computer, causes the implementation of a method (100) for detecting digital forgery according to any one of claims 1 to 12.

14. Electronic entity for verifying the conformity of a digital copy (DD) of an official document (OD) comprising a program memory (23) and a processing unit (21), respectively adapted to record and execute the instructions of a computer program product (P) according to claim 13.

15. System (1) comprising a first electronic entity (20) adapted according to claim 14 for verifying (20) the conformity of a digital copy (DD) of an official document (OD), a second electronic entity (40) operated by a state or a service provider, said second electronic entity (40) being arranged to: - transmit a digital copy (DD) of an official document (OD) whose conformity must be analyzed by said first electronic entity (20); - exploit the compliance result (SCM, ASC) produced by said first electronic entity (20).

Images