EP3903236A1 - Procédé de génération d'un exemple adversaire pour un modèle d'apprentissage automatique, procédé de robustification, dispositif de génération et programme produit ordinateur associés - Google Patents
Procédé de génération d'un exemple adversaire pour un modèle d'apprentissage automatique, procédé de robustification, dispositif de génération et programme produit ordinateur associésInfo
- Publication number
- EP3903236A1 EP3903236A1 EP19828796.3A EP19828796A EP3903236A1 EP 3903236 A1 EP3903236 A1 EP 3903236A1 EP 19828796 A EP19828796 A EP 19828796A EP 3903236 A1 EP3903236 A1 EP 3903236A1
- Authority
- EP
- European Patent Office
- Prior art keywords
- input data
- machine learning
- modified
- initial
- learning model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/09—Supervised learning
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/094—Adversarial learning
Definitions
- the present invention relates to a method for generating an adversary example for a machine learning model.
- the present invention also relates to a method for robustification of the machine learning model, an associated generation device and a computer program product.
- the invention relates to the field of artificial intelligence and more particularly machine learning models which are based on statistical approaches in order to improve the performance of the model in solving tasks without being explicitly programmed for each. of these tasks.
- Machine learning has two phases.
- the first phase consists in defining a model from data present in a database, called observations. Estimating the model involves performing a practical task, such as translating a speech, estimating a probability density, recognizing the presence of an object in an image or classifying a digital file, for example. This so-called learning phase is generally carried out before the practical use of the model.
- the second phase corresponds to the use of the model: the model being defined, new data outside the database can then be submitted to the model in order to obtain the result corresponding to the desired task.
- An adversary example is an input data item for which the associated result is different from the result associated with another neighboring input data item.
- the adversary example is therefore a particular problematic case for the machine learning model. Indeed, a slight variation in the input data should lead to a similar prediction if the model was sufficiently robust. For example, changing the brightness of an image should not change the objects identified on the image or adding a few words to a file should not change whether it contains a virus or no.
- these adversary examples are not entirely satisfactory, in particular in the field of cyber security.
- the object of the present invention is to propose a method making it possible to further improve the relevance of the adversary examples of the machine learning model.
- the subject of the invention is a method of generating at least one adversary example for a machine learning model, the machine learning model being associated with a database, the machine learning model associating a prediction of at least one input data, each input data comprising at least one digital component, the generation method comprising, for at least one initial input data chosen from the input data from the database data, the following steps:
- the generation process comprises one or more of the following characteristics, taken alone or in any technically possible combination:
- each input data item is associated with a digital file comprising a series of characters, the generation method comprising a step of converting the series of characters into the digital component or components of the input data item;
- the method further comprises before the generation steps, a step of defining a domain constraint to the variation function, so that each component of the modified input data belongs to a predetermined associated definition domain;
- the method further comprises, before the generation steps, a step of defining a threshold constraint on the variation function, so that the difference between each component of the modified input data and the corresponding component of the initial input data is less than a predetermined associated threshold;
- the generation of the adversary example is carried out by a gradient-based optimization algorithm.
- the subject of the invention is also a method of robustification of the machine learning model, comprising the steps of the tekl generation method defined above, the robustification method further comprising the following steps:
- the robustification method comprises a second iteration of all the steps of the robustification method as defined above from the enriched database.
- the invention also relates to a device for generating at least one adversary example for a machine learning model, the machine learning model being associated with a database, the machine learning model associating with at least an input data a prediction, each input data comprising at least one digital component, the electronic generation device being, for at least one initial input data chosen from the input data coming from the database, specific to :
- the invention also relates to a computer program product comprising software instructions which, when executed by a computer, implement a generation process as defined above.
- FIG. 1 is a schematic representation of an electronic assembly comprising an electronic generation device according to the invention.
- FIG. 2 is a flowchart of a robustification method according to the invention.
- the electronic assembly 10 comprises at least one database 12, an electronic modeling device 14, an electronic generation device 16 and an electronic robustification device 18.
- the database 12 is able to store a plurality of data.
- Each piece of data is for example a digital file, an image, a video, a sound, etc.
- the modeling device 14 is suitable for executing a machine learning model.
- the modeling device 14 is suitable for applying the machine learning model to at least one input data item X in order to obtain an associated prediction Y.
- the machine learning model is, for example, a neural network well known to those skilled in the art.
- the associated prediction Y is for example a text describing at least one object present on the image.
- the prediction Y is for example a translation into another language of the text of the file, a classification of the content of the file or an anti-virus analysis of the file.
- Each input data X is a vector or an array comprising at least one component Xi. At least one of the components Xi is numeric. Advantageously, all the components Xi of the input data X are digital.
- the modeling device 14 is in the form of an independent computer further comprising at least one processor and a memory.
- the modeling device 14 is at least partially in the form of a programmable logic circuit of FPGA type (from English "Field-Programmable Gate Array") and / or software stored in the computer's memory and executable by the computer's processor.
- the modeling device 14 is implemented by another computer.
- the modeling device 14 is at least partially in the form of software stored in a memory of such a computer and executable by a processor thereof.
- the generation device 16 is capable of generating at least one adversary example EA for the machine learning model.
- the generation device 16 comprises a first acquisition module 20, a first processing module 22 and a first output module 24.
- the first acquisition module 20 makes it possible to acquire all of the external data from the generation device 16 necessary for the operation thereof.
- the first acquisition module 20 is suitable for receiving input data X from the database 12.
- the first processing module 22 is able to process the data acquired by the acquisition module 20 using the modeling device 14 in order to implement a method 200 of generating the at least one adversary example EA for the machine learning model.
- the first output module 24 is able to transmit all of the data processed by the first processing module 22 to the robustification device 18.
- the generation device 16 is in the form of an independent computer further comprising at least one processor and a memory.
- the aforementioned modules are at least partially in the form of programmable logic circuits of FPGA type (from the English “Field-Programmable Gate Array”) and / or software stored in the memory of the computer and executable by the processor of it.
- the generation device 16 is implemented by another computer.
- the aforementioned modules are at least partially in the form of software stored in a memory of such a computer and executable by a processor thereof.
- the robustification device 18 is suitable for implementing the robustification of the machine learning model.
- Robustness of the model means an optimization of the model by which the model is made more robust and less sensitive to the effects of variability in the data and of the model's input parameters.
- the robustification device 18 comprises a second acquisition module 26, a second processing module 28 and a second output module 30.
- the second acquisition module 26 makes it possible to acquire all of the external data from the robustification device 18 necessary for the operation thereof.
- the second acquisition module 26 is capable of receiving at least one adversary example EA from the generation device 16.
- the second processing module 28 is able to process the data acquired by the second acquisition module 26 in order to implement a method 300 for learning the machine learning model from the database 12.
- the second output module 30 is able to transmit all of the data processed by the second processing module 28, and in particular to transmit the robust machine learning model, to the modeling device 14.
- the robustification device 18 is in the form of an independent computer further comprising at least one processor and a memory.
- the aforementioned modules are at least partially in the form of programmable logic circuits of FPGA type (from the English “Field-Programmable Gate Array”) and / or software stored in the memory of the computer and executable by the processor of it.
- modeling devices 14, generation 16 and robustification 18 are alternatively in the form of a single computer.
- the robustification device 18 is implemented by another computer.
- the aforementioned modules are at least partially in the form of software stored in a memory of such a computer and executable by a processor thereof.
- a robustification method 100 will now be explained with reference to FIG. 2 presenting a flow diagram of its steps.
- the robustification method 100 includes the generation method 200 followed by the learning method 300.
- an optional initial conversion step 210 at least one data item included in the database 12 is converted into an input data item X comprising at least one digital component Xi.
- each file is transformed into a digital matrix indicating for each character its place in the fragment of text analyzed.
- the rest of the robustification method 100 is described for an initial input data X chosen from the input data X from the database 12. Those skilled in the art will understand that the robustification method 100 is applicable to n ' any input data X from database 12.
- the generation device 16 defines a variation function F of the initial input data X defining a plurality of modified input data X ’from the initial input data X:
- the function F makes it possible to obtain a plurality of modified input data X ’in the vicinity of the initial input data X by slightly varying the latter.
- the generation method 200 comprises a step 230 of defining a direction constraint to the variation function F so that the difference between at least component Xi 'of each modified input data X' and the component Xi corresponding to it of the initial input data X is of predetermined sign Si.
- the predetermined sign Si is positive or negative.
- the direction constraint at the variation function F is defined so that the difference between each component Xi 'of each modified input data X' and the component Xi corresponding to it of the initial input data X is of sign predetermined If.
- the direction constraint is defined so that the difference between each component Xi 'of the modified input data X' and the component Xi corresponding to it of the initial input data X is of sign If positive:
- Corresponding component is understood to mean the same component of the modified input data item X 'as that of the initial input data item X.
- component corresponding to it means the component having the same index .
- the direction constraint Cd is particularly advantageous when the input data X comes from a digital file and represents a number of elements contained in this file. Indeed, removing data from the file could break the file structure and make this file corrupt. Thus, with the direction constraint, only adding content is possible in the file, which guarantees the integrity of the file.
- the direction constraint Cd is defined so that the difference between each component Xi ’of the modified input data X’ and the component Xi corresponding to it of the initial input data X is negative.
- the method 200 comprises a step 240 of defining a domain constraint to the variation function F, so that each component Xi 'of the modified input data item X' belongs to a domain of associated predetermined definition.
- each component Xi is for example a numerical interval comprised between a minimum value ai and a maximum value b ⁇ :
- the definition domain is a discrete set of numerical values.
- the definition domain corresponds to a physical limitation of the input data such as a necessarily positive light intensity, to a structural limitation such as a maximum file size or a predetermined preference limitation.
- the method 200 comprises a step 250 of defining a threshold constraint on the variation function F, so that the absolute value of the difference between each component Xi 'of the modified input data X 'and the component Xi corresponding to it of the initial input data X is less than a predetermined associated threshold Ai:
- the threshold Ai is of value less than 5% of the value of the digital component Xi.
- variations in light intensity limited to 5% will generally be difficult to detect for the human eye.
- the method 200 comprises a step 260 of generating a plurality of modified input data X ′ from the initial input data X by the function of variation F constrained by the direction constraint and advantageously the domain constraint and the threshold constraint.
- the variation function is defined as follows:
- G is for example equal to the tangent hyperbolic function, to the sine function or a function resulting from the sigmoid function.
- the generation method 200 then comprises a step 270 of generating the adversary example EA corresponding to the modified input data X 'among the plurality of modified input data X' maximizing the probability P that a modified prediction Y ' obtained by applying the machine learning model to this modified input data X 'is different from an initial prediction Y obtained by applying the machine learning model to the initial input data X:
- the adversary example EA is the modified input data X ’for which the probability P that the associated prediction Y’ is different from the initial prediction Y is maximum.
- the generation of the adversary example EA is carried out by a gradient-based optimization algorithm.
- the optimization algorithm is iterative by varying the variable r and proceeds by successive improvements following the gradient until the maximum probability is found.
- the algorithm calculates the modified input data X 'as a function of the input data X.
- the algorithm calculates the modified prediction Y' associated by means of the modeling device 14 , then the algorithm compares the modified prediction Y 'with the initial prediction Y.
- the algorithm determines the gradient of variation of the probability that the modified prediction Y' is different from the initial prediction Y, causes the variable r to vary following the positive gradient and repeats the calculations until the maximum probability is reached.
- the generation process 200 is followed by the learning process 300 which will now be explained with reference to FIG. 2, in order to allow the robustness of the machine learning model.
- the learning method 300 comprises a first step 310 of adding the adversary example EA to the database 12 in order to obtain an enriched database 12.
- the enriched database 12 thus includes all the data initially present as well as the adversary example EA added.
- the robustification device 18 implements the learning of the machine learning model from the enriched database 12.
- the learning of the machine learning model is carried out by learning from the data from the database 12, for example by means of an iterative algorithm by error correction well known to those skilled in the art. .
- the machine learning model is readjusted according to the adversary example EA so that the modified prediction Y 'associated with the adversary example EA is no longer different from the initial prediction Y.
- the machine learning model is readjusted so that the model detects the same objects on the image of the adversary example EA and the image of the initial input data.
- the robustification module 18 transmits the adjusted machine learning model to the modeling module 14 which is thus updated with the improved version of the model.
- the robustification method 100 advantageously comprises a second iteration of all the steps of the methods 200 and 300 from the enriched database 12.
- the robustification method thus comprises the generation of at least one second adversary example EA different from that of the first iteration, the addition of this second adversary example EA to the enriched database 12 and the training of the model with this base. of data 12 further enriched.
- the invention improves the robustness of the machine learning model. Indeed, the generation of an EA adversary example taking into account the direction constraint makes it possible to enrich the database 12 with more relevant examples.
- the management constraint is particularly advantageous in the field of cybersecurity where the data are digital files.
- the direction constraint makes it possible to avoid obtaining malformed files which would be useless for the machine learning model.
- the method according to the invention thus makes it possible to make the machine learning model more robust in order to improve its operation and to prevent possible computer attacks against this model.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Machine Translation (AREA)
Abstract
Description
Claims
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| FR1874213A FR3091390A1 (fr) | 2018-12-27 | 2018-12-27 | Procédé de génération d’un exemple adversaire pour un modèle d’apprentissage automatique, procédé de robustification, dispositif de génération et programme produit ordinateur associés |
| PCT/EP2019/087106 WO2020136264A1 (fr) | 2018-12-27 | 2019-12-27 | Procédé de génération d'un exemple adversaire pour un modèle d'apprentissage automatique, procédé de robustification, dispositif de génération et programme produit ordinateur associés |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| EP3903236A1 true EP3903236A1 (fr) | 2021-11-03 |
Family
ID=67441207
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| EP19828796.3A Pending EP3903236A1 (fr) | 2018-12-27 | 2019-12-27 | Procédé de génération d'un exemple adversaire pour un modèle d'apprentissage automatique, procédé de robustification, dispositif de génération et programme produit ordinateur associés |
Country Status (3)
| Country | Link |
|---|---|
| EP (1) | EP3903236A1 (fr) |
| FR (1) | FR3091390A1 (fr) |
| WO (1) | WO2020136264A1 (fr) |
-
2018
- 2018-12-27 FR FR1874213A patent/FR3091390A1/fr active Pending
-
2019
- 2019-12-27 EP EP19828796.3A patent/EP3903236A1/fr active Pending
- 2019-12-27 WO PCT/EP2019/087106 patent/WO2020136264A1/fr not_active Ceased
Also Published As
| Publication number | Publication date |
|---|---|
| WO2020136264A1 (fr) | 2020-07-02 |
| FR3091390A1 (fr) | 2020-07-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| EP3301617B1 (fr) | Procédés d'apprentissage sécurisé de paramètres d'un réseau de neurones à convolution, et de classification sécurisée d'une donnée d'entrée | |
| US10878318B2 (en) | Adaptive artificial neural network selection techniques | |
| WO2016102153A1 (fr) | Representation semantique du contenu d'une image | |
| FR2994495A1 (fr) | Procede et systeme pour detecter des evenements sonores dans un environnement donne | |
| WO2023005386A1 (fr) | Procédé et appareil d'entraînement de modèle | |
| CN110321562B (zh) | 一种基于bert的短文本匹配方法及装置 | |
| EP3588301A1 (fr) | Determination automatique et auto-optimisee des parametres d'execution d'une application logicielle sur une plateforme de traitement de l'information | |
| CN107133951A (zh) | 图像篡改检测方法及装置 | |
| FR3123748A1 (fr) | Apprentissage automatique sans annotation ameliore par regroupements adaptatifs en ensemble ouvert de classes | |
| CN114332564A (zh) | 车辆分类方法、设备及存储介质 | |
| EP3903236A1 (fr) | Procédé de génération d'un exemple adversaire pour un modèle d'apprentissage automatique, procédé de robustification, dispositif de génération et programme produit ordinateur associés | |
| EP3712775A1 (fr) | Procédé et dispositif de détermination de la taille mémoire globale d'une zone mémoire globale allouée aux données d'un réseau de neurones compte tenu de sa topologie | |
| US20230056131A1 (en) | Server and method for classifying entities of a query | |
| EP3729768B1 (fr) | Procédé de construction automatique de scénarios d'attaques informatiques, produit programme d'ordinateur et système de construction associés | |
| FR3076935A1 (fr) | Procedes d'apprentissage de parametres d'un reseau de neurones a convolution, et de classification d'une donnee d'entree | |
| FR3101715A1 (fr) | Procédé et système multimodal d’identification d’objets | |
| EP4078416B1 (fr) | Procédé de contrôle d'un système électronique par compteurs de performance bas niveaux et comprenant au moins un ensemble d'applicatif(s) logiciels non maîtrisés s'exécutant sur un processeur et un dispositif de contrôle | |
| EP3038007B1 (fr) | Procédé de mise en oeuvre d'un algorithme de détermination de sûreté de fichiers informatiques, module de détermination, procédé de construction d'une unité de contrôle et produit programme d'ordinateur associés | |
| Aruna et al. | Byte Segmented Convolutional Neural Network for Network Traffic Classification | |
| EP4432175A1 (fr) | Procédé de paramétrage d'une chaîne de traitement de données | |
| WO2023180257A1 (fr) | Procédé et dispositif d'entraînement d'une structure de neurones artificiels et programme d'ordinateur associé | |
| Othman et al. | Insuring android software security with a machine learning approach for better malware detection | |
| Iliadis | Advancements in Machine Learning Techniques for Detecting Structured Query Language Injection Attacks–a Systematic Literature Review | |
| WO2024003016A1 (fr) | Procédé et dispositif d'estimation de l'authenticité d'un contenu audio ou vidéo et programme d'ordinateur associé | |
| EP4538933A1 (fr) | Procédé de compilation d'un réseau de neurones artificiels entraîné |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: UNKNOWN |
|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: THE INTERNATIONAL PUBLICATION HAS BEEN MADE |
|
| PUAI | Public reference made under article 153(3) epc to a published international application that has entered the european phase |
Free format text: ORIGINAL CODE: 0009012 |
|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: REQUEST FOR EXAMINATION WAS MADE |
|
| 17P | Request for examination filed |
Effective date: 20210624 |
|
| AK | Designated contracting states |
Kind code of ref document: A1 Designated state(s): AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR |
|
| DAV | Request for validation of the european patent (deleted) | ||
| DAX | Request for extension of the european patent (deleted) | ||
| P01 | Opt-out of the competence of the unified patent court (upc) registered |
Effective date: 20230516 |
|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: EXAMINATION IS IN PROGRESS |
|
| 17Q | First examination report despatched |
Effective date: 20240820 |