Privacy-preserving feature processing method, feature prediction model training method and device

Abstract

The embodiments of this specification provide a privacy-protected feature processing method, a feature prediction model training method and a device, the method comprising: obtaining first-dimensional encrypted information sent by a client, comprising a plurality of first ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the first dimension; obtaining first-dimensional feature ciphertexts by performing a first homomorphic operation on the plurality of first ciphertexts; obtaining second-dimensional encrypted information sent by the client, comprising a plurality of second ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the second dimension; obtaining second-dimensional feature ciphertexts by performing a second homomorphic operation on the plurality of second ciphertexts; inputting the first-dimensional feature ciphertexts and the second-dimensional feature ciphertexts into a pre-trained feature prediction model, the feature prediction model outputting predicted third-dimensional feature ciphertexts; the third dimension is one of the three dimensions consisting of subject, environment and event, and the first dimension and the second dimension are the other two dimensions of the above three dimensions.

Description

Technical Field

[0001] One or more embodiments of this specification relate to the field of data security and privacy protection, and in particular, to a privacy-protecting feature processing method, a feature prediction model training method, and a device. Background Art

[0002] Many business systems often need to collect various types and dimensions of user data from their clients when providing services. In recent years, as user data privacy has received increasing attention, the technical requirements for protecting the privacy and security of user data while utilizing it to provide business services have also become increasingly stringent. Summary of the Invention

[0003] The embodiments in this specification aim to provide a method that can mine the correlation between customer data of different dimensions, thereby more effectively utilizing customer data, and effectively protect customer privacy data from being leaked when the privacy sensitivity of customer data changes, thereby addressing the shortcomings of the existing technology.

[0004] According to a first aspect, a privacy-preserving feature processing method is provided, comprising:

[0005] Obtaining first-dimensional encrypted information sent by the client, including a plurality of first ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the first dimension; and obtaining a first-dimensional feature ciphertext by performing a first homomorphic operation on the plurality of first ciphertexts.

[0006] Obtaining second-dimensional encrypted information sent by the client, including a plurality of second ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the second dimension; and obtaining a second-dimensional feature ciphertext by performing a second homomorphic operation on the plurality of second ciphertexts.

[0007] The first dimension feature ciphertext and the second dimension feature ciphertext are input into a pre-trained feature prediction model, and the feature prediction model outputs a predicted third dimension feature ciphertext; the third dimension is one of the three dimensions consisting of the client-related subject, environment and event, and the first dimension and the second dimension are the other two dimensions of the above three dimensions.

[0008] In one possible implementation, the method further includes:

[0009] Obtaining third-dimensional encrypted information sent by the client, including a plurality of third ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the third dimension; and obtaining actual third-dimensional feature ciphertexts by performing a third homomorphic operation on the plurality of third ciphertexts;

[0010] Based on the predicted third-dimensional feature ciphertext, it is determined whether the actual third-dimensional feature ciphertext has anomalies.

[0011] In one possible implementation, the method further includes:

[0012] If the result is yes, the predicted third-dimensional feature ciphertext is sent to the client to determine whether there is an anomaly in the attribute value of the third dimension.

[0013] In one possible implementation, the method further includes:

[0014] The predicted third-dimensional feature ciphertext is sent to the client for predicting the attribute value of the third dimension on the client.

[0015] In one possible implementation, the subject includes one of a user and a device; when the subject is a user, the attributes of the subject include one or more of work, age, ID card, and household registration; when the subject is a device, the attributes of the subject include one or more of inherent applications on the device and device identification.

[0016] In a possible implementation, the attributes of the environment include one or more of location information and network address.

[0017] In a possible implementation, the attributes of the event include one or more of face-swiping payment, face-swiping authentication, face-swiping application login, fingerprint payment, fingerprint authentication, and fingerprint application login.

[0018] According to a second aspect, a method for training a feature prediction model is provided, comprising:

[0019] Obtaining first-dimensional encrypted information, including a plurality of first ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the first dimension; and obtaining a first-dimensional feature ciphertext by performing a first homomorphic operation on the plurality of first ciphertexts;

[0020] Obtaining second-dimensional encrypted information, including a plurality of second ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the second dimension; and obtaining a second-dimensional feature ciphertext by performing a second homomorphic operation on the plurality of second ciphertexts;

[0021] Obtaining third-dimensional encrypted information, including a plurality of third ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the third dimension; obtaining actual third-dimensional feature ciphertexts by performing a third homomorphic operation on the plurality of third ciphertexts; the third dimension being one of three dimensions consisting of a subject, an environment, and an event corresponding to the client, the first dimension and the second dimension being the other two of the three dimensions;

[0022] The first-dimensional feature ciphertext and the second-dimensional feature ciphertext are input into a feature prediction model. The feature prediction model obtains a predicted third-dimensional feature ciphertext based on the first-dimensional feature ciphertext and the second-dimensional feature ciphertext. With the predicted third-dimensional feature ciphertext tending towards the actual third-dimensional feature ciphertext, the parameters of the feature prediction model are updated.

[0023] In one possible implementation, the feature prediction model obtains a predicted third-dimensional feature ciphertext based on the first-dimensional feature ciphertext and the second-dimensional feature ciphertext, and updates the parameters of the feature prediction model with the goal of making the predicted third-dimensional feature ciphertext approach the actual third-dimensional feature ciphertext, including:

[0024] The feature prediction model obtains a predicted third-dimensional feature ciphertext based on a homomorphic operation on the input ciphertext, and updates the parameters of the feature prediction model with the plaintext operation result corresponding to the predicted third-dimensional feature ciphertext tending towards the plaintext operation result corresponding to the actual third-dimensional feature ciphertext as the goal.

[0025] According to a third aspect, a feature processing device for protecting privacy is provided, comprising:

[0026] The first-dimensional feature acquisition unit is configured to acquire first-dimensional encrypted information sent by a client, including a plurality of first ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the first dimension; and obtain first-dimensional feature ciphertexts by performing a first homomorphic operation on the plurality of first ciphertexts;

[0027] The second-dimensional feature acquisition unit is configured to acquire second-dimensional encrypted information sent by the client, including a plurality of second ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the second dimension; and obtain the second-dimensional feature ciphertext by performing a second homomorphic operation on the plurality of second ciphertexts;

[0028] The third-dimensional feature prediction unit is configured to input the first-dimensional feature ciphertext and the second-dimensional feature ciphertext into a pre-trained feature prediction model, and the feature prediction model outputs a predicted third-dimensional feature ciphertext; the third dimension is one of the three dimensions consisting of the client-related subject, environment and event, and the first dimension and the second dimension are the other two dimensions of the above three dimensions.

[0029] In a possible implementation, the device further includes a feature abnormality judgment unit configured to:

[0030] Obtaining third-dimensional encrypted information sent by the client, including a plurality of third ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the third dimension; and obtaining actual third-dimensional feature ciphertexts by performing a third homomorphic operation on the plurality of third ciphertexts;

[0031] Based on the predicted third-dimensional feature ciphertext, it is determined whether the actual third-dimensional feature ciphertext has anomalies.

[0032] In a possible implementation, the feature abnormality judgment unit is further configured to:

[0033] If the result is yes, the predicted third-dimensional feature ciphertext is sent to the client to determine whether there is an anomaly in the attribute value of the third dimension.

[0034] In a possible implementation, the device further includes:

[0035] The feature sending unit is configured to send the predicted third-dimensional feature ciphertext to the client for predicting the attribute value of the third dimension on the client.

[0036] In one possible implementation, the subject includes one of a user and a device; when the subject is a user, the attributes of the subject include one or more of work, age, ID card, and household registration; when the subject is a device, the attributes of the subject include one or more of inherent applications on the device and device identification.

[0037] In a possible implementation, the attributes of the environment include one or more of location information and network address.

[0038] In a possible implementation, the attributes of the event include one or more of face-swiping payment, face-swiping authentication, face-swiping application login, fingerprint payment, fingerprint authentication, and fingerprint application login.

[0039] According to a fourth aspect, a training device for a feature prediction model is provided, comprising:

[0040] The first-dimensional feature acquisition unit is configured to acquire first-dimensional encrypted information, including a plurality of first ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the first dimension; and obtain the first-dimensional feature ciphertext by performing a first homomorphic operation on the plurality of first ciphertexts;

[0041] The second-dimensional feature acquisition unit is configured to acquire second-dimensional encrypted information, including a plurality of second ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the second dimension; and obtain the second-dimensional feature ciphertext by performing a second homomorphic operation on the plurality of second ciphertexts;

[0042] a third-dimensional feature acquisition unit configured to acquire third-dimensional encrypted information, including a plurality of third ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the third dimension; and obtain actual third-dimensional feature ciphertexts by performing a third homomorphic operation on the plurality of third ciphertexts; the third dimension being one of three dimensions corresponding to the client's subject, environment, and event, and the first dimension and the second dimension being the other two of the three dimensions;

[0043] The parameter updating unit is configured to input the first-dimensional feature ciphertext and the second-dimensional feature ciphertext into a feature prediction model. The feature prediction model obtains a predicted third-dimensional feature ciphertext based on the first-dimensional feature ciphertext and the second-dimensional feature ciphertext, and updates the parameters of the feature prediction model with the goal of making the predicted third-dimensional feature ciphertext tend towards the actual third-dimensional feature ciphertext.

[0044] In one possible implementation, the parameter updating unit is further configured to input the first-dimensional feature ciphertext and the second-dimensional feature ciphertext into a feature prediction model. The feature prediction model obtains a predicted third-dimensional feature ciphertext based on a homomorphic operation on the input ciphertext. The parameters of the feature prediction model are updated with the goal of making the plaintext operation result corresponding to the predicted third-dimensional feature ciphertext tend towards the plaintext operation result corresponding to the actual third-dimensional feature ciphertext.

[0045] According to a fifth aspect, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed in a computer, the computer is caused to execute the methods described in the first and second aspects.

[0046] According to the sixth aspect, a computing device is provided, comprising a memory and a processor, wherein executable code is stored in the memory, and when the processor executes the executable code, the methods described in the first and second aspects are implemented.

[0047] By utilizing one or more of the methods, apparatuses, computing devices, and storage media in the above aspects, the association relationships between customer data of different dimensions can be mined, thereby effectively utilizing customer privacy data and protecting customer privacy data from being leaked. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0049] Figure 1A schematic diagram illustrating a principle of a feature processing method for protecting privacy according to an embodiment of this specification;

[0050] Figure 2 A flowchart of a feature processing method for protecting privacy according to an embodiment of this specification is shown.

[0051] Figure 3 A flowchart of a method for training a feature prediction model according to an embodiment of this specification is shown.

[0052] Figure 4 A structural diagram of a feature processing device for protecting privacy according to an embodiment of this specification is shown.

[0053] Figure 5 A structural diagram of a training device for a feature prediction model according to an embodiment of this specification is shown. DETAILED DESCRIPTION

[0054] The solution provided in this invention will be described below with reference to the accompanying drawings.

[0055] As mentioned above, with the increasing importance of user data privacy, ensuring the privacy and security of user data is currently a hot topic for regulators and the industry. Many business systems often need to collect various types of user data from the client. Currently, some vendors typically use the following solution to protect the privacy and security of user data extracted from the client. First, the extracted user data is classified into high-sensitivity and low-sensitivity data based on privacy sensitivity. High-sensitivity data is then encrypted and the ciphertext is sent to the server for subsequent ciphertext processing. Low-sensitivity data is not encrypted and sent to the server in plaintext for subsequent plaintext processing. However, this solution has the following problems. First, because the server can access a large amount of plaintext user data (low-sensitivity data), this user data may be leaked during data processing. In particular, the sensitivity of user data is not static. Some user data may not be highly sensitive at one time, but may later become highly sensitive due to technological advances. For example, user facial data may not be very sensitive for a long time. However, with the widespread development of facial recognition technology, user facial information has become highly sensitive and private data. However, early leaks of facial data cannot be recovered or destroyed, potentially posing a risk to future facial recognition services. The aforementioned solutions cannot effectively eliminate the potential risks posed by this shift in data sensitivity. Furthermore, they cannot effectively explore the inherent connections between data of different dimensions, failing to fully capture the data's potential value.

[0056] In order to solve the above technical problems, the embodiments of this specification provide a feature processing method for protecting privacy. Figure 1 A schematic diagram showing the principle of a feature processing method for protecting privacy according to an embodiment of this specification is shown. Figure 1 , which explains the main idea of ​​this method. This method is mainly divided into two stages:

[0057] The first stage is the feature prediction model training stage. First, the various user data (or various attributes) that can be extracted by the client are divided into three dimensions: subject, environment, and event. The values ​​of the various attributes in the three dimensions are homomorphically encrypted to obtain the corresponding ciphertext of each attribute. The server performs homomorphic operations based on the attribute ciphertexts of the three dimensions to obtain the actual feature ciphertexts of the three dimensions, respectively. Then, the actual feature ciphertexts of two of the dimensions can be used as input data, and the actual feature ciphertexts of the third dimension can be used as the label of the input data to train a feature prediction model for predicting the feature ciphertext of the third dimension.

[0058] The second stage is the feature prediction stage. The client obtains attribute values ​​for two of the three dimensions: subject, environment, and event. It will be appreciated that, in different embodiments, these two dimensions can be different dimensions, such as one of the subject and environment dimensions, the subject and event dimension, or the environment and event dimension. For ease of description, the following uses the extraction of attribute values ​​for the subject and event dimensions as an example. After obtaining the attribute values ​​for the subject and event dimensions, the values ​​of the various attributes in these two dimensions are homomorphically encrypted, and the encrypted ciphertext is then sent to the server. After receiving the attribute ciphertexts for these two dimensions, the server performs homomorphic operations on the ciphertexts for the two dimensions, thereby obtaining feature ciphertexts for the subject and event dimensions, respectively. The feature ciphertexts for these two dimensions are input into the feature prediction model trained in the first stage to obtain the predicted feature ciphertext for the third dimension, namely the predicted environmental feature ciphertext. The predicted environmental feature ciphertext can, for example, be used to determine whether there are any anomalies in the client's attributes in the environmental dimension. Specifically, in one embodiment, the server can send the predicted environmental feature ciphertext back to the client. After the client decrypts the environmental feature ciphertext, it can determine whether there is any abnormality in the actual environmental attribute it extracted based on the decrypted environmental feature plaintext.

[0059] The above method has the following advantages: First, all attribute data uploaded by the client, regardless of its current privacy sensitivity, is homomorphically encrypted. However, the server typically does not possess the client's decryption key, so it cannot obtain the true attribute values ​​of any client attribute data it obtains. This ensures that true user privacy data is not leaked during the server's data processing, regardless of whether it is currently high-sensitivity or low-sensitivity. In other words, even if some low-sensitivity data becomes high-sensitivity in the future, it will not pose a security risk of privacy leakage. Second, the attribute ciphertext extracted by the client is uploaded to the server according to different dimensions. The server then extracts features from each dimension and trains a prediction model based on these features to learn the correlations between features in these dimensions. The trained prediction model can then be used to identify anomalies in the true attributes of different dimensions. Essentially, this method can uncover the inherent connections between user data in different dimensions and, based on these connections, determine whether anomalies exist in the user data.

[0060] The detailed process of this method is described below. Figure 2 FIG. 1 is a flow chart showing a feature processing method for protecting privacy according to an embodiment of this specification. Figure 2 Said method comprises at least the following steps:

[0061] First, in step 21, first-dimensional encrypted information sent by the client is obtained, which includes several first ciphertexts obtained by homomorphically encrypting several attribute values ​​of the first dimension; and the first-dimensional feature ciphertext is obtained by performing a first homomorphic operation on the several first ciphertexts.

[0062] In this step, the server receives a plurality of attribute ciphertexts corresponding to a plurality of attribute values ​​of the first dimension sent by the client. In different embodiments, the first dimension can be any one of the three dimensions of subject, environment, and event. The client can divide the various client data (i.e., various attributes) obtained according to the three dimensions of subject, environment, and event, and then homomorphically encrypt the attribute values ​​of the three dimensions and send them to the server.

[0063] Among them, the attributes of the subject dimension refer to the attributes of the client subject. The client subject mainly refers to the initiator of the client event or action. In different embodiments, the client subject can be different specific types of subjects. In a specific embodiment, for example, it can be one of a user and a device. In one example, the client subject is a user, and the attributes of the subject may include one or more of work, age, ID card, and household registration. In another example, the client subject is a device, and the attributes of the subject may include one or more of the inherent applications on the device and the device identification. The inherent applications may include the device's operating system, pre-installed applications, client version and other information. In other examples, the attributes of the subject may also include other device-related attribute information.

[0064] The attributes of the environment dimension refer to the attributes of the client environment. The client environment primarily refers to the environment in which client events or actions occur. In different embodiments, the client environment may have different specific attributes. In a specific embodiment, for example, it may be one or more of the following: location information, network address (IP address, Wi-Fi address, etc.).

[0065] The attributes of the event dimension refer to the attributes of the event that occurs on the client. In different embodiments, the events that occur on the client can be different specific attributes of different types of specific events. In a specific embodiment, the event attributes include, for example, attributes that identify the event type, specifically including one or more of face payment, face authentication, face login application, fingerprint payment, fingerprint authentication, and fingerprint login application.

[0066] In this step, the server obtains several attribute ciphertexts of the first dimension (first ciphertexts), and can directly perform homomorphic operations based on the ciphertexts to obtain the ciphertext operation results, that is, the first dimension feature ciphertexts. This step utilizes the homomorphic nature of the homomorphic encryption algorithm, that is, performing operations on the plaintext before encryption and performing corresponding operations on the ciphertext after encryption, the results are equivalent. For example, using the same public key PK to encrypt plaintext a and b, we get En PK (a) and E PK (b) If: or Then it is considered that the encryption algorithm satisfies homomorphism, where En is the encryption operation, Dec is the decryption operation, and PrK is the decryption private key. Corresponding to operations on plaintext and ciphertext respectively. When ⊙ represents addition, the encryption is homomorphic encryption, and when ⊙ represents multiplication, the encryption is multiplicative homomorphic encryption. If both the addition homomorphism and multiplication homomorphism properties are satisfied, and any number of addition and multiplication operations can be performed, then the encryption is fully homomorphic encryption. In different embodiments, the homomorphic encryption can be one of homomorphic addition, homomorphic multiplication, or fully homomorphic encryption, and this specification does not limit this.

[0067] A first homomorphic operation for the first dimension can be designed based on the homomorphic encryption algorithm used and the relationship between the multiple attribute values ​​of the first dimension. For example, in one example, the first homomorphic operation can be a weighted homomorphic sum of multiple first ciphertext items. In other examples, the first homomorphic operation can also be a homomorphic operation supported by other corresponding homomorphic encryption algorithms.

[0068] Then, in step 22, the second-dimensional encrypted information sent by the client is obtained, which includes a plurality of second ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the second dimension; and a second-dimensional feature ciphertext is obtained by performing a second homomorphic operation on the plurality of second ciphertexts.

[0069] Specifically, the obtained ciphertexts of the second dimension may be ciphertexts of any one of the three dimensions of subject, environment, and event except the first dimension in step 21 in different embodiments.

[0070] In this step, the encryption method of the obtained ciphertext and the process of obtaining the single-dimensional attribute characteristics through homomorphic operation are similar to the encryption method of the ciphertext obtained in step 21 and the process of obtaining the single-dimensional attribute characteristics through homomorphic operation. Please refer to the detailed description of step 21 and will not be repeated here. It should be understood that the operation method of the second homomorphic operation can be designed based on the homomorphic encryption algorithm and the association relationship between the multiple attribute values ​​of the second dimension. It can be the same as or different from the operation method of the first homomorphic operation, and this is not limited here.

[0071] Then, in step 23, the first-dimensional feature ciphertext and the second-dimensional feature ciphertext are input into a pre-trained feature prediction model, and the feature prediction model outputs a predicted third-dimensional feature ciphertext based on a homomorphic operation on the input ciphertext.

[0072] In this step, the first dimension feature ciphertext and the second dimension feature ciphertext obtained in steps 21 and 22 are input into a pre-trained feature prediction model to obtain a predicted feature ciphertext of a third dimension of the three dimensions of subject, environment, and event, in addition to the first and second dimensions. It will be understood that in different embodiments, the feature prediction model can be a prediction model for feature ciphertexts of different dimensions. For example, in one example, if the first dimension and the second dimension in steps 21 and 22 are the subject and event dimensions, respectively, the feature prediction model can be a prediction model for feature ciphertexts of the environment dimension. In other examples, depending on the specific first and second dimensions, the feature prediction model can also be a prediction model for feature ciphertexts of the environment dimension or the subject dimension.

[0073] The predicted characteristic ciphertext of the third dimension can be used to determine whether the actual characteristic ciphertext obtained based on the third dimension attributes actually extracted by the client contains an anomaly, thereby determining whether the attribute values ​​of the third dimension obtained by the client contain an anomaly. Therefore, in one embodiment, third-dimensional encrypted information sent by the client can be obtained, including multiple third ciphertexts obtained by homomorphically encrypting multiple attribute values ​​of the third dimension; the actual third-dimensional characteristic ciphertext is obtained by performing a third homomorphic operation on the multiple third ciphertexts; and based on the predicted third-dimensional characteristic ciphertext, it is determined whether the actual third-dimensional characteristic ciphertext contains an anomaly.

[0074] In another embodiment, the predicted third-dimensional feature ciphertext can also be sent to the client, which can then determine whether the attribute value of the third dimension is abnormal. Specifically, for example, the predicted third-dimensional feature can be sent to the client, which can then decrypt it using its private decryption key to obtain plaintext feature information. Based on the plaintext feature information, the client can then determine whether the attribute value of the third dimension is abnormal.

[0075] Specifically, in one embodiment, the client can, for example, send the ciphertext of the client's subject attributes (for example, several device attributes), the ciphertext of event attributes (for example, face recognition + login to application A) and the ciphertext of environment attributes (for example, WIFI address) to the server. The server obtains the actual subject feature ciphertext, event feature ciphertext and environment feature ciphertext respectively through homomorphic operations based on the subject attribute ciphertext, event attribute ciphertext and environment attribute ciphertext.

[0076] In one example, the server inputs the subject feature ciphertext and event feature ciphertext into the prediction model to obtain the predicted environmental feature ciphertext. The server then compares the actual environmental feature ciphertext with the predicted one to determine if there are any anomalies in the actual environmental feature ciphertext. If so, the predicted environmental feature ciphertext is sent to the client. The client decrypts the predicted environmental feature ciphertext using a decryption key to obtain the predicted environmental feature plaintext. Based on the predicted plaintext, the client ultimately determines whether the actual local environmental attributes (e.g., Wi-Fi address) are indeed abnormal.

[0077] For example, in one example, a client device B usually logs in to application C through WIFI address 1 or address 2. On one occasion, the device logs in to application C through WIFI address 3, which it has never used before. Based on the ciphertext of the subject feature (device B) and the ciphertext of the event attribute (logging in to application C), the prediction model predicts that the environmental feature ciphertext is the ciphertext of address 1 or address 2, which is different from the ciphertext of the actual environmental feature (address 3) this time. Based on this, the server determines that there may be an anomaly in the environmental feature ciphertext and sends it back to the client. Based on the fact that the feature plaintext of address 3 (the actual feature plaintext obtained) is different from the plaintext feature of address 1 or 2 (the plaintext of the predicted feature obtained by decryption), the client can determine that address 3 is an abnormal address, that is, an abnormal environmental attribute.

[0078] In another example, the server inputs the subject feature ciphertext and the environment feature ciphertext into the prediction model to obtain the predicted event feature ciphertext. The server then compares the actual event feature ciphertext with the predicted event feature ciphertext to determine whether there are any anomalies in the actual event feature ciphertext. If there are any anomalies, the predicted event feature ciphertext is sent to the client for decryption and further confirmation.

[0079] In another example, the server inputs the environmental and event feature ciphertexts into a prediction model to generate a predicted subject feature ciphertext. The server then compares the actual subject feature ciphertext with the predicted subject feature ciphertext to determine if there are any anomalies in the actual subject feature ciphertext. If there are any anomalies, the predicted subject feature ciphertext is sent to the client for decryption to generate the predicted feature plaintext, which is then used for further verification.

[0080] Since the predicted feature plain text is essentially predicted based on the historical correlation between different dimensional attribute values, it can represent the general characteristics of a dimensional attribute. Therefore, based on the difference between it and the attribute characteristics obtained at a specific time, it can be judged whether the attribute characteristics obtained at that time are abnormal.

[0081] In another embodiment, the client only sends the first dimension feature information and the second dimension feature information; after the server predicts the third dimension feature ciphertext, it does not judge the abnormality of the third dimension, but directly sends the predicted third dimension feature ciphertext to the client for predicting the attribute value of the third dimension on the client. For example, the client can obtain the plaintext features of the attribute to be predicted in a similar manner to the above embodiment. Since the plaintext features are essentially the result of plaintext operations based on the plaintext attributes, the predicted value of the attribute to be predicted can be obtained based on the plaintext features through the reverse operation of the plaintext operations. In some application scenarios, when the client cannot immediately obtain the actual value of the required attribute due to a specific fault, for example, the predicted value can be used instead of the actual value.

[0082] Figure 3FIG. 1 is a flow chart showing a method for training a feature prediction model according to an embodiment of this specification. Figure 3 Said method comprises at least the following steps:

[0083] First, in step 31, first-dimensional encrypted information is obtained, including a plurality of first ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the first dimension; and a first-dimensional feature ciphertext is obtained by performing a first homomorphic operation on the plurality of first ciphertexts.

[0084] In step 32, second-dimensional encrypted information is obtained, including a plurality of second ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the second dimension; and a second-dimensional feature ciphertext is obtained by performing a second homomorphic operation on the plurality of second ciphertexts.

[0085] In step 33, third-dimensional encrypted information is obtained, including a plurality of third ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the third dimension; an actual third-dimensional feature ciphertext is obtained by performing a third homomorphic operation on the plurality of third ciphertexts; the third dimension is one of the three dimensions corresponding to the subject, environment, and event of the client, and the first dimension and the second dimension are the other two dimensions of the above three dimensions.

[0086] In the above steps 31-33, the encryption method of the obtained ciphertext and the process of obtaining dimensional features through homomorphic operations are similar to the encryption method of the ciphertext obtained in step 21 and the process of obtaining dimensional features through homomorphic operations. Please refer to the detailed description of step 21 and will not be repeated here.

[0087] In addition, in different embodiments, in steps 31, 32, and 33, the encrypted information of the encrypted information of each dimension obtained can be obtained from the client, or can be obtained by homomorphic encryption based on the pre-acquired attribute values ​​of each dimension included in the local training sample of the server.

[0088] Then, in step 34, the first dimension feature ciphertext and the second dimension feature ciphertext are input into a feature prediction model. The feature prediction model obtains a predicted third dimension feature ciphertext based on the first dimension feature ciphertext and the second dimension feature ciphertext. With the predicted third dimension feature ciphertext tending towards the actual third dimension feature ciphertext, the parameters of the feature prediction model are updated.

[0089] The process of training a feature prediction model is essentially about learning the relationship between features of different dimensions, and then using the feature prediction model to predict features of the third dimension based on features of the two dimensions. However, by directly obtaining the feature plaintext based on the plaintext of the user attribute data and then learning based on the feature plaintext, the server side knows the user attributes and the plaintext of their features, making it difficult to ensure the security of these user privacy data. By utilizing the properties of homomorphic operations, based on the homomorphic encrypted ciphertexts of attributes of different dimensions, features of different dimensions are extracted through homomorphic operations. Then, the feature prediction model is trained to learn the intrinsic correlation between features of different dimensions. Specifically, during training, the parameters of the feature prediction model can be updated so that the predicted third-dimensional feature ciphertext output by the model based on the first-dimensional feature ciphertext and the second-dimensional feature ciphertext tends to the actual third-dimensional feature ciphertext. In different embodiments, the feature prediction model can be based on different types of neural networks, and this specification does not limit this.

[0090] The prediction model can also be trained with the goal of ensuring that the plaintext calculation results corresponding to the predicted third-dimensional feature ciphertexts approach the plaintext calculation results corresponding to the actual third-dimensional feature ciphertexts. Therefore, in one embodiment, the first-dimensional feature ciphertexts and the second-dimensional feature ciphertexts can be input into the feature prediction model. The feature prediction model performs homomorphic operations on the input ciphertexts to obtain predicted third-dimensional feature ciphertexts. The parameters of the feature prediction model are updated with the goal of ensuring that the plaintext calculation results corresponding to the predicted third-dimensional feature ciphertexts approach the plaintext calculation results corresponding to the actual third-dimensional feature ciphertexts. In a specific embodiment, the server obtains attribute ciphertexts for each dimension from the client. To protect the data security of the attributes, the server does not know the decryption key used for the homomorphic encryption. The server can send the predicted third-dimensional feature ciphertexts to the client. The client uses the decryption key to decrypt the predicted feature plaintexts, determines the feature difference between the predicted feature plaintexts and the actual feature plaintexts, and sends the feature difference to the server. The server can then update its network parameters based on the feature difference.

[0091] In another embodiment, the server side has the above-mentioned training samples, that is, the server side knows the original attribute values ​​and the decryption key of the homomorphic encryption. Therefore, the server side can directly use the decryption key to decrypt the predicted feature plaintext based on the predicted feature ciphertext, and determine the feature difference between the predicted feature plaintext and the actual feature plaintext, and then update its network parameters based on the feature difference. It can be understood that in this embodiment, since the decryption key is only set for training the model, it may not be the real user private key, and only the training data is used in the training, not the actual user-side private data, so it will not cause the leakage of user privacy data.

[0092] In summary, the method provided in this specification can enhance the privacy protection of all acquired user data and learn the intrinsic connections between user data of different dimensions, which can be used for abnormal judgment or data prediction of user data.

[0093] Another embodiment of the present specification also provides a feature processing device for protecting privacy. Figure 4 A structural diagram of a feature processing device for protecting privacy according to an embodiment of this specification is shown as follows: Figure 4 As shown, the apparatus 400 includes:

[0094] The first-dimensional feature acquisition unit 41 is configured to acquire first-dimensional encrypted information sent by the client, including a plurality of first ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the first dimension; and obtain first-dimensional feature ciphertexts by performing a first homomorphic operation on the plurality of first ciphertexts.

[0095] The second-dimensional feature acquisition unit 42 is configured to acquire the second-dimensional encrypted information sent by the client, including a plurality of second ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the second dimension; and obtain the second-dimensional feature ciphertext by performing a second homomorphic operation on the plurality of second ciphertexts.

[0096] The third-dimensional feature prediction unit 43 is configured to input the first-dimensional feature ciphertext and the second-dimensional feature ciphertext into a pre-trained feature prediction model, and the feature prediction model outputs a predicted third-dimensional feature ciphertext; the third dimension is one of the three dimensions consisting of the subject, the environment and the event, and the first dimension and the second dimension are the other two dimensions of the above three dimensions.

[0097] In one embodiment, the device may further include a feature abnormality judgment unit configured to:

[0098] Obtaining third-dimensional encrypted information sent by the client, including a plurality of third ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the third dimension; and obtaining actual third-dimensional feature ciphertexts by performing a third homomorphic operation on the plurality of third ciphertexts;

[0099] Based on the predicted third-dimensional feature ciphertext, it is determined whether the actual third-dimensional feature ciphertext has anomalies.

[0100] In one embodiment, the feature abnormality judgment unit may be further configured to:

[0101] If the result is yes, the predicted third-dimensional feature ciphertext is sent to the client to determine whether there is an anomaly in the attribute value of the third dimension.

[0102] In one embodiment, the apparatus may further include:

[0103] The feature sending unit is configured to send the predicted third-dimensional feature ciphertext to the client for predicting the attribute value of the third dimension on the client.

[0104] In one embodiment, the subject may include one of a user and a device; when the subject is a user, the subject's attributes include one or more of work, age, ID card, and household registration; when the subject is a device, the subject's attributes include one or more of inherent applications on the device and device identification;

[0105] In one embodiment, the attributes of the environment may include one or more of address information and network address;

[0106] In one embodiment, the attributes of the event may include one or more of face-swiping payment, face-swiping authentication, face-swiping application login, fingerprint payment, fingerprint authentication, and fingerprint application login.

[0107] According to another embodiment of the present specification, a training device for a feature prediction model is provided. Figure 5 A structural diagram of a training device for a feature prediction model according to an embodiment of this specification is shown. Figure 5 As shown, the device 500 includes:

[0108] The first dimension feature acquisition unit 51 is configured to acquire first dimension encrypted information, including a plurality of first ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the first dimension; and obtain first dimension feature ciphertexts by performing a first homomorphic operation on the plurality of first ciphertexts.

[0109] The second-dimensional feature acquisition unit 52 is configured to acquire second-dimensional encrypted information, including a plurality of second ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the second dimension; and obtain the second-dimensional feature ciphertext by performing a second homomorphic operation on the plurality of second ciphertexts.

[0110] The third-dimensional feature acquisition unit 53 is configured to acquire third-dimensional encrypted information, including a plurality of third ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the third dimension; and obtain actual third-dimensional feature ciphertexts by performing a third homomorphic operation on the plurality of third ciphertexts; the third dimension is one of the three dimensions corresponding to the client's subject, environment, and event, and the first dimension and the second dimension are the other two dimensions of the aforementioned three dimensions;

[0111] The parameter updating unit 54 is configured to input the first-dimensional feature ciphertext and the second-dimensional feature ciphertext into a feature prediction model. The feature prediction model obtains a predicted third-dimensional feature ciphertext based on the first-dimensional feature ciphertext and the second-dimensional feature ciphertext, and updates the parameters of the feature prediction model with the goal of making the predicted third-dimensional feature ciphertext tend towards the actual third-dimensional feature ciphertext.

[0112] In one embodiment, the parameter updating unit can be further configured to input the first-dimensional feature ciphertext and the second-dimensional feature ciphertext into a feature prediction model, wherein the feature prediction model obtains a predicted third-dimensional feature ciphertext based on a homomorphic operation on the input ciphertext, and updates the parameters of the feature prediction model with the plaintext operation result corresponding to the predicted third-dimensional feature ciphertext tending towards the plaintext operation result corresponding to the actual third-dimensional feature ciphertext as the goal.

[0113] On another aspect, the present specification provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed in a computer, the computer is caused to execute any one of the above methods.

[0114] On another aspect, the present specification provides a computing device, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, any one of the above methods is implemented.

[0115] It should be understood that the descriptions such as “first” and “second” in this article are only used to distinguish similar concepts for the sake of simplicity of description and do not have any other limiting effect.

[0116] Those skilled in the art will appreciate that, in one or more of the above examples, the functions described herein may be implemented using hardware, software, firmware, or any combination thereof. When implemented using software, these functions may be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium.

[0117] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made on the basis of the technical solution of the present invention should be included in the scope of protection of the present invention.

Claims

1. A privacy-preserving feature processing method, comprising: Obtaining first-dimension encrypted information sent by the client, including a plurality of first ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the first dimension; Obtaining a first dimension feature ciphertext by performing a first homomorphic operation on the plurality of first ciphertexts; Obtaining second-dimensional encrypted information sent by the client, including a plurality of second ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the second dimension; and obtaining a second-dimensional feature ciphertext by performing a second homomorphic operation on the plurality of second ciphertexts. Inputting the first dimension feature ciphertext and the second dimension feature ciphertext into a pre-trained feature prediction model, the feature prediction model outputting a predicted third dimension feature ciphertext; the third dimension is one of three dimensions consisting of client-related subjects, environments, and events, and the first dimension and the second dimension are the other two dimensions of the above three dimensions; Obtain third-dimensional encrypted information sent by the client, including several third ciphertexts obtained by homomorphically encrypting several attribute values ​​of the third dimension; obtain actual third-dimensional feature ciphertexts by performing third homomorphic operations on the several third ciphertexts; and determine whether the actual third-dimensional feature ciphertexts have anomalies based on the predicted third-dimensional feature ciphertexts.

2. The method according to claim 1, further comprising: If the result is yes, the predicted third-dimensional feature ciphertext is sent to the client to determine whether there is an anomaly in the attribute value of the third dimension.

3. The method according to claim 1, further comprising: The predicted third-dimensional feature ciphertext is sent to the client for predicting the attribute value of the third dimension.

4. The method according to claim 1, wherein The subject includes one of a user and a device; when the subject is a user, the attributes of the subject include one or more of work, age, ID card, and household registration; when the subject is a device, the attributes of the subject include one or more of inherent applications on the device and device identification.

5. The method according to claim 1, wherein The attributes of the environment include one or more of location information and network address.

6. The method according to claim 1, wherein The attributes of the event include one or more of face-swiping payment, face-swiping authentication, face-swiping application login, fingerprint payment, fingerprint authentication, and fingerprint application login.

7. A privacy-preserving feature processing device, comprising: The first-dimensional feature acquisition unit is configured to acquire first-dimensional encrypted information sent by a client, including a plurality of first ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the first dimension; and obtain first-dimensional feature ciphertexts by performing a first homomorphic operation on the plurality of first ciphertexts; The second-dimensional feature acquisition unit is configured to acquire second-dimensional encrypted information sent by the client, including a plurality of second ciphertexts obtained by homomorphically encrypting a plurality of attribute values ​​of the second dimension; and obtain the second-dimensional feature ciphertext by performing a second homomorphic operation on the plurality of second ciphertexts; A third-dimensional feature prediction unit is configured to input the first-dimensional feature ciphertext and the second-dimensional feature ciphertext into a pre-trained feature prediction model, wherein the feature prediction model outputs a predicted third-dimensional feature ciphertext; the third dimension is one of three dimensions consisting of a subject, an environment, and an event related to the client, and the first dimension and the second dimension are the other two dimensions of the above three dimensions; The feature anomaly judgment unit is configured as follows: Obtain third-dimensional encrypted information sent by the client, including several third ciphertexts obtained by homomorphically encrypting several attribute values ​​of the third dimension; obtain actual third-dimensional feature ciphertexts by performing third homomorphic operations on the several third ciphertexts; and determine whether the actual third-dimensional feature ciphertexts have anomalies based on the predicted third-dimensional feature ciphertexts.

8. The device according to claim 7, wherein the feature abnormality judgment unit is further configured to: If the result is yes, the predicted third-dimensional feature ciphertext is sent to the client to determine whether there is an anomaly in the attribute value of the third dimension.

9. The apparatus according to claim 7, further comprising: The feature sending unit is configured to send the predicted third-dimensional feature ciphertext to the client for predicting the attribute value of the third dimension on the client.

10. The device according to claim 7, wherein The subject includes one of a user and a device; when the subject is a user, the attributes of the subject include one or more of work, age, ID card, and household registration; when the subject is a device, the attributes of the subject include one or more of inherent applications on the device and device identification.

11. The device according to claim 7, wherein The attributes of the environment include one or more of location information and network address.

12. The device according to claim 7, wherein The attributes of the event include one or more of face-swiping payment, face-swiping authentication, face-swiping application login, fingerprint payment, fingerprint authentication, and fingerprint application login.

13. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method according to any one of claims 1 to 6.

14. A computing device comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method according to any one of claims 1 to 6 is implemented.

Images