A method for generating training data for an artificial intelligence of a central electronic computing device, a computer program product,

Abstract

A method for generating training data 28 for an artificial intelligence 14 of a central electronic computing device 12 comprises providing raw data 30 of a motor vehicle by an electronic computing dev

Description

[0001] The present invention relates to the field of automobiles. More specifically, the present invention relates to a method for generating training data for an artificial intelligence of a central electronic computing device according to the pending claim 1. Furthermore, the present invention relates to a corresponding computer program product, to a corresponding non-transitory computer-readable storage medium, as well as to a corresponding system. BACKGROUND INFORMATION

[0002] Data privacy regulations limit the use of personally identifiable information outside the confines of a motor vehicle, unless it is either anonymized or encrypted. Machine learning models can be trained directly on anonymized data but not on encrypted data.

[0003] Different data anonymization techniques are currently available, such as data masking, when values are changed based on some transformation rules, data generalization, when values are mapped to ranges, data perturbation, when random noise is added to the data, and data shuffling, when values are randomly swapped. SUMMARY OF THE INVENTION

[0004] It is an object of the present invention to provide a method, a corresponding computer program product, a corresponding non-transitory computer-storage medium, as well as a corresponding system, by which training data for an artificial intelligence can be generated in an improved manner.

[0005] This object is solved by a method, a corresponding computer program product, a corresponding non-transitory computer-readable storage medium, as well as by a corresponding system according to the independent claims. Advantageous embodiments are presented in the dependent claims.

[0006] One aspect of the invention relates to a method for generating training data for an artificial intelligence of a central electronic computing device by a system. Raw data of a motor vehicle are provided by an electronic computing device of the motor vehicle. The raw data are encrypted by an encoder of an autoencoder of the electronic computing device. The encrypted raw data are transmitted as first encrypted data to the central electronic computing device of the system by a transmitting device of the motor vehicle. The training data are generated depending on the encrypted first data by the central electronic computing device.

[0007] Therefore, an improved method for generating training data for the artificial intelligence of the central electronic computing device, which may be for example provided in a cloud or a backend server, can be provided.

[0008] In particular, by encrypting the data the data is transformed in such a way that the data is unrecognizable, but the data still comprises real numbers that can be red by a human, not like other encrypted data which is just gibberish.

[0009] In particular, most of the known anonymization techniques according to the state of the art are either changing the underlying behavior of the data or they do not provide sufficient protection against attempts to recover the original data. The autoencoder-based anonymization tries to solve these two main problems. With this approach, the relations in the data are preserved, and the large number of model parameters that are learned during training provides a strong protection against attempts of reverse engineering.

[0010] The autoencoder model consists of a fully connected neural network with a specific architecture, which is mainly driven by the number of input features that need to be anonymized and the level of the dimensionality reduction.

[0011] The autoencoder architecture consists of two parts, an encoder and a decoder. The encoder takes the original data and passes it through multiple transformations and reduction layers to reach the desired encoding level. The data dimensionality reduction is dictated by an accuracy level that needs to be attained during the decoding process. The data is encoded in such a way that there is no real connection with original data. The decoder does the same process in reverse, learning to recover the original data based on the encoded values.

[0012] The anonymization approach may unavoidably introduce a level of errors during the decoding process. These errors can be minimized by choosing the right model architecture and tuning the model. In case of GPS location data, where small errors translate into big changes in decoded positions, an additional transformation is introduced. Instead of using and inputting the raw GPS positions, both latitudes and longitudes are converted into their respective degree, minute and second representations. This increases the dimensionality of the data by using six values instead of two, which, however, improves the accuracy of the decoded positions.

[0013] The main advantage is that this approach is very general and can be applied to a wide range of data sources. Also, machine learning models can be trained directly on the anonymized data without a significant loss of accuracy. From a security perspective, since everything can be done in the motor vehicle, the possibility of recovering the original data is minimized.

[0014] The autoencoder model can be trained in the motor vehicle. Thus, both the encoder and the decoder will never leave the context of the motor vehicle. The only thing that may be transferred from the motor vehicle is the encoded data. If someone gains access to the encoded data, it may be almost impossible to guess the right decoder model to be able to recover the original data. To be able to do this, someone may have to guess the right network architecture and the right model weights, which are usually in the order of one or several hundreds to one or several thousands, if not more.

[0015] In particular, as already mentioned, this anonymized data is then used as input to train another model in the central electronic computing device. The encoded data is used to train a multitude of artificial intelligence models. When it comes to large language models, encoded data can be used as input to the model, although most of the data used to train these large language models is mainly text. Model fine-tuning is also possible, and most likely may happen in different situations. The fine-tuning may happen in the cloud on encoded data, however, it may also happen in the motor vehicle, directly on raw data, wherein the choice may depend on the use case considered. Furthermore, prompt engineering is again specific to large language models, and may be needed when dealing with such models. This is completely decoupled from the autoencoder stage.

[0016] According to an embodiment the artificial intelligence is trained with the training data.

[0017] In another embodiment the first encrypted data is further used to be analyzed by the artificial intelligence.

[0018] The analyzed data is transformed to encrypted second data by the central electronic computing device.

[0019] In another embodiment the encrypted second data is transmitted to the transmitting device of the received encrypted second data and the received encrypted second data is decrypted by a decoder of the autoencoder.

[0020] In another embodiment the autoencoder is trained by using the raw data of the motor vehicle.

[0021] In particular, the present invention is a computer-implemented method. Therefore, another aspect of the invention relates to a computer program product comprising program code means for performing a method according to the preceding aspect.

[0022] Furthermore, the present invention relates to a non-transitory computer-readable storage medium comprising at least the computer program product according to the preceding aspect.

[0023] Furthermore, the present invention relates to a system for generating training data for an artificial intelligence of a central electronic computing device, wherein the system is configured for performing a method according the preceding aspect. In particular, the method is performed by the system.

[0024] Furthermore, the system may comprise at least in part means of the motor vehicle in order to generate the training data. Therefore, the system may gain access of the raw data of the motor vehicle in order to generate the training data for the artificial intelligence.

[0025] Advantageous embodiments of the method are to be regarded as advantageous embodiments of the computer program product, the non-transitory computer-readable storage medium, as well as the system.

[0026] A computing unit / electronic computing device may in particular be understood as a data processing device, which comprises processing circuitry. The computing unit can therefore in particular process data to perform computing operations. This may also include operations to perform indexed accesses to a data structure, for example a look-up table, LUT.

[0027] In particular, the computing unit may include one or more computers, one or more microcontrollers, and / or one or more integrated circuits, for example, one or more application-specific integrated circuits, ASIC, one or more field-programmable gate arrays, FPGA, and / or one or more systems-on-a-chip, SoC. The computing unit may also include one or more processors, for example one or more microprocessors, one or more central processing units, CPU, one or more graphics processing units, GPU, and / or one or more signal processors, in particular one or more digital signal processors, DSP. The computing unit may also include a physical or a virtual cluster of computers or other of said units.

[0028] In various embodiments, the computing unit includes one or more hardware and / or software interfaces and / or one or more memory units.

[0029] A memory unit may be implemented as a volatile data memory, for example a dynamic random access memory, DRAM, or a static random access memory, SRAM, or as a non-volatile data memory, for example a read-only memory, ROM, a programmable read-only memory, PROM, an erasable programmable read-only memory, EPROM, an electrically erasable programmable read-only memory, EEPROM, a flash memory or flash EEPROM, a ferroelectric random access memory, FRAM, a magnetoresistive random access memory, MRAM, or a phase-change random access memory, PCRAM.

[0030] Further advantages, features, and details of the invention derive from the following description of preferred embodiments as well as from the drawing. The features and feature combinations previously mentioned in the description as well as the features and feature combinations mentioned in the following description of the figure and / or shown in the figure alone can be employed not only in the respectively indicated combination but also in any other combination or taken alone without leaving the scope of the invention. BRIEF DESCRIPTION OF THE DRAWING

[0031] The novel features and characteristic of the disclosure are set forth in the appended claims. The accompanying drawing, which is incorporated in and constitutes a part of this disclosure, illustrates exemplarily an embodiment and together with the description, serves to explain the disclosed principles. The same numbers are used throughout the figure to reference like features and components. Some embodiments of system and / or methods in accordance with the embodiment of the present subject matter are now described below, by way of example only, and with reference to the accompanying figure.

[0032] Therefore, the only Fig. 1 shows a schematic block diagram according to an embodiment of a system.

[0033] In the figures the same elements are comprising the same reference signs. DETAILED DESCRIPTION

[0034] In the present document, the word "exemplary" is used herein to mean "serving as an example, instance, or illustration". Any embodiment or implementation of the present subject matter described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments.

[0035] While the disclosure is susceptible to various modifications and alternative forms, specific embodiments thereof have been shown by way of example in the drawing and will be described in detail below. It should be understood, however, that it is not intended to limit the disclosure to the particular forms disclosed, but on the contrary, the disclosure is to cover all modifications, equivalents, and alternatives falling within the scope of the disclosure.

[0036] The terms “comprises”, “comprising”, or any other variations thereof, are intended to cover a non-exclusive inclusion so that a setup, device or method that comprises a list of components or steps does not include only those components or steps but may include other components or steps not expressly listed or inherent to such setup or device or method. In other words, one or more elements in a system or apparatus preceded by “comprises” or “comprise” does not or do not, without more constraints, preclude the existence of other elements or additional elements in the system or method.

[0037] In the following detailed description of the embodiment of the disclosure, reference is made to the accompanying drawing that forms part hereof, and in which is shown by way of illustration a specific embodiment in which the disclosure may be practiced. This embodiment is described in sufficient detail to enable those skilled in the art to practice the disclosure, and it is to be understood that other embodiments may be utilized and that changes may be made without departing from the scope of the present disclosure. The following description is, therefore, not to be taken in a limiting sense.

[0038] Fig. 1 shows a schematic block diagram according to an embodiment of a system 10. The system 10 comprises at least one central electronic computing device 12, which is in particular configured as a cloud server. An artificial intelligence 14 is provided by the central electronic computing device 12. Furthermore, a motor vehicle 16 is shown. The motor vehicle 16 may also comprise an electronic computing device 18. The electronic computing device 18 may comprise at least an autoencoder 20 comprising an encoder 22 and a decoder 24. Furthermore, the motor vehicle 16 may comprise a transmitting device 26.

[0039] In particular, a method for generating training data 28 for the artificial intelligence 14 is provided. Raw data 30 are provided by the electronic computing device 16. The raw data 30 is encrypted to first encrypted data 32 by the encoder 22 and transmitted to the central electronic computing device 12. The training data 28 is generated depending on the encrypted first data 32 by the central electronic computing device 12.

[0040] Furthermore, the artificial intelligence 14 is trained by the generated training data 28. Furthermore, the first data is further used to be analyzed by the artificial intelligence 14. Furthermore, the analyzed data is transformed to encrypted second data 34 by the central electronic computing device 12. The encrypted second data 34 may be transmitted to the transmitting device 26 and the received encrypted second data 34 is decrypted by the decoder 24 of the autoencoder 20. The autoencoder 20 is trained by using the raw data 30 of the motor vehicle 16.

[0041] In particular, Fig. 1 shows that first, the autoencoder 20, in particular both encoder 22 and decoder 24, are trained in the motor vehicle 16 where personalized data resides. The data is then encoded in the motor vehicle 16 and the encoded, in particular anonymized data as the encrypted first data 32, is transferred out of the motor vehicle 16 to the central electronic computing device 12. Currently, some of the personal data, for example location, user data, or vehicle VIN, is not transferred to the central electronic computing device 12 at all. Some other data, such as vehicle velocity, acceleration, is anonymized using a data generalization approach where values are mapped to range, while other data is either shifted and / or scaled by some factors.

[0042] Embeddings are usually created once the data is transferred to the cloud, which represents a next stage of data processing, and depends on the machine learning use case. Some artificial intelligence models use / need embeddings, other models do not.

[0043] Technically, any data recorded or generated in the motor vehicle 16 can be anonymized using the autoencoder 20. Also multiple autoencoders 20 may be used, each targeting a specific data feature set.

[0044] The anonymized data is then used as input to train the artificial intelligence 14 in the central electronic computing device 12. The encoded data can be used to train a multitude of artificial intelligence models. When it comes to for example large language models, encoded data can be used as input to the model, although most of the data to train these large language models is mainly text.

[0045] Furthermore, a model fine-tuning is possible, and most likely may happen in different situations. The fine-tuning may happen in the cloud on encoded data, however may also happen in the motor vehicle 16, directly on raw data 30, wherein the choice may depend on the use case considered.

[0046] Prompt engineering is again specific to large language models and may be needed when dealing with such models. This is completely decoupled from the autoencoder stage.

[0047] The output of the artificial intelligence 14 is then sent back to the motor vehicle 16 and decoded in the motor vehicle 16. With this approach, both parts of the autoencoder 20, the encoder 22 and the decoder 24, never leave the motor vehicle 16, so that even if someone might intercept the data, there is no way that it can be probably decoded.

[0048] To train the artificial intelligence 14 first there is a need to encode the raw data 32 in the motor vehicle 16 and transfer it to the central electronic computing device 12. This encoded data is then used to train the artificial intelligence model in the cloud, in particular the encoded trained data contains the model input features, such as position, speed, acceleration, and the target data features, which the artificial intelligence model learns to predict, such as particular functions in the motor vehicle 16. Once the artificial intelligence model is trained, it can be used to perform predictions on new input data, which again is encoded, which is how the artificial intelligence model is trained. In return, the output of the artificial intelligence model may be the target data value that the model learned to predict. This predicted value is the encoded value of the raw value in the motor vehicle 16. This encoded value that the artificial intelligence model predicts is then sent back to the motor vehicle 16 and decoded, in particular using the decoder 24 of the autoencoder 20, to map it to the raw value that the motor vehicle 16 or the user can act on.

[0049] This approach may enable the central electronic computing device 12 to train personalized artificial intelligence models in the cloud without the need of a complex data encryption system. This may also allow the central electronic computing device 12 to use data that currently is not available outside the motor vehicle 16. This translates into the potentially better artificial intelligence models.

[0050] Furthermore, the encoding of the GPS data is a slight addition to the whole anonymization process. The main drawback when encoding road GPS positions is that very small decoding errors translate into huge position errors, which renders the autoencoder-based anonymization process useless for the particular type of data. One solution is to use the degree, minute, second representation of the GPS positions as input to the autoencoder 20. In this way, the GPS data can be added into the mix and encoded alongside other data features to be used in the artificial intelligence model.

[0051] The architecture of the autoencoder 20 may mainly depend on the set of data features that is needed to be bundled together. As an example, an autoencoder 20 is trained to anonymize together vehicle dynamic signals such as velocity, acceleration, inclination angles and furthermore. The number of data features that are bundled together determines the number of nodes in the input and output layers of the autoencoder network. The rest of the architecture, such as the number of hidden layers, the dimensionality reduction level is mainly driven by minimizing the decoding error of the network. These parameters are usually obtained by tuning the model to minimize the decoding error.

[0052] The encoder part of the autoencoder 20 can be used alone, and there might be use cases when this is all that is needed. For example, the artificial intelligence model is trained to predict a value, such as driving mode, based on vehicle dynamic signals, such as velocity or acceleration. The vehicle dynamic signals may have to be encoded before sending them to the central electronic computing device 12. The drive mode signal is not required to be encoded since it contains no personal identifiable data. In this situation, only the encoder 22 is used to anonymize the vehicle data. Since the predicted value is the raw value, there is no need to use a decoder 24.

[0053] Depending on the artificial intelligence model that needs to be trained and depending on how many data bundles are needed, situations such as using the encoder parts of several autoencoders 20 and the decoder part of a specific autoencoder 20 may happen quite often. This means that there may be situations when a mix of encoders 22 and decoders 24 from different autoencoders 20 are used together as a part of the general workflow. Reference Signs system central electronic computing device artificial intelligence motor vehicle electronic computing device autoencoder encoder decoder transmitting device training data raw data encrypted first data encrypted second data

Claims

1. A method for generating training data (28) for an artificial intelligence (14) of a central electronic computing device (12) by a system (10), comprising the steps of: - providing raw data (30) of a motor vehicle (16) by an electronic computing device (18) of the motor vehicle (16);- encrypting the raw data (30) by an encoder (22) of an autoencoder (20) of the electronic computing device (18);- transmitting the encrypted raw data as encrypted first data (32) to the central electronic computing device (12) of the system (10) by a transmitting device (26) of the motor vehicle (16); and- generating the training data (28) depending on the encrypted first data (32) by the central electronic computing device (12).

2. The method according to claim 1, characterized in thatthe artificial intelligence (14) is trained with the training data (28).

3. The method according to claim 1 or 2, characterized in thatthe first data (32) is further used to be analyzed by the artificial intelligence (14).

4. The method according to claim 3, characterized in thatthe analyzed data is transmitted as encrypted second data (34) to the transmittingdevice (26) and the received encrypted second data (34) is decrypted by a decoder (24) of the autoencoder (20).

5. The method according to any one of claims 1 to 4, characterized in thatthe autoencoder (20) is trained by using the raw data (30) of the motor vehicle (16).

6. A computer program product comprising program code means for performing a method according to any one of claims 1 to 5.

7. A non-transitory computer-readable storage medium comprising at least the computer program product according to claim 6.

8. A system (10) for generating training data (28) for an artificial intelligence (14) of a central electronic computing device (12), wherein the system (10) is configured for performing a method according to any one of claims 1 to 5.

Images