本公開提供了安全的特徵工程方法和裝置，使得合作方均不必洩漏自己的隱私資料明文，但最終能完成特徵工程處理。例如，第一設備可向第二設備傳送第一資料集，第一資料集包含多個資料對象的標籤資訊的密文。第二設備可對包含該多個資料對象的特徵資料的第二資料集進行特徵工程處理以生成第二資料集的子集。第二設備可根據第二資料集的子集來生成第一資料集之中相應資料對象的標籤資訊的密文子集，將密文子集的密文發送給第一設備。第一設備可解密該密文子集並生成統計資訊，將該統計資訊發送給第二設備。第二設備可使用該統計資訊來計算特徵工程指標。
根據本公開的一個實施例，提供了一種特徵工程方法，其包括：接收第一資料集，第一資料集包含多個資料對象的標籤資訊的密文；對第二資料集進行特徵工程處理以生成第二資料集的子集，第二資料集包含所述多個資料對象的特徵資料；根據所述第二資料集的子集中所包含的資料對象來生成第一資料集之中相應資料對象的標籤資訊的密文子集；將所述密文子集的密文發送給第一設備；從第一設備接收在解密所述密文子集的密文後生成的統計資訊；以及使用所述統計資訊來計算對第二資料集進行的特徵工程處理的特徵工程指標。
在一方面，該方法還包括：根據所述特徵工程指標來生成特徵工程結果；以及將所述特徵工程結果發送給第一設備。
在一方面，所述特徵工程處理包括對第二資料集進行特徵選擇以生成第二資料集的特徵資料子集，其中計算特徵工程指標包括利用所述特徵資料子集與所述統計資訊來計算特徵選擇指標。
在一方面，所述特徵工程處理包括對第二資料集進行分箱以生成多個分箱段，並且所述密文子集包括與所述多個分箱段相對應的多個密文子集，其中每個密文子集包括第一資料集之中的落在相應分箱段中的資料對象的標籤資訊密文，其中計算特徵工程指標包括利用所述第二資料集的分箱段與所述統計資訊來計算分箱指標。
在一方面，第一資料集的標籤資訊的密文是使用同態加密對所述標籤資訊進行加密來生成的，所述標籤資訊為二元值，並且將所述密文子集的密文發送給第一設備包括：對所述密文子集中的密文進行同態加法以得到密文總和並將所述密文總和發送給第一設備。
在一方面，第一資料集和第二資料集之中的資料對象分別具有相關聯的識別符，所述方法還包括在生成所述第二資料集的子集之前對第一資料集和第二資料集進行識別符匹配，以使得第一資料集的標籤資訊與第二資料集的特徵資料對應於相同的多個資料對象。
在一方面，所述統計資訊包括所述密文子集的正樣本數、及/或負樣本數。
在一方面，對第二資料集進行特徵工程處理是在接收第一資料集之前、之時、或之後發生的。
在一方面，所述特徵工程指標包括以下至少一者：所述第二資料集的子集的正樣本率、所述第二資料集的子集的負樣本率、所述第二資料集的子集與所述密文子集中的標籤資訊的相關度、所述第二資料集的子集的資訊價值、所述第二資料集的子集的證據權重。
根據本公開的另一個實施例，提供了一種特徵工程方法，其包括：提供第一資料集，第一資料集包含多個資料對象的標籤資訊的密文；接收第一資料集的密文子集的密文，所述密文子集是根據對第二資料集進行的特徵工程處理來生成的，第二資料集包含所述多個資料對象的特徵資料，其中對第二資料集進行的特徵工程處理生成第二資料集的子集，其中所述密文子集包含第一資料集之中的與所述第二資料集的子集相同的資料對象的標籤資訊密文；解密所述密文子集的密文並生成所述密文子集的統計資訊；以及提供所述統計資訊以用於計算對第二資料集的特徵工程處理的特徵工程指標。
在一方面，所述方法還包括：接收根據所述特徵工程指標生成的第一特徵工程結果。
在一方面，所述方法還包括：使用所述第一資料集的標籤資訊和第三資料集的特徵資料來執行所述特徵工程處理以生成第二特徵工程結果；以及組合第一特徵工程結果和第二特徵工程結果以獲得組合的特徵工程結果。
在一方面，所述特徵工程處理包括對第二資料集進行特徵選擇以生成第二資料集的特徵資料子集，其中計算特徵工程指標包括利用所述特徵資料子集與所述統計資訊來計算特徵選擇指標。
在一方面，所述特徵工程處理包括對第二資料集進行分箱以生成多個分箱段，並且所述密文子集包括與所述多個分箱段相對應的多個密文子集，其中每個密文子集包括第一資料集之中的落在相應分箱段中的資料對象的標籤資訊密文，所述計算特徵工程指標包括利用所述第二資料集的分箱段與所述統計資訊來計算分箱指標。
在一方面，第一資料集的標籤資訊的密文是使用同態加密對所述標籤資訊進行加密來生成的，所述標籤資訊為二元值，所述密文子集的密文包括將所述密文子集中的密文進行同態加法得到的密文總和。
在一方面，第一資料集和第二資料集之中的資料對象分別具有相關聯的識別符，所述方法還包括在生成所述第二資料集的子集之前對第一資料集和第二資料集進行識別符匹配，以使得第一資料集的標籤資訊與第二資料集的特徵資料對應於相同的多個資料對象。
在一方面，所述統計資訊包括所述密文子集的正樣本數、及/或負樣本數。
在一方面，所述特徵工程指標包括以下至少一者：所述第二資料集的子集的正樣本率、所述第二資料集的子集的負樣本率、所述第二資料集的子集與所述密文子集中的標籤資訊的相關度、所述第二資料集的子集的資訊價值、所述第二資料集的子集的證據權重。
根據本公開的另一個實施例，提供了一種特徵工程裝置，其包括：傳輸模組，其接收第一資料集，第一資料集包含多個資料對象的標籤資訊的密文；特徵工程處理模組，其對第二資料集進行特徵工程處理以生成第二資料集的子集，第二資料集包含所述多個資料對象的特徵資料；以及密文子集生成模組，其根據所述第二資料集的子集中所包含的資料對象來生成第一資料集之中相應資料對象的標籤資訊的密文子集；其中所述傳輸模組將所述密文子集的密文發送給第一設備並從第一設備接收在解密所述密文子集的密文後生成的統計資訊；並且所述特徵工程處理模組使用所述統計資訊來計算對第二資料集進行的特徵工程處理的特徵工程指標。
在一方面，所述特徵工程處理模組根據所述特徵工程指標來生成特徵工程結果；以及所述傳輸模組將所述特徵工程結果發送給第一設備。
在一方面，所述特徵工程處理包括特徵選擇或分箱。
在一方面，所述統計資訊包括所述密文子集的正樣本數、及/或負樣本數。
根據本公開的另一個實施例，提供了一種特徵工程裝置，其包括：傳輸模組，其向第二設備提供第一資料集，第一資料集包含多個資料對象的標籤資訊的密文，所述傳輸模組還從第二設備接收第一資料集的密文子集的密文，所述密文子集是根據對第二資料集進行的特徵工程處理來生成的，第二資料集包含所述多個資料對象的特徵資料，其中對第二資料集進行的特徵工程處理生成第二資料集的子集，其中所述密文子集包含第一資料集之中的與所述第二資料集的子集相同的資料對象的標籤資訊密文；加解密模組，其解密所述密文子集的密文；以及處理模組，其生成所述密文子集的統計資訊，其中所述傳輸模組向第二設備提供所述統計資訊以供第二設備用於計算對第二資料集的特徵工程處理的特徵工程指標。
在一方面，其中所述傳輸模組從第二設備接收根據所述特徵工程指標生成的第一特徵工程結果，並且所述特徵工程裝置還包括：特徵工程處理模組，其使用所述第一資料集的標籤資訊和第三資料集的特徵資料來執行所述特徵工程處理以生成第二特徵工程結果，以及組合第一特徵工程結果和第二特徵工程結果以獲得組合的特徵工程結果。
在一方面，所述特徵工程處理包括特徵選擇或分箱。
在一方面，所述統計資訊包括所述密文子集的正樣本數、及/或負樣本數。
根據本公開的另一個實施例，提供了一種特徵工程系統，其包括：處理器；用於儲存處理器可執行指令的儲存器，其中所述處理器被配置成執行所述處理器可執行指令以實現如上所述的方法。
根據本公開，合作方均不必洩漏自己的隱私資料明文，但最終能完成特徵工程處理，從而提高了特徵工程處理的安全性。
The present disclosure provides a secure feature engineering method and device, so that the cooperating parties do not need to leak the plaintext of their private data, but the feature engineering processing can be completed in the end. For example, the first device may transmit a first data set to the second device, and the first data set includes the cipher text of the tag information of the multiple data objects. The second device can perform feature engineering processing on the second data set containing the feature data of the multiple data objects to generate a subset of the second data set. The second device may generate the ciphertext subset of the tag information of the corresponding data object in the first data set according to the subset of the second data set, and send the ciphertext of the ciphertext subset to the first device. The first device can decrypt the ciphertext subset and generate statistical information, and send the statistical information to the second device. The second device can use the statistical information to calculate feature engineering indicators.
According to an embodiment of the present disclosure, there is provided a feature engineering method, which includes: receiving a first data set, the first data set containing a ciphertext of tag information of a plurality of data objects; performing feature engineering processing on the second data set Generate a subset of a second data set, the second data set includes characteristic data of the plurality of data objects; generate corresponding data objects in the first data set according to the data objects included in the subset of the second data set The ciphertext subset of the tag information; sending the ciphertext of the ciphertext subset to the first device; receiving from the first device the statistical information generated after decrypting the ciphertext of the ciphertext subset; and using the statistical information To calculate the feature engineering index of the feature engineering processing of the second data set.
In one aspect, the method further includes: generating a feature engineering result according to the feature engineering index; and sending the feature engineering result to the first device.
In one aspect, the feature engineering processing includes performing feature selection on a second data set to generate a feature data subset of the second data set, wherein calculating the feature engineering index includes using the feature data subset and the statistical information to calculate Feature selection index.
In one aspect, the feature engineering processing includes binning a second data set to generate a plurality of binning segments, and the ciphertext subset includes a plurality of ciphertext subsets corresponding to the plurality of binning segments, Each ciphertext subset includes the label information ciphertext of the data objects in the corresponding binning segment in the first data set, and calculating the feature engineering index includes using the binning segment of the second data set and the Statistics information is used to calculate binning indicators.
In one aspect, the ciphertext of the tag information of the first data set is generated by encrypting the tag information using homomorphic encryption, the tag information is a binary value, and the ciphertext of the ciphertext subset is sent Providing the first device includes: performing a homomorphic addition on the ciphertext in the ciphertext subset to obtain a ciphertext sum, and sending the ciphertext sum to the first device.
In one aspect, the data objects in the first data set and the second data set respectively have associated identifiers, and the method further includes comparing the first data set and the second data set before generating the subset of the second data set. The two data sets perform identifier matching, so that the tag information of the first data set and the characteristic data of the second data set correspond to the same multiple data objects.
In one aspect, the statistical information includes the number of positive samples and/or the number of negative samples of the ciphertext subset.
In one aspect, the feature engineering processing of the second data set occurs before, during, or after the first data set is received.
In one aspect, the characteristic engineering index includes at least one of the following: a positive sample rate of a subset of the second data set, a negative sample rate of a subset of the second data set, and a The correlation between the subset and the tag information in the ciphertext subset, the information value of the subset of the second data set, and the weight of evidence of the subset of the second data set.
According to another embodiment of the present disclosure, a feature engineering method is provided, which includes: providing a first data set, the first data set including a ciphertext of tag information of a plurality of data objects; and receiving a ciphertext subset of the first data set The ciphertext of the ciphertext subset is generated based on the feature engineering processing of the second data set, the second data set contains the feature data of the multiple data objects, and the feature engineering performed on the second data set Processing to generate a subset of the second data set, wherein the ciphertext subset includes the tag information ciphertext of the same data object in the first data set as the subset of the second data set; decrypting the ciphertext subset And generate statistical information of the ciphertext subset; and provide the statistical information for calculating the feature engineering index of the feature engineering processing of the second data set.
In one aspect, the method further includes: receiving a first feature engineering result generated according to the feature engineering index.
In one aspect, the method further includes: using the tag information of the first data set and the feature data of the third data set to perform the feature engineering process to generate a second feature engineering result; and combining the first feature engineering result And the second feature engineering result to obtain a combined feature engineering result.
In one aspect, the feature engineering processing includes performing feature selection on a second data set to generate a feature data subset of the second data set, wherein calculating the feature engineering index includes using the feature data subset and the statistical information to calculate Feature selection index.
In one aspect, the feature engineering processing includes binning a second data set to generate a plurality of binning segments, and the ciphertext subset includes a plurality of ciphertext subsets corresponding to the plurality of binning segments, Each ciphertext subset includes the label information ciphertext of the data objects in the corresponding binning segment in the first data set, and the calculated feature engineering index includes using the binning segment and all the binning segments of the second data set. The statistical information is used to calculate the binning index.
In one aspect, the ciphertext of the tag information of the first data set is generated by encrypting the tag information using homomorphic encryption, the tag information is a binary value, and the ciphertext of the ciphertext subset includes all The sum of ciphertexts obtained by homomorphic addition of ciphertexts in the ciphertext subset.
In one aspect, the data objects in the first data set and the second data set respectively have associated identifiers, and the method further includes comparing the first data set and the second data set before generating the subset of the second data set. The two data sets perform identifier matching, so that the tag information of the first data set and the characteristic data of the second data set correspond to the same multiple data objects.
In one aspect, the statistical information includes the number of positive samples and/or the number of negative samples of the ciphertext subset.
In one aspect, the characteristic engineering index includes at least one of the following: a positive sample rate of a subset of the second data set, a negative sample rate of a subset of the second data set, and a The correlation between the subset and the tag information in the ciphertext subset, the information value of the subset of the second data set, and the weight of evidence of the subset of the second data set.
According to another embodiment of the present disclosure, there is provided a feature engineering device, which includes: a transmission module that receives a first data set, the first data set includes a ciphertext of tag information of a plurality of data objects; a feature engineering processing module Group, which performs feature engineering processing on a second data set to generate a subset of the second data set, the second data set includes the feature data of the plurality of data objects; and a ciphertext subset generation module, which is based on the first 2. The data objects contained in the subset of the data set are used to generate the ciphertext subset of the label information of the corresponding data object in the first data set; wherein the transmission module sends the ciphertext of the ciphertext subset to the first device And receiving from the first device the statistical information generated after decrypting the ciphertext of the ciphertext subset; and the feature engineering processing module uses the statistical information to calculate the feature engineering of the feature engineering process performed on the second data set index.
In one aspect, the feature engineering processing module generates a feature engineering result according to the feature engineering index; and the transmission module sends the feature engineering result to the first device.
In one aspect, the feature engineering process includes feature selection or binning.
In one aspect, the statistical information includes the number of positive samples and/or the number of negative samples of the ciphertext subset.
According to another embodiment of the present disclosure, there is provided a feature engineering device, which includes: a transmission module that provides a first data set to a second device, the first data set including ciphertext of tag information of a plurality of data objects, The transmission module also receives the ciphertext of the ciphertext subset of the first data set from the second device, the ciphertext subset is generated based on the feature engineering processing performed on the second data set, and the second data set contains all The feature data of the multiple data objects, wherein the feature engineering process performed on the second data set generates a subset of the second data set, wherein the ciphertext subset includes the first data set and the second data set The tag information ciphertext of the same data object in the subset; an encryption and decryption module that decrypts the ciphertext of the ciphertext subset; and a processing module that generates statistical information of the ciphertext subset, wherein the transmission module The group provides the statistical information to the second device for the second device to use to calculate the feature engineering index of the feature engineering processing of the second data set.
In one aspect, the transmission module receives the first feature engineering result generated according to the feature engineering index from the second device, and the feature engineering device further includes: a feature engineering processing module that uses the first feature engineering The label information of the data set and the feature data of the third data set are used to perform the feature engineering process to generate a second feature engineering result, and to combine the first feature engineering result and the second feature engineering result to obtain a combined feature engineering result.
In one aspect, the feature engineering process includes feature selection or binning.
In one aspect, the statistical information includes the number of positive samples and/or the number of negative samples of the ciphertext subset.
According to another embodiment of the present disclosure, there is provided a feature engineering system, which includes: a processor; a storage for storing processor-executable instructions, wherein the processor is configured to execute the processor-executable instructions To achieve the method described above.
According to the present disclosure, the partners do not need to disclose the plaintext of their own private data, but can finally complete the feature engineering processing, thereby improving the security of the feature engineering processing.
下面結合具體實施例和圖式對本公開作進一步說明，但不應以此限制本公開的保護範圍。
本公開提供了安全的特徵工程方法和裝置。在聯合建模場景中，一般存在其中一方擁有特徵資料，另一方擁有標籤資訊而沒有或有少量不齊全的特徵資料的情況，這就需要建模合作方的樣本資料共同協作來完成特徵工程處理。根據本公開的一個實施例，第一設備可向第二設備傳送第一資料集，第一資料集包含多個資料對象的標籤資訊的密文。第二設備可對包含該多個資料對象的特徵資料的第二資料集進行特徵工程處理以生成第二資料集的子集。第二設備可根據第二資料集的子集來生成第一資料集之中相應資料對象的標籤資訊的密文子集，將密文子集的密文發送給第一設備。第一設備可解密該密文子集並生成統計資訊，將該統計資訊發送給第二設備。第二設備可使用該統計資訊來計算特徵工程指標。
本公開的分布式特徵工程和指標計算使得合作方均不必洩漏自己的隱私資料明文，但最終能完成特徵工程處理，實現了安全的特徵工程方法和裝置。
圖1為根據本公開一個實施例的特徵工程方法的流程圖。該方法可以是在例如第二設備處執行的。
步驟102：第二設備接收第一資料集，第一資料集包含多個資料對象的標籤資訊的密文。例如，第一設備可請求第二設備進行特徵工程處理，並且第一設備可將第一資料集發送給第二設備，其中所發送的第一資料集包含多個資料對象的標籤資訊的密文。在另一示例中，第二設備需要進行特徵工程處理並且可請求第一設備提供包含標籤資訊的第一資料集。在任一種情形中，第二設備可具有包含特徵資料的第二資料集，並且可將第一資料集和第二資料集相結合地進行特徵工程處理，例如特徵選擇、特徵提取、特徵離散化(如分箱)等。
特徵資料可以表示對象的特性或行為等資訊，例如用戶的年齡、身高、消費記錄等。標籤資訊可以識別對象的類別，例如用戶是否為學生、用戶是否為失信者等等。作為示例而非限定，標籤資訊可以為二元值，例如用1、0分別表示正樣本和負樣本，或者用0、-1分別表示正樣本和負樣本，或以其他方式區分正樣本和負樣本。
在一些情形中，雖然第二設備可將第一資料集的標籤資訊和第二資料集的特徵資料相結合地進行特徵工程處理，但擁有第一資料集的第一設備可能不希望將第一資料集的標籤資訊的具體值洩露給第二設備。由此，根據一個實施例，第一設備發送給第二設備的第一資料集包含多個資料對象的標籤資訊的密文，而不是這些標籤資訊的明文資訊。
第一設備可以利用密鑰對第一資料集之中的各個資料對象的標籤資訊逐個進行加密以生成標籤資訊的密文，或者第一設備可以接收包含由第三方加密後的標籤資訊密文的第一資料集。較佳地，第一設備能夠解密第一資料集的密文，而第二設備不能解密第一資料集的密文。因此，即使第一設備將第一資料集的標籤資訊密文發送給第二設備，第二設備也不知曉每個資料對象的具體標籤資訊值。第一設備可以直接將包含密文的第一資料集發送給第二設備，也可以經由中間設備將第一資料集轉發給第二設備。例如，第一設備可以將第一資料集上傳到雲端(比如阿里oss或者開源sftp等)供第二設備下載，也可以透過網路進行大文件的點對點傳輸提供給第二設備。第二設備可以從中間設備(例如，雲端)接收第一資料集，或者接收來自匿名方的第一資料集，或者以其他方式獲得包含密文的第一資料集。
步驟104：第二設備對第二資料集進行特徵工程處理以生成第二資料集的子集。如上所述，第二資料集可包含多個資料對象的特徵資料。第二設備可以對第二資料集之中所包含的多個資料對象的特徵資料進行資料清理、資料篩選、特徵離散化、特徵提取、特徵選擇等特徵工程處理，以生成第二資料集的子集，該子集可包括第二資料集之中的一些資料對象的特徵資料。
對第二資料集的特徵工程處理可以在步驟102之前或之後執行。例如，擁有第二資料集的第二設備可以在任何時間自主地對第二資料集進行特徵工程處理，也可以在從第一設備接收到特徵工程處理請求(及/或第一資料集)之後對第二資料集進行特徵工程處理。
步驟106：第二設備根據第二資料集的子集中所包含的資料對象來生成第一資料集之中相應資料對象的標籤資訊的密文子集。
如上所述，第一資料集(例如，包含標籤資訊)可與第二資料集(例如，包含特徵資料)聯合地用於特徵工程處理。然而，第二設備所接收到的第一資料集之中的標籤資訊是經加密的(即，為密文形式)，並且第二設備可能無法對第一資料集的密文進行解密，由此不知曉第一資料集的標籤資訊的具體值。以特徵資料和經加密的標籤資訊為例，第二設備由於不知曉標籤資訊的具體值，因此不能直接用第二資料集的特徵資料和第一資料集之中的經加密的標籤資訊來聯合地進行特徵工程處理。
根據本公開的一個實施例，第一資料集的標籤資訊與第二資料集的特徵資料可對應於相同的多個資料對象。作為示例而非限定，第一設備擁有的標籤資訊和第二設備擁有的特徵資料可各自具有識別符(id)，第一設備和第二設備可進行識別符匹配以分別在第一設備處生成第一資料集並在第二設備處生成第二資料集，使得第一資料集的標籤資訊與第二資料集的特徵資料對應於相同的多個資料對象。此後，第一設備可在步驟102將第一資料集發送給第二資料集，第一資料集可包括該多個資料對象的標籤資訊的密文和相關聯的識別符。第二設備可以根據第二資料集的子集中所包含的資料對象按照識別符來生成第一資料集之中相應資料對象的標籤資訊的密文子集。
作為進一步示例，如果在第一設備處生成的第一資料集和在第二設備處生成的第二資料集針對相同的多個資料對象，則還可以例如根據識別符按相同順序來排列雙方的資料，例如使得第一資料集和第二資料集的每條資料彼此對應，其中每條資料可對應於不同的對象。如果第一資料集和第二資料集的資料對象按相同的方式排序，則第一設備發送給第二設備的第一資料集可以不必包含識別符。第二設備可以按照兩個資料集之間的對應關係，根據第二資料集的子集中所包含的資料對象來生成第一資料集之中相應資料對象的標籤資訊的密文子集。
在另一實施例中，如果第二設備在接收第一資料集之前沒有與第一設備進行識別符匹配，即，第一資料集可能包含與第二資料集不同的資料對象，則第二設備可在接收第一資料集之後對第一資料集和第二資料集進行識別符匹配，由此剔除一方獨有的資料對象，以使得第一資料集的標籤資訊和第二資料集的特徵資料對應於相同的多個資料對象。此後，第二設備可在步驟106根據第二資料集的子集中所包含的資料對象來生成第一資料集之中相應資料對象的標籤資訊的密文子集。
以特徵選擇作為示例，對第二資料集的特徵工程處理可包括選擇第二資料集的特徵資料子集，並且第二設備可生成與該特徵資料子集中所包含的資料對象相對應的第一資料集之中的標籤資訊的密文子集。作為另一示例，對第二資料集的特徵工程處理可包括對特徵資料進行分箱以生成多個分箱段，由此第二設備可以生成第一資料集的與該多個分箱段相對應的多個密文子集，其中每個密文子集包括第一資料集之中的落在相應分箱段中的資料對象的標籤資訊密文。即，第二設備可以根據對第二資料集的特徵資料的分箱來確定第一資料集之中相應的資料對象分別落在哪個分箱段中，並按分箱段來生成第一資料集之中的標籤資訊的密文子集。在步驟106中，第二設備可以生成第一資料集之中的標籤資訊的密文子集，並且可以知曉該密文子集中的標籤資訊的總數。然而如上所述，第一資料集的標籤資訊是加密的，第二設備不知曉這些標籤資訊的值，由此無法完成特徵工程處理的指標計算。例如，以分箱為例，第二設備不知道該密文子集中哪些標籤資訊為正樣本，哪些標籤資訊為負樣本，由此不知曉每個分箱段中的標籤資訊的正樣本數和負樣本數，無法計算分箱指標，如IV(Information Value，資訊價值)、Woe(Weight of Evidence，證據權重)等。
步驟108：第二設備將該密文子集的密文發送給第一設備，並從第一設備接收在解密該密文子集的密文後生成的統計資訊。第二設備發送給第一設備的該密文子集的密文可以是該密文子集中的各個標籤資訊的密文集合，或者對這些標籤資訊的密文進行某種操作或運算(例如，組合、級聯、代數運算等)之後得到的密文。
第一設備可以是第一資料集的擁有方並且能夠對標籤資訊的密文進行解密。由此，第一設備可以對該密文子集的密文進行解密並生成統計資訊(例如，該密文子集的正樣本數及/或負樣本數)，並且可以將該統計資訊發送給第二設備。應理解，第一設備發送的統計資訊不同於第一資料集的標籤資訊密文的解密值，由此不會洩露第一資料集的標籤資訊值。
以分箱為例，第二設備可以將與多個分箱段相對應的多個密文子集發送給第一設備，每個密文子集的密文可包括第一資料集之中的落在相應分箱段中的標籤資訊密文的集合。第一設備可以解密出每個分箱段對應的標籤資訊密文，並獲得每個分箱段對應的標籤資訊的統計資訊(例如，每個分箱段對應的標籤資訊的正樣本數及/或負樣本數)作為上述統計資訊發送給第二設備。對於其他特徵工程處理(例如，特徵選擇等)，第一設備可類似地解密每個密文子集的密文並生成特徵工程處理所需的統計資訊發送給第二設備。
在一個示例中，第一資料集的標籤資訊的密文可以是使用同態加密(Homomorphic Encryption)對標籤資訊進行加密來生成的。同態加密允許對經同態加密後的密文進行特定的代數運算得到仍然是加密的運算結果，將其解密所得到的運算結果與對明文進行同樣的運算得到的運算結果一樣。換言之，這項技術令人們可以在經加密的資料中進行諸如檢索、比較等操作，得出正確的結果，而在整個處理過程中無需對資料進行解密。如果標籤資訊為二元值並使用同態加密來加密，則第二設備可以對密文子集中的密文進行同態加法以得到密文總和，並將該密文總和發送給第一設備。第一設備解密該密文總和得到的值等同於第一設備解密該密文子集中的每個標籤資訊密文並求和得到的值。由此，第一設備可以解密該密文總和，並且解密結果可指示正樣本數及/或負樣本數。
步驟110：第二設備使用該統計資訊來計算對第二資料集的特徵工程處理的特徵工程指標。特徵工程指標可用於評價特徵工程處理產生的結果。特徵工程指標可根據具體的特徵工程處理進行選擇，例如第二資料集的子集的正樣本率、第二資料集的子集的負樣本率、第二資料集的子集與密文子集中的標籤資訊的相關度、第二資料集的子集的資訊價值、第二資料集的子集的證據權重等。由此，第二設備可以根據第二資料集的子集和對應的統計資訊來獲得該子集的特徵工程指標。以分箱為例，該統計資訊可以是每個分箱段的關於標籤資訊的統計值，第二設備可以利用從第一設備接收的統計值並結合第二設備在對特徵資料進行分箱時獲得的分箱段資訊一起進行分箱指標計算。如果特徵工程指標達到要求(例如，閾值)，則第二設備可以生成特徵工程處理結果並完成該特徵工程處理。所生成的特徵工程處理結果(例如，特徵資料子集)可用於各種目的，例如模型訓練、目標預測等。在可選的實施例中，第二設備可以將特徵工程結果發送給第一設備。在進一步的實施例中，如果特徵工程指標不理想(例如，未達到閾值)，第二設備可重複步驟104-110，直到獲得符合要求的特徵工程指標或者直到達成終止特徵工程處理的條件。
根據上述方法，合作方均不必洩漏自己的隱私資料明文，但最終能完成特徵工程處理。以下結合圖2-4進一步詳細描述本公開的應用場景。
圖2為根據本公開一個實施例的特徵工程方法的示意圖。作為示例而非限定，以第一設備擁有標籤資訊Y(第一資料集)，第二設備擁有特徵資料X(第二資料集)的特徵選擇場景為例進行說明。為簡化描述，假設標籤資訊Y(第一資料集)和特徵資料X(第二資料集)對應於相同的多個資料對象(例如，透過進行識別符匹配)。特徵選擇的目標是尋找最優特徵子集。特徵選擇能剔除不相關(irrelevant)或冗餘(redundant)的特徵，從而達到減少特徵個數、降低資料集維度、提高模型精確度、減少運行時間的目的。另一方面，特徵選擇可用於選取出真正相關的特徵簡化模型，協助理解資料產生的過程。
在步驟202，第一設備可加密第一資料集之中的Y樣本並生成經加密的Y’樣本。在步驟204，第一設備可將經加密的Y’樣本發送給第二設備。作為示例而非限定，第一設備可在發送之前進一步壓縮經加密的Y’樣本，並將壓縮後的加密Y’樣本發送給第二設備。若Y’樣本資料量大，可考慮將密文文件上傳到雙方可存取的安全雲端(例如oss等)。
在一個實施例中，第一設備可對第一資料集之中的Y樣本進行逐個加密以生成Y樣本的密文Y’。第一設備可以利用密鑰以任何合適的加密演算法對第一資料集之中的Y樣本逐個進行加密，並將Y’樣本(例如，不壓縮或壓縮後)發送給第二設備。作為示例而非限定，可以採用同態加密(Homomorphic Encryption)來對第一資料集之中的Y樣本進行逐個全量同態加密。
在步驟206，第二設備可接收經加密的Y’樣本。例如，第二設備可從第一設備或其他中間方(例如，oss等雲端)獲取經加密的Y’樣本。如果Y’樣本是經壓縮的，則第二設備可解壓Y’樣本。
在步驟208，第二設備可選擇本地擁有的X樣本(例如，特徵資料)的子集。應理解，第二設備可以按各種準則或組合來選擇X樣本子集。此外，第二設備可以選擇一個或多個X樣本子集並依次或並行地評價這一個或多個X樣本子集。
雖然圖2中示出了步驟208在步驟206之後執行，但是應理解，步驟208可以在步驟202-206中的任一個步驟之前或之後執行，也可以與步驟202-206中的任一個步驟併發地執行。
在步驟210，第二設備可根據所選擇的X樣本子集來讀取相對應的Y’樣本子集。類似於以上所述，X樣本的id和Y樣本的id可以相匹配，以使得X樣本和Y樣本具有對應關係。由此，第二設備可根據所選擇的X樣本子集來讀取相對應id的Y’樣本子集。如上所述，由於Y’樣本是加密的，因此第二設備不知曉Y’樣本的值，從而該Y’樣本子集可包括經加密的Y’樣本(包含密文)。
在步驟212，第二設備可以將Y’樣本子集的密文發送給第一設備進行解密。在一個示例中，在步驟212發送的Y’樣本子集的密文可包括Y’樣本子集所包含的各個Y’樣本的集合。
在步驟214，第一設備解密Y’樣本子集的密文。例如，第一設備可解密該子集中包含的每個Y’樣本以獲得經解密的Y樣本。
在步驟216，第一設備在解密Y’樣本子集的密文之後可以生成關於特徵選擇的統計資訊。例如，第一設備可解密出各個Y’樣本的值，並由此獲得用於評價特徵選擇子集的統計資訊(例如正樣本數、負樣本數等)。
在另一個示例中，如果Y樣本為二元值並使用同態加密來生成Y’樣本，則第二設備可以對Y’樣本子集的各個Y’樣本進行同態加法以得到密文總和，並在步驟212將該密文總和發送給第一設備，而不必將該Y’樣本子集完整地發送給第一設備。第一設備可在步驟214解密該密文總和並直接生成統計資訊。例如，如果用1、0分別表示正樣本和負樣本，則Y’樣本子集的同態密文總和即為所有1的密文總和，該密文總和被解密後即為所有正樣本的數量。反之，如果用1、0分別表示負樣本和正樣本，則解密Y’樣本子集的同態密文總和可以得到負樣本數量。
在步驟218，第一設備可將Y’樣本子集的統計資訊發送給第二設備。應理解，第一設備發送Y’樣本子集的統計資訊不會洩露樣本Y的值。
在步驟220，第二設備可根據該統計資訊來計算特徵選擇指標，如X樣本子集的正樣本率、X樣本子集的負樣本率、X樣本子集與預測目標(例如，Y樣本為正或負)的相關度、X樣本子集的資訊價值、X樣本子集的證據權重等。作為示例而非限定，第二設備可從步驟220返回到步驟208以選擇下一組的一個或多個X樣本子集進行評價。
在步驟222，在獲得了符合要求的(例如，最優的)特徵子集的情況下，第二設備可以生成特徵選擇結果並完成特徵選擇。
在可選步驟224，第二設備可將特徵選擇結果發送給第一設備。相應地，在可選步驟226，第一設備可從第二設備接收特徵選擇結果。
圖3為根據本公開另一個實施例的特徵工程方法的示意圖。作為示例而非限定，以第一設備(例如，需求方)擁有標籤資訊Y(第一資料集)，第二設備(例如，資料方)擁有特徵資料X(第二資料集)的分箱場景為例進行說明。為簡化描述，假設標籤資訊Y(第一資料集)和特徵資料X(第二資料集)對應於相同的多個資料對象。
分箱是一種常用的資料預處理方法，其按照屬性值劃分的子區間，如果一個屬性值處於某個子區間範圍內，就稱把該把待處理的資料(某列屬性值)按照一定的規則放進一些箱子中，考察每一個箱子中的資料，採用某種方法分別對各個箱子中的資料進行處理。分箱可包括等深分箱法、等寬分箱法、最小熵法和用戶自定義區間法等。
圖3中的步驟302-306與圖2中的步驟202-206相似並且不再贅述。
在步驟308，第二設備可對本地擁有的X樣本(例如，特徵資料)進行分箱。雖然圖3中示出了步驟308在步驟306之後執行，但是應理解，步驟308的分箱可以在步驟302-306中的任一個步驟之前或之後執行，也可以與步驟302-306中的任一個步驟併發地執行。例如，步驟308處對X樣本的分箱可以是預先(例如，在步驟302之前)執行的。在其他實施例中，對X樣本的分箱可以在步驟306之後(或者，在接收到來自第一設備的分箱請求之後)執行。
在步驟310，第二設備以X樣本的分箱段為單位讀取相應的Y’樣本子集。
根據本公開的一個實施例中，可在步驟308之前的任何時間對X樣本的id和Y樣本的id進行匹配，以使得X樣本和Y樣本具有對應關係。作為示例而非限定，第一設備發送給第二設備的加密Y’樣本具有相關聯的識別符(id)，第二設備本地的X樣本具有相關聯的識別符，其中X樣本和Y樣本的識別符是匹配的(例如，對應於相同的一組對象)。由此，第二設備可以根據X樣本的分箱來確定相應的id分別落在哪個分箱段中，並按分箱段來讀取各id對應的Y’樣本。在一個實施例中，第一設備發送給第二設備的加密Y’樣本的識別符可以被加密(例如，進行雜湊散列)或不被加密。在另一實施例中，如果第一設備和第二設備先進行了匹配，以使得X樣本和Y’樣本按特定方式排序(例如使得每項資料一一對應)，則第一設備發送給第二設備的加密Y’樣本可以不必包括id資訊，第二設備可以直接根據X樣本的分箱來讀取相對應的Y’樣本。
在步驟312，第二設備可以將各個分箱段對應的Y’樣本子集的密文發送給第一設備。在一個實施例中，可在步驟312發送多個Y’樣本子集的密文，其中每個Y’樣本子集的密文可包括相應分箱段對應的Y’樣本的集合。在另一個實施例中，如果Y’樣本是採用同態加密來加密的，則在步驟312發送的每個Y’樣本子集的密文可包括將該分箱段對應的Y’樣本進行同態加法得到的密文總和。
在步驟314，第一設備解密每個分箱段對應的Y’樣本子集的密文。第一設備可以是曾對Y進行加密的設備並且能夠解密Y’，或者即使加密不是由第一設備執行的，但第一設備具有密鑰權限並由此能夠解密Y’。例如，第一設備可解密Y’樣本子集中包含的每個Y’樣本，或者在該子集包含同態密文總和的情況下解密該同態密文總和。如果Y’樣本是二元值且採用同態加密來加密，則解密該密文總和可以直接得出正樣本數或負樣本數。
在步驟316，第一設備在解密每個分箱段對應的Y’樣本子集的密文之後可以生成關於每個分箱段的統計資訊。例如，第一設備可解密出每個分箱段的各個Y’樣本的值，並由此獲得每個分箱段的正樣本數、負樣本數、及/或總樣本數等參數。在另一個實施例中，在步驟312處發送的Y’樣本子集的密文包含同態密文總和的情況下解密同態密文總和提供了各個分箱段的正樣本數或負樣本數。例如，如果用1、0分別表示正樣本和負樣本，則同態密文總和即為所有正樣本的數量。
在步驟318，第一設備可將各個分箱段的統計資訊(例如Y樣本統計資訊，如正樣本數、負樣本數、及/或總樣本數等)發送給第二設備。應理解，第一設備發送各個分箱段的統計資訊不會洩露樣本Y的值。
在步驟320，第二設備可根據該統計資訊來計算分箱指標。例如，第二設備可以利用從第一設備接收的關於每個分箱段的統計值並結合第二設備在對特徵資料X進行分箱時獲得的其他資訊一起進行分箱指標計算，如各X樣本分箱段的正樣本率、各X樣本分箱段的負樣本率、各X樣本分箱段與預測目標(例如，Y樣本為正或負)的相關度、各X樣本分箱段的IV(Information Value，資訊價值)、各X樣本分箱段的Woe(Weight of Evidence，證據權重)等。
在步驟322，在分箱指標有效(例如，達到閾值)的情況下，第二設備可以生成分箱結果並完成分箱。在分箱指標不理想的情況下，第二設備可重複步驟308-320以重新執行分箱並進行分箱指標計算。
在可選步驟324，第二設備可將分箱結果發送給第一設備。相應地，在可選步驟326，第一設備可以從第二設備接收分箱結果。
圖4為根據本公開另一個實施例的特徵工程方法的示意圖。在一個實施例中，一方(例如，第一設備)可以具有特徵資料X2和標籤資訊Y，另一方(例如，第二設備)具有特徵資料X1。特徵資料X1和特徵資料X2可以彼此相同、部分相同、或者彼此不同。為簡化描述，假設標籤資訊Y(第一資料集)、特徵資料X1(第二資料集)、和特徵資料X2(第三資料集)對應於相同的多個資料對象。第一設備可以如圖3所述地將經加密的標籤資訊Y’發送給第二設備以結合特徵資料X1進行分箱，還可以在本地針對特徵資料X2和標籤資訊Y進行本地分箱。
在步驟402-422中，第一設備和第二設備可以如參考圖3中的步驟302-322所述地根據特徵資料X1和標籤資訊Y’進行分箱並生成第一分箱結果。
在步驟424，第二設備可以將第一分箱結果發送給第一設備。
在步驟426，第一設備可在本地針對特徵資料X2和未加密的標籤資訊Y進行分箱以獲得第二分箱結果。應注意，第一設備在步驟426進行的本地分箱可以在步驟402-424中的任一個步驟之前或之後執行，或與其中任一個步驟並行地執行。
在步驟428，第一設備可以組合第一分箱結果和第二分箱結果，以得到組合的分箱結果。
上文以特徵工程處理中的特徵選擇和分箱為例介紹了分布式特徵工程和指標計算，使得合作方均不必洩漏自己的隱私資料明文，但最終能完成特徵工程處理。本領域技術人員可以將本公開的分布式特徵工程方法和裝置廣泛地應用於需要把各方的建模樣本資料融合在一起進行處理而不想洩露樣本資料明文資訊的場合，比如資料清理、特徵離散化、特徵提取、特徵選擇等。
圖5示出了根據本公開一個實施例的特徵工程裝置的方塊圖。例如，第一設備510可執行以上所描述的關於第一設備的操作，第二設備520可執行以上所描述的關於第二設備的操作。
如圖5所示，第一設備510可包括加解密模組512、傳輸模組514、處理模組516。加解密模組512可對第一資料集的標籤資訊進行加密以生成密文。傳輸模組514可將包含標籤資訊密文的第一資料集傳送給第二設備520或其他設備或雲端。
第二設備520可包括傳輸模組522、特徵工程處理模組524和密文子集生成模組526。傳輸模組522可接收包含標籤資訊密文的第一資料集。此外，第二設備520可具有第二資料集(例如，儲存在儲存器中)，第二資料集包含多個資料對象的特徵資料。特徵工程處理模組524可對第二資料集進行特徵工程處理以生成第二資料集的子集。第一資料集的標籤資訊與第二資料集的特徵資料可對應於相同的多個資料對象。密文子集生成模組526可根據第二資料集的子集中所包含的資料對象來生成第一資料集之中相應資料對象的標籤資訊的密文子集。傳輸模組522可進一步將該密文子集的密文發送給第一設備。
在第一設備510中，傳輸模組514可接收該密文子集的密文並將其傳遞給加解密模組512，加解密模組512可解密所述密文子集的密文。處理模組516可根據該密文子集的密文被解密後的資料生成用於特徵工程處理的統計資訊，例如正樣本數、負樣本數、總樣本數等。傳輸模組514可將所述統計資訊提供給第二設備520。
在第二設備520中，傳輸模組522接收所述統計資訊並將其提供給特徵工程處理模組524。特徵工程處理模組524可使用所述統計資訊來計算對第二資料集進行的特徵工程處理的特徵工程指標並完成所述特徵工程處理。可選地，在完成特徵工程處理後，傳輸模組522還可以將特徵工程處理結果發送給第一設備510。
在進一步的實施例中，第一設備510可任選地包括特徵工程處理模組518，其可使用第一資料集的未加密的標籤資訊和第三資料集(例如，包含特徵資料X2)來執行特徵工程處理以生成第二特徵工程結果。第一資料集的標籤資訊與第三資料集的特徵資料可對應於相同的多個資料對象。第一設備510還可以組合由第二設備520生成的上述特徵工程結果和第二特徵工程結果以獲得組合的特徵工程結果。在一個實施例中，處理模組516和特徵工程處理模組518可以一起實現或分開實現。
根據本公開，聯合建模的資料合作各方(例如，第一設備510和第二設備520)不對外輸出隱私資料明文，但最終得到特徵工程處理結果，防止了私有資料洩漏，解決了資料合作過程中的隱私資料洩漏和資料信任問題。
以上描述的特徵工程方法和裝置的各個步驟和模組可以用硬體、軟體、或其組合來實現。如果在硬體中實現，結合本公開描述的各種說明性步驟、模組、以及電路可用通用處理器、數位信號處理器(DSP)、專用積體電路(ASIC)、現場可程式化閘陣列(FPGA)、或其他可程式化邏輯組件、硬體組件、或其任何組合來實現或執行。通用處理器可以是處理器、微處理器、控制器、微控制器、或狀態機等。如果在軟體中實現，則結合本公開描述的各種說明性步驟、模組可以作為一條或多條指令或代碼儲存在電腦可讀媒體上或進行傳送。實現本公開的各種操作的軟體模組可駐留在儲存媒體中，如RAM、快閃記憶體、ROM、EPROM、EEPROM、暫存器、硬碟、可移除碟、CD-ROM、雲儲存等。儲存媒體可耦接到處理器以使得該處理器能從/向該儲存媒體讀寫資訊，並執行相應的程式模組以實現本公開的各個步驟。而且，基於軟體的實施例可以透過適當的通訊手段被上載、下載或遠端地存取。這種適當的通訊手段包括例如網際網路、全球資訊網、內聯網、軟體應用、電纜(包括光纖電纜)、磁通訊、電磁通訊(包括RF、微波和紅外通訊)、電子通訊或者其他這樣的通訊手段。
還應注意，這些實施例可能是作為被描繪為流程圖、流圖、結構圖、或方塊圖的過程來描述的。儘管流程圖可能會把諸操作描述為順序過程，但是這些操作中有許多操作能夠並行或併發地執行。另外，這些操作的次序可被重新安排。
所公開的方法、裝置和系統不應以任何方式被限制。相反，本公開涵蓋各種所公開的實施例(單獨和彼此的各種組合和子組合)的所有新穎和非顯而易見的特徵和方面。所公開的方法、裝置和系統不限於任何具體方面或特徵或它們的組合，所公開的任何實施例也不要求存在任一個或多個具體優點或者解決特定或所有技術問題。
上面結合圖式對本公開的實施例進行了描述，但是本公開並不局限於上述的具體實施方式，上述的具體實施方式僅僅是示意性的，而不是限制性的，本領域的普通技術人員在本公開的啟示下，在不脫離本公開宗旨和請求項所保護的範圍情況下，還可做出很多更改，這些均落在本公開的保護範圍之內。
The present disclosure will be further described below in conjunction with specific embodiments and drawings, but the protection scope of the present disclosure should not be limited by this.
The present disclosure provides safe feature engineering methods and devices. In a joint modeling scenario, there is generally a situation where one party has feature data and the other party has label information but no or a small amount of incomplete feature data. This requires the collaboration of the sample data of the modeling partners to complete the feature engineering process . According to an embodiment of the present disclosure, the first device may transmit a first data set to the second device, and the first data set includes the cipher text of the tag information of a plurality of data objects. The second device can perform feature engineering processing on the second data set containing the feature data of the multiple data objects to generate a subset of the second data set. The second device may generate the ciphertext subset of the tag information of the corresponding data object in the first data set according to the subset of the second data set, and send the ciphertext of the ciphertext subset to the first device. The first device can decrypt the ciphertext subset and generate statistical information, and send the statistical information to the second device. The second device can use the statistical information to calculate feature engineering indicators.
The distributed feature engineering and index calculation of the present disclosure make it unnecessary for the cooperating parties to leak the plaintext of their own private data, but can finally complete feature engineering processing, realizing a safe feature engineering method and device.
Fig. 1 is a flowchart of a feature engineering method according to an embodiment of the present disclosure. The method may be executed at, for example, the second device.
Step 102: The second device receives the first data set, the first data set including the cipher text of the tag information of the multiple data objects. For example, the first device can request the second device to perform feature engineering processing, and the first device can send the first data set to the second device, where the sent first data set contains the ciphertext of the tag information of multiple data objects . In another example, the second device needs to perform feature engineering processing and can request the first device to provide a first data set containing tag information. In any case, the second device can have a second data set containing feature data, and can combine the first data set and the second data set to perform feature engineering processing, such as feature selection, feature extraction, and feature discretization ( Such as sub-box) and so on.
The characteristic data can represent information such as the characteristics or behavior of the object, such as the user's age, height, consumption history, and so on. The tag information can identify the category of the object, such as whether the user is a student, whether the user is a dishonest person, and so on. As an example and not a limitation, the label information can be a binary value, for example, use 1, 0 to represent positive samples and negative samples, or use 0, -1 to represent positive samples and negative samples, or distinguish positive samples and negative samples in other ways sample.
In some cases, although the second device may combine the tag information of the first data set with the feature data of the second data set to perform feature engineering processing, the first device that owns the first data set may not want to transfer the first data set to the first device. The specific value of the tag information of the data set is leaked to the second device. Thus, according to one embodiment, the first data set sent by the first device to the second device contains the ciphertext of the tag information of the multiple data objects, rather than the plaintext information of the tag information.
The first device can use the key to encrypt the label information of each data object in the first data set one by one to generate the ciphertext of the label information, or the first device can receive the ciphertext of the label information encrypted by a third party The first data set. Preferably, the first device can decrypt the ciphertext of the first data set, but the second device cannot decrypt the ciphertext of the first data set. Therefore, even if the first device sends the tag information cipher text of the first data set to the second device, the second device does not know the specific tag information value of each data object. The first device may directly send the first data set containing the ciphertext to the second device, or may forward the first data set to the second device via the intermediate device. For example, the first device may upload the first data set to the cloud (for example, Ali OSS or open source sftp, etc.) for the second device to download, or it may perform point-to-point transmission of large files to the second device through the network. The second device may receive the first data set from the intermediate device (for example, the cloud), or receive the first data set from the anonymous party, or obtain the first data set containing the ciphertext in other ways.
Step 104: The second device performs feature engineering processing on the second data set to generate a subset of the second data set. As described above, the second data set may include characteristic data of multiple data objects. The second device can perform feature engineering processing such as data cleaning, data screening, feature discretization, feature extraction, and feature selection on the feature data of multiple data objects included in the second data set to generate a subset of the second data set The subset may include characteristic data of some data objects in the second data set.
The feature engineering processing of the second data set can be performed before or after step 102. For example, the second device with the second data set can autonomously perform feature engineering processing on the second data set at any time, or after receiving a feature engineering processing request (and/or the first data set) from the first device Perform feature engineering processing on the second data set.
Step 106: The second device generates a ciphertext subset of the label information of the corresponding data object in the first data set according to the data objects included in the subset of the second data set.
As described above, the first data set (e.g., containing tag information) can be used in combination with the second data set (e.g., containing feature data) for feature engineering processing. However, the tag information in the first data set received by the second device is encrypted (that is, in the form of ciphertext), and the second device may not be able to decrypt the ciphertext of the first data set. The specific value of the label information of the first data set is unknown. Take characteristic data and encrypted label information as an example. Since the second device does not know the specific value of the label information, it cannot directly combine the characteristic data of the second data set with the encrypted label information in the first data set. Carry out feature engineering processing.
According to an embodiment of the present disclosure, the tag information of the first data set and the feature data of the second data set can correspond to the same multiple data objects. As an example and not a limitation, the tag information possessed by the first device and the characteristic data possessed by the second device may each have an identifier (id), and the first device and the second device may perform identifier matching to generate respectively at the first device The first data set and the second data set are generated at the second device, so that the label information of the first data set and the characteristic data of the second data set correspond to the same multiple data objects. Thereafter, the first device may send the first data set to the second data set in step 102, and the first data set may include the ciphertext of the tag information of the plurality of data objects and the associated identifiers. The second device may generate a ciphertext subset of the tag information of the corresponding data object in the first data set according to the identifiers of the data objects contained in the subset of the second data set.
As a further example, if the first data set generated at the first device and the second data set generated at the second device are for the same multiple data objects, for example, the two data sets can be arranged in the same order according to the identifiers. The data, for example, makes each data in the first data set and the second data set correspond to each other, and each data can correspond to a different object. If the data objects of the first data set and the second data set are sorted in the same way, the first data set sent by the first device to the second device may not need to include an identifier. The second device can generate the ciphertext subset of the label information of the corresponding data object in the first data set according to the corresponding relationship between the two data sets and based on the data objects contained in the subset of the second data set.
In another embodiment, if the second device does not match the identifier with the first device before receiving the first data set, that is, the first data set may contain a data object different from the second data set, the second device After receiving the first data set, the first data set and the second data set can be matched with identifiers, thereby eliminating the unique data object of one party, so that the label information of the first data set and the characteristic data of the second data set Correspond to the same multiple data objects. Thereafter, the second device may generate a ciphertext subset of the tag information of the corresponding data object in the first data set according to the data objects contained in the subset of the second data set in step 106.
Taking feature selection as an example, the feature engineering processing of the second data set can include selecting a feature data subset of the second data set, and the second device can generate a first data object corresponding to the data objects contained in the feature data subset. The ciphertext subset of the label information in the data set. As another example, the feature engineering processing of the second data set may include binning the feature data to generate multiple binning segments, so that the second device can generate the first data set and the multiple binning segments. Corresponding multiple ciphertext subsets, wherein each ciphertext subset includes the tag information ciphertext of the data objects in the corresponding binning segment in the first data set. That is, the second device can determine which binning segment the corresponding data object in the first data set falls in according to the binning of the feature data of the second data set, and generate the first data set according to the binning segment. The ciphertext subset of the tag information in. In step 106, the second device can generate a ciphertext subset of the tag information in the first data set, and can know the total number of tag information in the ciphertext subset. However, as described above, the tag information of the first data set is encrypted, and the second device does not know the value of the tag information, and thus cannot complete the index calculation of the feature engineering process. For example, taking binning as an example, the second device does not know which label information in the ciphertext subset is a positive sample and which label information is a negative sample, and thus does not know the number of positive samples and negative label information in each binning segment. The number of samples, the binning index cannot be calculated, such as IV (Information Value), Woe (Weight of Evidence, weight of evidence), etc.
Step 108: The second device sends the ciphertext of the ciphertext subset to the first device, and receives the statistical information generated after decrypting the ciphertext of the ciphertext subset from the first device. The ciphertext of the ciphertext subset sent by the second device to the first device may be the ciphertext set of each tag information in the ciphertext subset, or a certain operation or operation (for example, combination, The ciphertext obtained after cascading, algebraic operations, etc.).
The first device may be the owner of the first data set and can decrypt the ciphertext of the tag information. Thus, the first device can decrypt the ciphertext of the ciphertext subset and generate statistical information (for example, the number of positive samples and/or the number of negative samples of the ciphertext subset), and can send the statistical information to the second equipment. It should be understood that the statistical information sent by the first device is different from the decrypted value of the tag information ciphertext of the first data set, so that the tag information value of the first data set will not be disclosed.
Taking binning as an example, the second device may send multiple ciphertext subsets corresponding to multiple binning segments to the first device, and the ciphertext of each ciphertext subset may include the ciphertext in the first data set. The collection of tag information ciphertexts in the corresponding binning segment. The first device can decrypt the ciphertext of the label information corresponding to each binning segment, and obtain the statistical information of the label information corresponding to each binning segment (for example, the number of positive samples of the label information corresponding to each binning segment and/ Or negative sample number) as the above statistical information sent to the second device. For other feature engineering processing (for example, feature selection, etc.), the first device can similarly decrypt the ciphertext of each ciphertext subset and generate statistical information required for feature engineering processing to send to the second device.
In an example, the ciphertext of the tag information of the first data set may be generated by encrypting the tag information using Homomorphic Encryption. Homomorphic encryption allows a specific algebraic operation to be performed on the ciphertext after homomorphic encryption to obtain the result of the operation that is still encrypted, and the operation result obtained by decrypting it is the same as the operation result obtained by performing the same operation on the plaintext. In other words, this technology allows people to perform operations such as retrieval and comparison in the encrypted data to get the correct results, without the need to decrypt the data during the entire process. If the tag information is a binary value and is encrypted using homomorphic encryption, the second device may perform homomorphic addition on the ciphertext in the ciphertext subset to obtain the ciphertext sum, and send the ciphertext sum to the first device. The value obtained by the first device decrypting the ciphertext sum is equivalent to the value obtained by the first device decrypting and summing each tag information ciphertext in the ciphertext subset. Thus, the first device can decrypt the ciphertext sum, and the decryption result can indicate the number of positive samples and/or the number of negative samples.
Step 110: The second device uses the statistical information to calculate the feature engineering index of the feature engineering process of the second data set. Feature engineering indicators can be used to evaluate the results of feature engineering processing. Feature engineering indicators can be selected according to specific feature engineering processing, such as the positive sample rate of the subset of the second data set, the negative sample rate of the subset of the second data set, the subset of the second data set and the ciphertext subset The relevance of the label information, the information value of the subset of the second data set, the weight of evidence of the subset of the second data set, etc. Thus, the second device can obtain the feature engineering index of the subset according to the subset of the second data set and the corresponding statistical information. Taking binning as an example, the statistical information can be the statistical value of label information for each binning segment. The second device can use the statistical value received from the first device and combine it with the second device when binning the characteristic data. The obtained binning segment information is calculated together with binning index. If the feature engineering index meets the requirement (for example, the threshold), the second device can generate a feature engineering processing result and complete the feature engineering processing. The generated feature engineering processing results (for example, a subset of feature data) can be used for various purposes, such as model training, target prediction, and so on. In an alternative embodiment, the second device may send the feature engineering result to the first device. In a further embodiment, if the feature engineering index is not ideal (for example, the threshold value is not reached), the second device may repeat steps 104-110 until the required feature engineering index is obtained or the condition for terminating the feature engineering process is reached.
According to the above method, the partners do not need to disclose the plaintext of their private data, but they can finally complete the feature engineering processing. The following describes the application scenarios of the present disclosure in further detail with reference to FIGS. 2-4.
Fig. 2 is a schematic diagram of a feature engineering method according to an embodiment of the present disclosure. As an example and not a limitation, a feature selection scenario in which the first device has tag information Y (first data set) and the second device has feature data X (second data set) is taken as an example for description. To simplify the description, it is assumed that the tag information Y (first data set) and the characteristic data X (second data set) correspond to the same multiple data objects (for example, by performing identifier matching). The goal of feature selection is to find the optimal feature subset. Feature selection can eliminate irrelevant or redundant features, so as to reduce the number of features, reduce the dimensionality of the data set, improve the accuracy of the model, and reduce the running time. On the other hand, feature selection can be used to select a simplified model of truly relevant features to help understand the process of data generation.
In step 202, the first device may encrypt the Y samples in the first data set and generate encrypted Y'samples. In step 204, the first device may send the encrypted Y'sample to the second device. As an example and not a limitation, the first device may further compress the encrypted Y'sample before sending, and send the compressed encrypted Y'sample to the second device. If the amount of Y’ sample data is large, consider uploading the ciphertext file to a secure cloud (such as oss, etc.) accessible by both parties.
In an embodiment, the first device may encrypt the Y samples in the first data set one by one to generate the ciphertext Y'of the Y samples. The first device can use the key to encrypt the Y samples in the first data set one by one with any suitable encryption algorithm, and send the Y'samples (for example, uncompressed or compressed) to the second device. As an example and not a limitation, homomorphic encryption (Homomorphic Encryption) can be used to perform full homomorphic encryption on the Y samples in the first data set one by one.
In step 206, the second device may receive the encrypted Y'sample. For example, the second device may obtain the encrypted Y'sample from the first device or other intermediate parties (for example, cloud such as oss). If the Y'sample is compressed, the second device can decompress the Y'sample.
In step 208, the second device may select a subset of X samples (for example, characteristic data) owned locally. It should be understood that the second device may select the X sample subset according to various criteria or combinations. In addition, the second device may select one or more subsets of X samples and evaluate the one or more subsets of X samples sequentially or in parallel.
Although FIG. 2 shows that step 208 is performed after step 206, it should be understood that step 208 can be performed before or after any of steps 202-206, or concurrently with any of steps 202-206.地Execute.
In step 210, the second device may read the corresponding Y'sample subset according to the selected X sample subset. Similar to the above, the id of the X sample and the id of the Y sample can be matched, so that the X sample and the Y sample have a corresponding relationship. Thus, the second device can read the Y'sample subset of the corresponding id according to the selected X sample subset. As described above, since the Y'samples are encrypted, the second device does not know the value of the Y'samples, so the Y'sample subset may include encrypted Y'samples (including ciphertext).
In step 212, the second device may send the ciphertext of the Y'sample subset to the first device for decryption. In an example, the ciphertext of the Y'sample subset sent in step 212 may include each set of Y'samples included in the Y'sample subset.
In step 214, the first device decrypts the ciphertext of the Y'sample subset. For example, the first device may decrypt each Y'sample included in the subset to obtain a decrypted Y sample.
In step 216, the first device may generate statistical information about feature selection after decrypting the ciphertext of the Y'sample subset. For example, the first device can decrypt the value of each Y'sample, and thereby obtain statistical information (such as the number of positive samples, the number of negative samples, etc.) for evaluating the feature selection subset.
In another example, if the Y samples are binary values and homomorphic encryption is used to generate Y'samples, the second device can perform homomorphic addition on each Y'sample of the Y'sample subset to obtain the ciphertext sum, And in step 212, the ciphertext sum is sent to the first device, instead of sending the Y′ sample subset to the first device in its entirety. The first device can decrypt the ciphertext sum in step 214 and directly generate statistical information. For example, if 1 and 0 are used to represent positive samples and negative samples, the sum of homomorphic ciphertexts of the Y'sample subset is the sum of ciphertexts of all 1s, and the sum of the ciphertexts is the number of all positive samples after being decrypted . Conversely, if 1 and 0 are used to represent negative samples and positive samples, the sum of homomorphic ciphertexts of the Y'sample subset can be decrypted to obtain the number of negative samples.
In step 218, the first device may send the statistical information of the Y'sample subset to the second device. It should be understood that the statistical information of the sample subset of Y'sent by the first device will not reveal the value of sample Y.
In step 220, the second device may calculate feature selection indicators based on the statistical information, such as the positive sample rate of the X sample subset, the negative sample rate of the X sample subset, the X sample subset and the prediction target (for example, the Y sample is Positive or negative) correlation degree, information value of X sample subset, evidence weight of X sample subset, etc. As an example and not a limitation, the second device may return from step 220 to step 208 to select one or more subsets of X samples of the next group for evaluation.
In step 222, in the case that a qualified (for example, optimal) feature subset is obtained, the second device may generate a feature selection result and complete the feature selection.
In optional step 224, the second device may send the feature selection result to the first device. Accordingly, in optional step 226, the first device may receive the feature selection result from the second device.
Fig. 3 is a schematic diagram of a feature engineering method according to another embodiment of the present disclosure. As an example and not a limitation, take a binning scenario where the first device (for example, the requester) has label information Y (first data set), and the second device (for example, the data party) has characteristic data X (second data set) Take it as an example. To simplify the description, it is assumed that the label information Y (first data set) and the characteristic data X (second data set) correspond to the same multiple data objects.
Binning is a commonly used method of data preprocessing, which is divided into sub-intervals according to attribute values. If an attribute value is within a certain sub-interval range, it is said that the data to be processed (a column of attribute values) should be processed according to certain rules. Put it into some boxes, examine the data in each box, and use a certain method to process the data in each box separately. The binning can include equal depth binning method, equal width binning method, minimum entropy method and user-defined interval method.
Steps 302-306 in FIG. 3 are similar to steps 202-206 in FIG. 2 and will not be described in detail.
In step 308, the second device may bin the locally owned X samples (for example, characteristic data). Although step 308 is shown in FIG. 3 to be performed after step 306, it should be understood that the binning of step 308 can be performed before or after any of steps 302-306, or it can be performed with any of steps 302-306. One step is executed concurrently. For example, the binning of X samples at step 308 may be performed in advance (for example, before step 302). In other embodiments, the binning of the X samples may be performed after step 306 (or after receiving a binning request from the first device).
In step 310, the second device reads the corresponding subset of Y'samples in units of binning segments of X samples.
According to an embodiment of the present disclosure, the id of the X sample and the id of the Y sample can be matched at any time before step 308, so that the X sample and the Y sample have a corresponding relationship. As an example and not a limitation, the encrypted Y'sample sent by the first device to the second device has an associated identifier (id), and the local X sample of the second device has an associated identifier. The identifiers are matched (for example, corresponding to the same set of objects). Therefore, the second device can determine which binning segment the corresponding id falls in according to the binning of the X samples, and read the Y'samples corresponding to each id according to the binning segment. In one embodiment, the identifier of the encrypted Y'sample sent by the first device to the second device may be encrypted (e.g., hashed) or unencrypted. In another embodiment, if the first device and the second device are matched first, so that the X samples and Y'samples are sorted in a specific manner (for example, each item of data corresponds to one by one), then the first device sends the first device to the first device. The encrypted Y'sample of the second device does not need to include id information, and the second device can directly read the corresponding Y'sample according to the binning of the X sample.
In step 312, the second device may send the ciphertext of the Y'sample subset corresponding to each binning segment to the first device. In one embodiment, the ciphertext of a plurality of Y'sample subsets may be sent in step 312, wherein the ciphertext of each Y'sample subset may include a set of Y'samples corresponding to the corresponding binning segment. In another embodiment, if the Y'samples are encrypted using homomorphic encryption, the ciphertext of each Y'sample subset sent in step 312 may include performing the same for the Y'samples corresponding to the binning segment. The sum of ciphertexts obtained by state addition.
In step 314, the first device decrypts the ciphertext of the Y'sample subset corresponding to each bin segment. The first device may be a device that has encrypted Y and can decrypt Y', or even if the encryption is not performed by the first device, the first device has key authority and can thus decrypt Y'. For example, the first device may decrypt each Y'sample included in the Y'sample subset, or decrypt the homomorphic ciphertext sum if the subset contains the homomorphic ciphertext sum. If the Y'sample is a binary value and is encrypted by homomorphic encryption, the number of positive samples or the number of negative samples can be directly obtained by decrypting the sum of the ciphertext.
In step 316, the first device can generate statistical information about each binning segment after decrypting the ciphertext of the Y'sample subset corresponding to each binning segment. For example, the first device can decrypt the value of each Y'sample of each bin segment, and thereby obtain parameters such as the number of positive samples, the number of negative samples, and/or the total number of samples of each bin segment. In another embodiment, when the ciphertext of the Y'sample subset sent at step 312 contains the sum of homomorphic ciphertext, decrypting the sum of homomorphic ciphertext provides the number of positive samples or the number of negative samples of each bin segment. . For example, if 1 and 0 are used to represent positive samples and negative samples, the sum of homomorphic ciphertexts is the number of all positive samples.
In step 318, the first device may send statistical information of each bin segment (for example, Y sample statistical information, such as the number of positive samples, the number of negative samples, and/or the total number of samples, etc.) to the second device. It should be understood that the statistical information of each bin segment sent by the first device will not reveal the value of the sample Y.
In step 320, the second device may calculate a binning index based on the statistical information. For example, the second device can use the statistical value of each binning segment received from the first device and combine with other information obtained by the second device when binning the characteristic data X to calculate binning indicators, such as each X The positive sample rate of the sample binning segment, the negative sample rate of each X sample binning segment, the correlation between each X sample binning segment and the prediction target (for example, the Y sample is positive or negative), the X sample binning segment IV (Information Value, information value), Woe (Weight of Evidence, weight of evidence) of each X sample bin segment, etc.
In step 322, in the case that the binning indicator is valid (for example, reaching a threshold), the second device may generate a binning result and complete binning. In the case that the binning index is not ideal, the second device may repeat steps 308-320 to re-execute the binning and calculate the binning index.
In optional step 324, the second device may send the binning result to the first device. Accordingly, in optional step 326, the first device may receive the binning result from the second device.
Fig. 4 is a schematic diagram of a feature engineering method according to another embodiment of the present disclosure. In one embodiment, one party (for example, the first device) may have characteristic data X2 and tag information Y, and the other party (for example, the second device) may have characteristic data X1. The feature data X1 and the feature data X2 may be the same, partially the same, or different from each other. To simplify the description, it is assumed that the label information Y (first data set), the characteristic data X1 (the second data set), and the characteristic data X2 (the third data set) correspond to the same multiple data objects. The first device may send the encrypted label information Y'to the second device as shown in FIG. 3 to combine the characteristic data X1 for binning, and may also perform local binning for the characteristic data X2 and the label information Y locally.
In steps 402-422, the first device and the second device may perform binning according to the characteristic data X1 and the tag information Y'as described with reference to steps 302-322 in FIG. 3 and generate the first binning result.
In step 424, the second device may send the first binning result to the first device.
In step 426, the first device may perform binning on the characteristic data X2 and the unencrypted tag information Y locally to obtain a second binning result. It should be noted that the local binning performed by the first device in step 426 may be performed before or after any of steps 402-424, or performed in parallel with any of the steps.
In step 428, the first device may combine the first binning result and the second binning result to obtain a combined binning result.
In the above, the feature selection and binning in feature engineering processing are used as examples to introduce distributed feature engineering and index calculation, so that the partners do not have to leak their private data plaintext, but the feature engineering processing can be completed in the end. Those skilled in the art can apply the distributed feature engineering method and device of the present disclosure to a wide range of situations where modeling sample data of all parties needs to be fused together for processing, such as data cleaning and feature discreteness. Transformation, feature extraction, feature selection, etc.
Fig. 5 shows a block diagram of a feature engineering device according to an embodiment of the present disclosure. For example, the first device 510 may perform the above-described operations on the first device, and the second device 520 may perform the above-described operations on the second device.
As shown in FIG. 5, the first device 510 may include an encryption and decryption module 512, a transmission module 514, and a processing module 516. The encryption and decryption module 512 can encrypt the tag information of the first data set to generate a ciphertext. The transmission module 514 can transmit the first data set containing the tag information ciphertext to the second device 520 or other devices or the cloud.
The second device 520 may include a transmission module 522, a feature engineering processing module 524, and a ciphertext subset generation module 526. The transmission module 522 can receive the first data set containing the tag information ciphertext. In addition, the second device 520 may have a second data set (for example, stored in a memory), and the second data set includes characteristic data of a plurality of data objects. The feature engineering processing module 524 can perform feature engineering processing on the second data set to generate a subset of the second data set. The label information of the first data set and the feature data of the second data set may correspond to the same multiple data objects. The ciphertext subset generating module 526 can generate the ciphertext subset of the tag information of the corresponding data object in the first data set according to the data objects contained in the subset of the second data set. The transmission module 522 may further send the ciphertext of the ciphertext subset to the first device.
In the first device 510, the transmission module 514 can receive the ciphertext of the ciphertext subset and pass it to the encryption and decryption module 512, and the encryption and decryption module 512 can decrypt the ciphertext of the ciphertext subset. The processing module 516 can generate statistical information for feature engineering processing, such as the number of positive samples, the number of negative samples, and the total number of samples, based on the decrypted data of the ciphertext of the ciphertext subset. The transmission module 514 can provide the statistical information to the second device 520.
In the second device 520, the transmission module 522 receives the statistical information and provides it to the feature engineering processing module 524. The feature engineering processing module 524 can use the statistical information to calculate the feature engineering index of the feature engineering process performed on the second data set and complete the feature engineering process. Optionally, after the feature engineering process is completed, the transmission module 522 may also send the feature engineering process result to the first device 510.
In a further embodiment, the first device 510 may optionally include a feature engineering processing module 518, which may use the unencrypted tag information of the first data set and the third data set (for example, including the feature data X2) to Perform feature engineering processing to generate a second feature engineering result. The label information of the first data set and the characteristic data of the third data set may correspond to the same multiple data objects. The first device 510 may also combine the above-mentioned feature engineering result generated by the second device 520 and the second feature engineering result to obtain a combined feature engineering result. In one embodiment, the processing module 516 and the feature engineering processing module 518 can be implemented together or separately.
According to the present disclosure, the data cooperation parties of the joint modeling (for example, the first device 510 and the second device 520) do not output the plaintext of private data, but finally obtain the result of the feature engineering processing, which prevents the leakage of private data and solves the problem of data cooperation. The issue of privacy information leakage and information trust in the process.
The various steps and modules of the feature engineering method and device described above can be implemented by hardware, software, or a combination thereof. If implemented in hardware, in combination with the various illustrative steps, modules, and circuits described in the present disclosure, general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASICs), field programmable gate arrays ( FPGA), or other programmable logic components, hardware components, or any combination thereof. A general-purpose processor may be a processor, microprocessor, controller, microcontroller, or state machine, etc. If implemented in software, the various illustrative steps and modules described in conjunction with the present disclosure can be stored on a computer-readable medium or transmitted as one or more instructions or codes. The software modules that implement the various operations of the present disclosure can reside in storage media, such as RAM, flash memory, ROM, EPROM, EEPROM, scratchpad, hard disk, removable disk, CD-ROM, cloud storage, etc. . The storage medium can be coupled to the processor so that the processor can read and write information from/to the storage medium, and execute corresponding program modules to implement each step of the present disclosure. Moreover, software-based embodiments can be uploaded, downloaded, or accessed remotely through appropriate communication means. Such appropriate communication means include, for example, the Internet, World Wide Web, Intranet, software applications, cables (including fiber optic cables), magnetic communications, electromagnetic communications (including RF, microwave, and infrared communications), electronic communications, or other such communications. Means of communication.
It should also be noted that these embodiments may be described as processes depicted as flowcharts, flow diagrams, structure diagrams, or block diagrams. Although the flowchart may describe the operations as sequential processes, many of these operations can be performed in parallel or concurrently. In addition, the order of these operations can be rearranged.
The disclosed methods, devices and systems should not be restricted in any way. On the contrary, the present disclosure covers all novel and non-obvious features and aspects of the various disclosed embodiments (individually and in various combinations and subcombinations with each other). The disclosed methods, devices and systems are not limited to any specific aspects or features or their combinations, and any disclosed embodiment does not require any one or more specific advantages or to solve specific or all technical problems.
The embodiments of the present disclosure are described above in conjunction with the drawings, but the present disclosure is not limited to the above-mentioned specific embodiments. The above-mentioned specific embodiments are only illustrative and not restrictive. Those of ordinary skill in the art are Under the enlightenment of this disclosure, many changes can be made without departing from the scope of protection of the purpose and claims of this disclosure, and these all fall within the scope of protection of this disclosure.
