ABSTRACT
Abstract
A method for evaluating data is based on a computational model, the computational model comprising model data, a training function and a prediction function. The method includes training the computational model by: receiving training data and training result data for training the computational model, and computing the model data from the training data and the training result data with the training function. The method includes predicting result data by: receiving field data for predicting result data; and computing the result data from the field data and the model data with the prediction function. The training data may be plaintext and the training result data may be encrypted with a homomorphic encryption algorithm, wherein the model data may be computed in encrypted form from the training data and the encrypted training result data with the training function. The field data may be plaintext, wherein the result data may be computed in encrypted form from the field data and the encrypted model data with the prediction function.
Description
FIELD OF THE INVENTION
The invention relates to the field of homomorphic encryption. In particular, the invention relates to a method, a computer program and a system for evaluating data based on a computational model.
BACKGROUND OF THE INVENTION
The computation of key performance indicators and of trend analysis requires to use machine learning techniques on various types of data, such as sensor data or weather data. Machining learning techniques may comprise linear regression, support vector machines, k-nearest neighbors, to name but a few. With those algorithms, a prediction model may be trained with the aid of training data and then this prediction model may be used to perform predictions on new data. In order to improve the precision of the predictions, it is desirable to train the model on a training data set that is as large as possible. This usually requires large amounts of computational resources.
Cloud Computing may offer such resources in an economically attractive way by providing access to a shared pool of configurable and inexpensive computing and storage resources to store and process data in third-party data centers. However, some of the data input and/or produced by these algorithms may be confidential.
Sending business-relevant data to a third-party cloud provider requires trust. In order to circumvent the trust issue, it may be possible to encrypt the confidential data before sending it to the cloud. If the data should also be processed in the cloud, then commonly used encryptions schemes, such as AES, DES, blowfish, require the data to be decrypted before it can be processed. In these cases, either the user stores the encryption keys in the cloud, where the cloud provider may use the keys to decrypt the data, or the user downloads the encrypted data from the cloud, decrypts it, and processes it locally, so the computational resources of the cloud are not utilized.
Another approach is to use homomorphic encryption schemes, which enable the processing of data in the cloud without providing decryption keys to the cloud provider. With homomorphic encryption it is possible to carry out computations, such as machine learning techniques, on data that is stored in encrypted form. However, known approaches for such computations on encrypted data slow down the computations to such an extent that the computations become impractical.
DESCRIPTION OF THE INVENTION
It is an objective of the invention to provide secure and computational feasible machine learning techniques that may be executed in a cloud computing system.
This objective is achieved by the subject-matter of the independent claims. Further exemplary embodiments arc evident from the dependent claims and the following description.
An aspect of the invention relates to a method for evaluating data based on a computational model, the computational model comprising model data, a training function and a prediction function. With the training function, the model may be trained with training data, and with the prediction function, predictions may be made based on new data. The computational model may be seen as a machine learning technique, since it encodes its knowledge about the training data in the model data. The computational model may model a real system, such as a wind park or a plurality of solar cell panels at distributed locations. For example, the computational model may predict the power output of a wind park or the plurality of solar cell panels based on weather data.
In general, the training of the computational model and the predictions may be computed mainly on an evaluation system with big computing resources, such as a cloud computing facility. It may be possible that some of the calculations, which need a decryption of data, are performed on a client device. However, most of the computations may be provided by the evaluation system. Thus, the computations âmainlyâ may be performed in the evaluation system. In the evaluation system, no encryption or decryption of data may take place.
According to an embodiment of the invention, the method comprises training the computational model by: receiving training data and training result data for training the model, and computing the model data from the training data and the training result data with the training function. The training data may be data representing data that may be generated by devices, which determine specific parameters relating to a real system, such as devices acquiring weather data. The training result data may be data representing the result the computational model should generate when the training data is applied to the prediction function.
For example, a client device, which may be communicatively connected with the evaluation device, may send the training data and the training result data to the evaluation system, where the model data is mainly computed.
According to an embodiment of the invention, the method comprises predicting result data by: receiving field data for predicting result data, and computing the result data from the field data and the model data with the prediction function. While the training data and the training result data is used for training the model, i.e. computing the model data, the field data is data used for computing the result data based on the model data. Field data may be real data collected by external devices.
The field data may be data corresponding to the training data, i.e. may have the format and/or an analogous information content. For example, the field data may be collected by one or a plurality of devices, which may be seen as collection device, and may be sent to the evaluation system, where the predictions are mainly calculated. For example, the collection device(s) may be device(s) of weather forecast provider(s). The result data may be data corresponding to the training result data, i.e. may have the format and/or an analogous information content. For example, the result data may encode the power output of wind parks and/or solar panels, which has been predicted from weather forecast data as field data.
The result data may be sent to the client device, which has provided the training data and the training result data, or to another client device.
The training and the prediction may take place on encrypted data, which is processed based on homomorphic encryption. Homomorphic encryption provides the possibility that calculations may be based on plaintext data and encrypted data simultaneously. There may be one or more functions mapping a plaintext first data value and an encrypted second data value to an encrypted result data value, which are compatible with the encryption. When the result data value is decrypted, it may have the same value, as when a specific other function, such as addition or multiplication, is applied to the first data value and the encrypted second data value.
With such functions compatible with homomorphic encryption, it may be possible that not all of the three sets of data during training and prediction, i.e. training data or field data, training result data or result data, and model data, are encrypted but only two of the three sets of data. In such a way, computational power may be saved while preserving data confidentiality.
In general, with the method, computations for machine learning algorithms may be performed on encrypted data to achieve a level of performance that makes their application on encrypted data practically feasible. With the method, only a part of the data is encrypted. This may be applicable in settings where a part of the data is publicly available, such as weather and/or wind data, and other data is confidential, for example, the power output of a wind farm.
According to an embodiment of the invention, the training data is plaintext and the training result data is encrypted with a homomorphic encryption algorithm, wherein the model data is computed in encrypted form from the training data and encrypted training result data with the training function; and the field data is plaintext, wherein the result data is computed in encrypted form from the field data and the encrypted model data with the prediction function. In other words, the input data to the training function and the prediction data is plaintext, while the model data and the output data of training function and the prediction function is encrypted.
Since the input data, i.e. training data and field data, is usually the larger data set, this may additionally save computational power for encryption. Furthermore, the collection of devices need not provide the ability to encrypt data.
According to an embodiment of the invention, the training data and the training result data are encrypted with a homomorphic encryption algorithm, wherein the model data is computed in plaintext from the encrypted training data and the encrypted training result data with the training function; and the field data is encrypted with the homomorphic encryption algorithm, wherein the result data is computed in encrypted form from the encrypted field data and the plaintext model data with the prediction function. Here, only the input data and the output data of the training function and the prediction function are encrypted, while the model may stay in plaintext. This may have the additional advantage that only encrypted data has to be transmitted via communication lines, such as the internet.
According to an embodiment of the invention, the homomorphic encryption algorithm is additively homomorphic and multiplicative homomorphic for multiplication with a plaintext factor. For example, the Paillier cryptosystem, which is a homomorphic encryption algorithm, has such a property. Multiplication of two encrypted values may result in an encrypted value representing the sum of the two corresponding plaintext values. An encrypted value raised to the power with a plaintext exponent, results in an encrypted value representing the multiplication of the exponent with the corresponding plaintext value.
According to an embodiment of the invention, the training function is a polynomial in at least one of the training data and the training result data and/or the prediction function is a polynomial in at least one of the field data and model data. At least for polynomial functions, which only comprise additions and multiplications, a calculation scheme may be found, in which the plaintext data is only multiplied with the encrypted data and the multiplicative homomorphism for multiplication with a plaintext factor may be used.
According to an embodiment of the invention, each of the training data and the field data is a set of vectors or matrices of data values. Each of the training result data and the result data may be a vector of result data values. The model data may be a vector of model data values. In general, the training data and the field data may be provided as a set of m samples, wherein each sample comprises a vector of n values. The training result data and the result data may be provided as m values. The model data may be provided as a set of n values, which, for example, may be used for determining a value of the result data based on a vector of field data.
According to an embodiment of the invention, the model is a linear regression model, in which the prediction function is a linear function in the field data and the model data. A linear regression model is a model, in which a relation between the training data and field data and the corresponding result data is assumed to be linear. For linear regression, the model data is found that minimizes the deviation of the training data from a linear relationship.
The training function of the linear regression model may be based on minimizing a cost function, which quadratically minimizes a difference between the prediction function and the training result data.
For example, the training function may be the normal equation for inverting the prediction function based on minimizing the cost function. Thus, for calculating the model data, a in general rather big matrix has to be inverted.
The training function also may be a gradient descent function for iteratively inverting the prediction function based on minimizing the cost function. In this case, a in general rather large number of iterations have to be performed for calculating the model data.
According to an embodiment of the invention, before encryption, data values from the training data and/or the training result data are approximated by multiplication with an approximation factor and rounding to a closest integer. In such a way, rounding errors of the operations performed on encrypted data may be controlled. In particular, the homomorphic encryption algorithm may be based on a finite field based on the product of two prime numbers and the approximation factor is smaller than this product. One divided by the approximation factor may be seen as the smallest possible increment, which defines the precision. Furthermore, the product of the two prime numbers divided by the approximation factor may be seen as the largest possible value.
According to an embodiment of the invention, the training function for computing the model data is iteratively updated with the training data and the training result data. For example, this may be the case when gradient descent is used. The training function may be an equation in which a next value of the model data is computed by adding or subtracting a correction factor multiplied by a convergence factor to an actual value of the model data. The correction factor is calculated from the training result data and the actual model data. The convergence factor may comprise or may be a parameter divided by the number m of samples of the training data.
According to an embodiment of the invention, in each iteration, the training function is evaluated in an evaluation server device, the updated model data, i.e. the correction factor, is sent to a secure device, decrypted, multiplied with a convergence factor and encrypted in the secure device and sent back to evaluation server device.
The correction factor only may comprise additions and multiplications with values from the actual training data, training result data and model data, which all may be approximated to integers as described above. The convergence factor may be a rather small value, which would result in a loss of precision when multiplied in approximated form, i.e. multiplied with the approximation factor and rounded to integer to the correction factor. Thus, this multiplication may be performed in decrypted form on the secure device.
The server device may be a device with high computing power, but which may be rated as unsecure. The secure device may be a client device with much less computing power, but which is rated as secure. Thus, computational computing power may be provided by the server device, where most of the calculations are performed. The secure device may provide much less computational power than the server device. However, the much fewer operations, which need higher precision, may be performed with the secure device.
According to an embodiment of the invention, the training data and/or the training result data is provided by a client device communicatively interconnected with an evaluation server device, wherein the client device encrypts the training data and/or the training result data and/or decrypts the result data and wherein the evaluation server device at least partially computes the model data and/or the result data. In such a way, the client device, which may be a secure device, only or nearly only has to perform the encryption and decryption, while the computational more demanding operations may all be performed in a secure way on the server device.
According to an embodiment of the invention, the field data is provided by at least one or a plurality of devices communicatively interconnected with the evaluation server device. Also the field devices only may have to perform the encryption of field data. The one
FIELD OF THE INVENTION
The invention relates to the field of homomorphic encryption. In particular, the invention relates to a method, a computer program and a system for evaluating data based on a computational model.
BACKGROUND OF THE INVENTION
The computation of key performance indicators and of trend analysis requires to use machine learning techniques on various types of data, such as sensor data or weather data. Machining learning techniques may comprise linear regression, support vector machines, k-nearest neighbors, to name but a few. With those algorithms, a prediction model may be trained with the aid of training data and then this prediction model may be used to perform predictions on new data. In order to improve the precision of the predictions, it is desirable to train the model on a training data set that is as large as possible. This usually requires large amounts of computational resources.
Cloud Computing may offer such resources in an economically attractive way by providing access to a shared pool of configurable and inexpensive computing and storage resources to store and process data in third-party data centers. However, some of the data input and/or produced by these algorithms may be confidential.
Sending business-relevant data to a third-party cloud provider requires trust. In order to circumvent the trust issue, it may be possible to encrypt the confidential data before sending it to the cloud. If the data should also be processed in the cloud, then commonly used encryptions schemes, such as AES, DES, blowfish, require the data to be decrypted before it can be processed. In these cases, either the user stores the encryption keys in the cloud, where the cloud provider may use the keys to decrypt the data, or the user downloads the encrypted data from the cloud, decrypts it, and processes it locally, so the computational resources of the cloud are not utilized.
Another approach is to use homomorphic encryption schemes, which enable the processing of data in the cloud without providing decryption keys to the cloud provider. With homomorphic encryption it is possible to carry out computations, such as machine learning techniques, on data that is stored in encrypted form. However, known approaches for such computations on encrypted data slow down the computations to such an extent that the computations become impractical.
DESCRIPTION OF THE INVENTION
It is an objective of the invention to provide secure and computational feasible machine learning techniques that may be executed in a cloud computing system.
This objective is achieved by the subject-matter of the independent claims. Further exemplary embodiments arc evident from the dependent claims and the following description.
An aspect of the invention relates to a method for evaluating data based on a computational model, the computational model comprising model data, a training function and a prediction function. With the training function, the model may be trained with training data, and with the prediction function, predictions may be made based on new data. The computational model may be seen as a machine learning technique, since it encodes its knowledge about the training data in the model data. The computational model may model a real system, such as a wind park or a plurality of solar cell panels at distributed locations. For example, the computational model may predict the power output of a wind park or the plurality of solar cell panels based on weather data.
In general, the training of the computational model and the predictions may be computed mainly on an evaluation system with big computing resources, such as a cloud computing facility. It may be possible that some of the calculations, which need a decryption of data, are performed on a client device. However, most of the computations may be provided by the evaluation system. Thus, the computations âmainlyâ may be performed in the evaluation system. In the evaluation system, no encryption or decryption of data may take place.
According to an embodiment of the invention, the method comprises training the computational model by: receiving training data and training result data for training the model, and computing the model data from the training data and the training result data with the training function. The training data may be data representing data that may be generated by devices, which determine specific parameters relating to a real system, such as devices acquiring weather data. The training result data may be data representing the result the computational model should generate when the training data is applied to the prediction function.
For example, a client device, which may be communicatively connected with the evaluation device, may send the training data and the training result data to the evaluation system, where the model data is mainly computed.
According to an embodiment of the invention, the method comprises predicting result data by: receiving field data for predicting result data, and computing the result data from the field data and the model data with the prediction function. While the training data and the training result data is used for training the model, i.e. computing the model data, the field data is data used for computing the result data based on the model data. Field data may be real data collected by external devices.
The field data may be data corresponding to the training data, i.e. may have the format and/or an analogous information content. For example, the field data may be collected by one or a plurality of devices, which may be seen as collection device, and may be sent to the evaluation system, where the predictions are mainly calculated. For example, the collection device(s) may be device(s) of weather forecast provider(s). The result data may be data corresponding to the training result data, i.e. may have the format and/or an analogous information content. For example, the result data may encode the power output of wind parks and/or solar panels, which has been predicted from weather forecast data as field data.
The result data may be sent to the client device, which has provided the training data and the training result data, or to another client device.
The training and the prediction may take place on encrypted data, which is processed based on homomorphic encryption. Homomorphic encryption provides the possibility that calculations may be based on plaintext data and encrypted data simultaneously. There may be one or more functions mapping a plaintext first data value and an encrypted second data value to an encrypted result data value, which are compatible with the encryption. When the result data value is decrypted, it may have the same value, as when a specific other function, such as addition or multiplication, is applied to the first data value and the encrypted second data value.
With such functions compatible with homomorphic encryption, it may be possible that not all of the three sets of data during training and prediction, i.e. training data or field data, training result data or result data, and model data, are encrypted but only two of the three sets of data. In such a way, computational power may be saved while preserving data confidentiality.
In general, with the method, computations for machine learning algorithms may be performed on encrypted data to achieve a level of performance that makes their application on encrypted data practically feasible. With the method, only a part of the data is encrypted. This may be applicable in settings where a part of the data is publicly available, such as weather and/or wind data, and other data is confidential, for example, the power output of a wind farm.
According to an embodiment of the invention, the training data is plaintext and the training result data is encrypted with a homomorphic encryption algorithm, wherein the model data is computed in encrypted form from the training data and encrypted training result data with the training function; and the field data is plaintext, wherein the result data is computed in encrypted form from the field data and the encrypted model data with the prediction function. In other words, the input data to the training function and the prediction data is plaintext, while the model data and the output data of training function and the prediction function is encrypted.
Since the input data, i.e. training data and field data, is usually the larger data set, this may additionally save computational power for encryption. Furthermore, the collection of devices need not provide the ability to encrypt data.
According to an embodiment of the invention, the training data and the training result data are encrypted with a homomorphic encryption algorithm, wherein the model data is computed in plaintext from the encrypted training data and the encrypted training result data with the training function; and the field data is encrypted with the homomorphic encryption algorithm, wherein the result data is computed in encrypted form from the encrypted field data and the plaintext model data with the prediction function. Here, only the input data and the output data of the training function and the prediction function are encrypted, while the model may stay in plaintext. This may have the additional advantage that only encrypted data has to be transmitted via communication lines, such as the internet.
According to an embodiment of the invention, the homomorphic encryption algorithm is additively homomorphic and multiplicative homomorphic for multiplication with a plaintext factor. For example, the Paillier cryptosystem, which is a homomorphic encryption algorithm, has such a property. Multiplication of two encrypted values may result in an encrypted value representing the sum of the two corresponding plaintext values. An encrypted value raised to the power with a plaintext exponent, results in an encrypted value representing the multiplication of the exponent with the corresponding plaintext value.
According to an embodiment of the invention, the training function is a polynomial in at least one of the training data and the training result data and/or the prediction function is a polynomial in at least one of the field data and model data. At least for polynomial functions, which only comprise additions and multiplications, a calculation scheme may be found, in which the plaintext data is only multiplied with the encrypted data and the multiplicative homomorphism for multiplication with a plaintext factor may be used.
According to an embodiment of the invention, each of the training data and the field data is a set of vectors or matrices of data values. Each of the training result data and the result data may be a vector of result data values. The model data may be a vector of model data values. In general, the training data and the field data may be provided as a set of m samples, wherein each sample comprises a vector of n values. The training result data and the result data may be provided as m values. The model data may be provided as a set of n values, which, for example, may be used for determining a value of the result data based on a vector of field data.
According to an embodiment of the invention, the model is a linear regression model, in which the prediction function is a linear function in the field data and the model data. A linear regression model is a model, in which a relation between the training data and field data and the corresponding result data is assumed to be linear. For linear regression, the model data is found that minimizes the deviation of the training data from a linear relationship.
The training function of the linear regression model may be based on minimizing a cost function, which quadratically minimizes a difference between the prediction function and the training result data.
For example, the training function may be the normal equation for inverting the prediction function based on minimizing the cost function. Thus, for calculating the model data, a in general rather big matrix has to be inverted.
The training function also may be a gradient descent function for iteratively inverting the prediction function based on minimizing the cost function. In this case, a in general rather large number of iterations have to be performed for calculating the model data.
According to an embodiment of the invention, before encryption, data values from the training data and/or the training result data are approximated by multiplication with an approximation factor and rounding to a closest integer. In such a way, rounding errors of the operations performed on encrypted data may be controlled. In particular, the homomorphic encryption algorithm may be based on a finite field based on the product of two prime numbers and the approximation factor is smaller than this product. One divided by the approximation factor may be seen as the smallest possible increment, which defines the precision. Furthermore, the product of the two prime numbers divided by the approximation factor may be seen as the largest possible value.
According to an embodiment of the invention, the training function for computing the model data is iteratively updated with the training data and the training result data. For example, this may be the case when gradient descent is used. The training function may be an equation in which a next value of the model data is computed by adding or subtracting a correction factor multiplied by a convergence factor to an actual value of the model data. The correction factor is calculated from the training result data and the actual model data. The convergence factor may comprise or may be a parameter divided by the number m of samples of the training data.
According to an embodiment of the invention, in each iteration, the training function is evaluated in an evaluation server device, the updated model data, i.e. the correction factor, is sent to a secure device, decrypted, multiplied with a convergence factor and encrypted in the secure device and sent back to evaluation server device.
The correction factor only may comprise additions and multiplications with values from the actual training data, training result data and model data, which all may be approximated to integers as described above. The convergence factor may be a rather small value, which would result in a loss of precision when multiplied in approximated form, i.e. multiplied with the approximation factor and rounded to integer to the correction factor. Thus, this multiplication may be performed in decrypted form on the secure device.
The server device may be a device with high computing power, but which may be rated as unsecure. The secure device may be a client device with much less computing power, but which is rated as secure. Thus, computational computing power may be provided by the server device, where most of the calculations are performed. The secure device may provide much less computational power than the server device. However, the much fewer operations, which need higher precision, may be performed with the secure device.
According to an embodiment of the invention, the training data and/or the training result data is provided by a client device communicatively interconnected with an evaluation server device, wherein the client device encrypts the training data and/or the training result data and/or decrypts the result data and wherein the evaluation server device at least partially computes the model data and/or the result data. In such a way, the client device, which may be a secure device, only or nearly only has to perform the encryption and decryption, while the computational more demanding operations may all be performed in a secure way on the server device.
According to an embodiment of the invention, the field data is provided by at least one or a plurality of devices communicatively interconnected with the evaluation server device. Also the field devices only may have to perform the encryption of field data. The one or more devices may be field devices directly acquiring sensor data as field data. The client device providing the training data also may be the device providing the field data.
According to an embodiment of the invention, the training data and the field data encodes local wind speed and/or local sunshine intensity and the training result data and the result data encodes electrical power production of wind turbine facilities and/or solar energy facilities. One application of the method may be the secure computation of forecast data for electrical power networks. A forecast of the power output of a plurality of the renewable power sources may be calculated in a cloud computing facility without the danger of publishing training data and/or model data to third parties. Such forecasts may be important for predicting prices when electrical power is traded.
A further aspect of the invention relates to a computer program for evaluating data based on a computational model, which, when executed on an evaluation system, is adapted for performing the method as described above and in the following, as well as to a computer-readable medium on which such a computer program is stored. A computer-readable medium may be a floppy disk, a hard disk, an USB (Universal Serial Bus) storage device, a RAM (Random Access Memory), a ROM (Read Only Memory), an EPROM (Erasable Programmable Read Only Memory) or a FLASH memory. A computer-readable medium may also be a data communication network, e.g. the Internet, which allows downloading a program code. In general, the computer-readable medium may be a non-transitory or transitory medium.
A further aspect of the invention relates to an evaluation system adapted for performing the method as described above and in the following.
According to an embodiment of the invention, the system comprises an evaluation server device, which may be a cloud computing system/facility. The server device needs not be secure, i.e. restricted from access from third parties, since all of the operations may be performed in encrypted form in the server device. Additionally, the model data may be stored in the server device only in encrypted form.
The evaluation system furthermore may comprise a client device communicatively interconnected with the server device. This client device may be a secure device and may provide the training data and training result data to the server device. The client device may be a secure device adapted for encrypting and decrypting the training data, training result data and/or result data. The client device need not be adapted for performing calculations with high computational power.
The evaluation system furthermore may comprise a plurality of field devices communicatively interconnected with the server device, which may provide field data to the server device. These field devices may only be adapted for encrypting field data, but need not be adapted for decrypting field data and/or for performing calculations with high computational power.
It has to be understood that features of the method as described in the above and in the following may be features of the computer program, the computer-readable medium and evaluation system as described in the above and in the following, and vice versa.
These and other aspects of the invention will be apparent from and elucidated with reference to the embodiments described hereinafter.
BRIEF DESCRIPTION OF THE DRAWINGS
The subject-matter of the invention will be explained in more detail in the following text with reference to exemplary embodiments which are illustrated in the attached drawings.
FIG. 1 schematically shows a system evaluation system according to an embodiment of the invention and illustrates a method for evaluating data based on a computational model system according to an embodiment of the invention.
FIG. 2 schematically shows a system evaluation system according to a further embodiment of the invention and illustrates a method for evaluating data based on a computational model system according to a further embodiment of the invention.
The reference symbols used in the drawings, and their meanings, are listed in summary form in the list of reference symbols. In principle, identical parts are provided with the same reference symbols in the figures.
DETAILED DESCRIPTION OF EXEMPLARY EMBODIMENTS
FIG. 1 and FIG. 2 show an evaluation system 10 comprising a server device 12 , which may be a cloud computing facility. The system 10 furthermore comprises
several computing devices
14 , 16 , 18 and 20 , which are communicatively interconnected with the server device 12 , for example via Internet. The arrows in FIG. 1 illustrate method steps that will be described in more detail below. The direction of these arrows illustrates a possible data flow between the server device 12 and the
computing devices
16 , 18 , 20 .
The training device 14 may be adapted for encrypting data with a homomorphic encryption algorithm, for example based on the Paillier encryption scheme, and may be adapted for sending plaintext, i.e. unencrypted, data and encrypted data to the server device 12 .
In a first method step S 12 , the training device 14 sends training data 22 and training result data 24 to the server device 12 , which trains a computational model by receiving the training data 22 and training result data 24 and by computing model data 26 from the training data 22 and the training result 24 data with a training function.
The computational model may comprise the model data 26 , the training function used during a training phase and a prediction function used during a prediction phase of the method.
On the server device 12 , no data is encrypted and decrypted. Only homomorphic operations compatible with the homomorphic encryption scheme are applied to the
data
22 , 24 , 26 .
The evaluation system 10 may comprise a secure device 16 , which is adapted for encrypting and decrypting data with the homomorphic encryption algorithm. The term âsecureâ refers to the fact that the server device 12 needs not be secured with respect to third parties, i.e. it needs not be protected from an illegal access of third parties, which are interested in reading information from the data. Contrary to the server device 12 , the
devices
14 , 16 , 18 may be seen as âsecureâ, since they may store and/or process confidential data in plaintext form.
During the step S 12 , when the server device 12 computes the model data 26 , this may be performed iteratively and the server device 12 may send partial results in a step S 14 to the secure device 16 , which decrypts the partial results, performs calculations with the partial results, encrypts the calculation results and sends them back to the server device 12 . It has to be noted that most of the calculations, such as more than 95% or more than 99%, may be performed in the server device 12 and only the rest may be performed by the secure device 16 .
The secure device 16 and the training device 14 may be provided by the same computational hardware.
Step S 12 and S 14 refer to a training phase of the computational model. After the training phase, model data 26 may be stored, in for example encrypted or decrypted form, in the server device 12 .
The evaluation system 10 also may comprise one or more field devices 20 , which are adapted for sending field data 28 to the server device 12 . The field data may be encrypted or in plaintext, in the first case, the one or more field devices 20 may be adapted for encrypting field data 28 with the homomorphic encryption algorithm.
In step S 16 , the one or more field devices 20 send the field data 28 to the server device, which predicts result data 30 by receiving field data 28 and computing the result data 30 from the field data 28 and the model data with the prediction function.
In step S 18 , the result data 30 is sent from the server device 12 to a beneficiary device 18 , which is adapted for decrypting the result data 30 , which then may be further processed. The beneficiary device 18 may be provided by the same computational hardware as the training device 14 and/or the secure device 16 .
In both of the systems 10 of FIG. 1 and FIG. 2 , only a part of the
data
22 , 24 , 26 , 28 , 30 is encrypted to save computational power.
In FIG. 1 , the training data 22 is plaintext and the training result data 24 is encrypted with the homomorphic encryption algorithm, wherein the model data 26 is computed in encrypted form from the training data 22 and the encrypted training result data 24 with the training function. Furthermore, the field data 28 is plaintext, wherein the result data 30 is computed in encrypted form from the field data 28 and the encrypted model data 26 with the prediction function.
In the training phase, i.e. step S 12 and optionally S 14 encrypted model data 26 is computed in the server device 12 based on unencrypted, non-critical training data 22 , which may be old field data, and encrypted, confidential training result data 24 , which corresponds to the training data 22 . In the prediction phase, i.e. steps S 16 and S 18 , new field data 28 is sent to the server device 12 , for example from arbitrary devices 20 at arbitrary locations. The server device 12 uses the encrypted model data 26 to compute encrypted predictions/ result data 30 based on the new field data 28 .
In FIG. 2 , the training data 22 and the training result data 24 are encrypted with the homomorphic encryption algorithm, wherein the model data 26 is computed in plaintext from the encrypted training data 22 and the encrypted training result data 24 with the training function. Furthermore, the field data 28 is encrypted with the homomorphic encryption algorithm, wherein the result data 30 is computed in encrypted form from the encrypted field data 28 and the plaintext model data 26 with the prediction function.
The method and system of FIG. 1 is more suitable when the training data 22 and/or field data 28 is confidential as well. The model data 26 of the training phase, however, is not encrypted. This may not be a security issue as the computed parameters in the model data 26 only may make sense with respect to the
encrypted data
22 , 24 or 28 , 30 . Furthermore, the provider of the server device 12 cannot abuse its knowledge of the model data 26 because it cannot decrypt the result data 30 .
In both FIGS. 1 and 2 , the result data 30 may be sent to the client 18 , where they may be decrypted and consumed, for example archived, displayed, or further processed.
Typically, the training phase may be much more expensive in computational power. The model data 26 may be obtained once and then used several times. It also may be possible that new model data 26 is obtained periodically.
By only encrypting two out of the three parts, input data
22 , 28 , model data 26 and output data
24 , 30 substantial performance improvements are possible, which enable practical privacy-preserving model data 26 and result data 30 computations for remote service applications. For the computations in the server device 12 , a homomorphic encryption scheme can be used that usually only offers a small set of operations that can be carried out on encrypted data. However, these operations suffice to compute model data 26 and result data 30 as shown in the following.
If all three parts were encrypted, much more expensive, for example fully homomorphic encryption schemes may have to be used, degrading performance to the point where remote computation of model data 26 and result data 30 may become impractical.
In the following, an embodiment to linear regression as computational model is described that may be generalized to other computational models based on model data 26 , a training function and a prediction function.
The Computational Model
As already mentioned, the computational model comprises model data 26 , a training function Æ and a prediction function g.
During the training phase, i.e. steps S 12 , S 14 , the model data 26 is fitted to the training data 22 and the training result data 24 according to the training function Æ. The training data 22 may comprise m samples, where each sample i contains a vector x (i) of n features/data values. The training data 22 , as well as the field data 28 , may be seen as independent variables. The training result data 24 , as well as the result data 30 , may comprise m scalars y (i) , which also may be seen as a vector and/or may constitute the dependent variable. Let X and y denote the matrix and the vector of all independent and dependent variables, respectively. Then the model data 26 may be provided by a vector θ may be computed in the training phase by θ=Æ(X,y).
During the training phase, i.e. steps S 14 , S 16 , the model data 26 , θ is computed from known X and y, to predict the dependent variable for new independent variables. The field data 28 and the result data 30 may have the same format as the data
22 , 24 and also may be denoted by X and y. Then, the result data 30 , y may be computed through the prediction function g based on the vector X of independent variables and the model data θ:y=g(x,θ).
In general, the method may be performed when the functions Æ and g are or can be approximated by a bounded-degree polynomial.
In order to ensure that the server device 12 learns as little as possible during the course of the computation, data is encrypted before being transmitted. Note that when using asymmetric cryptography, only the secure device 16 and/or the beneficiary device 18 needs access to a secret key. The devices
14 and 20 solely may use a public key for encryption.
The homomorphic encryption scheme used i
CLAIMS
Claims ( 16 )
The invention claimed is:
1. A method for evaluating data based on a computational model, the computational model comprising model data, a training function and a prediction function, the method comprising:
training the computational model by:
receiving training data and training result data for training the computational model;
computing the model data from the training data and the training result data with the training function;
predicting result data by:
receiving field data for predicting result data; and
computing the result data from the field data and the model data with the prediction function;
wherein,
the training data is plaintext and the training result data is encrypted with a homomorphic encryption algorithm, the homomorphic encryption algorithm being additively homomorphic;
the model data is computed in encrypted form from the plaintext training data and the encrypted training result data with the training function;
the field data is plaintext, wherein the result data is computed in encrypted form from the plaintext field data and the encrypted model data with the prediction function, wherein two or less data types are encrypted at any given time, the datatypes including the training data, the model data, and the field data; and
wherein, the computational model is a linear regression model, in which the prediction function is a linear function in the field data and the model data;
wherein, the training function of the linear regression model is based on minimizing a cost function, which quadratically minimizes a difference between the prediction function and the training result data.
2. The method of claim 1 , wherein the training function is a polynomial in at least one of the training data and the training result data; and wherein the prediction function is a polynomial in at least one of the field data and model data.
3. The method of claim 1 , wherein each of the training data and the field data is a set of vectors or matrices of field data values; wherein each of the training result data and the result data is a vector of result data values; wherein the model data is a vector of model data values.
4. The method of claim 1 , wherein, before encryption, data values are approximated by multiplication with an approximation factor and rounding to a closest integer; and wherein the homomorphic encryption algorithm is based on a finite field based on the product of two prime numbers and the approximation factor is smaller than this product.
5. The method of claim 1 , wherein the training function for computing the model data is iteratively updated with the training data and the training result data; wherein in each iteration, the training function is evaluated in an evaluation server device, the updated model data is sent to a secure device, decrypted, multiplied with a convergence factor and encrypted in the secure device and sent back to the evaluation server device.
6. The method of claim 1 , wherein at least one of the training data and the training result data is provided by a client device communicatively interconnected with an evaluation server device, wherein the client device encrypts at least one of the training data and the training result data and decrypts the result data and wherein the evaluation server device at least partially computes at least one of the model data and the result data; and wherein the field data is provided by at least one or a plurality of devices communicatively interconnected with the evaluation server device.
7. The method of claim 1 , wherein the training data and the field data encodes at least one of local wind speed and local sunshine intensity and the training result data and the result data encodes at least one of electrical power production of wind turbine facilities and solar energy facilities.
8. A method for evaluating data based on a computational model, the computational model comprising model data, a training function and a prediction function, the method comprising:
training the computational model by:
receiving training data and training result data for training the computational model;
computing the model data from the training data and the training result data with the training function;
predicting result data by:
receiving field data for predicting result data; and
computing the result data from the field data and the model data with the prediction function;
wherein,
the training data and the training result data are encrypted with a homomorphic encryption algorithm, the homomorphic encryption algorithm being additively homomorphic;
the model data is computed in plaintext from encrypted training data and the encrypted training result data with the training function;
the field data is encrypted with homomorphic encryption algorithm, wherein the result data is computed in encrypted form from the encrypted field data and the plaintext model data with the prediction function, wherein two or less data types are encrypted at any given time, the data types including the training data, the model data, and the field data; and
wherein, the computational model is a linear regression model, in which the prediction function is a linear function in the field data and the model data;
wherein, the training function of the linear regression model is based on minimizing a cost function, which quadratically minimizes a difference between the prediction function and the training result data.
9. The method of claim 8 , wherein the training function is a polynomial in at least one of the training data and the training result data; and wherein the prediction function is a polynomial in at least one of the field data and model data.
10. The method of claim 8 , wherein each of the training data and the field data is a set of vectors or matrices of field data values, wherein each of the training result data and the result data is a vector of result data values; wherein the model data is a vector of model data values.
11. The method of claim 8 , wherein, before encryption, data values are approximated by multiplication with an approximation factor and rounding to a closest integer; and wherein the homomorphic encryption algorithm is based on a finite field based on the product of two prime numbers and the approximation factor is smaller than this product.
12. The method of claim 8 , wherein the training function for computing the model data is iteratively updated with the training data and the training result data; wherein in each iteration, the training function is evaluated in an evaluation server device, the updated model data is sent to a secure device, decrypted, multiplied with a convergence factor and encrypted in the secure device and sent back to the evaluation server device.
13. The method of claim 8 , wherein at least one of the training data and the training result data is provided by a client device communicatively interconnected with an evaluation server device, wherein the client device encrypts at least one of the training data and the training result data and decrypts the result data and wherein the evaluation server device at least partially computes at least one of the model data and the result data; and wherein the field data is provided by at least one or a plurality of devices communicatively interconnected with the evaluation server device.
14. A computer-readable medium executing a computer program for evaluating data based on a computational model, the computational model comprising model data, a training function and a prediction function, which, when executed on an evaluation system, comprises:
training the computational model by:
receiving training data and training result data for training the computational model;
computing the model data from the training data and the training result data with the training function;
predicting result data by:
receiving field data for predicting result data; and
computing the result data from the field data and the model data with the prediction function;
wherein,
the training data is plaintext and the training result data is encrypted with a homomorphic encryption algorithm, the homomorphic encryption algorithm being additively homomorphic;
the model data is computed in encrypted form from the plaintext training data and the encrypted training result data with the training function;
the field data is plaintext, wherein the result data is computed in encrypted form from the plaintext field data and the encrypted model data with the prediction function, wherein two or less data types are encrypted at any given time, the data types including the training data, the model data, and the field data; and
wherein, the computational model is a linear regression model, in which the prediction function is a linear function in the field data and the model data;
wherein, the training function of the linear regression model is based on minimizing a cost function, which quadratically minimizes a difference between the prediction function and the training result data.
15. An evaluation system for evaluating data based on a computational model, the computational model comprising model data, a training function and a prediction function, which when the computational model is executed on the evaluation system, the system operable to:
train the computational model by:
receive training data and training result data for training the computational model;
compute the model data from the training data and the training result data with the training function;
predict result data by:
receive field data for predicting result data; and
compute the result data from the field data and the model data with the prediction function;
wherein,
the training data is plaintext and the training result data is encrypted with a homomorphic encryption algorithm, the homomorphic encryption algorithm being additively homomorphic;
the model data is computed in encrypted form from the plaintext training data and the encrypted training result data with the training function;
the field data is plaintext, wherein the result data is computed in encrypted form from the plaintext field data and the encrypted model data with the prediction function, wherein two or less data types are encrypted at any given time, the data types including the training data, the model data, and the field data; and
wherein, the computational model is a linear regression model, in which the prediction function is a linear function in the field data and the model data;
wherein, the training function of the linear regression model is based on minimizing a cost function, which quadratically minimizes a difference between the prediction function and the training result data.
16. An evaluation system for evaluation data based on a computational model, the computational model comprising model data, a training function and a prediction function, which when the computational model is executed on the evaluation system, the system operable to:
train the computational model by:
receive training data and training result data for training the computational model;
compute the model data from the training data and the training result data with the training function;
predict result data by:
receive field data for predicting result data; and
compute the result data from the field data and the model data with the prediction function;
wherein,
the training data and the training result data are encrypted with a homomorphic encryption algorithm, the homomorphic encryption algorithm being additively homomorphic;
the model data is computed in plaintext from the encrypted training data and the encrypted training result data with the training function;
the field data is encrypted with the homomorphic encryption algorithm, wherein the result data is computed in encrypted form from the encrypted field data and the plaintext model data with the prediction function, wherein two or less data types are encrypted at any given time, the data types including the training data, the model data, and the field data; and
wherein, the computational model is a linear regression model, in which the prediction function is a linear function in the field data and the model data;
wherein, the training function of the linear regression model is based on minimizing a cost function, which quadratically minimizes a difference between the prediction function and the training result data.
US16/055,705
2016-02-04
2018-08-06
Machine learning based on homomorphic encryption
Active
2040-04-12
US11551035B2
( en )
Applications Claiming Priority (7)
Application Number
Priority Date
Filing Date
Title
EP16154245
2016-02-04
EP16154245.1
2016-02-04
EP16154245
2016-02-04
EP16163018.1
2016-03-30
EP16163018.1A
EP3203679A1
( en )
2016-02-04
2016-03-30
Machine learning based on homomorphic encryption
EP16163018
2016-03-30
PCT/EP2017/052458
WO2017134269A1
( en )
2016-02-04
2017-02-03
Machine learning based on homomorphic encryption
Related Parent Applications (1)
Application Number
Title
Priority Date
Filing Date
PCT/EP2017/052458
Continuation
WO2017134269A1
( en )
2016-02-04
2017-02-03
Machine learning based on homomorphic encryption
Publications (2)
Publication Number
Publication Date
US20180349740A1
US20180349740A1 ( en )
2018-12-06
US11551035B2
true
US11551035B2 ( en )
2023-01-10
Family
ID=56112805
Family Applications (1)
Application Number
Title
Priority Date
Filing Date
US16/055,705
Active
2040-04-12
US11551035B2
( en )
2016-02-04
2018-08-06
Machine learning based on homomorphic encryption
Country Status (3)
Country
Link
US
( 1 )
US11551035B2
( en )
EP
( 2 )
EP3203679A1
( en )
WO
( 1 )
WO2017134269A1
( en )
Cited By (2)
* Cited by examiner, â Cited by third party
Publication number
Priority date
Publication date
Assignee
Title
US20210203484A1
( en )
*
2017-12-04
2021-07-01
Koninklijke Philips N.V.
Nodes and methods of operating the same
US20250150256A1
( en )
*
2023-11-02
2025-05-08
Crypto Lab Inc.
Server device for providing homomorphic encryption ai model and method thereof
Families Citing this family (51)
* Cited by examiner, â Cited by third party
Publication number
Priority date
Publication date
Assignee
Title
US20190391295A1
( en )
*
2016-11-07
2019-12-26
Schlumberger Technology Corporation
Seismic Data Processing Artificial Intelligence
US11196541B2
( en )
*
2017-01-20
2021-12-07
Enveil, Inc.
Secure machine learning analytics using homomorphic encryption
US10554390B2
( en )
*
2017-06-12
2020-02-04
Microsoft Technology Licensing, Llc
Homomorphic factorization encryption
EP3461054A1
( en )
*
2017-09-20
2019-03-27
Universidad de Vigo
System and method for secure outsourced prediction
CN107886009B
( en )
*
2017-11-20
2020-09-08
å京大å¦
Big data generation method and system for preventing privacy disclosure
US10198399B1
( en )
*
2018-03-06
2019-02-05
KenSci Inc.
Cryptographically secure machine learning
US11586743B2
( en )
*
2018-03-22
2023-02-21
Via Science, Inc.
Secure data processing
KR102272928B1
( en )
2018-03-27
2021-07-02
ìì¸ëíêµì°ííë ¥ë¨
Operating method for machine learning model using encrypted data and apparatus based on machine learning model
US11087223B2
( en )
2018-07-11
2021-08-10
International Business Machines Corporation
Learning and inferring insights from encrypted data
US11657409B2
( en )
*
2018-12-31
2023-05-23
Tatat Consultancy Services Limited
System and method for product demand transfer estimation through machine learning
US11201745B2
( en )
*
2019-01-10
2021-12-14
International Business Machines Corporation
Method and system for privacy preserving biometric authentication
US11469878B2
( en )
*
2019-01-28
2022-10-11
The Toronto-Dominion Bank
Homomorphic computations on encrypted data within a distributed computing environment
CN109886417B
( en )
*
2019-03-01
2024-05-03
æ·±å³åæµ·å¾®ä¼é¶è¡è¡ä»½æéå ¬å¸
Model parameter training method, device, equipment and medium based on federal learning
EP3716122A1
( en )
*
2019-03-27
2020-09-30
Siemens Healthcare GmbH
Ai based system and method for secure data processing - applications in clinical workflow deployment and model training
CN110210233B
( en )
*
2019-04-19
2024-05-24
å¹³å®ç§æï¼æ·±å³ï¼æéå ¬å¸
Combined construction method and device of prediction model, storage medium and computer equipment
CN110135467A
( en )
*
2019-04-23
2019-08-16
å京æ·çä¿¡æ¯ç§ææéå ¬å¸
A kind of model training method, device, system and recording medium based on data splicing
CN110084063B
( en )
*
2019-04-23
2022-07-15
ä¸å½ç§å¦ææ¯å¤§å¦
A Gradient Descent Computation Method for Protecting Privacy Data
EP3959839A1
( en )
*
2019-04-23
2022-03-02
OneSpan NV
Methods and systems for privacy preserving evaluation of machine learning models
CN110163008B
( en )
*
2019-04-30
2023-03-07
åæ°å è¿ææ¯æéå ¬å¸
Method and system for security audit of deployed encryption model
US12143465B2
( en )
2019-05-17
2024-11-12
International Business Machines Corporation
Searching over encrypted model and encrypted data using secure single-and multi-party learning based on encrypted data
CN110233730B
( en )
*
2019-05-22
2022-05-03
æ¨å大å¦
A privacy information protection method based on K-means clustering
CN112149174B
( en )
*
2019-06-28
2024-03-12
å京ç¾åº¦ç½è®¯ç§ææéå ¬å¸
Model training methods, devices, equipment and media
CN112149141B
( en )
*
2019-06-28
2023-08-29
å京ç¾åº¦ç½è®¯ç§ææéå ¬å¸
Model training method, device, equipment and medium
CN112149706B
( en )
*
2019-06-28
2024-03-15
å京ç¾åº¦ç½è®¯ç§ææéå ¬å¸
Model training methods, devices, equipment and media
CN112149834B
( en )
*
2019-06-28
2023-11-07
å京ç¾åº¦ç½è®¯ç§ææéå ¬å¸
Model training methods, devices, equipment and media
CN112182594B
( en )
*
2019-07-02
2023-08-04
å京ç¾åº¦ç½è®¯ç§ææéå ¬å¸
A data encryption method and device
WO2021087073A1
( en )
2019-10-30
2021-05-06
Via Science, Inc.
Secure outsourcing of a multiplication
US12212651B2
( en )
2019-11-15
2025-01-28
International Business Machines Corporation
Enabling boosting protocols on encrypted data
CN110889447B
( en )
*
2019-11-26
2022-05-17
æ¯ä»å®(æå·)ä¿¡æ¯ææ¯æéå ¬å¸
Method and device for checking model feature saliency based on multi-party secure calculation
FR3105684B1
( en )
*
2019-12-20
2022-12-23
Idemia France
CRYPTOGRAPHIC PROCESSING METHOD, ELECTRONIC DEVICE AND ASSOCIATED COMPUTER PROGRAM
CN111143878B
( en )
*
2019-12-20
2021-08-03
æ¯ä»å®(æå·)ä¿¡æ¯ææ¯æéå ¬å¸
A method and system for model training based on private data
CN112668016B
( en )
*
2020-01-02
2023-12-08
åæ§æ¸ 交信æ¯ç§æ(å京)æéå ¬å¸
Model training method and device and electronic equipment
US12099997B1
( en )
2020-01-31
2024-09-24
Steven Mark Hoffberg
Tokenized fungible liabilities
US11593711B2
( en )
*
2020-02-03
2023-02-28
Intuit Inc.
Method and system for adaptively reducing feature bit-size for homomorphically encrypted data sets used to train machine learning models
US11727125B2
( en )
2020-03-31
2023-08-15
General Electric Company
Emergent language based data encryption
DE102020109896A1
( en )
2020-04-08
2021-10-14
Endress + Hauser Process Solutions Ag
Method for encrypting data from a field device
KR20210130048A
( en )
2020-04-21
2021-10-29
ì¼ì±ìì¤ëìì¤ ì£¼ìíì¬
Method and apparatus for training analysis model
CN111371544B
( en )
*
2020-05-27
2020-09-08
æ¯ä»å®(æå·)ä¿¡æ¯ææ¯æéå ¬å¸
Prediction method, device, electronic device and storage medium based on homomorphic encryption
KR102475273B1
( en )
*
2020-06-15
2022-12-08
주ìíì¬ í¬ë¦½í ë©
Simulation apparatus for homomorphic encryption system and method thereof
US11599806B2
( en )
2020-06-22
2023-03-07
International Business Machines Corporation
Depth-constrained knowledge distillation for inference on encrypted data
CN111898145B
( en )
*
2020-07-22
2022-11-25
èå·æµªæ½®æºè½ç§ææéå ¬å¸
Neural network model training method, device, equipment and medium
CN113139475B
( en )
*
2021-01-27
2024-09-06
æ¦æ±æå·¡æºè½ç§ææéå ¬å¸
Intelligent camera model terminal training method, device, equipment and medium
CN113553610B
( en )
*
2021-09-22
2021-12-31
åå°æ»¨å·¥ä¸å¤§å¦(æ·±å³)(åå°æ»¨å·¥ä¸å¤§å¦æ·±å³ç§æåæ°ç 究é¢)
Multi-party privacy protection machine learning method based on homomorphic encryption and trusted hardware
US12386785B2
( en )
2021-10-15
2025-08-12
Lognovations Holdings, Llc
Encoding / decoding system and method
CN114186669B
( en )
*
2021-12-10
2023-08-18
å京ç¾åº¦ç½è®¯ç§ææéå ¬å¸
Training method, device, equipment and storage medium of neural network model
FR3131649B1
( en )
*
2022-01-05
2024-09-06
Technip Energies France
Method of monitoring an installation
EP4293456A1
( en )
2022-06-13
2023-12-20
Abb Schweiz Ag
Method and system for anonymization and negotiation for predictive maintenance
US20260119986A1
( en )
*
2022-10-04
2026-04-30
Lenovo (Beijing) Limited
Path trajectory functional encryption
WO2024076638A1
( en )
*
2022-10-07
2024-04-11
CoinCircle, Inc.
Cryptographic key generation using machine learning
CN115801474B
( en )
*
2023-02-13
2023-06-20
天èå°å(èå·)ç§æè¡ä»½æéå ¬å¸
Power trading method and system based on privacy computing, power consumer and power generator
CN118573784B
( en )
*
2024-07-31
2024-09-27
æ¹åç´«å æ°åç§ææéå ¬å¸
Embedding method and system of confidential mechanism in Winograd algorithm
Citations (4)
* Cited by examiner, â Cited by third party
Publication number
Priority date
Publication date
Assignee
Title
US20050049990A1
( en )
*
2003-08-29
2005-03-03
Milenova Boriana L.
Support vector machines processing system
US20160350648A1
( en )
*
2014-11-07
2016-12-01
Microsoft Technology Licensing, Llc.
Neural networks for encrypted data
US20170016430A1
( en )
*
2015-07-17
2017-01-19
General Electric Company
Systems and methods for improved wind power generation
US20190036678A1
( en )
*
2015-01-12
2019-01-31
Morphology, LLC
Systems and methods for implementing an efficient, scalable homomorphic transformation of encrypted data with minimal data expansion and improved processing efficiency
2016
2016-03-30
EP
EP16163018.1A
patent/EP3203679A1/en
not_active
Withdrawn
2017
2017-02-03
EP
EP17702412.2A
patent/EP3412000B1/en
active
Active
2017-02-03
WO
PCT/EP2017/052458
patent/WO2017134269A1/en
not_active
Ceased
2018
2018-08-06
US
US16/055,705
patent/US11551035B2/en
active
Active
Patent Citations (4)
* Cited by examiner, â Cited by third party
Publication number
Priority date
Publication date
Assignee
Title
US20050049990A1
( en )
*
2003-08-29
2005-03-03
Milenova Boriana L.
Support vector machines processing system
US20160350648A1
( en )
*
2014-11-07
2016-12-01
Microsoft Technology Licensing, Llc.
Neural networks for encrypted data
US20190036678A1
( en )
*
2015-01-12
2019-01-31
Morphology, LLC
Systems and methods for implementing an efficient, scalable homomorphic transformation of encrypted data with minimal data expansion and improved processing efficiency
US20170016430A1
( en )
*
2015-07-17
2017-01-19
General Electric Company
Systems and methods for improved wind power generation
Non-Patent Citations (13)
* Cited by examiner, â Cited by third party
Title
Aono et al., " Fast and Secure Linear Regression and Biometric Authentication with Security Update, " International Association for Cryptologic Research, vol. 20150730:012111, Jul. 30, 2015, pp. 1-29.
Chen et al., " Highly Efficient Linear Regression Outsourcing to a Cloud, " IEEE Transactions on Cloud Computing, vol. 2, No. 4, Oct.-Dec. 2014, pp. 499-508.
De Cock et al., " Fast, Privacy Preserving Linear Regression over Distributed Datasets based on Pre-Distributed Data, " Proceedings of the 8th ACM Workshop on Artificial Intelligence and Security, 2015, 12 pp.
European Patent Office, Extended Search Report issued in corresponding Application No. 16163018.1, dated Apr. 18, 2017, 10 pp.
European Patent Office, International Preliminary Report on Patentability issued in corresponding Application No. PCT/EP2017/052458, dated Apr. 30, 2018, 21 pp.
European Patent Office, International Search Report & Written Opinion issued in corresponding Application No. PCT/EP2017/052458, dated Apr. 18, 2017, 16 pp.
European Patent Office, Office Action issued in corresponding Application No. 17702412.2, dated Jun. 17, 2020, 9 pp.
European Patent Office, Written Opinion of the International Preliminary Examining Authority issued in corresponding Application No. PCT/EP2017/052458, dated Feb. 5, 2018, 9 pp.
Gong, Yanmin, Yuguang Fang, and Yuanxiong Guo. " Privacy-preserving collaborative learning for mobile health monitoring. " In 2015 IEEE Global Communications Conference (GLOBECOM), pp. 1-6. IEEE, 2015. (Year: 2015).
*
Graepel et al., " ML Confidential: Machine Learning on Encrypted Data, " Information Security and Cryptology ICISC 2012, Berlin, Heidelberg, pp. 1-21.
Nikolaenko et al., " Privacy-Preserving Ridge Regression on Hundreds of Millions of Records, " 2013 IEEE Symposium on Security and Privacy, May 19, 2013, pp. 334-348.
Xie, Pengtao, Misha Bilenko, Tom Finley, Ran Gilad-Bachrach, Kristin Lauter, and Michael Naehrig. " Crypto-nets: Neural networks over encrypted data. " arXiv preprint arXiv:1412.6181 (2014). (Year: 2014).
*
Zhang, Qingchen, Laurence T. Yang, and Zhikui Chen. " Privacy preserving deep computation model on cloud for big data feature learning. " IEEE Transactions on Computers 65, No. 5 (2015): 1351-1362. (Year: 2015).
*
Cited By (3)
* Cited by examiner, â Cited by third party
Publication number
Priority date
Publication date
Assignee
Title
US20210203484A1
( en )
*
2017-12-04
2021-07-01
Koninklijke Philips N.V.
Nodes and methods of operating the same
US11818249B2
( en )
*
2017-12-04
2023-11-14
Koninklijke Philips N.V.
Nodes and methods of operating the same
US20250150256A1
( en )
*
2023-11-02
2025-05-08
Crypto Lab Inc.
Server device for providing homomorphic encryption ai model and method thereof
Also Published As
Publication number
Publication date
US20180349740A1
( en )
2018-12-06
WO2017134269A1
( en )
2017-08-10
EP3412000A1
( en )
2018-12-12
EP3412000B1
( en )
2026-05-20
EP3203679A1
( en )
2017-08-09
Similar Documents
Publication
Publication Date
Title
US20180349740A1
( en )
2018-12-06
Machine learning based on homomorphic encryption
US20190294956A1
( en )
2019-09-26
Secure data processing
US20160036584A1
( en )
2016-02-04
Privacy-preserving ridge regression using partially homomorphic encryption and masks
CN114696990B
( en )
2022-08-26
Multi-party computing method, system and related equipment based on fully homomorphic encryption
JP6730740B2
( en )
2020-07-29
Processing device, processing method, processing program, and cryptographic processing system
CN113221153B
( en )
2022-12-27
Graph neural network training method and device, computing equipment and storage medium
CN106170943A
( en )
2016-11-30
Privacy Preserving Ridge Regression Using Partially Homomorphic Encryption and Masking
CN113904808B
( en )
2024-07-19
Private key distribution and decryption method, device, equipment and medium
CN109190395A
( en )
2019-01-11
A kind of full homomorphic cryptography method and system based on data transformation
CN113434878B
( en )
2023-07-07
Modeling and application method, device, equipment and storage medium based on federal learning
CN116644778A
( en )
2023-08-25
Construction method of quantum homomorphic neural network and encrypted image classification method
EP3843319B1
( en )
2025-12-03
Apparatus and method for set intersection operation
CN117978361A
( en )
2024-05-03
Cloud-based privacy computing method and device, electronic equipment and readable medium
JP7205016B2
( en )
2023-01-16
Confidential Information Processing System and Confidential Information Processing Method
CN112913184A
( en )
2021-06-04
Computing the key rotation period of a block cipher-based encryption scheme system and method
Park et al.
2023
Hyperdimensional computing as a rescue for efficient privacy-preserving machine learning-as-a-service
US20250192983A1
( en )
2025-06-12
Privacy-preserving and non-interactive training of regression trees
JP2018205511A
( en )
2018-12-27
Parameter conversion method, parameter conversion device, parameter conversion program, pairing operation method, pairing operation device and pairing operation program
CN118368046A
( en )
2024-07-19
Ciphertext noise reduction method and device, electronic equipment and storage medium
JP2019101083A
( en )
2019-06-24
Encryption system
JP2022122660A
( en )
2022-08-23
Secret calculation method, secret calculation system, and secret calculation management device
CN118278035B
( en )
2024-10-08
Privacy calculation method, device, equipment and storage medium
US20260135689A1
( en )
2026-05-14
Method for processing a digital content using a homomorphic encryption protocol
CN119299160B
( en )
2025-10-21
Cloud computing-oriented secure outsourcing data processing method and device
EP4133679B1
( en )
2025-11-26
A method and computation entity for performing computation of a function over encrypted data
Legal Events
Date
Code
Title
Description
2018-08-06
FEPP
Fee payment procedure
Free format text : ENTITY STATUS SET TO UNDISCOUNTED (ORIGINAL EVENT CODE: BIG.); ENTITY STATUS OF PATENT OWNER: LARGE ENTITY
2019-02-04
STPP
Information on status: patent application and granting procedure in general
Free format text : DOCKETED NEW CASE - READY FOR EXAMINATION
2021-05-17
AS
Assignment
Owner name : ABB SCHWEIZ AG, SWITZERLAND
Free format text : ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:SCHNEIDER, JOHANNES;REEL/FRAME:056258/0009
Effective date : 20210331
2021