LEARNING METHOD A NEURAL NETWORK AND COMPUTER PROGRAM FOR IMPLEMENTING THIS METHOD

DE602020041592T2Active Publication Date: 2024-11-20BULL SA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
DE602020041592
Authority / Receiving Office
DE · DE
Patent Type
Patents
Current Assignee / Owner
Filing Date
2020-06-02
Publication Date
2024-11-20
Estimated Expiration
2040-06-02

AI Technical Summary

Technical Problem

Current methods to enhance neural network robustness against adversarial attacks, such as adversarial learning, often degrade performance on non-adversarial images.

Method used

A method that trains neural networks using a learning base comprising both adversarial and non-adversarial images, where adversarial images are paired with target data and non-adversarial images are used to maintain performance on non-adversarial data, allowing for improved robustness without degrading performance on non-adversarial images.

Benefits of technology

This approach enhances neural network robustness against adversarial attacks while maintaining or improving performance on non-adversarial images, effectively addressing the trade-off in existing adversarial learning methods.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader
Need to check novelty before this filing date? Find Prior Art

Description

[0001] The present invention relates to a method for training a neural network used for image processing. It also relates to a computer program configured to implement such a method.

[0002] The field of the invention is the field of neural networks used for image processing. State of the art

[0003] Nowadays, neural networks are widely used in the field of automated image processing, mainly for automated image classification, or for automated object recognition in images. To do this, a neural network is first trained, during a learning phase, on a base of images, called the learning base, then its performance is tested on a base of images, called the test base: the latter can be partially or totally different from the learning base.

[0004] However, sometimes an image intended to be processed by a neural network includes a modification, sometimes barely visible, called an "adversarial attack", introduced into the image deliberately with the aim of disrupting the neural network. When an adversarial attack aims to direct the response of the neural network towards a given target, then it is called a "targeted adversarial attack".

[0005] The robustness of a neural network against adversarial attacks is measured by comparing the performance of this neural network obtained on images not containing an adversarial attack, called non-adversarial images, to the performance of said neural network obtained on these same images in which an adversarial attack has previously been introduced, these images being called “adversarial images”.

[0006] The traditional solution to improve the robustness of a neural network to adversarial attacks is to use a training base comprising images containing adversarial attacks: such training is called "adversarial learning". However, adversarial learning has the disadvantage of degrading the performance of the neural network on non-adversarial images.

[0007] An aim of the present invention is to remedy this drawback.

[0008] Another aim of the present invention is to propose a method for learning a neural network used for image processing making it possible to improve the robustness of said neural network against adversary attacks while avoiding, or at least limiting, the degradation of the performance of said neural network on non-adversary images.

[0009] Hongyi Zhang ET AL: "mixup: Beyond Empirical Risk Minimization", discloses an adversarial learning method for an image classifier. Statement of the invention

[0010] The invention makes it possible to achieve at least one of these aims by a method according to claim 1.

[0011] In this document, “image” means a digital image or digital picture, and in particular digital data representing an image.

[0012] An "adversarial attack" refers to a modification, sometimes barely visible, introduced into an image with the aim of disrupting a neural network that processes this image.

[0013] A "targeted adversarial attack" refers to an adversarial attack whose goal is to steer the neural network's response toward a target. For example, an adversarial attack targeting "tree" is designed to trick the neural network into thinking that the processed image is a tree.

[0014] An "adversary image" means an image containing an adversary attack.

[0015] A "non-adversary image" means an image that does not contain an adversary attack.

[0016] By "adversarial learning" we mean training a neural network using an image database including adversarial images and, possibly, non-adversarial images.

[0017] By "adversarial learning step" we mean a learning step using an adversarial image.

[0018] By "non-adversarial learning step" we mean a learning step using a non-adversarial image.

[0019] For at least one adversary image, the result data and the target data can be stored together in the same data, called adversary label, indicated to said neural network during the adversary learning step using said adversary image.

[0020] In this case, the provision of the result data and the target data is carried out simultaneously in a single operation. Indeed, the result data and the target data are indicated to the neural network simultaneously, in a single operation, indicating to the neural network the adversary label which includes both data.

[0021] Alternatively, for at least one adversary image, the result data and the target data can be stored individually. In this case, the result data and the target data can be indicated to the neural network simultaneously, or in turn.

[0022] The method according to the invention may further comprise at least one learning step which provides as input to the neural network, an image, called non-adversarial, not comprising an adversarial attack.

[0023] Such a learning step can be called a non-adversarial learning step.

[0024] Such a training step does not provide target data to the neural network since the non-adversarial image does not include an adversarial attack.

[0025] In particular, such a non-adversarial learning step provides the neural network with only result data associated with the non-adversarial image provided to the neural network during said learning step.

[0026] For at least one non-adversarial image, the result data can be called a non-adversarial label.

[0027] According to another aspect of the present invention, there is provided an image base according to claim 4.

[0028] The learning base according to the invention makes it possible to carry out learning of a neural network, used for image processing, making it possible to improve the robustness of said neural network against adversary attacks while avoiding, or at least limiting, the degradation of the performance of said neural network on non-adversary images.

[0029] According to one embodiment, the learning base according to the invention may comprise only adversarial images.

[0030] According to another embodiment, the learning base according to the invention may comprise: at least one image, called non-adversary, not comprising an adversary attack; and for at least one, in particular each, non-adversary image, a first data item, called result data item, stored in association with said non-adversary image and intended to indicate to said neural network the expected result for said non-adversary image.

[0031] Such a base allows for better learning of the neural network, and to maintain better performance on non-adversarial images when using the neural network.

[0032] Two adversarial images may include the same adversarial attack, that is, the same targeted modification aimed at directing the neural network towards the same erroneous result, or the same target, for each of the said images.

[0033] In this case, these two adversary images can have the same content, that is, the same result data, or different result data.

[0034] Alternatively, or in addition, two adversarial images may include different adversarial attacks, i.e., different targeted modifications aimed at directing the neural network toward different erroneous outcomes, or different targets.

[0035] In this case, these two adversary images can have the same content, that is, the same result data, or different result data.

[0036] According to another aspect of the present invention, there is provided a computer program comprising instructions, which when executed by an electronic and / or computer device, implement the learning method according to the invention.

[0037] The computer program can be coded with any type of computer language, such as for example in C, C++, JAVA, Python, etc.

[0038] The computer program can be stored in a computer or electronic device.

[0039] Alternatively, the computer program may be stored on a medium readable by a computer or electronic device, such as a memory card or a USB key for example. In this case, the invention also relates to the medium storing said computer program.

[0040] According to another aspect of the present invention, which is not covered by the claims, there is provided a neural network trained by the learning method according to the invention.

[0041] The neural network can be any type of neural network that can be used for image processing and can be trained in a supervised manner.

[0042] In particular, the neural network can be a feed forward neural network, for example a single perceptron or multi-layer perceptron, a recurrent neural network, a resonant neural network, etc. Description of figures and embodiments

[0043] Other advantages and characteristics will appear on examining the detailed description of a non-limiting embodiment, and the attached drawings in which: there FIGURE 1 is a schematic representation of a non-limiting exemplary embodiment of an adversarial learning step that can be implemented in a method according to the invention; FIGURE 2 is a schematic representation of a non-limiting exemplary embodiment of a non-adversarial learning step that can be implemented in a method according to the invention; and the FIGURE 3is a schematic representation of a non-limiting exemplary embodiment of a method according to the invention.

[0044] It is understood that the embodiments which will be described below are in no way limiting. In particular, it is possible to imagine variants of the invention comprising only a selection of characteristics described below isolated from the other characteristics described, if this selection of characteristics is sufficient to confer a technical advantage or to differentiate the invention compared to the state of the prior art. This selection includes at least one preferably functional characteristic without structural details, or with only a part of the structural details if this part is only sufficient to confer a technical advantage or to differentiate the invention compared to the state of the prior art.

[0045] In the figures, elements common to several figures retain the same reference.

[0046] There FIGURE 1 is a schematic representation of a non-limiting exemplary embodiment of an adversarial learning step that can be implemented in a method according to the invention.

[0047] Learning step 100, shown in the FIGURE 1 , allows training a neural network 102 with an adversary image 104.

[0048] The neural network 102 may be any type of neural network that can be used for image processing and can be trained by supervised learning.

[0049] The adversary image 104 can be any type of image. It includes content, which in the present example is a dog. Of course, the content can be anything else. The content of the adversary image 104 can represent an object, or anything else such as a color or a shape. The content is indicated by a data item 106, called content data, stored in association with the adversary image 104. In the example shown, the content data 106 indicates for example "dog" because the adversary image 104 represents a dog.

[0050] In addition, the adversary image 104 includes a modification 108, called an adversary attack, visible or not to the eye. This adversary attack 108 can be a modification of any kind: modification of a color of the image, deletion of a part of the image, addition of information in the image, etc. This adversary attack 108 aims to disrupt the neural network 102 and to direct the response of the neural network 102 towards a given target, which is not the expected result for this image. For example, this adversary attack 108 aims to direct the response of the neural network 102 towards “tree” while the content of the adversary image 104 is a dog.

[0051] Furthermore, according to the invention, a piece of data 110, called target data, is stored in association with the adversary image 104 to indicate the target towards which the adversary attack 108 aims to orient the neural network 102. In other words, this target data 110 makes it possible to indicate to the neural network 102 an erroneous result that the neural network must not return when it receives the adversary image 104 as input.

[0052] The supervised adversarial learning step 100 of the FIGURE 1 includes an operation 112 of providing the adversary image 104 as input to the neural network 102.

[0053] In addition, the supervised adversarial learning step 100 further comprises an operation 114 of providing, to said neural network 102, the result data indicating the expected result for said adversarial image 104.

[0054] Finally, the supervised adversarial learning step 100 comprises an operation 116 of providing, to said neural network 102, the target data 110 indicating, to the neural network 102, the target of the adversarial attack used to modify the image 104.

[0055] The operations 112-116 can be performed in turn or simultaneously. Indeed, the adversary image 104, the result data 106 and the target data 110 can be provided as input to the neural network 102, simultaneously, for example as input parameters.

[0056] In particular, the operations 114-116 of providing the result data 106 and target data 110 can be performed simultaneously. In this case, according to one embodiment, the result data 106 and the target data 110 can be stored individually, for example respectively as a result label and a target label. Alternatively, the result data 106 and the target data 110 can be stored together, for example in a single label.

[0057] There FIGURE 2 is a schematic representation of a non-limiting exemplary embodiment of a non-adversarial learning step that can be implemented in a method according to the invention.

[0058] Learning step 200, shown in the FIGURE 2 , allows training the neural network 102 with a non-adversarial image 204.

[0059] The non-adversarial image 204 can be any type of image. It includes content, which in the present example is a dog. Of course, the content of the non-adversarial image 204 can represent an object, or anything else such as a color or a shape. The content is indicated by a data item 206, called content data, stored in association with the non-adversarial image 204. In the example shown, the content data 206 indicates for example "dog" because the adversarial image 104 represents a dog.

[0060] In particular, in the example shown the non-adversary image 204 corresponds to the adversary image 104 of the FIGURE 1 , without the adversary attack 108. Of course, this example is in no way limiting and the content of the non-adversary image may be different from the content of the adversary image.

[0061] The supervised non-adversarial learning step 200 of the FIGURE 2includes an operation 212 of providing the non-adversarial image 204 as input to the neural network 102.

[0062] In addition, the supervised non-adversarial learning step 200 further comprises an operation 214 of providing, to said neural network 102, the result data indicating the expected result for said non-adversarial image 104.

[0063] Operations 212-214 may be performed in turn or simultaneously. Indeed, the non-adversarial image 204 and the result data 206 may be provided as input to the neural network 102, simultaneously, for example as input parameters.

[0064] There FIGURE 3 is a schematic representation of a non-limiting exemplary embodiment of a learning method according to the invention.

[0065] The 300 process of the FIGURE 3is used to train a neural network in a supervised manner, such as for example the 102 neural network of FIGURES 1 and 2

[0066] To do this, the method uses an image base 302 according to the invention comprising adversary images, and possibly non-adversary images.

[0067] In the example shown, the image base 302 comprises adversary images, such as for example the adversary image 104 of the FIGURE 1 , and non-adversarial images, such as for example the non-adversarial image 204 of the FIGURE 2 .

[0068] The method 300 comprises one or more iterations of an adversarial learning step 304, each iteration being carried out with an adversarial image stored in the image base 302.

[0069] The adversarial training step 304 may correspond to the adversarial training step 100 of the FIGURE 1 .

[0070] The method 300 further comprises one or more iterations of a non-adversarial learning step 306, each iteration being performed with a non-adversarial image stored in the image database 302.

[0071] The non-adversarial learning step 306 may correspond to the non-adversarial learning step 200 of the FIGURE 2 .

[0072] The iteration(s) of the adversarial learning step 304 and the iteration(s) of the non-adversarial learning step 306 may be performed in any order, for example alternating, or in turn, etc. Alternatively, all of the iterations of the adversarial learning step 304 may be performed before the iteration(s) of the non-adversarial learning step 306.

[0073] Generally, the iterations of the adversarial learning step 304 and the non-adversarial learning step 306 may be performed and chained according to any other scheme than those indicated.

[0074] The number of iteration(s) of the adversarial learning step 304 may be the same as, or different from, the number of iteration(s) of the non-adversarial learning step 306.

[0075] Of course, the invention is not limited to the examples detailed above.

[0076] In particular, the images used are not limited to the examples described.

[0077] According to alternatives not shown, the method according to the invention may not include a non-adversarial learning step.

[0078] According to alternatives not shown, the base according to the invention may not include a non-adversarial image.

Claims

1. A supervised adversarial learning method (300) for a neural network (102), comprising at least one iteration of a learning step (100;304), called adversarial learning step, comprising the following operations: - supplying, to said neural network (102), an image (104), called adversarial image, containing a modification (108), called adversarial attack, provided to orient said neural network (102) towards a result of a classification of the image or a recognition of objects in the image, the result being called target, different from an expected result, and corresponding to the response towards which the adversarial attack is intended to orient the neural network and - supplying, to said neural network (102), a first data item, called result data item, indicating the expected result for said adversarial image (104), said expected result corresponding to the desired response associated with this adversarial image; characterized in that said step (100;304) of adversarial learning also comprises supplying, to said neural network (102), a second data item associated with said image (104), called target data item, indicating said target to said neural network (102), the target forming an incorrect result of a classification of the image or a recognition of objects in the image that the neural network (102) should not return when it receives said adversarial image (104) as input.

2. The method (300) according to the preceding claim, characterized in that, for at least one adversarial image (104), the result data item and the target data item are stored together in one and the same data item, called adversarial label, supplied to said neural network (102) during the adversarial learning step (100;304) utilizing said adversarial image (104).

3. The method (300) according to any one of the preceding claims, characterized in that it also comprises at least one learning step (200;306) that supplies as input of the neural network (102) an image, called non-adversarial image, not containing an adversarial attack.

4. A set of images (302), stored on a storage medium, provided to be utilized in a supervised adversarial learning method of a neural network, comprising: - at least one image (104), called adversarial image, containing a modification (108), called adversarial attack, provided to orient said neural network (102) towards a result of a classification of the image or a recognition of objects in the image, the result being called target, different from an expected result, and corresponding to the response towards which the adversarial attack is intended to orient the neural network; and - for at least one adversarial image (104), a first data item, called result data item, stored in association with said adversarial image (104), provided to be supplied to said neural network (102) and indicate to said neural network (102) the expected result for said adversarial image (104), said expected result corresponding to the desired response associated with this adversarial image; characterized in that it comprises, for at least one adversarial image (104), a second data item, called target data item, stored in association with said adversarial image (104), provided to be supplied to said neural network (102) and indicate said target to said neural network (102), the target forming an incorrect result of a classification of the image or a recognition of objects in the image that the neural network (102) should not return when it receives said adversarial image (104) as input.

5. The set according to the preceding claim, characterized in that it comprises only adversarial images.

6. The set (302) according to claim 4, characterized in that it also comprises: - at least one image (204), called non-adversarial image, not containing an adversarial attack; and - for at least one non-adversarial image (204), a first data item, called result data item, stored in association with said non-adversarial image (204), and provided to indicate to said neural network (102) the expected result for said non-adversarial image (204).

7. The set (302) according to any one of claims 4 to 6, characterized in that two adversarial images comprise the same adversarial attack, or two different adversarial attacks.

8. A computer program comprising instructions, which when they are executed by an electronic and / or computerized appliance, implement the learning method (300) according to any one of claims 1 to 3.