Method and device for unlearning a trained facial recognition model for vehicles

Abstract

The present invention relates to a method and device for unlearning a trained vehicle facial recognition model implemented as a neural network. The method generates (52) a first subset of labeled training data comprising labeled training data from said labeled training dataset that correspond to said at least one label represented by the received label data, and a second subset of labeled training data comprising labeled training data from said labeled training dataset that is not included in the first subset of labeled training data.The method adjusts (54) each parameter of the facial recognition model according to the importance of these parameters calculated (53) for the first and second labeled training data subsets and replaces (55) the training dataset with the second labeled training data subset. Figure for the abstract: Figure 5.

Description

Title of the invention: Method and device for unlearning a trained facial recognition model for vehicles. Technical field

[0001] The invention relates to methods and devices for unlearning a trained facial recognition model for vehicles, including but not limited to motor vehicles. Technological background

[0002] Modern vehicles incorporate various driver assistance systems, known as ADAS (Advanced Driver-Assistance System), to assist the driver of such a vehicle in controlling the driving of the vehicle, or even to replace the driver in an autonomous driving mode of the vehicle.

[0003] Some of these ADAS systems rely on predictive models implemented by neural networks. These predictive models need to be trained in a learning phase before they can be used in the vehicle. The learning phase of a neural network implementing a predictive model requires a large amount of data so that the neural network can then function in all the real-life situations encountered by the vehicle. The quality of the training data used in the learning phase is important for improving the relevance of the neural network and the predictions it makes. It is therefore necessary that the training data reflect as closely as possible all the real-life situations the vehicle will encounter once it is in operation.To this end, car manufacturers use data collected from test vehicles as well as from private vehicles driving on public roads in real-world conditions. For test vehicles, the data captured by the vehicle's various sensors (radar, lidar, cameras) is stored on dashcams, which are retrieved at the end of the tests to extract the data. For private vehicles, the data captured by the various onboard sensors is retrieved from one or more cloud servers to which the vehicles are connected wirelessly, for example, via a wireless cellular network.

[0004] To reduce the bandwidth required for transmitting training data acquired by private vehicles, it may be possible to implement the training of the neural network parameters directly in the vehicle, thus avoiding the transmission of the data acquired by the vehicle. to a remote device implementing the learning process. The implementation of learning within the vehicle is then made possible by using a pre-trained neural network, by validating the training data acquired on board the vehicle via feedback from a vehicle occupant, and by implementing a supervised learning method that does not require backpropagation of errors. This reduces the volume of computation required to train the neural network and thus reduces the need for computing resources (processors and memory).

[0005] Neural networks can be used to implement a facial recognition model embedded in a vehicle. This facial recognition model can be used, for example, by ADAS systems and / or to identify a driver before the vehicle starts.

[0006] Facial recognition is a sensitive issue with regard to the confidentiality of personal data. Since neural networks trained to recognize facial images process personal data, they are subject to the European General Data Protection Regulation (GDPR). For example, the GDPR stipulates that "In order to ensure that personal data are not kept longer than necessary, time limits must be set by the controller for erasure or for periodic review. Every reasonable step must be taken to ensure that inaccurate personal data are rectified or erased."Article 17 of the GDPR guarantees the "right to erasure: the data subject shall have the right to obtain from the controller the erasure of personal data concerning him or her within a reasonable time and the controller shall have the obligation to erase personal data within a reasonable time".

[0007] As a consequence of the GDPR, car manufacturers must be able to delete personal data at the request of any customer, in all its forms. Specifically, in the context of facial recognition software that uses neural networks, car manufacturers must be able to remove knowledge of a given facial identity from the neural network. However, erasing a facial identity from a neural network cannot be done as easily as deleting it from a dataset. This is because deep neural networks store knowledge about their training dataset in connection weights, and this knowledge is generally not located in a single connection, but distributed across a large number of them.This difficult question gave rise to the research topic of machine unlearning (Liu (2024). Machine Unlearning in 2024. Ken Ziyu Liu's blog - Stanford Computer Science)). Unlearning is a technological challenge that... This remains to be solved, particularly when trained neural networks are used as a facial recognition model embedded in a vehicle.

[0008] Several neural network unlearning methods exist to remove targeted knowledge from a neural network. These methods generally require retraining the neural network. They require labeled training data. This can be problematic because retraining the neural network is not always possible. Indeed, the network may have been pre-trained with labeled training data that is not available at the time of retraining, and even if this labeled training data is available, retraining may modify the neural representations used for the facial identities that are not removed, which can disrupt all downstream software functions that rely on these neural representations.

[0009] It is therefore necessary to define a way of unlearning certain neural representations of targeted facial identities from the neural network, while affecting as little as possible the neural representations of other non-targeted facial identities, so as not to impair the performance of facial recognition for these other facial identities.

[0010] Loss-Free Selective Synaptic Dampening (LFSSD) is a machine unlearning technique proposed by Foster et al. (Loss-Free Machine Unlearning. International Conference on Learning and Representations (ICLR) 2024) that can be implemented after training and does not require retraining of the neural network or labeled training data. Basically, the LFSSD method seeks to identify the neural network parameters (connection weights) that are important for a first subset of data representing facial identities to be deleted but not for a second subset of data representing facial identities not to be deleted, and then to adjust these parameters according to their importance. Like other existing machine unlearning methods (Tarun, Chundawat, Mandai, Kankanhalli (2023) Fast Yet Effective Machine Unlearning).IEEE Transactions on Neural Networks and Learning Systems, Choi (2024). Towards Machine Unlearning Benchmarks: Forgetting the Personal Identities in Facial Recognition Systems. Association for the Advancement of Artificial Intelligence (AAAI) 2024. link: https: / / arxiv.org / pdf / 2311.02240), the LFSSD method has been tested in the field of facial recognition, but it has not been applied to facial identification on peripheral devices, and a fortiori not to facial recognition software embedded in a vehicle, which is one of the applications envisaged by the present invention. Summary of the present invention

[0011] One object of the present invention is to solve at least one of the problems of the technological background described above.

[0012] Another object of the present invention is to unlearn a neural network implementing an easy recognition model, certain targeted facial identities, while affecting any other facial representation as little as possible.

[0013] Another object of the present invention is to unlearn facial identities from a neural network without retraining the neural network.

[0014] Another object of the present invention is to unlearn facial identities locally, so that no data leaves the vehicle, which promotes data confidentiality.

[0015] According to a first aspect, the present invention relates to a method for unlearning a trained facial recognition model for vehicles implemented by a neural network, said facial recognition model being trained by refining a set of parameters of said pre-trained facial recognition model, said refining being implemented from a labeled training dataset comprising a label vector associated with face image data, said method being implemented by at least one computer embedded in said vehicle, said method comprising the following steps: - receiving label data representative of at least one label from said label vector; - generation of a first subset of labeled training data comprising labeled training data from said labeled training dataset that correspond to said at least one label represented by the received label data, and of a second subset of labeled training data comprising labeled training data from said labeled training dataset that are not included in the first subset of labeled training data; - calculation of a first data representative of an importance of the parameters of said set of parameters of said trained facial recognition model for the labeled training data of said first subset of labeled training data and calculation of a second data representative of an importance of the parameters of said set of parameters of said trained facial recognition model for the labeled training data of said second subset of labeled training data; - adjusting each parameter of said parameter set based on a comparison of the first and second calculated data points; and - replacing the labeled training dataset with the second labeled training subset.

[0016] The method makes it possible to remove, at the request of a vehicle occupant, the knowledge of facial identities used by a facial recognition model implemented by a deep neural network and embedded in a vehicle.

[0017] The process adapts the LFSSD method to the facial recognition model so that the deletion of a facial identity can be carried out directly on board the vehicle without any information leaving the vehicle, which improves compliance with the GDPR.

[0018] The method is advantageous because this suppression of targeted facial identities affects the knowledge of other facial identities very little and maintains a satisfactory recognition accuracy of these other facial identities.

[0019] The method makes it possible to unlearn the trained vehicle facial recognition model without retraining this model in any way, which saves computational retraining costs.

[0020] According to a particular and non-limiting embodiment of the present invention, the first subset of labeled training data is supplemented by labeled training data obtained from the labeled training data initially forming the first subset of labeled training data.

[0021] According to a particular and non-limiting embodiment of the present invention, said calculation of the first and second data is implemented by at least one processor of said at least one computer when a load level of said at least one processor induced by processing tasks different from said calculation of the first and second data by said at least one processor is less than a determined threshold.

[0022] According to a particular and non-limiting embodiment of the present invention, the calculation of the first and second data representing the importance of a parameter of said set of parameters of said trained facial recognition model is based on the calculation of a partial derivative with respect to this parameter of an output data of the neural network.

[0023] According to a particular and non-limiting embodiment of the present invention, the partial derivative is calculated using a method without backpropagation of error.

[0024] According to a particular and non-limiting embodiment of the present invention, each parameter of said parameter set of said trained facial recognition model is adjusted as follows:

[0025] Q fi >aÜ ri

[0026] with „ _ • / , ja indicates the adjusted parameter, a, A and fi are P — min[ LJ, parameters and Q / j, Q, , respectively indicate the first and second data.

[0027] According to a second aspect, the present invention relates to a device for unlearning a trained vehicle facial recognition model, the device comprising a memory associated with a processor configured for implementing the steps of the process according to the first aspect of the present invention.

[0028] According to a third aspect, the present invention relates to a vehicle, for example of the automobile type, comprising a device as described above according to the second aspect of the present invention.

[0029] According to a fourth aspect, the present invention relates to a computer program which includes instructions adapted for carrying out the steps of the process according to the first aspect of the present invention, in particular when the computer program is executed by at least one processor.

[0030] Such a computer program may use any programming language, and be in the form of source code, object code, or an intermediate code between source code and object code, such as in a partially compiled form, or in any other desirable form.

[0031] According to a fifth aspect, the present invention relates to a computer-readable recording medium on which is recorded a computer program comprising instructions for carrying out the steps of the process according to the first aspect of the present invention.

[0032] On the one hand, the recording medium can be any entity or device capable of storing the program. For example, the medium can include a storage means, such as a ROM, a CD-ROM or a microelectronic circuit-type ROM, or a magnetic recording means or a hard disk drive.

[0033] On the other hand, this recording medium can also be a transmissible medium such as an electrical or optical signal, such a signal being able to be transmitted via an electrical or optical cable, by conventional or radio frequency, by self-directing laser beam, or by other means. The computer program according to the present invention can, in particular, be downloaded from an Internet-type network.

[0034] Alternatively, the recording medium may be an integrated circuit in which the computer program is incorporated, the integrated circuit being adapted to execute or to be used in the execution of the process in question. Brief description of the figures

[0035] Other features and advantages of the present invention will become apparent from the description of the particular and non-limiting embodiments of the present invention below, with reference to the attached Figures 1 to 5, in which:

[0036] [Fig-1] schematically illustrates part of a vehicle passenger compartment, according to a example of a particular embodiment of the present invention;

[0037] [Fig.2] schematically illustrates a learning process of a network of neurons implementing a facial recognition model in the vehicle of [Fig.1], according to a particular embodiment of the present invention;

[0038] [Fig.3] schematically illustrates a process of unlearning a network of neurons implementing a facial recognition model in the vehicle of [Fig.1] and trained according to the process of [Fig.2], according to a particular embodiment of the present invention;

[0039] [Fig.4] schematically illustrates a device configured for learning and the unlearning of a trained neural network implementing a facial recognition model in the vehicle of [Fig.1], according to a particular and non-limiting embodiment of the present invention.

[0040] [Fig. 5] illustrates a flowchart of the different stages of a process of unlearning of a neural network implementing a facial recognition model in the vehicle of [Fig.1] and trained according to the process of [Fig.2], according to a particular and non-limiting embodiment of the present invention. Description of examples of achievements

[0041] A method and device for learning and a method and device for unlearning a facial recognition model for vehicles implemented by a neural network will now be described in what follows with joint reference to Figures 1 to 5. The same elements are identified with the same reference signs throughout the description that follows.

[0042] The terms "first," "second" (or "firsts," "seconds"), etc., are used in this document by arbitrary convention to allow for the identification and distinction of different elements (such as operations, means, etc.) implemented in the embodiments described below. Such elements may be distinct or correspond to a single element, depending on the embodiment.

[0043] Fig. 1 schematically illustrates part of the passenger compartment of a vehicle 10, according to a particular and non-limiting embodiment of the present invention.

[0044] Vehicle 10 corresponds, for example, to a vehicle with an internal combustion engine, with electric motor(s), or even a hybrid vehicle with an internal combustion engine and one or more electric motors. Vehicle 10 thus corresponds, for example to a land vehicle comprising a passenger compartment or at least equipped with a display screen, for example a car, a truck, a bus.

[0045] The vehicle 10 advantageously carries one or more cameras 13 configured for the acquisition of image data of the passenger compartment of the vehicle 10, in particular for the acquisition of images of the faces of the driver and any passengers of the vehicle 10.

[0046] The camera 13 is, for example, arranged in the passenger compartment of the vehicle 10 at the level of the interior rearview mirror. Such a camera 13 has a field of vision corresponding to the front of the passenger compartment including the front seats and possibly one or more rear seats.

[0047] According to one embodiment, the vehicle 10 further comprises another camera arranged on the dashboard, for example in a space behind the steering wheel. Such a camera is configured for acquiring images of the face of the driver of the vehicle 10. Such a camera corresponds, for example, to the camera of a driver attention monitoring system, known as a DMS (Driver Monitoring System).

[0048] A camera includes, for example, the following elements: - a photosensitive sensor corresponding for example to a matrix of photoreceptors associated for example with a Bayer filter; - an optical assembly arranged in front of the sensor with respect to the scene to be acquired by the sensor, the optical assembly comprising, for example, an arrangement of one or more lenses; and - optionally one or more computers associated with memory and configured for processing images acquired by the sensor.

[0049] The image data received or obtained from each camera, in particular camera 13, are thus representative of one or more images of the interior of the vehicle 10, this image data corresponding for example to data representative of a pixel matrix, color values ​​being for example associated with each pixel, for example according to one or more color channels; for example, the pixel data are coded in the form of RGB values ​​(from the English "Red, Green, Blue" or "Rouge, vert, bleu" in French).

[0050] The vehicle 10 further incorporates a display system comprising one or more control units controlling one or more display devices belonging to the display system. The display system includes, for example, a touchscreen 12 and a control unit configured to control the display of content(s) from a graphical HMI on the touchscreen 12, for example, integrated into the dashboard 11.

[0051] The computers controlling the display system (such as the computer of the infotainment system, known as the IVI computer (from the English "In-Vehicle Infotainment" or in French "Infodivertissement étoilé")), the various components of the vehicle 10 and a set of AD AS type driving assistance systems form, for example, a multiplexed architecture for the implementation of various services useful for the proper functioning of the vehicle.Computers communicate and exchange data with each other via one or more computer buses, for example a CAN (Controller Area Network), CAN FD (Controller Area Network Flexible Data-Rate), FlexRay (according to ISO 17458), LIN (Local Interconnect Network), or Ethernet (according to ISO / IEC 802-3) type communication bus.

[0052] The vehicle 10 also includes, for example, a set of driver assistance systems, known as AD AS systems (from the English "Advanced Driver-Assistance System" or in French "Système d'aide à la conduite avancé").

[0053] A learning process for a vehicle facial recognition model implemented by a neural network is carried out by one or more devices embedded in the vehicle 10, for example by one or more computers of the vehicle's embedded network 10, i.e. by one or more processors of the computer(s) in association with one or more memories, for example a memory of the computer(s). Examples of implementations of such a process are described opposite [Fig. 2] below.

[0054] [Fig.2] schematically illustrates a learning process of a neural network implementing a facial recognition model in the vehicle of [Fig.1], according to a particular embodiment of the present invention.

[0055] In a first operation 21 of the learning process, image data representative of a set of images 201 of the passenger compartment of the vehicle 10 are received.

[0056] The images are acquired by one or more cameras 13 of the vehicle 10. The images forming the image set 201 include one or more faces of one or more occupants of the vehicle 10, for example the driver and one or more passengers sitting in the seats of the vehicle 10.

[0057] Image data is received by the computer(s) implementing the process directly from the camera 13 as images are acquired by the camera 13 or from a vehicle memory 10 in which the image data is stored after acquisition for processing as explained below.

[0058] This image data feeds into a pre-trained facial recognition model implemented by a neural network, for example a neural network comprising several layers of neurons. The neural network corresponds, for example, to a convolutional neural network or convolutional neural network, called CNN (from the English "Convolutional Neural Network").

[0059] The neural network implementing the pre-trained facial recognition model was pre-trained (i.e., trained beforehand) for its implementation in the learning process of [Fig. 1] using a large amount of training data, for example, face image data and other image data. The pre-trained facial recognition model, in other words, the pre-trained neural network, corresponds to a model trained for face recognition in images or to a more generic model trained, for example, for object recognition in images, including faces.

[0060] The pre-trained facial recognition model processes the image data 201 provided as input to the pre-trained model to classify these images and associate one or more labels with them (a label being associated with each class of a set of classes forming the output of the pre-trained model).

[0061] An example of a pre-trained facial recognition model is described in the article entitled "MixFaceNets: Extremely Efficient Face Recognition Networks", by Fadi Boutros et al., published on July 27, 2021 under the reference arXiv:2107.13046vl.

[0062] The parameters of such a pre-trained model, i.e. the parameters of the pre-trained neural network, have been previously loaded into the memory of the vehicle 10 to which the computer(s) implementing the learning process have access.

[0063] The pre-trained facial recognition model detects the faces present in the images 201 and associates with each of these faces (or predicts for each of these faces) a label identifying each of these faces, the label being initially predicted randomly before converging with the acquisition of several images of the same face.

[0064] The output of the pre-trained facial recognition model is obtained from representative data of a set of faces 202 and their associated labels 203, these data 202, 203 forming or feeding a training dataset.

[0065] In a second operation 22 of the learning process, the display of the images of the faces 202 and the associated labels 203 is controlled so as to display each face and each label (i.e. a face identifier) ​​associated with that face on a screen of the vehicle 10, for example screen 12, to have the association between the displayed face and its label / identifier validated or corrected by the driver or a passenger of the vehicle 10.

[0066] Each face and associated label pair is, for example, displayed as a particular graphical content of a graphical human-machine interface of vehicle 10.

[0067] A display control of graphic content or any graphic object (text, pictogram, icon, etc.) includes a rendering of the graphic content or graphic object, such rendering corresponding to a set of operations performed by one or more processors on the pixels of one or more images of the graphic content to be displayed on the screen 12. For example, rendering consists of associating to a set of pixels of an image pixel data (for example color data expressed in an RGB (Red, Green, Blue) type space) associated with each graphic object.

[0068] The control of displaying graphic content or an image thus includes the transmission of control signals to the touch screen 12 to modify the values ​​associated with the pixels of the touch screen 12 at the location intended to display the graphic content or the image.

[0069] In a third operation 23 of the learning process, validation data representative of a validation or correction of the label associated with each image displayed on the screen 12 are received.

[0070] These validation data are for example received from the touch interface of the touch screen 12 when the occupant of the vehicle 10, viewing the images of faces and the associated labels, validates or corrects (by assigning another label to the displayed image, i.e. by assigning another identifier to the displayed image) the information displayed, i.e. the label associated with the image of the displayed face.

[0071] According to another embodiment, this validation data is received from a microphone arranged in the passenger compartment of the vehicle 10 when the occupant of the vehicle 10, validating or correcting the displayed information, uses a voice communication interface of the vehicle 10.

[0072] In a fourth operation 24 of the learning process, a labeled training dataset is generated from the training dataset and the validation data, the labeled training dataset comprising a label vector 204 associated with the face image data 202.

[0073] Such an operation 24 makes it possible to correct the training dataset formed from the image data 202 and the associated labels 203 by correcting the erroneous labels to generate a new set of labels 204 (for example in the form of a vector) associated with the face image data 202 on the basis of the corrections (or validations) made by the person who viewed the face image / associated label pairs displayed on the screen 12.

[0074] The labeled training dataset is stored in a memory of the vehicle 10 and forms a database that can be used to implement supervised learning of the parameters of the neural network forming the facial recognition model that will be used in the production or inference phase.

[0075] The representative data of the labeled training dataset are for example recorded in a compressed form, a compression operation being applied to this data according to any method known to a person skilled in the art.

[0076] When this data is recorded in compressed format, a data decompression operation is implemented before any processing operation is carried out on this data, for example during supervised learning as described in the fifth operation 25.

[0077] In a fifth operation 25 of the learning process, supervised learning of the parameters (or weights) of the neural network implementing the facial recognition model that will be used in the production or inference phase is implemented by refining the parameters (or connection weights) of the pre-trained facial recognition model.

[0078] Supervised learning corresponds to supervised fine-tuning, known as SFT (from the English "Supervised Fine-Tune" or in French "affinage surveillé").

[0079] Supervised learning is implemented based on the labeled training dataset, i.e. on the face images 202 and associated labels 203 or 204.

[0080] The method used for supervised learning corresponds to a method requiring fewer resources (in particular processor computing resources and memory resources) than a prior art method implemented on a server or computer comprising higher resources than those available on a vehicle computer (for example, low memory budget and limited number of operations (FLOPS, from the English "Floating-point Operations Per Second") available).

[0081] The method implemented in the present process corresponds to a method that does not implement the usual error backpropagation operation well known to the person skilled in the art.

[0082] The implemented method replaces error backpropagation by using a different so-called random matrix for each layer of the neural network. A matrix is ​​said to be random when its constituent elements are randomly generated. Thus, a different random matrix is ​​associated with each layer of the neural network.

[0083] A target label vector is used for refining the parameters of each neural network layer. The target label vector used for a given neural network layer is obtained by multiplying the label vector of the 204 labeled training dataset by the random matrix associated with the data layer of the neural network. Thus, a different target label vector is obtained for each neural network layer. The target label vector comprises the labels, called targets, that the parameters of the considered layer of the neural network must predict from the data that this layer receives as input.

[0084] Several refinement iterations are implemented, for example until the error between the predicted labels and the target labels is less than a threshold, for example for each layer. In other words, the supervised learning operation 25 is repeated until a percentage error between a set of labels obtained from the facial recognition model from the image data 202 and a set of labels from the label vector 204 becomes less than a threshold.

[0085] An example of such a method without backpropagation of error is described in more detail in the article entitled “Learning with Local Gradients at the Edge”, by Michael Lomnitz et al., published on September 16, 2022 under the reference arXiv:2208.08503v2.

[0086] According to a particular embodiment, the operations of the process, in particular but not exclusively the fifth supervised learning operation 25, is implemented by the processor(s) of the computer(s) in charge of the process only when the load level of these processors induced or due to the processing of tasks other than that relating to supervised learning is less than a determined threshold (for example when the load level is less than 25, 30, 40 or 50%).

[0087] This particular embodiment allows the processor(s) to be allocated to supervised learning when they are not heavily used for their assigned routine tasks (for example, controlling and executing one or more functions of an AD / AS system). This ensures that the processors are not used for supervised learning when they are being used to implement the functions necessary for the proper operation of the vehicle 10. It also ensures that the resource required for supervised learning is available when the computations associated with supervised learning are being performed.

[0088] According to another particular embodiment, the process further includes receiving confirmation data indicating the end of supervised learning. This data is, for example, received from the touchscreen interface of screen 12 when the driver (or a passenger) confirms and validates the completion of the facial recognition model's training for production use of the model, for example by pressing a virtual button displayed on the screen.

[0089] According to this other particular embodiment, the process as described in operations 21 to 25 is repeated at the request or command of the driver (or a passenger), for example for the acquisition and recognition of new faces.

[0090] The neural network obtained from the supervised learning described above is, for example, implemented by one or more AD AS systems of the vehicle 10, for example by a system controlling the identity of the person sitting in the seat driver of vehicle 10, for example to control the starting of the vehicle (only for the authorized person(s)) and / or to control the adjustment of the driver's seat, mirrors or any personalized element of vehicle 10 according to control parameters stored in memory and associated with the person thus identified.

[0091] A process for unlearning a facial recognition model for a trained vehicle, according, for example, to the process in [Fig. 2], is implemented by one or more devices embedded in the vehicle 10, for example by one or more computers of the vehicle's embedded network 10, i.e., by one or more processors of the computer(s) in association with one or more memories, for example, a memory of the computer(s). Examples of implementations of such a process are described opposite [Fig. 3] below.

[0092] Fig. 3 schematically illustrates a process of unlearning a neural network implementing a facial recognition model in the vehicle of Fig. 1 and trained according to the process of Fig. 2, according to a particular embodiment of the present invention.

[0093] In a first operation 31 of the unlearning process, label data 301 representative of at least one label of the label vector (203 or 204) are received.

[0094] Label data 301 is, for example, received from the touchscreen interface 12 when the vehicle occupant 10 views images of faces and their associated labels. The occupant may be asked to confirm the selection of a face image and its label because the unlearning operation is permanent.

[0095] According to another embodiment, this tag data 301 is received from a microphone arranged in the passenger compartment of the vehicle 10 when the occupant of the vehicle 10, viewing images of faces and their associated tags, uses a voice communication interface of the vehicle 10.

[0096] In a second operation 32 of the unlearning process, a first subset of training data labeled 302 and a second subset of training data labeled 303 are generated from the labeled training data set (203 or 204) and the labeled data 301.

[0097] The first subset of labeled training data 302 comprises labeled training data from the labeled training dataset (203 or 204) that correspond to said at least one label represented by the received labeled data 301, and the second subset of labeled training data 303 comprises the labeled training data from the dataset labeled training data (203 or 204) that are not included in the first subset of labeled training data 302.

[0098] According to a particular and non-limiting embodiment of the present invention, the first subset of labeled training data 302 can be supplemented by labeled training data obtained from the labeled training data initially forming the first subset of labeled training data 302.

[0099] For example, the first subset of labeled training data 302 can be supplemented with labeled training data obtained by shifting, adding noise and / or rotating the labeled training data initially forming the first subset of labeled training data 302.

[0100] According to another example, the first subset of training data labeled 302 can be supplemented with data from more sophisticated techniques such as, for example, those described in by Jabberi et al. (Generative Data Augmentation applied to Face Recognition. International Conference on Information Networking (ICOIN), Bangkok, Thailand, 2023, pp. 242-247, doi: 10.1109 / ICOIN56518.2023.10049052, https: / / ieeexplore.ieee.org / document / 10049052) or that described by Liu et al. (Approximate Nullspace Augmented Finetuning for Robust Vision Transformers. https: / / arxiv.org / abs / 2403.10476).

[0101] This last example, however, has the disadvantage of increasing computational and memory requirements. Indeed, storing the generative model requires memory, generating new faces requires computation and memory to store them, and then unlearning must be performed on a larger initial subset 302, which also requires more computation.

[0102] In a third operation 33 of the unlearning process, a first data 304 and a second data 305 are calculated from the first subset of training data labeled 302 and the second subset of training data labeled 303 respectively.

[0103] The first data 304 is representative of the importance of the parameters of the parameter set of the trained facial recognition model for the labeled training data of the first subset of labeled training data 302, and the second data 305 is representative of the importance of the parameters of the parameter set of the trained facial recognition model for the labeled training data of the second subset of labeled training data 303.

[0104] According to a particular and non-limiting embodiment of the present invention, a data point representative of the importance of a parameter of The set of parameters of the trained facial recognition model for the labeled training data of a subset 5 = {Xk} of N labeled training data can be calculated by equation (1) below: [°105] n ^4 N ;| 9¾

[0106] In essence, the calculation of the parameter data is the partial derivative with respect to that parameter of the norm of a data / (¾ 0) output of the network of neurons, averaged over all the labeled training data from set 5.

[0107] The first data point 304, also denoted ü / ,, is obtained from equation (1) considering the set s as being the first subset of training data labeled 302 and the second data 305, also noted &rj, is obtained from equation (1) considering the set s as being the second subset of training data labeled 303.

[0108] This third operation 33 requires the most resources compared to the other operations of the unlearning process because it requires the calculation of partial derivatives for all parameters 0(, for all training data of the first and second subset of labeled training data.

[0109] According to a particular and non-limiting embodiment of the present invention, said calculation of the first (304) and second (305) data is implemented by at least one processor of said at least one computer when a load level of said at least one processor induced by processing tasks different from said calculation of the first (304) and second (305) data by said at least one processor is less than a determined threshold.

[0110] According to a particular and non-limiting embodiment of the present invention, to reduce the resources required for the execution of the third operation 33, it may also be advantageous to estimate the partial derivatives using fast algorithms such as the error-free backpropagation method described in the article entitled "Leaming with Local Gradients at the Edge", by Michael Lomnitz et al., published on September 16, 2022 under arXiv:2208.08503v2.

[0111] According to a particular and non-limiting embodiment of the present invention, the data for a parameter 0, of the set of parameters of the trained facial recognition model for the labeled training data of a subset (first 302 or second 303) of labeled training data, can be calculated on the basis of a Fisher information matrix. However, this calculation is then more demanding in terms of computational resources.

[0112] In a fourth operation 34 of the unlearning process, each parameter of the parameter set of the trained facial recognition model is adjusted according to a comparison of the first data 304 and the second data 305 calculated.

[0113] According to a particular and non-limiting embodiment of the present invention, the previously mentioned LFSSD method is used for the implementation of the fourth operation 34.

[0114] Basically, it is decided whether each parameter S, should be reduced as a function of the first data Qy / (304) and the second data / (305), that is to say as a function of the data representative of the importance of this parameter in each of the first 302 and second 303 labeled training data subset.

[0115] If a parameter is estimated to be less important for the first subset 302 than for the second subset, then that parameter remains unchanged. Conversely, if it is estimated to be more important for the first subset 302 than for the second subset, then the parameter is attenuated by a damping factor fi, which may be adjustable.

[0116] More formally, each parameter is adjusted as follows:

[0117]

[0118] with : (HAS ;

[0119] a and 2 are parameters that may be adjustable.

[0120] In the context of unlearning representative face data, the parameters a and 2 can be respectively equal to 10 and 1.

[0121] According to a particular and non-limiting embodiment of the present invention, the parameters a and 2 can be sampled multiple times with different values ​​in a search grid, in which case only the best values ​​can be retained for the final unlearning. This embodiment multiplies the computations by the number of pairs (®,2) that were sampled for this grid search, which can significantly increase the computational resource requirements.

[0122] The final neural network thus lost knowledge of the facial identity or identities corresponding to the labeled training data of the labeled training dataset that correspond to said at least one label represented by the received label data 301, through attenuation of the most important parameters for this labeled training data.

[0123] In a fifth operation 35 of the unlearning process, the labeled training dataset is replaced by the second labeled training dataset subset 303.

[0124] The facial recognition model as a whole has now lost all knowledge of the identity or identities corresponding to the label data 301, by the direct deletion of the labeled training data from the first subset 302, as well as by lossless selective synaptic damping in the neural network (operation 34).

[0125] For example, the second subset of training data labeled 303 is stored in a memory of the vehicle 10 and forms a new database usable for implementing supervised learning of the parameters of the neural network implementing the facial recognition model that will be used in the production or inference phase.

[0126] Fig. 4 schematically illustrates a device 4 configured for learning and unlearning a vehicle facial recognition model implemented in the form of a neural network, according to examples of embodiments of the invention.

[0127] Device 4 advantageously corresponds to a data processing device embedded in a vehicle, for example a computer.

[0128] Device 4 is, for example, configured to carry out the operations described opposite Figures 1 to 2 and / or the steps of the process described opposite [Fig. 5]. Examples of such a device 4 include, but are not limited to, embedded electronic equipment such as a vehicle's on-board computer or an electronic control unit such as an ECU (Electronic Control Unit). The elements of device 4, individually or in combination, may be integrated into a single integrated circuit, into several integrated circuits, and / or into discrete components. Device 4 may be implemented in the form of electronic circuits or software (or computer) modules, or a combination of electronic circuits and software modules.

[0129] The device 4 comprises one (or more) processor(s) 40 configured to execute instructions for carrying out the steps of the process and / or for executing instructions from the software embedded in the device 4. The processor 40 may include integrated memory, an input / output interface, and various circuits known to those skilled in the art. The device 4 further comprises at least one memory 41, for example, volatile and / or non-volatile memory, and / or includes a memory storage device that may include volatile and / or non-volatile memory, such as EEPROM, ROM, PROM, RAM, DRAM, SRAM, flash, magnetic disk, or optical disk.

[0130] The computer code of the embedded software(s) including the instructions to be loaded and executed by the processor is for example stored on memory 41.

[0131] According to various particular and non-limiting embodiments, the device 4 is coupled in communication with other similar devices or systems and / or with communication devices, for example a TCU (Telematic Control Unit), for example via a communication bus or through dedicated input / output ports.

[0132] According to a particular and non-limiting embodiment, the device 4 includes a block 42 of interface elements for communicating with external devices. The interface elements of the block 42 include one or more of the following interfaces: - radio frequency RF interface, for example of the Wi-Fi® type (according to IEEE 802.11), for example in the 2.4 or 5 GHz frequency bands, or of the Bluetooth® type (according to IEEE 802.15.1), in the 2.4 GHz frequency band, or of the Sigfox type using UBN (Ultra Narrow Band) radio technology, or LoRa in the 868 MHz frequency band, LTE (Long-Term Evolution), LTE-Advanced; - USB interface (from the English "Universal Serial Bus" or "Universal Serial Bus" in French); - HDMI interface (from the English "High Definition Multimedia Interface", or "High Definition Multimedia Interface" in French); - LIN interface (from the English "Local Interconnect Network", or in French "Réseau interconnecté local").

[0133] According to another particular and non-limiting embodiment, the device 4 includes a communication interface 43 which enables communication with other devices (such as other computers in the embedded system) via a communication channel 430. The communication interface 43 corresponds, for example, to a transmitter configured to transmit and receive information and / or data via the communication channel 430. The communication interface 43 corresponds, for example, to a wired network of the CAN (Controller Area Network), CAN FD (Controller Area Network Flexible Data-Rate), FlexRay (standardized by ISO 17458) or Ethernet (standardized by ISO / IEC 802-3) type.

[0134] Internal parameters (weights) of the pre-trained neural network can for example be received via the communication interface 43.

[0135] According to a particular and non-limiting embodiment, the device 4 can provide output signals to one or more external devices, such as a display screen 440, touch or not, one or more loudspeakers 450 and / or other peripherals 460 (projection system) via output interfaces 44, 45 and 46 respectively. According to a variant, one or more of the external devices is integrated into the device 4.

[0136] Figure 5 illustrates a flowchart of the different steps in a process for unlearning a neural network implementing a facial recognition model in the vehicle of Figure 1 and trained according to the process of Figure 2, according to a particular and non-limiting embodiment of the present invention. The process is, for example, implemented by device 4 of Figure 4.

[0137] In a first step 51, label data 301 representative of at least one label of the label vector 204 are received.

[0138] In a second step 52, a first subset of training data labeled 302 and a second subset of training data labeled 303 are generated from the labeled training dataset and the labeled data 301.

[0139] In a third step 53, a first data 304 and a second data 305 are calculated from the first subset of training data labeled 302 and the second subset of training data labeled 303.

[0140] In a fourth step 54, each parameter of the parameter set of the trained facial recognition model is adjusted according to a comparison of the first data 304 and the second data 305 calculated.

[0141] In a fifth step 55, the labeled training dataset is replaced by the second labeled training dataset subset 303.

[0142] According to one variant, the variants and examples of the operations described in relation to [Fig.1], 2 and / or 3 apply to the steps of the process in [Fig.5].

[0143] Of course, the present invention is not limited to the embodiments described above but extends to a method for learning and / or unlearning a trained facial recognition model for vehicles, which would include secondary steps without falling outside the scope of the present invention. The same would apply to a device configured for implementing such a method.

[0144] The present invention also relates to a vehicle, for example an automobile or more generally an autonomous land-powered vehicle, comprising the device 4 of [Fig.4].

Claims

1. Demands Method for unlearning a trained vehicle facial recognition model implemented by a neural network, said facial recognition model being trained by refining a set of parameters of said pre-trained facial recognition model, said refining being implemented from a labeled training dataset comprising a label vector (203, 204) associated with face image data (202), said method being implemented by at least one computer embedded in said vehicle, said method comprising the following steps: - receiving (51) label data (301) representative of at least one label of said label vector (203, 204); - generation (52) of a first subset of labeled training data (302) comprising labeled training data of said labeled training data set (203, 204) which correspond to said at least one label represented by the received label data (301), and of a second subset of labeled training data (303) comprising labeled training data of said labeled training data set (203, 204) which are not included in the first subset of labeled training data (302); - calculation (53) of a first data (304) representative of an importance of the parameters of said parameter set of said trained facial recognition model for the labeled training data of said first subset of labeled training data (302) and calculation of a second data (305) representative of an importance of the parameters of said parameter set of said trained facial recognition model for the labeled training data of said second subset of labeled training data (303); - adjustment (54) of each parameter of said parameter set based on a comparison of the first and second calculated data; and - replacement (55) of the labeled training dataset with the second labeled training data subset.

2. A method according to claim 1, wherein the first subset of labeled training data (302) is supplemented by labeled training data obtained from the labeled training data initially forming the first subset of labeled training data (302).

3. A method according to any one of the preceding claims, wherein said calculation of the first (304) and second (305) data is implemented by at least one processor of said at least one computer when a load level of said at least one processor induced by processing tasks other than said calculation of the first (304) and second (305) data by said at least one processor is less than a determined threshold.

4. A method according to any one of the preceding claims, wherein the calculation (53) of the first (304) and second (305) data representing an importance of a parameter of said parameter set of said trained facial recognition model is based on the calculation of a partial derivative with respect to that parameter of an output data of the neural network.

5. A method according to claim 4, wherein the partial derivative is calculated using a method without backpropagation of error.

6. A method according to any one of the preceding claims, wherein each parameter 0t of said parameter set of said trained facial recognition model is adjusted as follows: 3t, with indicates the adjusted parameter, a, and 3 are parameters and ûy / , fl, ; indicate respectively the first and second data.

7. A computer program comprising instructions for carrying out the method according to any one of the preceding claims, when such instructions are executed by at least one processor.

8. Computer-readable recording medium on which is recorded a computer program comprising instructions for carrying out the steps of the process according to any one of claims 1 to 6.

9. 23 Device (4) for unlearning a trained vehicle facial recognition model, said device (4) comprising a memory (41) associated with at least one processor (40) configured to implement at least one step of the method according to any one of claims 1 to 6.

10. Vehicle comprising device (4) according to claim 9.

Images