Method for implementing an electronic data-processing computer having at least one artificial intelligence model protected against distillation attacks, associated computer program and computer

By adding primary and secondary AI functions to the inputs and outputs of AI models and connecting processing modules for inverse alterations, the method enhances protection against distillation attacks, ensuring the models remain inaccessible to attackers.

WO2026233911A1PCT designated stage Publication Date: 2026-11-12THALES SA +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/EP2026/062771
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-05
Filing Date
2026-05-04
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Existing methods for protecting artificial intelligence models are not optimal against distillation attacks, which exploit the searchable and usable nature of these models by potential attackers.

Method used

Implementing a method that modifies initial AI models by adding primary and secondary AI functions to their inputs and outputs, respectively, making them inaccessible to external entities, and connecting processing modules to perform inverse alterations to compensate for these changes, thereby creating a protected AI model.

Benefits of technology

The modified AI models are rendered inaccessible to attackers, effectively thwarting distillation attacks by rendering input and output data as noise, while maintaining model functionality through inverse alterations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure EP2026062771_12112026_PF_FP_ABST
    Figure EP2026062771_12112026_PF_FP_ABST
Patent Text Reader

Abstract

The invention relates to a method for implementing an electronic computer (5), which method comprises: - training at least one initial artificial intelligence model (22); - modifying the initial model in order to obtain at least one modified model (20), by adding a primary function (30) to the input of the initial model in order to perform an inverse primary alteration, and / or by adding a secondary function (32) to the output of the initial model in order to perform a secondary alteration; and - generating a software application (15), by connecting a primary processing module (24) to the input of the modified model to perform a primary alteration (34) equal to an inverse operation of the inverse primary alteration, and / or by connecting a secondary processing module (26) to the output of the modified model to perform an inverse secondary alteration (36) equal to an inverse operation of the secondary alteration; - storing the software application on a computing platform (10) to form the computer.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] Method for implementing an electronic data processing computer with at least one artificial intelligence model protected against distillation attacks, computer program and associated computer

[0002] The present invention relates to a method of implementing an electronic data processing computer, the method being implemented by computer.

[0003] The invention also relates to a computer program, comprising software instructions which, when executed by a computer, implement such an implementation method.

[0004] The invention also relates to such an electronic data processing computer.

[0005] The invention relates to the field of cybersecurity, particularly in the avionics field.

[0006] The invention aims in particular to protect at least one artificial intelligence model against distillation attacks, these attacks being facilitated when the artificial intelligence model is searchable and usable by potential attackers.

[0007] Document EP 4369229 A2 describes various techniques for protecting artificial intelligence models.

[0008] However, these protection techniques are not always optimal against distillation attacks.

[0009] The aim of the invention is therefore to propose a method for implementing an electronic data processing computer comprising at least one artificial intelligence model benefiting from better protection against distillation attacks.

[0010] To this end, the invention relates to a method for implementing an electronic data processing computer, the method being implemented by computer and comprising the following steps:

[0011] - learning with training data from at least one initial artificial intelligence model, each having an input and an output;

[0012] - modification of at least one initial artificial intelligence model to obtain at least one modified artificial intelligence model, via the addition of at least one function chosen from:

[0013] a primary artificial intelligence function added to the input of a respective initial model and configured to perform a reverse primary alteration, the respective modified model then being an aggregation of the primary function and the respective initial model, and the input of the respective initial model becoming inaccessible to an electronic entity external to the respective modified model; and

[0014] a secondary artificial intelligence function added to the output of the respective initial model and configured to perform a secondary alteration, the respective modified model then being an aggregation of the respective initial model and the secondary function, and the output of the respective initial model becoming inaccessible to the electronic entity external to the respective modified model; and

[0015] - generation of a processing software application, by connecting at least one processing module chosen from among the following to each modified model:

[0016] if the respective modified model includes the addition of the primary artificial intelligence function to the input of the respective initial model: a primary processing module connected to the input of the respective modified model and configured to perform a primary alteration equal to an inverse operation of the inverse primary alteration; and

[0017] if the respective modified model includes the addition of the secondary artificial intelligence function to the output of the respective initial model: a secondary processing module connected to the output of the respective modified model and configured to perform an inverse secondary alteration equal to an inverse operation of the secondary alteration;

[0018] the software application comprising at least one processing module and at least one modified model;

[0019] - storage of the processing software application on a computer platform capable of running said application, to form the electronic data processing computer.

[0020] With the method according to the invention, at least one initial artificial intelligence model is protected against distillation attacks by adding at least one artificial intelligence function to the input and / or output of the initial model, thereby rendering the input and / or output inaccessible to a potential attacker. The at least one artificial intelligence function performs an alteration to protect the input and / or output of the initial model. Finally, at least one processing module performs a reverse alteration to compensate for said alteration, which is unknown to the potential attacker.

[0021] According to other advantageous aspects of the invention, the implementation method comprises one or more of the following features, taken individually or in all technically possible combinations:

[0022] - during the modification step, the primary artificial intelligence function and the secondary artificial intelligence function are added to the initial artificial intelligence model to obtain the modified artificial intelligence model, and during the generation step, the primary processing module and the secondary processing module are connected to the modified model; the software application comprising the primary processing module, the modified model, and the secondary processing module;

[0023] - during the modification step, a first and a second initial artificial intelligence models are modified to obtain a first and a second modified artificial intelligence models; the first modified model being an aggregation of the first initial model and a first secondary artificial intelligence function, the first secondary function being added to the output of the first initial model and configured to perform a first secondary alteration, and the output of the first initial model becoming inaccessible to the electronic entity external to the first modified model;the second modified model being an aggregation of a second primary artificial intelligence function and the second initial model, the second primary function being added to the input of the second initial model and configured to perform a second inverse primary alteration, and the input of the second initial model becoming inaccessible to an electronic entity external to the second modified model;

[0024] During the generation step, a first secondary processing module is connected to the output of the first modified model and configured to perform a first inverse secondary alteration equal to an inverse operation of the first secondary alteration; a second primary processing module is connected to the input of the second modified model and configured to perform a second primary alteration equal to an inverse operation of the second inverse primary alteration; the software application comprising the first modified model, the first secondary processing module, the second primary processing module, and the second modified model; the first secondary processing module and the second primary processing module being preferably integrated within a single intermediate processing module and connected to each other by a link inaccessible to an electronic entity external to said intermediate module;

[0025] - during the modification step, the first modified model is an aggregation of a first primary artificial intelligence function, the first initial model, and the first secondary function, the first primary function being further added to the input of the first initial model and configured to perform a first inverse primary alteration, and the input of the first initial model becoming inaccessible to an electronic entity external to the first modified model; and

[0026] During the generation stage, a first primary processing module is connected to the input of the first modified model and configured to perform a first primary alteration equal to an inverse operation of the first inverse primary alteration; the software application further comprising the first primary processing module;

[0027] - during the modification step, the second modified model is an aggregation of the second primary artificial intelligence function, the second initial model, and a second secondary function, the second secondary function being further added to the output of the second initial model and configured to perform a second secondary alteration, and the output of the second initial model becoming inaccessible to an electronic entity external to the second modified model; and

[0028] During the generation step, a second secondary processing module is connected to the output of the second modified model and configured to perform a second inverse secondary alteration equal to an inverse operation of the second secondary alteration; the software application further comprising the second secondary processing module;

[0029] - during the modification step, a single function from among the primary artificial intelligence function and the secondary artificial intelligence function is added to the initial artificial intelligence model to obtain the modified artificial intelligence model, and

[0030] During the generation stage, a single module from among the primary processing module and the secondary processing module is connected to the modified model; the software application comprising said processing module and the modified model;

[0031] - where each alteration is a reversible operation chosen from the group including: a bit operation, a mixing operation, a matrix operation, a vector operation, and an encryption operation;

[0032] - the primary processing module(s) is configured to perform in addition a primary processing chosen from the group including: a processing of numerical value(s), a processing of textual data(s), an image processing, a time series processing, and a digital filtering processing;

[0033] - The secondary processing module(s) is configured to perform, in addition, a secondary processing operation chosen from the group comprising: an inverse operation of a operation performed by a respective primary processing module, a classification operation, a binarization operation, a dimensionality reduction operation, a three-dimensional reconstruction operation, a decoding operation, a conversion operation, an aggregation operation, a merging operation, a filtering operation, a correction operation, and a geometric transformation operation. The invention also relates to a computer program comprising software instructions which, when executed by a computer, implement an implementation method as defined above.

[0034] The invention also relates to an electronic data processing computer, comprising a computer platform and a processing software application, stored on the platform capable of executing said application, the software application comprising at least one processing module and at least one modified artificial intelligence model;

[0035] - at least one modified model having been obtained by modifying at least one initial artificial intelligence model, each having one input and one output; the modification being carried out by adding at least one function chosen from:

[0036] a primary artificial intelligence function added to the input of a respective initial model and configured to perform a reverse primary alteration, the respective modified model then being an aggregation of the primary function and the respective initial model, and the input of the respective initial model becoming inaccessible to an electronic entity external to the respective modified model; and

[0037] a secondary artificial intelligence function added to the output of the respective initial model and configured to perform a secondary alteration, the respective modified model then being an aggregation of the respective initial model and the secondary function, and the output of the respective initial model becoming inaccessible to the electronic entity external to the respective modified model;

[0038] at least one initial model having been previously trained using training data; and

[0039] - the software application having been generated by connecting to each modified model at least one processing module chosen from:

[0040] if the respective modified model includes the addition of the primary artificial intelligence function to the input of the respective initial model: a primary processing module connected to the input of the respective modified model and configured to perform a primary alteration equal to an inverse operation of the inverse primary alteration; and

[0041] If the respective modified model includes the addition of the secondary artificial intelligence function to the output of the respective initial model: a secondary processing module connected to the output of the respective modified model and configured to perform an inverse secondary alteration equal to an inverse operation of the secondary alteration. These features and advantages of the invention will become clearer upon reading the following description, given solely by way of non-limiting example, and made with reference to the accompanying drawings, in which:

[0042] - Figure 1 is a schematic representation of an electronic data processing computer comprising a computer platform and a processing software application, according to a first embodiment of the invention where the software application includes a single modified artificial intelligence model;

[0043] - Figure 2 is a view similar to that of Figure 1, according to a second embodiment of the invention where the software application includes several, such as two, modified artificial intelligence models;

[0044] - Figure 3 is a schematic representation of a method for obtaining the modified artificial intelligence model from Figure 1 or 2, starting from an initial artificial intelligence model; and

[0045] - Figure 4 is a schematic representation of a flowchart of a method for implementing the electronic processing computer according to the invention.

[0046] In Figure 1, an electronic data processing computer 5 comprises a computer platform 10 and a processing software application 15 stored on the platform 10, the platform 10 being capable of executing said application 15.

[0047] As is known, the computer platform 10 includes, for example, an information processing unit typically consisting of a memory and a processor associated with the memory, not shown. The processing software application 15 is then stored in the memory and executed by the processor.

[0048] In the example in Figure 1, the processing software application 15 includes a modified artificial intelligence model 20, obtained via the modification - described in more detail later with reference to Figures 3 and 4 - of a respective initial artificial intelligence model 22; a primary processing module 24 connected as input to the modified artificial intelligence model 20, also called the modified model 20; and a secondary processing model 26 connected as output to the modified model 20.

[0049] In this example in Figure 1, the processing software application 15 then includes a single modified model 20.

[0050] Alternatively, the processing software application 15 includes several modified models 20.

[0051] According to this variant, in the example of Figure 2, the processing software application 15 comprises two modified artificial intelligence models 20, namely a first 20A and a second 20B modified artificial intelligence models. In the example of Figure 2, the processing software application 15 further comprises a first primary processing module 24A connected to the input of the first modified model 20A, and a first secondary processing module 26A connected to the output of the first modified model 20A. The processing software application 15 further comprises a second primary processing module 24B connected to the input of the second modified model 20B, and a second secondary processing module 26B connected to the output of the second modified model 20B.

[0052] Advantageously, in the example of Figure 2, the first secondary processing module 26A and the second primary processing module 24B are further merged into a single intermediate processing module 28 connected between the first modified model 20A and the second modified model 20B.

[0053] In the following description, when general reference is made to the modified artificial intelligence processing model, whether it is the single modified model 20 of Figure 1 or one of the first 20A and second 20B modified models of Figure 2, the modified model will be designated by the general reference 20. Similarly, when general reference is made to the primary processing module, whether it is the single primary processing module 24 of Figure 1 or one of the first 24A and second 24B primary processing modules, the primary processing module will be designated by the general reference 24.Similarly, when general reference is made to the secondary processing module, whether it is the single secondary processing module 26 of Figure 1 or one of the first 26A and second 26B secondary processing modules, the secondary processing module will be designated by the general reference 26.

[0054] In the example in Figure 1, on the one hand, the modified model 20 is the aggregation of the initial model 22 and a primary artificial intelligence function 30 at the input of the initial model 22, such that the input of the initial model 22 becomes inaccessible to an unrepresented electronic entity external to the modified model 20, the primary function 30 being configured to perform an alteration called inverse primary alteration. On the other hand, the modified model 20 is the aggregation of the initial model 22 and a secondary artificial intelligence function 32 at the output of the initial model 22, such that the output of the initial model 22 becomes inaccessible to an electronic entity external to the modified model 20, the secondary function 32 being configured to perform an alteration called secondary alteration.

[0055] In the example in Figure 1, the primary processing module 24 connected to the input of the modified model 20 is then configured to perform a primary alteration 34 equal to an inverse operation of the inverse primary alteration subsequently performed by the primary function 30. The secondary processing module 26 connected to the output of the modified model 20 is then configured to perform an inverse secondary alteration 36 equal to an inverse operation of the secondary alteration previously performed by the secondary function 32.

[0056] In the example in Figure 2, on the one hand, the first modified model 20A is the aggregation of the first initial model 22A and a first primary artificial intelligence function 30A at the input of the first initial model 22A, such that the input of the first initial model 22A becomes inaccessible to an unrepresented electronic entity external to the first modified model 20A, the first primary function 30A being configured to perform an alteration called the first inverse primary alteration. On the other hand, the first modified model 20A is the aggregation of the first initial model 22A and a first secondary artificial intelligence function 32A at the output of the first initial model 22A, such that the output of the first initial model 22A becomes inaccessible to the electronic entity external to the first modified model 20A, the first secondary function 32A being configured to perform an alteration called the first secondary alteration.

[0057] In the example in Figure 2, the first primary processing module 24A, connected to the input of the first modified model 20A, is configured to perform a first primary alteration 34A, which is the inverse of the first inverse primary alteration subsequently performed by the first primary function 30A. The first secondary processing module 26A, connected to the output of the first modified model 20A, is then configured to perform a first inverse secondary alteration 36A, which is the inverse of the first secondary alteration previously performed by the first secondary function 32A.

[0058] In the example in Figure 2, similarly, on the one hand, the second modified model 20B is the aggregation of the second initial model 22B and a second primary function 30B of artificial intelligence at the input of the second initial model 22B, so that the input of the second initial model 22B becomes inaccessible to an electronic, unrepresented entity external to the second modified model 20B, the second primary function 30B being configured to perform an alteration called the second inverse primary alteration.On the other hand, the second modified model 20B is the aggregation of the second initial model 22B and a second secondary artificial intelligence function 32B at the output of the second initial model 22B, such that the output of the second initial model 22B becomes inaccessible to the electronic entity external to the second modified model 20B, the second secondary function 32B being configured to perform an alteration called the second secondary alteration. In the example in Figure 2, the second primary processing module 24B connected to the input of the second modified model 20B is then configured to perform a second primary alteration 34B equal to an inverse operation of the inverse second primary alteration subsequently performed by the second primary function 30B.The second secondary processing module 26B connected to the output of the second modified model 20B is then configured to perform a second inverse secondary alteration 36B equal to an inverse operation of the second secondary alteration previously performed by the second secondary function 32B.

[0059] Advantageously, the first secondary processing module 26A and the second primary processing module 24B are integrated within the intermediate processing module 28 and connected to each other by a link inaccessible to an electronic entity external to said intermediate module 28. The intermediate processing module 28 is then connected between the first modified model 20A and the second modified model 20B. The intermediate processing module 28 is then configured to perform, on the one hand, the first inverse secondary alteration 36A following the first secondary alteration previously performed by the first secondary function 32A to whose output it is connected, and on the other hand, the second primary alteration 34B prior to the second inverse primary alteration which will then be performed by the second primary function 30B to whose input it is connected.

[0060] In the following description, when primary alteration is referred to in general terms, this will mean both the single primary alteration in the example in Figure 1, and the first or second primary alteration in the example in Figure 2. Similarly, when secondary alteration is referred to in general terms, this will also mean both the single secondary alteration in the example in Figure 1, and the first or second secondary alteration in the example in Figure 2. More generally still, when alteration is referred to in general terms, this will mean both primary alteration in the general sense and secondary alteration in the general sense.

[0061] Advantageously, the alteration performed by the secondary function 32, namely the secondary alteration, is distinct from the alteration performed by the primary function 30, namely the inverse primary alteration. According to this advantage, the security of the modified model is further enhanced by having an alteration performed at its input that is distinct from that performed at its output.

[0062] Alternatively, the secondary alteration is identical to the inverse primary alteration.

[0063] The modified model 20 is of the same type as the initial model 22. The initial model 22 of artificial intelligence includes, for example, an artificial neural network or ANN (from the English Artificial Neural Network), and the modified model 20 then also includes an artificial neural network.

[0064] The initial model 22 was previously trained by learning with training data used as input and output of said initial model 22. In the case where the initial model 22 and the modified model 20 are each in the form of an artificial neural network with synaptic weights, the learning of the weights of the initial model 22 is for example carried out via backpropagation of the gradient.

[0065] Each alteration is typically a reversible operation chosen from the group including: a bit operation, a mixing operation, a matrix operation, a vector operation, and an encryption operation.

[0066] The bit operation is, for example, an operation chosen from the group including: constant or conditional XOR Shift; constant or conditional Left Bit Rotation; constant or conditional Right Bit Rotation; permutation of bits according to a predefined pseudo-random pattern; Bitwise NOT; bit shuffling; random permutation of bytes; lookup table between ciphertext and plaintext bytes; and interleaving bits of a predefined value.

[0067] The mixing operation is, for example, an operation chosen from the group including: reversible mixing of the original data into a large volume of data; split and recombine; addition of a known predictable pseudo-random suppressible noise pattern; and pattern replacement.

[0068] The matrix or vector operation is, for example, a matrix or vector operation F satisfying the relation F-1(F(X)) = X, where X denotes an input.

[0069] The encryption operation is an operation chosen from the group including: alteration by encryption and decryption; XTEA operation (from the English extended TEA); and AES operation (from the English Advanced Encryption Standard).

[0070] As an optional addition, the primary processing module(s) 24 is configured to perform an additional primary processing step, also called preprocessing. The primary processing step is typically chosen from the following group: processing of numerical value(s), processing of textual data, processing of image(s), processing of time series, and processing of digital filtering.

[0071] Numerical value processing, also called general preprocessing, is for example a processing step chosen from the group including: data normalization; data standardization; dimensionality reduction (e.g., PCA); detection and elimination of outliers; model-based logarithmic or functional transformation; data hybridization by adding new sources of information; conversion of raw data into usable formats (e.g., JSON to array); transformation of dates into numerical features (e.g., day of the week); analog-to-digital conversion of signals (ADC); channeling of heterogeneous data into a uniform format; data binarization; data smoothing; extrapolation or generation of missing values ​​(polynomial features); data undersampling / oversampling;pre-filtering of data to remove inconsistent values; pre-filtering of noise in the data; encoding of categorical variables (e.g., one-hot encoding); and combination of inputs to generate more appropriate inputs (e.g., XY, X; 2 Y, XY 2 ...).

[0072] Text data processing is, for example, a processing chosen from the group including: tokenization of texts; removal of stop words; vectorization of texts (e.g., TF-IDF); extraction of word roots (stemming or lemmatization); rescaling of text vectors (Word2Vec, GloVe embeddings); automatic language detection; conversion of abbreviations / acronyms or other into usable data; language switching; and addition of a system prompt.

[0073] Image processing is, for example, a chosen processing from the group including: contrast enhancement, white balance of images; conversion of image formats (for example, RGB to greyscale, or jpg to array...); resizing of image formats, change of resolution (for example, 4K to HD...); edge detection in images; application of Gaussian or Laplacian filters; mini-batch cutting of the original data for multi-application on a smaller AI model; detection of key points in images using local descriptors (for example SIFT or SURF); image modification by rotation, translation and / or scaling; temporal delta between two images; generation of additional layers (for example, R layer for red, G layer for green, B layer for blue, Time layer); removal and / or addition of masking patches.

[0074] Time series processing, for example, is a process chosen from the group including: processing a time series to make it stationary; detrending in time series; Kalman filtering for smoothing trajectories in time series data; and time registration or shifting. Digital filtering processing, for example, is a process chosen from the group including: converting a digital signal (e.g., audio) into another format (e.g., visual); Fourier transform for signal processing; frequency filtering (for noise removal in audio signals); filtering frequencies of interest in the audio signal; normalizing spectrograms for audio analysis; and digital frequency and power filtering.

[0075] Feature extraction and transformation processing is, for example, a processing chosen from the group including: feature extraction; transformation into polar or Cartesian representation; gridding of spatial data; geodetic transformations (e.g., WGS84 in 2D); data formatting (unit conversion, change of space...); and data masking (anonymization...).

[0076] In addition, the primary processing module or modules 24 are configured to perform a combination of primary processing from among the aforementioned processing, the processing combination then forming, for example, a complete functional chain acting as preprocessing for human interaction, such as interactions on keys, a touchpad, a mouse, a joystick, or with a camera or an interaction sensor.

[0077] As an optional complement, the secondary processing module(s) 26 is configured to perform an additional secondary processing step, also called post-processing. The secondary processing step is typically chosen from the group comprising: an inverse processing step of a step performed by a respective primary processing module 24, a classification processing step, a binarization processing step, a dimensionality reduction processing step, a three-dimensional reconstruction processing step, a decoding processing step, a conversion processing step, an aggregation processing step, a merging processing step, a filtering processing step, a correction processing step, and a geometric transformation processing step.

[0078] The classification treatment is for example a treatment chosen from the group including: application of mask on predictions for image segmentation; and separation of results into distinct categories (from the English clustering).

[0079] Binarization processing is, for example, a processing by thresholding relative to a threshold.

[0080] Dimensionality reduction processing is, for example, an inverse PCA process. Three-dimensional reconstruction processing is, for example, a three-dimensional reconstruction from two-dimensional coordinate predictions. Decoding processing is, for example, decoding predicted labels (converting indices into original values).

[0081] The conversion treatment is for example a treatment chosen from the group including: application of an inverse operation (from the English detokenization) of the rearrangement with token (from the English tokenizatiori); inverse conversion of the Fourier transform to obtain contemporary forms from frequency predictions.

[0082] Aggregation processing is, for example, a processing chosen from the group including: aggregation of predictions (e.g., averaging of results); aggregation of temporal accuracy from multiple models (from the English word stacking); aggregation of decisions made over several time windows (in the context of time series).

[0083] Fusion processing is, for example, a processing chosen from the group including: fusion of results from different models (from the English ensembling); fusion of feature maps to improve object segmentation in AI vision; application of predictive ensemble fusion to combine several models (from the English bagging, boosting); data compression.

[0084] The filtering treatment is, for example, a treatment chosen from the group including: filtering of results according to selection or acceptance criteria / thresholds; post-prediction filtering based on the confidence of a respective model; detection and removal of outliers in regression predictions.

[0085] The correction treatment is, for example, a treatment chosen from the group including: regularization by Kalman filtering; real-time output correction by feedback; and application of explainability metrics and associated monitoring.

[0086] Geometric transformation processing is, for example, a processing chosen from the group including: transformation and merging of bounding box coordinates (for object detection); combination of geometric transformations; application of overlap suppression mechanisms in object detection with tracking and automatic classification; and edge refinement in image segmentation.

[0087] In addition, secondary processing is chosen, for example, from the group including: application of specific business rules to return to the functional domain; application of a hierarchical classification model to interpret the outputs of nested models; functional coding of the output data; human-machine interface; sorting of outputs by importance or probability; recombination of text segments (if initially segmented); and conversion of predictions into robotic commands. Furthermore, each secondary processing module is configured to perform a combination of secondary processing steps from among those mentioned above, the combination of processing then forming, for example, a complete functional chain acting as post-processing up to human interactions (2D and 3D graphics processing, sound, haptics, etc.).

[0088] The implementation of the electronic processing computer 5 according to the invention will now be described with reference to Figure 3, which represents a flowchart of a method for implementing the electronic processing computer 5 according to the invention, the method being implemented by a computer.

[0089] The implementation process includes an initial step 100 of training each respective initial artificial intelligence model 22 with training data. In other words, during the initial training step 100, each respective initial model 22 is trained with said training data.

[0090] After this initial learning step 100, the process includes a step 110 of modifying each respective initial model 22 of artificial intelligence, trained during the learning step 100, to obtain a respective modified model 20 of artificial intelligence via the addition to the respective initial model 22 of at least one of the primary function 30 and the secondary function 32 of artificial intelligence.

[0091] When the respective modified model 20 results from the addition of the primary function 30 to the respective initial model 22, said primary function 30 is added to the input of the respective initial model 22 and is configured to perform the inverse primary alteration described previously. The respective modified model 20 is then the aggregation of the primary function 30 and the respective initial model 22, and the input of the respective initial model 22 then becomes inaccessible to the electronic entity external to the respective modified model 20.

[0092] When the respective modified model 20 results from the addition of the secondary function 32 to the respective initial model 22, said secondary function 32 is added to the output of the respective initial model 22 and is configured to perform the secondary alteration described above. The respective modified model 20 is then the aggregation of the respective initial model 22 and the secondary function 32, and the output of the respective initial model 22 then becomes inaccessible to the electronic entity external to the respective modified model 20.

[0093] As an optional addition, when the respective modified model 20 results from the addition of both the primary function 30 and the secondary function 32 to the respective initial model 22, on the one hand, the primary function 30 configured to perform the inverse primary alteration is added to the input of the respective initial model 22, and on the other hand, the secondary function 32 configured to perform the secondary alteration is added to the output of the respective initial model 22. The respective modified model 20 is then the aggregation of the primary function 30, the respective initial model 22, and the secondary function 32, and both the input and output of the respective initial model 22 then become inaccessible to the electronic entity external to the respective modified model 20.

[0094] After this modification step 110, the process further includes a step 120 of generating the processing software application 15 by connecting to each respective modified model 20, obtained during the modification step 110, at least one of the primary processing module 24 and the secondary processing module 26.

[0095] During this generation step 120, if the respective modified model 20 includes the addition of the primary function 30 to the input of the respective initial model 22, then the primary processing module 24 is connected to the input of the respective modified model 20 to perform the primary alteration 34, which is equal to the inverse operation of the inverse primary alteration. The generated processing software application 15 then includes the primary processing module 24 and the respective modified model 20 connected to the output of said primary module 24.

[0096] During this generation step 120, if the respective modified model 20 includes the addition of the secondary function 32 to the output of the respective initial model 22, then the secondary processing module 26 is connected to the output of the respective modified model 20 to perform the inverse secondary alteration 36, which is equal to the inverse operation of the secondary alteration. The generated processing software application 15 then includes the respective modified model 20 and the secondary processing module 26 connected to the output of said modified model 20.

[0097] As an optional addition, during this generation step 120, if the respective modified model 20 includes the addition of both the primary function 30 at the input of the respective initial model 22 and the secondary function 32 at the output of said initial model 22, then the primary processing module 24 is connected to the input of the respective modified model 20 to perform the primary alteration 34, and the secondary processing module 26 is connected to the output of the respective modified model 20 to perform the inverse secondary alteration 36. The generated processing software application 15 then includes the primary processing module 24, the respective modified model 20, and the secondary processing module 26, with the modified model 20 being connected to the output of the primary module 24, and the secondary module 26 being connected to the output of the modified model 20.

[0098] After this generation step 120, the process finally includes a storage step 130 of the processing software application 15, generated during the generation step 120. During this storage step 130, the processing software application 15 is stored on the computer platform 10, which is capable of executing said application 15, to form the electronic processing computer 5 according to the invention. Figure 4 then illustrates an example of a method for modifying the respective initial model 22 into the respective modified model 20.

[0099] In the example in Figure 4, the modification method first involves converting the initial model 22 into an ONNX format model. ONNX (Open Neural Network Exchange) is an open format designed to facilitate interoperability between different artificial intelligence and machine learning software frameworks. It allows deep learning and machine learning models to be represented in a standardized format, so they can be trained in one environment and run in another.

[0100] When the modification involves adding the primary function 30 to the input of the initial model 22, the modification method then includes a first addition 210 of one or more alteration operators (opcodes) from the ONNX library to form the inverse primary alteration performed by the primary function 30. In other words, the alteration operator(s) added during this first addition 210 form the primary function 30.

[0101] When the modification involves adding the secondary function 32 to the output of the initial model 22, the modification method then includes a second addition 220 of one or more alteration operators (opcodes) from the ONNX library to form the secondary alteration performed by the secondary function 32. In other words, the alteration operator(s) added during this second addition 220 form the secondary function 32.

[0102] The modification method also includes obfuscation of the alteration operator(s) added during the first addition and / or the second addition, as well as parameters of the initial model, such as synaptic weights of the network when the initial model is a neural network. Obfuscation is performed, for example, using an encryption toolbox, such as the xorshif toolbox.

[0103] The modified model 20 is then delivered at the output of obfuscation 230.

[0104] In the examples in Figures 1 and 2, both the input and output of the modified model 20 are protected by the addition of, on the one hand, the primary function 30 and, on the other hand, the secondary function 32.

[0105] A person skilled in the art will understand that it is sufficient for the input or output of the modified model 20 to be protected by the addition of the primary function 30, or respectively the secondary function 32, for the modified model 20 according to the invention to be protected against distillation attacks. Indeed, an attacker will not be able to steal the modified model 20 using a distillation attack because:

[0106] - the input data injected at the input interface of the modified model 20 will be interpreted as noise by the modified model 20 after the inverse primary alteration performed by the primary function 30, and / or

[0107] - the output data captured at the output interface of the modified model 20 will be interpreted as noise by the new model being trained by an attacker attempting a distillation attack, due to the secondary alteration carried out by the secondary function 30.

[0108] Thus, with the method according to the invention, at least one initial artificial intelligence model 20 is protected against distillation attacks by adding at least one artificial intelligence function 30, 32 to the input and / or output of the initial model 22, the input and / or output of the initial model 22 then becoming inaccessible to a potential attacker. At least one artificial intelligence function 30, 32 performs an alteration to protect the input and / or output of the initial model 22. Finally, at least one processing module 24, 26 performs a reverse alteration to compensate for said alteration, which is unknown to the potential attacker.

[0109] It is thus conceived that the electronic processing computer 5 according to the invention includes at least one artificial intelligence model 20 benefiting from better protection against distillation attacks.

Claims

DEMANDS 1. Method for implementing an electronic data processing computer (5), the method being implemented by computer and comprising the following steps: - learning (100) with training data from at least one initial artificial intelligence model (22, 22A, 22B), each having an input and an output; - modification (110) of at least one initial artificial intelligence model (22, 22A, 22B) to obtain at least one modified artificial intelligence model (20, 20A, 20B), via the addition of a secondary artificial intelligence function (32, 32A, 32B) added to the output of the respective initial model (22, 22A, 22B) and configured to perform a secondary alteration, the respective modified model (20, 20A, 20B) then being an aggregation of the respective initial model (22, 22A, 22B) and the secondary function (32, 32A, 32B), and the output of the respective initial model (22, 22A, 22B) becoming inaccessible to an electronic entity external to the respective modified model (20, 20A, 20B); and - generation (120) of a processing software application (15), by connecting to each modified model (20, 20A, 20B) a secondary processing module (26, 26A, 26B) connected to the output of the respective modified model (20, 20A, 20B) and configured to perform an inverse secondary alteration (36) equal to an inverse operation of the secondary alteration; the software application (15) comprising the secondary processing module (26, 26A, 26B) and at least one modified model (20, 20A, 20B); - storage (130) of the processing software application (15) on a computer platform (10) capable of executing said application, to form the electronic data processing computer.

2. A method according to claim 1, wherein, during the modification step (110), a primary function (30, 30A, 30B) of artificial intelligence added to the input of the respective initial model (22, 22A, 22B) and configured to perform an inverse primary alteration, and the secondary function (32, 32A, 32B) of artificial intelligence are added to the initial model (22, 22A, 22B) of artificial intelligence to obtain the modified model (20, 20A, 20B) of artificial intelligence, the respective modified model (20, 20A, 20B) then being an aggregation of the primary function (30, 30A, 30B), the respective initial model (22, 22A, 22B) and the secondary function (32, 32A, 32B), the input and output of the initial model (22, 22A, 22B) respectively becoming inaccessible to the external electronic entity of the modified model (20, 20A, 20B) respectively, and during the generation step (120), a primary processing module (24, 24A, 24B) connected to the input of the modified model (20, 20A,20B) respective and configured to perform a primary alteration (34) equal to an inverse operation of the inverse primary alteration and the secondary processing module (26, 26A, 26B) are connected to the modified model (20, 20A, 20B); the software application (15) comprising the primary processing module (24, 24A, 24B), the modified model (20, 20A, 20B), and the secondary processing module (26, 26A, 26B).

3. Method according to claim 1 or 2, wherein, during the modification step (110), a first (22A) and a second (22B) initial artificial intelligence models are modified to obtain a first (20A) and a second (20B) modified artificial intelligence models; the first modified model (20A) being an aggregation of the first initial model (22A) and a first secondary function (32A) of artificial intelligence, the first secondary function (32A) being added to the output of the first initial model (22A) and configured to perform a first secondary alteration, and the output of the first initial model (22A) becoming inaccessible to the electronic entity external to the first modified model (20A);the second modified model (20B) being an aggregation of a second primary function (30B) of artificial intelligence and the second initial model (22B), the second primary function (30B) being added to the input of the second initial model (22B) and configured to perform a second inverse primary alteration, and the input of the second initial model (22B) becoming inaccessible to an electronic entity external to the second modified model (20B); during the generation step (120), a first secondary processing module (26A) is connected to the output of the first modified model (20A) and configured to perform a first inverse secondary alteration (36A) equal to an inverse operation of the first secondary alteration; a second primary processing module (24B) is connected to the input of the second modified model (20B) and configured to perform a second primary alteration (34B) equal to an inverse operation of the second inverse primary alteration; the software application (15) comprising the first modified model (20A), the first secondary processing module (26A), the second primary processing module (24B), and the second modified model (20B); the first secondary processing module (26A) and the second primary processing module (24B) preferably being integrated within a single intermediate processing module (28) and connected to each other by a link inaccessible to an electronic entity external to said intermediate module (28).

4. Method according to claim 3, wherein, during the modification step (110), the first modified model (20A) is an aggregation of a first primary function (30A) of artificial intelligence, the first initial model (22A) and the first secondary function (32A), the first primary function (30A) being further added to the input of the first initial model (22A) and configured to perform a first inverse primary alteration, and the input of the first initial model (22A) becoming inaccessible to an electronic entity external to the first modified model (20A); and during the generation step (120), a first primary processing module (24A) is connected to the input of the first modified model (20A) and configured to perform a first primary alteration (34A) equal to an inverse operation of the first inverse primary alteration; the software application (15) further comprising the first primary processing module (24A).

5. A method according to claim 3 or 4, wherein, during the modification step (110), the second modified model (20B) is an aggregation of the second primary artificial intelligence function (30B), the second initial model (22B) and a second secondary function (32B), the second secondary function (32B) being further added to the output of the second initial model (22B) and configured to perform a second secondary alteration, and the output of the second initial model (22B) becoming inaccessible to an electronic entity external to the second modified model (20B); and during the generation step (120), a second secondary processing module (26B) is connected to the output of the second modified model (20B) and configured to perform a second inverse secondary alteration (36B) equal to an inverse operation of the second secondary alteration; the software application (15) further comprising the second secondary processing module (26B).

6. A method according to any one of the preceding claims, wherein the alteration or each alteration (34, 34A, 34B, 36, 36A, 36B) is a reversible operation selected from the group comprising: a bit operation, a mixing operation, a matrix operation, a vector operation, and an encryption operation.

7. A method according to any one of claims 2 to 6, wherein the primary processing module(s) (24, 24A, 24B) is configured to further perform a primary processing operation selected from the group comprising: processing of numeric value(s), processing of text data(s), processing of image(s), processing of time series(s), and processing of digital filtering.21 8. A method according to any one of the preceding claims, wherein the secondary processing module or each secondary processing module (26, 26A, 26B) is configured to further perform a secondary processing selected from the group comprising: an inverse processing of a processing performed by a respective primary processing module (24, 24A, 24B), a classification processing, a binarization processing, a dimension reduction processing, a three-dimensional reconstruction processing, a decoding processing, a conversion processing, an aggregation processing, a merging processing, a filtering processing, a correction processing, and a geometric transformation processing.

9. Computer program, comprising software instructions which, when executed by a computer, implement a method according to any one of the preceding claims.

10. Electronic computer (5) for data processing, comprising a computer platform (10) and a processing software application (15), stored on the platform (10) capable of executing said application (15), the software application (15) comprising at least one processing module (26, 26A, 26B) and at least one modified model (20, 20A, 20B) of artificial intelligence; - at least one modified model (20, 20A, 20B) having been obtained by a modification of at least one initial model (22, 22A, 22B) of artificial intelligence, each having an input and an output; the modification being carried out by adding a secondary function (32, 32A, 32B) of artificial intelligence added to the output of the respective initial model (22, 22A, 22B) and configured to perform a secondary alteration, the respective modified model (20, 20A, 20B) then being an aggregation of the respective initial model (22, 22A, 22B) and the secondary function (32, 32A, 32B), and the output of the respective initial model (22, 22A, 22B) becoming inaccessible to the electronic entity external to the respective modified model (20, 20A, 20B); at least one initial model (22, 22A, 22B) having been previously trained using training data; and - the software application (15) having been generated by connecting to each modified model (20, 20A, 20B) a secondary processing module (26, 26A, 26B) connected to the output of the respective modified model (20, 20A, 20B) and configured to perform an inverse secondary alteration (36) equal to an inverse operation of the secondary alteration.