Method and device for detecting at least one image generated by artificial intelligence

Abstract

Method and device for detecting at least one image generated by artificial intelligence. The present invention relates to a method (40) for detecting at least one image generated by artificial intelligence, comprising: - a training phase (42) comprising the following steps: - separate training (44) of at least five distinct detection models using a first set of which each image is previously compressed according to a random compression ratio; - application of each of said at least five trained models to each image compressed according to a random compression ratio of a second set, and obtaining, at the output, an associated logit; - obtaining a trained metamodel corresponding to a machine learning tool trained with said logits as input;- an inference phase (48) applying said trained metamodel to an input test image and providing, as output, a result classifying said test image as generated by artificial intelligence or not. Figure for the abbreviation: Figure 2;

Description

Title of the invention: Method and device for detecting at least one image generated by artificial intelligence

[0001] The present invention relates to a method for detecting at least one image generated by artificial intelligence, the method being implemented by an electronic device for detecting at least one image generated by artificial intelligence.

[0002] The present invention also relates to a computer program comprising software instructions which, when executed by a computer, implement such a method of detecting at least one image generated by artificial intelligence.

[0003] The present invention also relates to said electronic device for detecting at least one image generated by artificial intelligence.

[0004] The present invention falls within the field of digital image processing and in particular their falsification via the use of Artificial Intelligence (AI).

[0005] The falsification of digital images has become an unavoidable reality, especially in the field of cybercrime. A growing number of fake or falsified images (deepfakes) created using Artificial Intelligence (AI) are used on social networks and fuel the spread of misinformation (fake news). These modifications can be relatively innocuous (e.g., retouching a person's appearance to remove skin imperfections), disturbing (e.g., removing flaws from an item for sale online), or have serious social repercussions, particularly with harmful effects on social networks (e.g., a montage showing an improbable meeting of political figures).

[0006] Falsified images or images generated by artificial intelligence techniques are also used, for example, to create a fake speech by a political leader or fake operations.

[0007] In other words, the dissemination of such falsified images or images generated by artificial intelligence techniques constitutes a new technological threat, and their effective detection is necessary, from a cybersecurity point of view, to limit serious consequences in the civilian sector and even beyond.

[0008] Several AI-generated image detection models have been developed in recent years. For example, there is the DNF model, as described by Zhang, Yichi, et al. in the 2023 article entitled "Diffusion noise feature: Accurate and fast generated image detection", or the SSP model as described by Chen, Jiaxuan, et al. in the article entitled "A single simple patch is all y ou needfor ai- generated image detection” of 2024, or the DCT model as described by Corvi, Riccardo, et al. in the article entitled “Intriguing properties of synthetic images: from generative adversarial networks to diffusion models” of 2023, or the CLIP model as described by Cozzolino, Davide, et al. in the article entitled “Raising the Bar of AI-generated Image Detection with CLIP” of 2024.

[0009] Each of these image detection models certainly has separate advantages, but also disadvantages, such as overfitting due to artifacts specific to each model, overfitting of the distribution of the training image set, overfitting of the JPEG (Joint Photography Experts Group) compression biases used in the training image set, semantic overfitting, colorimetric bias.

[0010] Currently, it is possible to reduce one of the aforementioned disadvantages, such as overfitting due to artifacts of the model used, or to limit overfitting on JPEG compression biases for another model, but not both at the same time, and no model makes it possible to effectively reduce all of the aforementioned disadvantages.

[0011] The aim of the invention is then to improve the detection of images generated (i.e. falsified) by artificial intelligence by further reducing the aforementioned disadvantages of the current state of the art techniques.

[0012] To this end, the invention relates to a method for detecting at least one image generated by artificial intelligence, the method being implemented by an electronic device for detecting at least one image generated by artificial intelligence, and comprising:

[0013] - a training phase comprising the following steps:

[0014] - separate training of at least five distinct detection models of at least an image generated by artificial intelligence, using, for each of said separate trainings of said at least five distinct models, the same first set of training images, each image of said first set of training images being, before being used as input to each of said at least five distinct models, compressed according to a compression rate randomly selected from a list of predetermined compression rates;

[0015] - compression of each image of a second set of training images, according to a compression rate randomly selected from said list of predetermined compression rates;

[0016] - application of each of said at least five trained models, to each image compressed said second set of training images, and obtaining, at the output of each of said at least five trained models, an associated logit;

[0017] - obtaining a trained metamodel corresponding to a learning tool automatic trained using, as input, for each compressed image of said second set of training images, said corresponding output logits of said at least five trained models;

[0018] - an inference phase applying said trained metamodel to a test image input and capable of providing, as output, a detection result classifying said test image as generated by artificial intelligence or not.

[0019] Thus, the present invention proposes to train separately at least five models of detection of at least one image generated by artificial intelligence, using as input, for each of these at least five distinct models, a first set of training images, each image of which is previously compressed according to a compression rate selected randomly from a list of predetermined compression rates.

[0020] Once these five models have been trained, they are combined into a metamodel. To do this, a second set of training images is used, and each image in this second set of training images is compressed using a compression ratio randomly selected from a predetermined list of compression ratios.

[0021] Each image from the second set of training images, thus compressed, is used as input to the five trained models. Each of the at least five trained models provides as output a logit associated with said compressed image.

[0022] In other words, for a compressed image of the second set of training images, we obtain at least five logits.

[0023] The metamodel according to the present invention is then trained using, as input to a machine learning tool, for each compressed image of the second set of training images, the logits associated with said compressed image considered.

[0024] Such a trained metamodel limits the aforementioned overfitting, because it combines the advantages of at least five models, previously trained separately with a first set of training images, cleverly using only their output logit, when applied to a second set of training images (with which they have not been trained), as training input for another machine learning tool.

[0025] Such a metamodel is also more robust, during inference, to compressed images, since its training was carried out by considering training images whose compression rate was chosen at random, which increases the variety of the training image sets (from the English datasef).

[0026] According to other advantageous aspects of the invention, the method for detecting at least one image generated by artificial intelligence comprises one or more of the following features, taken individually or in all technically possible combinations:

[0027] - each of the images of said first set of training images and of said second set of training images is in JPEG format, and in which said compression rate list includes the following set of compression rates: 40%, 50%, 60%, 70% and 96%;

[0028] - said at least five distinct detection models of at least one generated image by artificial intelligence belong to the group including at least:

[0029] - the DNF model based on the diffusion noise characteristic;

[0030] - the SSP model based on a simple and unique patch;

[0031] - the DCT model based on the discrete cosine transform;

[0032] - the CLIP model based on a contrastive language-image pre-training;

[0033] - an NF model without feature extraction;

[0034] - a DINO self-distillation model without a label.

[0035] - said first set of training images and said second set training images are distinct;

[0036] - said first set of training images is the set of images Genlmage training;

[0037] - said second set of training images is the set of images Synthbuster training;

[0038] - said machine learning tool is a reinforcement model called gradient boosting;

[0039] - said machine learning tool is a decision tree forest.

[0040] The invention also relates to a computer program comprising software instructions which, when executed by a computer, implement a method for detecting at least one image generated by artificial intelligence as defined above.

[0041] The invention also relates to an electronic device for detecting at least one image generated by artificial intelligence, the electronic device being configured to implement a method for detecting at least one image generated by artificial intelligence as defined above.

[0042] The invention will become clearer upon reading the following description, given solely by way of non-limiting example, and made with reference to the drawings in which:

[0043] [Fig.1] [Fig.1] is a schematic representation of an electronic device for detecting at least one image generated by artificial intelligence according to the present invention;

[0044] [Fig.2] [Fig.2] is a general flowchart of the method for detecting at least an image generated by artificial intelligence according to the present invention;

[0045] [Fig.3] [Fig.4] Figures 3 and 4 illustrate respectively two of the steps training of said process.

[0046] Fig. 1 first schematically illustrates a non-limiting example of an electronic device 10 for detecting at least one image generated by artificial intelligence according to the present invention.

[0047] According to the present invention, the electronic detection device 10 comprises firstly a storage module 12 configured to store:

[0048] - at least five distinct models to be trained to detect at least one image generated by artificial intelligence;

[0049] - said first set of training images and said second set training images;

[0050] - said list of compression ratios;

[0051] - said machine learning tool to be trained.

[0052] As an optional addition, which of the said at least five distinct models for detecting at least one image generated by artificial intelligence belong to the group comprising at least:

[0053] - the DNF model based on the diffusion noise characteristic;

[0054] - the SSP model based on a simple and unique patch;

[0055] - the DCT model based on the discrete cosine transform;

[0056] - the CLIP model based on a contrastive language-image pre-training;

[0057] - an NF model without feature extraction;

[0058] - a DINO self-distillation model without a label.

[0059] The electronic detection device 10 also includes a compression module 14 configured to compress each of the images of the first set of training images and the second set of training images with a compression rate randomly selected from said list of predetermined compression rates.

[0060] As an optional addition, each of the images in said first set of training images and said second set of training images is in JPEG format, and in which said compression rate list includes the following compression rates: 40%, 50%, 60%, 70%, and 96%. Note that the lower the compression rate, the more compressed the image. In other words, an image An image compressed at a rate of 40% is more compressed than an image compressed at a rate of 96%.

[0061] As an optional complement, said first set of training images and said second set of training images are distinct, which advantageously provides variety in the training datasets.

[0062] According to an example of this optional supplement, said first training image set is the Genlmage training image set as notably cited by M. Zhu et al. in the article entitled “Genlmage: A Million-Scale Benchmark for Detecting Al-Generated Image”, while said second training image set is the Synthbuster training image set as notably cited by Q. Bammey in the article entitled “Synthbuster: Towards Detection of Diffusion Model Generated Images”.

[0063] The electronic detection device 10 also includes a separate training module 16 suitable for separately training each of said at least five distinct detection models of at least one image generated by artificial intelligence, using, for each of said separate trainings, the same first set of training images, each image of said first set of training images being, before being used as input to each of said at least five distinct models, compressed, via said compression module 14, according to a compression rate selected randomly from a list of predetermined compression rates.

[0064] The electronic detection device 10 also includes an application module 18, configured to apply each of the at least five trained models, supplied as output by the separate training module 16, to each compressed image of said second set of training images supplied by said compression module.

[0065] As illustrated later by [Fig.4], the module 18 is configured to provide as output, for each compressed image of the second set of training images, the logits associated with said compressed image under consideration, each logit being provided by one of said at least five trained models.

[0066] The electronic detection device 10 also includes a retrieval module 20, configured to obtain a trained metamodel, said metamodel corresponding to (i.e. being) the machine learning tool trained using, as input, for each compressed image of said second set of training images, said corresponding output logits of said at least five trained models.

[0067] According to a first variant, said machine learning tool is a reinforcement model called gradient boosting (from the English Gradient Boosting).

[0068] According to another variant, said machine learning tool is a decision tree forest (from the English Random Forest).

[0069] The electronic detection device 10 also includes a module 22 configured to implement an inference phase by applying said trained metamodel provided by the acquisition module 20 to an input test image and capable of providing, at output, a detection result classifying said test image as generated by artificial intelligence or not.

[0070] As an optional complement, the electronic detection device 10 also includes a rendering module 24 configured to render said detection result classifying said test image as generated by artificial intelligence or not, via a sound rendering, and / or by display on a screen, and / or by transmission to another device (not shown) distinct from said detection device 10.

[0071] In the example of [Fig.1], the electronic device for detecting at least one image generated by artificial intelligence includes an information processing unit 26 formed for example of a memory 28 and a processor 30 associated with the memory 28.

[0072] In the example of [Fig. 1], the storage module 12, the compression module 14, the separate training module 16 for said at least five distinct detection models, the application module 18, the acquisition module 20, the module 22 configured to implement an inference phase, and, optionally, the rendering module, are each implemented as software, or a software component, executable by the processor. The memory of the electronic device for detecting at least one image generated by artificial intelligence is thus capable of storing storage software, compression software, separate training software, application software, acquisition software, and software for implementing the inference phase, as well as, optionally, rendering software.The processor is then capable of executing each of the following software programs: storage software, compression software, separate training software, application software, acquisition software, and inference phase implementation software, as well as, optionally, rendering software.

[0073] In an alternative not shown, the storage module, the compression module, the separate training module, the application module, the acquisition module, and the inference phase implementation module, as well as, optionally, the rendering module, are each implemented in the form of a programmable logic component, such as an FPGA (Field Programmable Gamut Arras). or even an integrated circuit, such as an ASIC (from the English Application Specifies Integrated Circuit).

[0074] When the electronic device for detecting at least one image generated by artificial intelligence is implemented in the form of one or more software programs, i.e., in the form of a computer program, also called a computer program product, it is further capable of being stored on a computer-readable medium (not shown). The computer-readable medium is, for example, a medium capable of storing electronic instructions and being connected to a bus of a computer system. For example, the readable medium is an optical disc, a magneto-optical disc, ROM, RAM, any type of non-volatile memory (e.g., FLASH or NVRAM), or a magnetic card. A computer program comprising software instructions is then stored on the readable medium.

[0075] A general embodiment of the operation of the electronic device for detecting at least one image generated by artificial intelligence in [Fig.1] is described below in relation to [Fig.2].

[0076] More specifically, the method 40 for detecting at least one image generated by artificial intelligence includes first of all phase 42 of training P_E.

[0077] This training phase P_E itself comprises two sets 44 and 46 of distinct steps, each subsequently described in more detail respectively in relation to figures 3 and 4.

[0078] The first set 44 of steps corresponds to (i.e. is) the separate E_S training of at least five distinct models for detecting at least one image generated by artificial intelligence, using, for each of said separate trainings of said at least five distinct models, the same first set Ei of training images, each image of said first set of training images being, before being used as input to each of said at least five distinct models, compressed according to a compression rate selected randomly from a list of predetermined compression rates.

[0079] The second set 46 is a set of steps leading to obtaining 0_M_E of a trained metamodel M_E.

[0080] The second set 46 is implemented after the implementation 44 of the separate E_S training of at least five distinct detection models, and comprises, as illustrated hereafter in relation to [Fig. 4], the compression of each image of a second set of training images E2, according to a compression ratio randomly selected from said list of predetermined compression ratios; then the application of each of said at least five trained models to each compressed image of said second set of training images, and obtaining, at the output of each of the said at least five trained models, of an associated logit; then obtaining, as such, a trained metamodel ME corresponding to a machine learning tool trained using, as input, for each compressed image of said second set of training images, the corresponding output logits of said at least five trained models.

[0081] Once the trained metamodel M_E is obtained, the process 40 includes an inference phase 48 P_I applying said trained metamodel M_E to an input test image IT and suitable for providing, as output, a detection result classifying (i.e. binary) said test image as generated by artificial intelligence or not.

[0082] More specifically, during this inference phase 48, the input test image IT is received, and then each of the at least five trained models is applied to it, to obtain an associated logit as output for each of the at least five trained models. These at least five associated logits are then used as input to the trained metamodel M_E, which then provides a detection result as output, classifying the test image as generated by artificial intelligence or not. For example, this result is also a logit (i.e., a logistic regression value), namely a real value. A positive logit, for example, would be representative of an image generated (i.e., falsified, a deepfake) by artificial intelligence, while, according to the same example, a negative logit would be representative of an authentic image (i.e., not generated or falsified by artificial intelligence).

[0083] As an optional complement, illustrated by the example of [Fig.2], the method 40 includes a step 50 of restitution R of said result, via a sound restitution, and / or by display on a screen, and / or by transmission to another device distinct from said detection device 10 implementing said method 40.

[0084] Such a rendering may optionally be accompanied by an alert or a step to block the dissemination of such a "deepfake" image once it has been detected as such.

[0085] Fig. 3 illustrates an example of a first set of 44 separate I_S training steps of at least five distinct detection models.

[0086] According to this example in [Fig.3], five distinct models for detecting at least one image generated by artificial intelligence are trained separately in five distinct and independent steps.

[0087] More specifically, the first set 44 of separate I_S training steps includes a step 52 of training a detection model of at least one AI-generated image corresponding to (i.e. being the, identical to) the DNF model based on the diffusion noise feature, as introduced by Zhang, Yichi, et al. in the aforementioned article entitled: "Diffusion noise feature: Accurate and fast generated image detection" of 2023.

[0088] To do this, step 52 includes substeps 54, 56, 58, 60, 62 and 64. Substep 54 is a retrieval substep of the first set Ei of training images, stored within the storage module 12 of [Fig.1].

[0089] The first set Ei of training images is used in common to separately train each of said at least five distinct detection models.

[0090] As an optional complement, said first set Ei of training images is the training image set Genlmage.

[0091] According to substep 56, each image in the first set Ei of training images is compressed according to a compression ratio randomly selected from a list of predetermined compression ratios.

[0092] As an optional complement, each of the images of said first set Ei of training images and of said second set of training images is in JPEG format, and in which said compression rate list includes the following set of compression rates: 40%, 50%, 60%, 70% and 96%.

[0093] At the end of substep 56, according to substep 58, we obtain a set Ei_C of compressed training images.

[0094] According to substep 60, a feature extraction 62 of Diffusion Noise Feature (DNF) is performed from each compressed image of said set Ei_C of compressed training images, and then said extracted DNF_F features are used, according to substep 64, as input to the ResNet-50 learning tool as introduced by H. Kaiming, et al. in the 2015 article entitled "Deep Residual Learning for Image Recognition", which provides as output a logit 68 (i.e., a logistic regression value) associated with each source training image of the first set Eh

[0095] Hereafter, "Logit" refers to a real value, a positive logit being, for example, representative of an image generated (i.e., falsified, a deepfake) by artificial intelligence, while, according to the same example, a negative logit is representative of an authentic image (i.e., not generated or falsified by artificial intelligence). In other words, the sign of the logit allows a digital image to be classified as a deepfake or not.

[0096] At the end of step 52, the DNF model based on the diffusion noise characteristic is trained when the set of images of said first training image set Ei has been used, so that the use of the DNF model achieves a predetermined success rate, for example on the order of 84%, for recognizing the images of said first training image set Ei that are indeed deepfakes (i.e. generated or falsified via artificial intelligence) from those which are not, the "truth" associated with each of these training images being known.

[0097] The first set 44 of separate E_S training steps includes another 70 training step of another detection model, distinct from the DNF model based on the diffusion noise characteristic, namely, as in the example of [Fig.3], the DCT model based on the discrete cosine transform as described by Corvi, Riccardo, et al. in the article entitled “Intriguing properties of synthetic images: from generative adversarial networks to diffusion models” of 2023.

[0098] To do this, step 70 includes substeps 72, 74, 76, 78, 80 and 82. Substep 72 is a retrieval substep of the first set Ei of training images, stored within the storage module 12 of [Fig.1].

[0099] According to substep 74, each image of the first set Ei of training images is compressed according to a compression rate randomly selected from the list of predetermined compression rates (including in particular the following set of compression rates: 40%, 50%, 60%, 70% and 96%), the rate being randomly identical or not to that which is used, according to substep 56, in the separate training 52 of the DNF model based on the diffusion noise characteristic.

[0100] In other words, the same image from the first set Ei is, according to a first example, compressed, according to step 56, with a compression rate of 40% while this image is compressed separately, according to step 74 with a compression rate of 60%, the rates of steps 56 and 74 being selected randomly (i.e. at random) from the same list of predetermined compression rates.

[0101] According to a second example, the same image from the first set Ei is compressed, according to step 56, with a compression ratio of 50%, while this image is compressed separately, according to step 74, with a compression ratio of 96%. The ratios in steps 56 and 74 are selected randomly from the same list of predetermined compression ratios.

[0102] According to a third example, the same image from the first set Ei is, according to a first example, compressed, according to step 56, with a compression rate of 70%, and randomly, identically according to step 74. The rates of steps 56 and 74 being selected randomly (i.e. at random) from the same list of predetermined compression rates.

[0103] At the end of substep 74, according to substep 76, we obtain a set Ei_C of randomly compressed training images.

[0104] According to substep 78, a feature extraction 80 DCT based on the discrete cosine transform is put implemented from each compressed image of said set Ei_C of compressed training images, then said extracted DCT_F features are used, according to substep 82, as input to the aforementioned ResNet-50 learning tool, which provides as output a logit 84 (i.e., a logistic regression value) associated with each source training image of the first set Eh

[0105] At the end of step 70, the DCT model based on the discrete cosine transform is trained when the set of images of said first set Ei of training images has been used, so that the use of the DCT model reaches a predetermined success rate, for example of the order of 76%, to recognize the images of said first set Ei of training images which are indeed deepfakes (i.e. generated or falsified via artificial intelligence) from those which are not, the "truth" associated with each of these training images being known.

[0106] The first set 44 of separate E_S training steps includes another 86 training step of another detection model, distinct from the DNF model based on the diffusion noise characteristic and distinct from the DCT model based on the discrete cosine transform, namely, according to the example in [Fig.3], the SSP model based on a single simple patch as described by Chen, Jiaxuan, et al. in the article entitled "A single simple patch is all you need for ai-generated image detection" of 2024.

[0107] To do this, step 86 includes substeps 88, 90, 92, 94, 96 and 98. Substep 88 is a retrieval substep of the first set Ei of training images, stored within the storage module 12 of [Fig.1].

[0108] According to substep 90, each image of the first set Ei of training images is compressed according to a compression rate randomly selected from the list of predetermined compression rates (including in particular the following set of compression rates: 40%, 50%, 60%, 70% and 96%), the rate being randomly identical or not to that which is used respectively, according to substeps 56 or 74, in the separate training 52 of the DNF model based on the diffusion noise characteristic or in the separate training 70 of the DCT model based on the discrete cosine transform.

[0109] At the end of substep 90, according to substep 92, an ELC set of randomly compressed training images is obtained.

[0110] According to substep 94, a single-simple patch (SSP) feature extraction is implemented from each compressed image of said Ei_C set of compressed training images, and then said extracted SSP_F features are used, according to substep 98, as input to the aforementioned ResNet-50 learning tool, which provides as output a logit 100 (i.e., a logistic regression value) associated with each source training image of the first set Eh

[0111] At the end of step 86, the SSP model based on a simple and unique patch is trained when the set of images of said first set Ei of training images has been used, so that the use of the SSP model reaches a predetermined success rate, for example of the order of 79%, to recognize the images of said first set Ei of training images which are indeed deepfakes (i.e. generated or falsified via artificial intelligence) from those which are not, the "truth" associated with each of these training images being known.

[0112] The first set 44 of separate E_S training steps includes another 102 training step of another detection model, distinct from the DNF, DCT and SSP models, namely, as in the example of [Fig.3], the CLIP model based on Contrastive Language-Image Pre-Training as described by Cozzolino, Davide, et al. in the article entitled “Raising the Bar of Al-generated Image Detection with CLIP” of 2024.

[0113] To do this, step 102 includes substeps 104, 106, 108, 110, 112 and 114. Substep 104 is a retrieval substep of the first set Ei of training images, stored within the storage module 12 of [Fig.1].

[0114] According to substep 106, each image of the first set Ei of training images is compressed according to a compression rate randomly selected from the list of predetermined compression rates (including in particular the following set of compression rates: 40%, 50%, 60%, 70% and 96%), the rate being randomly identical or not to that which is used respectively, according to substeps 56 or 74 or 90, in the separate trainings 52, 70, 86 respectively of the DNF, DCT and SSP models.

[0115] At the end of substep 106, according to substep 108, we obtain a set Ei_C of randomly compressed training images.

[0116] According to substep 110, a CLIP feature extraction based on contrastive language-image pretraining is implemented from each compressed image of said ELC set of compressed training images, and then said extracted CLIP_F features are used, according to substep 114, as input to a Multi-Layered Perceptron (MLP), which provides as output a logit (i.e., a logistic regression value) associated with each source training image of the first set Eb

[0117] At the end of step 102, the CLIP model based on contrastive language-image pre-training is trained when the set of images of said first set Ei of training images has been used, so that the use of the CLIP model reach a predetermined success rate, for example of the order of 88%, to recognize the images of said first set Ei of training images which are indeed deepfakes (i.e. generated or falsified via artificial intelligence) from those which are not, the "truth" associated with each of these training images being known.

[0118] The first set 44 of separate I_S training steps includes another step 118 of training another detection model, distinct from the DNF, DCT, SSP and CLIP models, namely, according to the example in [Fig.3], the NF model without feature extraction (from the English Non Feature) with respect to the previously described DNF, DCT, SSP and CLIP models.

[0119] To do this, step 118 includes substeps 120, 122, 124 and 126. Substep 118 is a retrieval substep of the first set Ei of training images, stored within the storage module 12 of [Fig.1].

[0120] According to substep 122, each image of the first set Ei of training images is compressed according to a compression rate randomly selected from the list of predetermined compression rates (including in particular the following set of compression rates: 40%, 50%, 60%, 70% and 96%), the rate being randomly identical or not to that which is used respectively, according to substeps 56 or 74 or 90 or even 106, in the separate trainings 52, 70, 86 and 118 respectively of the DNF, DCT and SSP models.

[0121] At the end of substep 122, according to substep 124, an ELC set of randomly compressed training images is obtained.

[0122] According to substep 126, each compressed image of said set Ei_C of compressed training images is then used directly (i.e., without feature extraction) as input to the aforementioned ResNet-50 learning tool, which provides as output a logit 128 (i.e., a logistic regression value) associated with each source training image of the first set Eb

[0123] At the end of step 126, the feature-free NF model is trained when the set of images of said first set Ei of training images has been used, so that the use of the trained NF model reaches a predetermined success rate, for example of the order of 79%, to recognize the images of said first set Ei of training images which are indeed deepfakes (i.e. generated or falsified via artificial intelligence) from those which are not, the "truth" associated with each of these training images being known.

[0124] Steps 52, 70, 86, 102 and 118 are carried out independently of each other. These steps can be carried out in parallel, or successively in any order, or spaced out over time, or with implementation work of two steps in parallel, then two more in parallel, then the last training step not yet carried out, or with implementation of three steps in parallel then two steps in parallel, etc., depending on the computing resources (i.e. computer) necessary for said training available.

[0125] It should also be noted that another set of distinct detection models could be used instead of the five DNF, DCT, SSP, CLIP and NF models in the example in [Fig. 3]. For example, six models could be trained separately by also adding training for the DINO self-distillation model without a label, as described by Mr. Oquab in the 2023 paper entitled “DINOv2: Learning Robust Visual Features without Supervision”.

[0126] As an alternative, the DINO model could be used instead of one of the five aforementioned models DNF, DCT, SSP, CLIP and NF.

[0127] Fig. 4 now illustrates the use of these trained models for the second set of 46 training steps leading to the obtaining of 0_M_E of a trained metamodel M_E.

[0128] First, the second set 46 of training steps includes a sub-step 130 of retrieval of the second set E2 of training images, stored within the storage module 12 of the [Fig.1].

[0129] The second set E2 of training images is optionally and advantageously distinct from the first training set Ei used in the previous set 44 of separate training steps of the distinct detection models as discussed previously in relation to [Fig.3].

[0130] As an optional complement, said second set E2 of training images is the Synthbuster training image set.

[0131] The second set 46 of training steps then includes a step 132 of compression of each image of the second set of training images, according to a compression rate randomly selected from said list of predetermined compression rates (including in particular the following set of compression rates: 40%, 50%, 60%, 70% and 96%).

[0132] At the end of substep 132, according to substep 134, we obtain a set E2_C of randomly compressed training images.

[0133] Then according to an application step 136, each of said at least five trained models is applied to each compressed image of said second set of training images.

[0134] Following the same example as that of [Fig.3], the trained DNF model obtained at the end of the set 44 of training steps of [Fig.3] is therefore applied 136 to each compressed image of said second set of training images compressed, to provide as output, according to step 138, a DNF logit associated with each compressed image of said second set of training images.

[0135] Similarly, the trained DCT model obtained at the end of the set 44 of training steps of [Fig.3] is therefore applied 136 to each compressed image of said second set of compressed training images, to provide as output, according to step 140, a DCT logit associated with each compressed image of said second set of training images.

[0136] Similarly, the trained SSP model obtained at the end of the set 44 of training steps of [Fig.3] is therefore applied 136 to each compressed image of said second set of compressed training images, to provide as output, according to step 142, an SSP logit associated with each compressed image of said second set of training images.

[0137] Similarly again, the trained CLIP model obtained at the end of the set 44 of training steps of [Fig.3] is therefore applied 136 to each compressed image of said second set of compressed training images, to provide as output, according to step 144, a CLIP logit associated with each compressed image of said second set of training images.

[0138] Finally, similarly, the trained NF model obtained at the end of the set 44 of training steps of [Fig.3] is therefore applied 136 to each compressed image of said second set of compressed training images, to provide as output, according to step 144, a CLIP logit associated with each compressed image of said second set of training images.

[0139] Then, according to step 148, a trained metamodel M_E is obtained, this trained metamodel M_E corresponding to an automatic machine learning tool O_A trained using, as input, for each compressed image of said second set of training images, said corresponding output logits of said at least five trained models. In other words, the machine learning tool O_A is trained with at least five times more inputs than there are images in the second set E2 of training images, an image of said second set E2 of training images being associated with at least five distinct logit inputs 138, 140, 142, 144 and 146.

[0140] According to a first option, said automatic O_A learning tool is a reinforcement model called gradient boosting. According to a second option, said automatic O_A learning tool is a decision tree forest (from the English Random Forest).

[0141] At the end of step 148, the metamodel M_E is trained when the set of images of said second set E2 of training images has been used, so that the use of the trained metamodel M_E reaches a predetermined success rate, for example, around 91% to recognize the images of said second set E2 of training images which are indeed deepfakes (i.e. generated or falsified via artificial intelligence) from those which are not, the "truth" associated with each of these training images being known.

[0142] A person skilled in the art will understand that the invention is not limited to the embodiments described, nor to the particular examples of the description, the embodiments and variants mentioned above being capable of being combined with each other to generate new embodiments of the invention.

[0143] The present invention thus makes it possible to combine several complementary detectors (i.e., distinct detection models), for example, the combination of the at least five aforementioned DNF, DCT, SSP, CLIP and NF models based on their confidence per instance (i.e., per image) in order to retain the most confident prediction and takes advantage, in terms of robustness, of the compression of the advantageously distinct training image sets, one being dedicated to the separate training of the detectors and the other to the training of the metamodel as such.

[0144] Such a metamodel makes it possible to take advantage of the performance of each of the detectors (e.g. the said at least five models DNF, DCT, SSP, CLIP and NF), while reducing the weaknesses to improve overall the prediction (i.e. detection) performance of whether an image is generated via artificial intelligence or not.

[0145] Compared to each detector taken separately, the proposed solution provides better detection quality for authentic images versus deepfakes. For example, by way of comparison, the DNF model performs 84% ​​on its own, the DCT model performs 76% on its own, the SSP model performs 79% on its own, the CLIP model performs 88% on its own, and the metamodel according to the present invention performs 91%.

Claims

1.

2. Demands Method (40) for detecting at least one image generated by artificial intelligence, the method being implemented by an electronic device for detecting at least one image generated by artificial intelligence, and comprising: - a training phase (42) comprising the following steps: - separate training (44) of at least five distinct models for detecting at least one image generated by artificial intelligence, using, for each of said separate trainings of said at least five distinct models, the same first set of training images, each image of said first set of training images being, before being used as input to each of said at least five distinct models, compressed according to a compression rate selected randomly from a list of predetermined compression rates; - compression (132) of each image of a second set (E2) of training images, according to a compression rate randomly selected from said list of predetermined compression rates; - application (136) of each of said at least five trained models, to each compressed image of said second set of training images, and obtaining, at the output of each of said at least five trained models, an associated logit; - obtaining (148) a trained metamodel (M_E) corresponding to an automatic learning tool (O_A) trained using, as input, for each compressed image of said second set of training images, said corresponding output logits of said at least five trained models; - a phase (48) of inference applying, said trained metamodel, to an input test image and capable of providing, in output, a detection result classifying said test image as generated by artificial intelligence or not. A detection method (40) according to claim 1, wherein each of the images of said first set of training images and of said second set of training images is in JPEG format, and wherein said compression rate list includes the following compression ratios: 40%, 50%, 60%, 70% and 96%.

3. A detection method (40) according to claim 1 or 2, wherein said at least five distinct detection models of at least one AI-generated image belong to the group comprising at least: - the DNF model based on the diffusion noise characteristic; - the SSP model based on a single, simple patch; - the DCT model based on the discrete cosine transform; - the CLIP model based on contrastive language-image pre-training; - an NF model without feature extraction; - a DINO self-distillation model without a label.

4. A detection method (40) according to any one of the preceding claims, wherein said first set of training images and said second set of training images are distinct.

5. A detection method (40) according to claim 4, wherein said first training image set is the Genlmage training image set.

6. A detection method (40) according to claim 4 or 5, wherein said second training image set is the Synthbuster training image set.

7. A detection method (40) according to any one of the preceding claims, wherein said machine learning tool is a reinforcement model called gradient boosting.

8. A detection method (40) according to any one of the preceding claims, wherein said machine learning tool is a decision tree forest.

9. A program comprising software instructions which, when executed by a computer, implement a method for detecting at least one image generated by artificial intelligence according to any one of the preceding claims.

10. Electronic device (10) for detecting at least one image generated by artificial intelligence, the electronic device (10) being configured to implement a method according to any one of claims 1 to 8.

Images