WO2021114922A1 - 多方联合训练针对IoT机具的风险评估模型的方法及装置 - Google Patents

多方联合训练针对IoT机具的风险评估模型的方法及装置 Download PDF

Info

Publication number
WO2021114922A1
WO2021114922A1 PCT/CN2020/124289 CN2020124289W WO2021114922A1 WO 2021114922 A1 WO2021114922 A1 WO 2021114922A1 CN 2020124289 W CN2020124289 W CN 2020124289W WO 2021114922 A1 WO2021114922 A1 WO 2021114922A1
Authority
WO
WIPO (PCT)
Prior art keywords
computing node
merchant
machine
iot
intermediate result
Prior art date
Application number
PCT/CN2020/124289
Other languages
English (en)
French (fr)
Inventor
郑霖
陆梦倩
傅欣艺
汲小溪
王维强
Original Assignee
支付宝(杭州)信息技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 支付宝(杭州)信息技术有限公司 filed Critical 支付宝(杭州)信息技术有限公司
Publication of WO2021114922A1 publication Critical patent/WO2021114922A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q20/00Payment architectures, schemes or protocols
    • G06Q20/38Payment protocols; Details thereof
    • G06Q20/40Authorisation, e.g. identification of payer or payee, verification of customer or shop credentials; Review and approval of payers, e.g. check credit lines or negative lists
    • G06Q20/401Transaction verification
    • G06Q20/4016Transaction verification involving fraud or risk level assessment in transaction processing
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q20/00Payment architectures, schemes or protocols
    • G06Q20/38Payment protocols; Details thereof
    • G06Q20/382Payment protocols; Details thereof insuring higher security of transaction
    • G06Q20/3823Payment protocols; Details thereof insuring higher security of transaction combining multiple encryption tools for a transaction

Definitions

  • One or more embodiments of this specification relate to the field of data security technology, and in particular to a method and device for multi-party joint training of a risk assessment model for IoT equipment.
  • IoT equipment is a smart terminal used for merchants to collect payments, which can facilitate merchants to collect payments, and at the same time can bring users a convenient and fast payment experience.
  • IoT devices can support facial payment, so that users do not need to use their own mobile terminals (such as smart phones or wearable devices) to make payments, thereby simplifying the payment process.
  • One or more embodiments of this specification describe a method for multi-party joint training of a risk assessment model for IoT equipment, which can ensure the security of multi-party data while making full use of effective data, training to obtain a risk assessment model with excellent performance, and then Through the use of this risk assessment model, a comprehensive and accurate risk assessment of IoT machines can be carried out.
  • a method for multi-party joint training of a risk assessment model for IoT equipment wherein the multiple parties include a first computing node, a second computing node, and a first payment platform, and each maintains some parameters in the risk assessment model;
  • the first computing node is associated with a first IoT machine, storing machine features determined based on the machine privacy data of the first IoT machine, and the second computing node is associated with a first merchant that binds the first IoT machine.
  • the method includes: determining a first intermediate result based on the characteristics of the machine tool and the first parameter maintained by the first computing node; using secure multi-party computing MPC technology to provide the first intermediate result for combining the second
  • the second intermediate result determined by the computing node based on the merchant s privacy data and the second parameter maintained by the merchant, and the third parameter determined by the first payment platform based on the payment privacy data and the third parameter maintained by the risk tag.
  • the training loss for the first IoT machine is determined; the training loss is obtained, and the training loss is used to adjust the first parameter.
  • the method before determining the first intermediate result based on the characteristics of the machine tool and the first parameter maintained by the first computing node and the first calculation formula, the method further includes: acquiring the first IoT machine tool The device privacy data stored in the device, and the device privacy data are accumulated or vectorized to obtain the device characteristics; or, the device characteristics are received from the first IoT device, and the device characteristics are It is obtained by the first IoT machine performing cumulative processing or vector characterization processing on the machine privacy data stored by the first IoT machine.
  • the MPC technology includes a homomorphic encryption technology
  • the first computing node and the second computing node also store a first public key generated based on the homomorphic encryption technology
  • the first payment platform The first public key and the corresponding first private key are also stored in the file; wherein using the secure multi-party computing MPC technology to provide the first intermediate result includes: using the first public key to perform the first intermediate result on the first intermediate result Encrypt to obtain a first encryption result; send the first encryption result to the second computing node, so that the second computing node encrypts the first encryption result and the second intermediate result obtained by encrypting the second intermediate result
  • the second encryption result performs the first homomorphic addition operation, so that the first payment platform performs the first operation result on the first operation result obtained by the first homomorphic addition operation and the third encryption result obtained by encrypting the third intermediate result.
  • Another method for multi-party joint training of a risk assessment model for IoT equipment where multiple parties include a first computing node, a second computing node, and a first payment platform, and each maintains some parameters in the risk assessment model;
  • the first computing node is associated with a first IoT machine, and stores machine features related to the first IoT machine
  • the second computing node is associated with a first merchant that binds the first IoT machine, and stores Merchant privacy data of the first merchant
  • the first payment platform stores payment privacy data related to the first merchant and a risk tag indicating the risk situation of the first IoT machine; the method is applied to the first 2.
  • the first computing node is associated with a first IoT machine, and stores machine features related to the first IoT machine
  • the second computing node is associated with a first merchant that binds the first IoT machine, and stores Merchant privacy data of the first merchant
  • the first payment platform stores payment privacy data related to the first merchant and a risk tag indicating the risk situation of the first I
  • the method includes: determining a second intermediate result based on the privacy data of the merchant and the second parameter maintained by the second computing node; using the secure multi-party computing MPC technology to provide the second intermediate result for combining the second parameter
  • a computing node determines a first intermediate result based on the characteristics of the machine tool and the first parameter maintained, and the first payment platform determines the third parameter based on the payment privacy data and the third parameter maintained by the risk tag.
  • the training loss for the first IoT machine is determined; the training loss is obtained, and the second parameter is adjusted by using the training loss.
  • the second computing node is a trusted computing node.
  • the method further includes : Generate a second public key and a second private key, and send the second public key to multiple merchants, including the first merchant; receive encrypted privacy data from the first merchant, The encrypted private data is obtained by the first merchant using the second public key to encrypt the merchant private data; the second private key is used to decrypt the encrypted private data to obtain the merchant privacy data.
  • the first computing node determines the first intermediate result based on the features of the machine and the first parameter maintained, and the second computing node determines the second intermediate result determined based on the merchant’s private data and the second parameter maintained Regarding the training loss of the first IoT machine; acquiring the training loss, and adjusting the third parameter by using the training loss.
  • a device for multi-party joint training of a risk assessment model for IoT equipment wherein the multiple parties include a first computing node, a second computing node, and a first payment platform, and each maintains some parameters in the risk assessment model;
  • the first computing node is associated with a first IoT machine, storing machine features determined based on the machine privacy data of the first IoT machine, and the second computing node is associated with a first merchant that binds the first IoT machine.
  • merchant privacy data of the first merchant is stored, and the first payment platform stores payment privacy data related to the first merchant and a risk tag indicating the risk situation of the first IoT machine.
  • the device is integrated in the first computing node, and the device includes: an intermediate result determining unit configured to determine a first intermediate result based on the characteristics of the implement and the first parameter maintained by the first computing node; an intermediate result providing unit , Configured to use the secure multi-party computing MPC technology to provide the first intermediate result, which is used in conjunction with the second intermediate result determined by the second computing node based on the merchant’s private data and the second parameter maintained by it, the first A payment platform determines the training loss for the first IoT device based on the payment privacy data and the third parameter maintained by the third parameter and the third intermediate result determined by the risk tag; a loss acquisition unit configured to acquire the training Loss; a parameter adjustment unit configured to adjust the first parameter by using the training loss.
  • another device for multi-party joint training of a risk assessment model for IoT equipment where the multiple parties include a first computing node, a second computing node, and a first payment platform, and each maintains some parameters in the risk assessment model;
  • the first computing node is associated with a first IoT machine, and stores machine features related to the first IoT machine
  • the second computing node is associated with a first merchant that binds the first IoT machine, and stores Merchant privacy data of the first merchant
  • the first payment platform stores payment privacy data related to the first merchant and a risk tag indicating the risk of the first IoT machine
  • the device is integrated in the first 2.
  • the device includes: an intermediate result determining unit configured to determine a second intermediate result based on the merchant’s private data and a second parameter maintained by the second computing node; the intermediate result providing unit is configured to use secure multi-party computing
  • the MPC technology provides the second intermediate result, which is used in conjunction with the first intermediate result determined by the first computing node based on the machine characteristics and the first parameters maintained by the first computing node, and the first payment platform is based on the payment privacy
  • the third parameter of the data and its maintenance and the third intermediate result determined by the risk tag determine the training loss for the first IoT machine; the loss acquisition unit is configured to acquire the training loss; the parameter adjustment unit is configured to Use the training loss to adjust the second parameter.
  • the multiple parties include a first computing node, a second computing node, and a first payment platform, and each maintains some parameters in the risk assessment model;
  • the first computing node is associated with a first IoT machine, and stores machine features determined based on the machine privacy data of the first IoT machine
  • the second computing node is associated with the first IoT machine that is bound to the first IoT machine.
  • Merchants are associated and store merchant privacy data of the first merchant
  • the first payment platform stores payment privacy data related to the first merchant and a risk tag indicating the risk situation of the first IoT machine.
  • the device is integrated in the first payment platform, and the device includes: an intermediate result determining unit configured to determine a third intermediate result based on the payment privacy data, a third parameter maintained by the first payment platform, and the risk label Result; an intermediate result providing unit configured to use the secure multi-party computing MPC technology to provide the third intermediate result for combining the first intermediate result determined by the first computing node based on the machine feature and its maintenance first parameter As a result, the second computing node determines the training loss for the first IoT machine based on the second intermediate result determined by the merchant’s private data and the second parameter maintained; the loss obtaining unit is configured to obtain the training Loss; a parameter tuning unit configured to use the training loss to adjust the third parameter.
  • an intermediate result determining unit configured to determine a third intermediate result based on the payment privacy data, a third parameter maintained by the first payment platform, and the risk label Result
  • an intermediate result providing unit configured to use the secure multi-party computing MPC technology to provide the third intermediate result for combining the first intermediate result determined by the first
  • a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed in a computer, the computer is caused to execute the method of the first aspect or the second aspect or the third aspect.
  • a computing device including a memory and a processor, the memory stores executable code, and when the processor executes the executable code, the first aspect or the second aspect or the first aspect is implemented. Three-dimensional approach.
  • using the method and device provided in the embodiments of this specification can ensure the security of multiple parties while achieving full and comprehensive utilization of valid data from all parties, training to obtain a risk assessment model with excellent performance, and then by using the risk assessment model, Conduct a comprehensive and accurate risk assessment of IoT equipment.
  • Fig. 1 shows a framework diagram of multi-party joint risk control of IoT equipment according to an embodiment.
  • Fig. 2 shows a schematic diagram of a construction process of a trusted computing node according to an embodiment.
  • Fig. 3 shows a framework diagram of a multi-party joint training risk assessment model according to an embodiment.
  • Fig. 4 shows a multi-party interaction diagram based on homomorphic encryption according to an embodiment.
  • Fig. 5 shows a schematic structural diagram of a multi-party joint training risk assessment model according to an embodiment.
  • Fig. 6 shows a structure diagram of an apparatus for multi-party joint training of a risk assessment model for IoT equipment according to an embodiment.
  • Fig. 7 shows a structure diagram of an apparatus for multi-party joint training of a risk assessment model for IoT equipment according to another embodiment.
  • FIG. 8 shows a structure diagram of an apparatus for multi-party joint training of a risk assessment model for IoT equipment according to another embodiment.
  • the merchant can be risked based on merchant data (such as merchant business license, business status, etc.) collected in a payment platform (such as Alipay) Identify and control, so as to realize the risk control of the IoT equipment used by the merchant.
  • merchant data such as merchant business license, business status, etc.
  • a payment platform such as Alipay
  • the risk assessment of a single transaction in the IoT machine can be carried out to realize the risk control of the IoT machine.
  • the transaction information (including buyer ID, transaction time, location, and amount, etc.) of the transaction can be obtained from the certain payment platform, and the transaction information can be obtained from the certain payment platform.
  • buyer information (including historical transaction records, etc.) of transaction buyers in a payment platform to realize the risk identification and control of the transaction.
  • this kind of user-based risk control solution is to identify risks from the perspective of a single user and a single transaction, and the obtained recognition results are directly applied to IoT devices, which will have the problems of low recognition accuracy and false interruptions.
  • the inventor also found that neither of the above two solutions uses special data of the IoT scene, such as the startup time of the IoT machine, the frequency of unbinding and rebinding of the merchant account, and the location information of the IoT machine.
  • FIG. 1 shows a framework diagram of multi-party joint risk control of IoT equipment according to an embodiment, as shown in FIG. 1, in which the equipment produced by the IoT equipment end (the cash register in FIG. 1) is integrated Privacy data, merchant privacy data of merchants that are bound to IoT devices, payment privacy data related to the merchant in the payment platform (including transaction information, buyer information, and merchant information), using MPC (Secure Multi-Party Computation), safe and multi-party Computing) technology to achieve a comprehensive and accurate risk assessment of IoT machines.
  • the equipment produced by the IoT equipment end is integrated Privacy data, merchant privacy data of merchants that are bound to IoT devices, payment privacy data related to the merchant in the payment platform (including transaction information, buyer information, and merchant information), using MPC (Secure Multi-Party Computation), safe and multi-party Computing) technology to achieve a comprehensive and accurate risk assessment of IoT machines.
  • MPC Secure Multi-Party Computation
  • Fusion of multi-party data for risk assessment is implemented based on a risk assessment model.
  • the embodiment of this specification discloses a method for multi-party joint training of a risk assessment model.
  • the multiple parties that perform joint training will be introduced first, and then the implementation process of the training method will be introduced.
  • the aforementioned parties include a first computing node, a second computing node, and a first payment platform.
  • the first thing to note is that the "first”, “second” and similar terms elsewhere in the text are only used to distinguish similar matters and do not have other restrictive effects.
  • the aforementioned first computing node is associated with the first IoT machine, and stores machine features determined based on the machine privacy data of the first IoT machine. It can be understood that there are actually multiple IoT devices, and the first IoT device may be any one of the multiple IoT devices.
  • there is a one-to-one correspondence between computing nodes and IoT devices that is, a computing node corresponding to each IoT device is constructed.
  • the first computing node is integrated with the first IoT machine.
  • the first IoT machine can be directly used as the first computing node.
  • the first computing node may obtain and process the stored device privacy data from the first IoT device to determine the characteristics of the device.
  • multiple IoT devices may share one computing node, which means that the first computing node is associated with multiple IoT devices. At this time, due to data security compliance and other requirements, the privacy data in the IoT machine is not allowed to be leaked.
  • the IoT machine can perform feature aggregation processing on the privacy data of the machine, and send the aggregated machine features to the first
  • the computing node needs to understand that it is usually difficult to restore the original private data based on the characteristics of the machine, so that the leakage of the private data of the machine can be effectively prevented.
  • the device privacy data may include: power-on time, power-off time, and location information of the first IoT device, and operation data generated by unbinding or rebinding the first IoT device.
  • the power-on time and the power-off time may include multiple moments in the historical time period. For example, the power-on time includes 6:00 AM on Monday, 7:00 AM on Tuesday, and so on.
  • the location information may include location information collected using LBS (Location Based Services), such as latitude and longitude information.
  • the operation data generated by unbinding and rebinding may include the operation time of unbinding and rebinding, operation frequency, and the number of merchants (or merchant accounts) involved.
  • the above-mentioned tool feature may include a cumulative feature or a feature vector.
  • the feature aggregation processing for the privacy data of the equipment may include accumulation processing, vector characterization processing, and the like.
  • the accumulation processing may include determining the accumulation feature corresponding to the privacy data of the machine as the machine feature.
  • the cumulative feature may include the average number of times the IoT device is switched on and off per day.
  • the vector characterization processing may include using a characterization learning algorithm, such as a neural network, to calculate a feature vector corresponding to the privacy data of the machine as the machine feature.
  • the above-mentioned second computing node is associated with the first merchant (or the merchant account of the first merchant) bound to the first IoT machine, and stores the merchant's private data of the first merchant.
  • the multiple transactions that the first merchant participates in usually involve multiple payment platforms, such as the current mainstream Alipay payment platform and WeChat payment platform, etc., and for the first merchant’s other payment platforms other than the first payment platform.
  • Transaction data is usually not available to the first payment platform.
  • the first merchant can provide it.
  • the merchant's privacy data may include: transaction information generated by the first merchant on payment platforms other than the first payment platform, which may specifically include: transaction amount, transaction location, transaction time, product type, and risk events.
  • the second computing node may be a node built by the first merchant. In this way, the second computing node may directly obtain the private data of the merchant in the first merchant for model training.
  • TEE Trusted Execution Environment
  • the inventor proposes that TEE (Trusted Execution Environment) can be introduced to help multiple merchants (including the first merchant) to encrypt their respective merchant privacy data to the TEE environment while ensuring the security of their own data. Realize the construction of a trusted computing node (as the second computing node).
  • FIG. 2 shows a schematic diagram of a construction process of a trusted computing node according to an embodiment.
  • the trusted computing node is a trusted computing circle Enclave implemented using SGX technology. Specifically, by providing a series of CPU instruction codes, allowing user codes to create private memory areas with high access rights to form calculations Encircle the Enclave. No merchant can access the data in the surrounding Enclave. Therefore, the private data stored in the Enclave cannot be stolen or tampered with.
  • step S21 the merchant requests a software operation report from the trusted computing circle (Enclave).
  • the Intel CPU in the Enclave will generate a public key, private key, and algorithm code signature based on the C++ code of the algorithm used to calculate the merchant's private data.
  • step S22 Enclave returns the generated public key and algorithm code signature to the merchant as the content in the software operation report, and the private key is stored in the Intel CPU to ensure that no merchant can use the public key to encrypt other merchants. Decrypt the private data of merchants.
  • step S23 the merchant sends the received software operation report to the authentication interface of Intel Corporation for a third-party authentication.
  • step S24 Intel can inform the merchant that the software operation report is credible, which means that the public key and algorithm code signature included in it are indeed generated by Intel CPU, and there is no tampering in the middle, which is credible. .
  • step S25 after receiving the result of successful authentication, the merchant uses the public key to encrypt its own merchant’s private data, and in step S26, the encrypted data is sent to the Enclave, so that the Enclave can use the private key pair
  • the encrypted data is decrypted, and the original merchant privacy data is obtained, which is used for model training. In this way, each merchant can provide their private data for model training while ensuring the security of their own data.
  • the above mainly introduces the relationship between the second computing node and the first merchant, and the merchant's private data.
  • the aforementioned first payment platform stores payment privacy data related to the first merchant and a risk tag indicating the risk situation of the first IoT machine.
  • the payment privacy data may include contract information between the first merchant and the first payment platform, transaction information generated in the first payment platform, and the transaction information specifically includes one or more of the following: Species: user information, transaction amount, transaction location, transaction time, commodity type, and risk events of the transaction user.
  • the contract information may include information such as the business license of the first merchant, the time of the contract, and the length of the contract.
  • the user information may include basic attribute information, transaction preferences, and historical transaction records of the user.
  • the basic attribute information may include gender, age, occupation, permanent residence, hobbies and so on.
  • the transaction preferences may include the most frequently purchased commodity types (such as electronic commodities), and the most frequently purchased time period (such as 21:00-22:00 in the evening).
  • the risk events may include high-risk events that have occurred in the first merchant, such as the sale of illegal products (such as gambling products).
  • the above-mentioned risk label may include risky and non-risky.
  • the aforementioned risk label may also include multiple risk levels. In a specific embodiment, it may include high risk, medium risk, and low risk.
  • the payment privacy data and risk tags stored in the first payment platform are mainly introduced.
  • the first computing node, the second computing node, and the first payment platform also maintain part of the parameters in the risk assessment model.
  • some of the parameters maintained by the three parties are different from each other.
  • part of the parameters maintained by each party is associated with the sample characteristics corresponding to the data provided by the party.
  • which part of the parameters in the risk assessment model each party specifically maintains can be determined by MPC technology.
  • the risk assessment model can be implemented by using a logistic regression algorithm, a decision tree algorithm, a neural network, and the like.
  • FIG. 3 shows a framework diagram of a multi-party joint training risk assessment model according to an embodiment.
  • the first computing node, the second computing node, and the first payment platform each use their own stored private data and model parameters to perform calculations to obtain their respective intermediate calculation results, and then use MPC technology Provide respective intermediate calculation results, realize the fusion and sharing of data, and complete the training of the risk assessment model.
  • the following first introduces the process by which each party calculates the intermediate results, and then the process by which each party uses MPC technology to perform data fusion and sharing, and adjusts the model parameters maintained by each party.
  • the first computing node determines the first intermediate result based on the machine characteristics and the first parameters maintained.
  • the first parameter is used to calculate the tool feature, and the first intermediate result can be obtained.
  • ⁇ 1 can be used to represent the first parameter
  • z 1 can be used to represent the machine feature determined based on the machine privacy data x 1 , so that the first intermediate result can be determined
  • each of the multiple parties also maintains a partial calculation formula for the loss function of the risk assessment model.
  • the first calculation node can use the first calculation formula and the first parameter maintained by the first calculation node to calculate the machine tool feature to obtain the first intermediate result.
  • ⁇ 1 can be used to represent the first parameter
  • z 1 can be used to represent machine characteristics
  • l 1 () can be used to represent the first calculation formula. From this, the first intermediate result l 1 (z 1 ; ⁇ can be determined. 1 ).
  • the training method may further include: the first computing node obtains the device privacy data stored in the first IoT device, and accumulates or processes the device privacy data.
  • Vector characterization processing to obtain the characteristics of the machine tool.
  • the training method may further include: receiving the tool feature from the first IoT tool, the tool feature being used by the first IoT tool on its own
  • the stored privacy data of the machine is obtained by performing accumulation processing or vector characterization processing.
  • the first calculation node can determine the first calculation result.
  • the second computing node shown in FIG. 3 may determine the second intermediate result based on the merchant's private data and the second parameter maintained by it.
  • the second parameter is used to calculate the merchant's private data
  • the second intermediate result can be obtained.
  • ⁇ 2 can be used to represent the second parameter
  • x 2 can be used to represent the private data of the merchant, so that the second intermediate result can be determined
  • each of the multiple parties also maintains a partial calculation formula for the loss function of the risk assessment model.
  • the second computing node can use the second calculation formula and the second parameter maintained by the second computing node to calculate the merchant's private data to obtain the second intermediate result.
  • ⁇ 1 can be used to represent the second parameter
  • x 2 can be used to represent the privacy data of the merchant
  • l 2 () can be used to represent the second calculation formula. From this, the second intermediate result l 2 (x 2 ; ⁇ 2 ).
  • the training method may further include: the second computing node obtains the merchant's privacy data stored in the first merchant.
  • the second computing node is a trusted computing node.
  • the training method may further include: the second computing node generates a second public key and a second private key, and, The second public key is sent to multiple merchants, and the multiple merchants include the first merchant; the second computing node receives encrypted privacy data from the first merchant, and the encrypted privacy data is transferred from the first merchant.
  • the merchant uses the second public key to encrypt the merchant private data to obtain; the second computing node uses the second private key to decrypt the encrypted private data to obtain the merchant private data.
  • the second calculation node can determine the second calculation result.
  • the first payment platform shown in FIG. 3 may determine the third intermediate result based on the payment privacy data and the third parameter and risk tag maintained by it.
  • the third parameter is used to calculate the payment privacy feature, and the calculated result is compared with the label to obtain the third intermediate result.
  • ⁇ 3 can be used to represent the third parameter
  • x 3 can be used to represent the payment privacy data
  • y can be used to represent the sample label, so that the third intermediate result can be determined
  • each of the multiple parties also maintains a partial calculation formula for the loss function of the risk assessment model.
  • the first payment platform can use the third calculation formula and the third parameter maintained by it to calculate the machine characteristics and risk labels to obtain the third intermediate result.
  • ⁇ 3 is used to represent the third parameter
  • x 3 is used to represent the payment privacy data
  • y is used to represent the sample label
  • l 3 () is used to represent the third calculation formula, so that the third intermediate result can be determined l 3 (x 3 ; y; ⁇ 3 ).
  • the first payment platform can determine the third calculation result.
  • the first computing node, the second computing node, and the first payment platform can each calculate the first intermediate result, the second intermediate result, and the third intermediate result.
  • MPC technology can be used to realize the integration and sharing of data.
  • the MPC technology used may include homomorphic encryption technology, secret sharing technology, obfuscated circuit technology, and the like.
  • a homomorphic encryption technology can be used.
  • the first computing node and the second computing node also store the first public key generated based on the homomorphic encryption technology
  • the first payment platform also stores the The first public key and the corresponding first private key.
  • the first public key and the first private key may be generated by the first payment platform.
  • the first public key and the first private key may be generated by a third-party trusted organization.
  • Fig. 4 shows a multi-party interaction diagram based on homomorphic encryption according to an embodiment. As shown in Figure 4, the multi-party interaction process may include the following steps:
  • Step S401 The first computing node encrypts the first intermediate result by using the first public key to obtain the first encryption result.
  • the first intermediate result Encryption you can get the first encryption result
  • the first encryption result Epk (l 1 (z 1 ; ⁇ 1 )) can be obtained.
  • Step S403 The second computing node encrypts the second intermediate result by using the first public key to obtain the second encryption result.
  • the second intermediate result Encryption you can get the second encryption result
  • the second encryption result Epk (l 2 (x 2 ; ⁇ 2 )) can be obtained.
  • Step S404 The second computing node performs a first homomorphic addition operation on the first encryption result and the second encryption result to obtain the first operation result.
  • the first homomorphic addition operation is a multiplication operation of the first encryption result and the second encryption result.
  • the first operation result obtained may be In another specific embodiment, the obtained first operation result may be Epk (l 1 (z 1 ; ⁇ 1 ))*E pk (l 2 (x 2 ; ⁇ 2 )).
  • step S405 the second computing node sends the result of the first operation to the first payment platform.
  • Step S406 The first payment platform encrypts the third intermediate result by using the first public key to obtain the third encryption result.
  • the third intermediate result Encryption you can get the third encryption result
  • the third encryption result Epk (l 3 (x 3 ; y; ⁇ 3 )) can be obtained.
  • Step S407 The first payment platform performs a second homomorphic addition operation on the first operation result and the third encryption result to obtain the second operation result.
  • the second homomorphic addition operation is a multiplication operation of the first operation result and the third encryption result.
  • the second operation result obtained is In another embodiment, the obtained second operation result may be E pk (l 1 (z 1 ; ⁇ 1 ))*E pk (l 2 (x 2 ; ⁇ 2 ))*E pk (l 3 (x 3 ; y; ⁇ 3 )).
  • the first payment platform uses the first private key to decrypt the second operation result to obtain the training loss for the first IoT machine.
  • the training loss obtained by decryption may be:
  • step S409 the first payment platform uses the training loss to adjust the third parameter maintained by it.
  • step S410 the first payment platform sends the training loss to the first computing node.
  • step S411 the first computing node uses the training loss to adjust the first parameter maintained by it.
  • step S412 the first payment platform sends the training loss to the second computing node.
  • step S413 the second computing node uses the training loss to adjust the second parameter maintained by it.
  • the second computing node performs the first homomorphic addition operation. It should be understood that the second computing node can also send the second encryption result to the first computing node, and the first computing node performs the first homomorphic operation. Add operation, and then send the result of the first operation to the first payment platform. In this way, the first computing node, the second computing node, and the first payment platform can each adjust part of the parameters of the risk assessment model maintained by the first computing node, the second computing node, and the first payment platform according to the training loss obtained by implementing data fusion based on the MPC technology.
  • a secret sharing method can also be used to realize data fusion sharing.
  • the specific implementation can be carried out with reference to the prior art, which will not be repeated here.
  • each maintains its own private data and partial model parameters, and calculates the intermediate results separately, and then combines MPC technology for data fusion and sharing.
  • the risk assessment model is jointly trained. After performing the above training process many times, a final trained risk assessment model can be obtained, which can be used for risk identification and control of IoT equipment.
  • FIG. 5 shows a schematic diagram of the architecture of a multi-party joint training risk assessment model according to an embodiment.
  • the privacy data of the machine tool is aggregated to obtain the characteristics of each machine tool; in the merchant computing node (see the second computing node above), each of the multiple merchants encrypts its own merchant privacy data into the TEE; in the payment platform In the computing node (see the above-mentioned first payment platform), the payment platform database stores payment privacy data related to multiple merchants.
  • the machine-side computing node will correspond to the machine characteristics of a batch of machines (such as 5 or 20 machines) Enter the sub-model 1 (including the above-mentioned first parameter) to obtain the first intermediate result; the merchant computing node decrypts the merchant encrypted data of a batch of merchants bound to the batch of machines, obtains the corresponding merchant decrypted data, and enters it In sub-model 2 (including the above-mentioned second parameter), the second intermediate result is obtained; the payment platform computing node obtains the payment privacy data corresponding to the batch of merchants, inputs it into sub-model 3 (including the above-mentioned third parameter), and combines it with the sub-model The output result and the acquired risk label corresponding to the batch of equipment determine the third intermediate result.
  • the machine-side computing node, the merchant-side computing node, and the payment platform computing node use MPC technologies such as homomorphic encryption or secret sharing to provide the first intermediate result, the second intermediate result, and the third intermediate result, respectively, and compare these three The intermediate results are fused to determine the training loss, and then the parameters in the sub-model 1, sub-model 2 and sub-model 3 maintained respectively are adjusted. It can be understood that sub-model 1, sub-model 2 and sub-model 3 together constitute a risk assessment model. In this way, after multiple trainings until convergence, a final trained risk assessment model can be obtained, which can be used for risk assessment of IoT equipment.
  • MPC technologies such as homomorphic encryption or secret sharing
  • the target machine-side computing node associated with the target machine can give the first risk score based on the machine characteristics of the target machine and some adjusted model parameters; and
  • the target merchant’s computing node associated with the target merchant that determines the target IoT machine can give a second risk score based on the merchant’s privacy data of the target merchant and another part of the adjusted model parameters; the first payment platform may be based on the The payment privacy data related to the target merchant and another part of the adjusted model parameters are given a third risk score.
  • the three parties can use MPC technology to provide the first risk score, the second risk score, and the third risk score respectively to obtain the final comprehensive risk score.
  • the three parties may send their respective determined risk scores to a third-party trusted structure, and the third-party trusted agency will aggregate them, and return the aggregated comprehensive scores to each of the three parties. .
  • the effective data provided by all parties can be fully utilized, and accurate and highly available risk assessment results can be obtained through the risk assessment model, thereby realizing precise control over IoT devices or IoT devices that are bound to merchants or transactions.
  • the risk assessment result indicates that the risk is low
  • the transaction is allowed to be completed through the IoT device.
  • disabling intervention on the IoT device such as showing that the transaction has failed, or even freezing the account of the merchant bound to the IoT device, or paying for the use of the IoT device The user account is frozen.
  • the method of multi-party joint training of the risk assessment model for IoT equipment disclosed in the embodiments of this specification can ensure the security of multi-party data while achieving full and comprehensive utilization of valid data from all parties, and training to obtain a risk assessment model with excellent performance. , And then use the risk assessment model to conduct a comprehensive and accurate risk assessment of IoT machines.
  • the embodiment of this specification also discloses a training device. details as follows:
  • Fig. 6 shows a structure diagram of a device for multi-party joint training of a risk assessment model for IoT equipment according to an embodiment, where the multiple parties include a first computing node, a second computing node, and a first payment platform, each of which maintains parts of the risk assessment model Parameters; the first computing node is associated with the first IoT machine, storing machine features determined based on the machine privacy data of the first IoT machine, and the second computing node is bound to the first IoT machine
  • the first merchant is associated and stores merchant privacy data of the first merchant, and the first payment platform stores payment privacy data related to the first merchant and a risk tag indicating the risk situation of the first IoT machine.
  • the device 600 is integrated in the first computing node. As shown in FIG.
  • the device 600 includes: an intermediate result determining unit 610 configured to be based on the characteristics of the machine tool and the first parameter maintained by the first computing node , Determine the first intermediate result.
  • the intermediate result providing unit 620 is configured to use the secure multi-party computing MPC technology to provide the first intermediate result for combining with the second intermediate result determined by the second computing node based on the merchant’s private data and the second parameter maintained by it.
  • the first payment platform determines the training loss for the first IoT device based on the payment privacy data and the third parameter maintained by it and the third intermediate result determined by the risk tag.
  • the loss acquiring unit 630 is configured to acquire the training loss; the parameter tuning unit 640 is configured to adjust the first parameter by using the training loss.
  • the device 600 further includes: a feature acquiring unit 650 configured to acquire the device privacy data stored in the first IoT device, and perform cumulative processing or vector characterization on the device privacy data Processing to obtain the machine tool feature; or configured to receive the machine tool feature from the first IoT machine, and the machine tool feature is processed by the first IoT machine on its own stored machine privacy data or The vector characterization process is obtained.
  • a feature acquiring unit 650 configured to acquire the device privacy data stored in the first IoT device, and perform cumulative processing or vector characterization on the device privacy data Processing to obtain the machine tool feature; or configured to receive the machine tool feature from the first IoT machine, and the machine tool feature is processed by the first IoT machine on its own stored machine privacy data or The vector characterization process is obtained.
  • the device privacy data includes one or more of the following: the power-on time, shutdown time, and location information of the first IoT device, and unbinding or rebinding of the first IoT device Operational data generated.
  • the MPC technology includes a homomorphic encryption technology
  • the first computing node and the second computing node also store a first public key generated based on the homomorphic encryption technology, and the first payment platform The first public key and the corresponding first private key are also stored in it.
  • the intermediate result providing unit 620 is specifically configured to: use the first public key to encrypt the first intermediate result to obtain a first encryption result; and send the first encryption result to the second computing node to Make the second computing node perform a first homomorphic addition operation on the first encryption result and the second encryption result obtained by encrypting the second intermediate result, so that the first payment platform can perform the first homomorphic addition operation Perform a second homomorphic addition operation on the first operation result obtained by the state addition operation and the third encryption result obtained by encrypting the third intermediate result, and use the private key to perform the second homomorphism addition operation on the second homomorphism addition operation
  • the operation result is decrypted to obtain the training loss.
  • the loss obtaining unit 630 is specifically configured to receive the training loss from the first payment platform.
  • FIG. 7 shows a structure diagram of an apparatus for multi-party joint training of a risk assessment model for IoT equipment according to another embodiment, wherein the multi-party includes a first computing node, a second computing node, and a first payment platform, each of which maintains the risk assessment model Part of the parameters; the first computing node is associated with the first IoT machine, storing machine features related to the first IoT machine, and the second computing node is related to the first merchant that binds the first IoT machine And store the merchant privacy data of the first merchant, and the first payment platform stores the payment privacy data related to the first merchant and a risk tag indicating the risk situation of the first IoT machine.
  • the device 700 is integrated in the second computing node.
  • the device 700 includes an intermediate result determining unit 710 configured to determine a second intermediate result based on the merchant’s privacy data and the second parameter maintained by the second computing node. result.
  • the intermediate result providing unit 720 is configured to use the secure multi-party computing MPC technology to provide the second intermediate result for combining the first intermediate result determined by the first computing node based on the first parameter of the machine tool feature and its maintenance
  • the first payment platform determines the training loss for the first IoT device based on the payment privacy data and the third parameter maintained by it and the third intermediate result determined by the risk tag.
  • the loss obtaining unit 730 is configured to obtain the training loss.
  • the parameter adjustment unit 740 is configured to adjust the second parameter by using the training loss.
  • the device 700 further includes a privacy data obtaining unit 750 configured to obtain the merchant's privacy data stored in the first merchant.
  • the second computing node is a trusted computing node
  • the device 700 further includes a privacy data acquisition unit 750 configured to generate a second public key and a second private key, and to transfer the second public key and the second private key to The second public key is sent to multiple merchants, including the first merchant; receiving encrypted privacy data from the first merchant, and the encrypted privacy data is used by the first merchant using the second public key
  • the merchant private data is encrypted to obtain; the second private key is used to decrypt the encrypted private data to obtain the merchant private data.
  • the merchant privacy data includes transaction information generated by the first merchant on a payment platform other than the first payment platform, specifically including one or more of the following: transaction amount, transaction Location, trading hours, commodity types, risk events.
  • the MPC technology includes a homomorphic encryption technology
  • the first computing node and the second computing node also store a first public key generated based on the homomorphic encryption technology, and the first payment platform The first public key and the corresponding first private key are also stored in it.
  • the intermediate result providing unit 720 is specifically configured to: use the first public key to encrypt the second intermediate result to obtain a second encryption result; The first encryption result obtained by encrypting the first intermediate result; performing a homomorphic addition operation on the first encryption result and the second encryption result to obtain the first operation result; sending the first operation result to the first A payment platform, so that the first payment platform performs a second homomorphic addition operation on the first operation result and the third encryption result obtained by encrypting the third intermediate result, and uses the private key to perform the second homomorphic addition operation on the The second operation result obtained by the second homomorphic addition operation is decrypted to obtain the training loss.
  • the loss obtaining unit 730 is specifically configured to receive the training loss from the first payment platform.
  • FIG. 8 shows a device structure diagram of a multi-party joint training of a risk assessment model for IoT equipment according to yet another embodiment, where the multi-party includes a first computing node, a second computing node, and a first payment platform, each of which maintains the risk assessment model Part of the parameters; the first computing node is associated with the first IoT machine, storing machine features determined based on the machine privacy data of the first IoT machine, and the second computing node is bound to the first IoT machine Is associated with the first merchant of, and stores merchant privacy data of the first merchant, and the first payment platform stores payment privacy data related to the first merchant and a risk tag indicating the risk situation of the first IoT machine.
  • the multi-party includes a first computing node, a second computing node, and a first payment platform, each of which maintains the risk assessment model Part of the parameters; the first computing node is associated with the first IoT machine, storing machine features determined based on the machine privacy data of the first IoT machine, and
  • the device 800 is integrated in the first payment platform, and the device 800 includes: an intermediate result determining unit 810 configured to be based on the payment privacy data, the third parameter maintained by the first payment platform, and the risk tag , Determine the third intermediate result.
  • the intermediate result providing unit 820 is configured to use the secure multi-party computing MPC technology to provide the third intermediate result for combining the first intermediate result determined by the first computing node based on the first parameter of the machine tool feature and its maintenance ,
  • the second computing node determines the training loss for the first IoT device based on the second intermediate result determined by the merchant's private data and the second parameter maintained by it.
  • the loss obtaining unit 830 is configured to obtain the training loss.
  • the parameter adjustment unit 840 is configured to adjust the third parameter by using the training loss.
  • the payment privacy feature includes the contract information between the first merchant and the first payment platform, and the transaction information generated in the first payment platform specifically includes one or more of the following Species: user information, transaction amount, transaction location, transaction time, commodity type, and risk events of the transaction user.
  • the MPC technology includes a homomorphic encryption technology
  • the first computing node and the second computing node also store a first public key generated based on the homomorphic encryption technology, and the first payment platform The first public key and the corresponding first private key are also stored in it.
  • the intermediate result providing unit 820 is specifically configured to receive a first operation result from the second computing node, and the first operation result performs a first operation on the second encryption result and the first encryption result received from the first computing node. Obtained by a homomorphic addition operation, the second encryption result is obtained by encrypting the second intermediate result using the first public key, and the first encryption result is obtained by using the first public key to The first intermediate result is encrypted and obtained; the third intermediate result is encrypted using the first public key to obtain a third encryption result; the first operation result and the third encryption result are second The homomorphic addition operation obtains the second operation result; wherein the loss obtaining unit 830 is specifically configured to decrypt the second operation result by using the private key to obtain the training loss.
  • the device for multi-party joint training of the risk assessment model for IoT equipment disclosed in the embodiments of this specification can ensure the security of multi-party data while achieving full and comprehensive use of valid data from all parties, and training to obtain a risk assessment model with excellent performance. , And then use the risk assessment model to conduct a comprehensive and accurate risk assessment of IoT machines.
  • a computer-readable storage medium with a computer program stored thereon, and when the computer program is executed in a computer, the computer is executed as described in conjunction with FIG. 3 or FIG. 4 or FIG. 5.
  • a computing device including a memory and a processor, the memory stores executable code, and when the processor executes the executable code, a combination of FIG. 3 or FIG. 4 is implemented. Or the method described in Figure 5.
  • the functions described in the present invention can be implemented by hardware, software, firmware, or any combination thereof.
  • these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on the computer-readable medium.

Landscapes

  • Business, Economics & Management (AREA)
  • Engineering & Computer Science (AREA)
  • Accounting & Taxation (AREA)
  • Computer Security & Cryptography (AREA)
  • Finance (AREA)
  • Strategic Management (AREA)
  • Physics & Mathematics (AREA)
  • General Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Financial Or Insurance-Related Operations Such As Payment And Settlement (AREA)

Abstract

一种多方联合训练针对IoT机具的风险评估模型的方法,其中多方包括机具端计算节点、商户端计算节点和支付平台计算节点,分别存储多个机具的机具隐私数据、绑定多个机具的多个商户的商户隐私数据,以及与多个商户相关的支付隐私数据,并且各自维护风险评估模型中的部分参数。此外,支付平台计算节点还存储机具风险标签。在该方法中,机具端计算节点和商户端计算节点基于各自存储的隐私数据和维护的部分参数,确定出中间计算结果,支付平台计算节点基于其存储的隐私数据、维护的部分参数和标签计算出中间计算结果。然后,三方基于安全多方计算MPC技术,提供各自计算出的中间计算结果,确定训练损失,再各自调整维护的部分模型参数。

Description

多方联合训练针对IoT机具的风险评估模型的方法及装置 技术领域
本说明书一个或多个实施例涉及数据安全技术领域,尤其涉及一种多方联合训练针对IoT机具的风险评估模型的方法及装置。
背景技术
IoT(Internet of Things,物联网)机具是一种应用于商户收款的智能终端,可以方便商户收款,同时可以带给用户方便、快捷的支付体验。比如说,IoT机具可以支持人脸支付,使得用户无需使用自己的移动终端(如智能手机或可穿戴设备等)进行付款,从而简化支付流程。
然而,在便捷支付的背后,给风险控制也带来了更多的挑战。比如说,因为用户无需利用移动终端即可完成支付操作,使得可用于风控的数据减少。又比如,出于安全监管合规等要求,对IoT机具中数据的安全性要求极高。
因此,迫切需要一种合理的方案,可以实现对IoT机具进行全面、准确地风险评估。
发明内容
本说明书一个或多个实施例描述了一种多方联合训练针对IoT机具的风险评估模型的方法,可以在保障多方数据安全的同时,实现充分利用有效数据,训练得到性能优良的风险评估模型,进而通过使用该风险评估模型,对IoT机具的进行全面、准确的风险评估。
根据第一方面,提供一种多方联合训练针对IoT机具的风险评估模型的方法,其中多方包括第一计算节点、第二计算节点和第一支付平台,各自维护风险评估模型中的部分参数;所述第一计算节点与第一IoT机具相关联,存储基于所述第一IoT机具的机具隐私数据而确定的机具特征,所述第二计算节点与绑定所述第一IoT机具的第一商户相关联,存储所述第一商户的商户隐私数据,所述第一支付平台存储与所述第一商户相关的支付隐私数据和指示所述第一IoT机具风险情况的风险标签;所述方法应用于所述第一计算节点。该方法包括:基于所述机具特征及所述第一计算节点维护的第一参数,确定第一中间结果;利用安全多方计算MPC技术,提供所述第一中间结果,用于结合所 述第二计算节点基于所述商户隐私数据及其维护的第二参数确定的第二中间结果,所述第一支付平台基于所述支付隐私数据及其维护的第三参数和所述风险标签确定的第三中间结果,确定针对所述第一IoT机具的训练损失;获取所述训练损失,并且,利用所述训练损失调整所述第一参数。
在一个实施例中,在基于所述机具特征及所述第一计算节点维护的第一参数和第一计算式,确定第一中间结果之前,所述方法还包括:获取所述第一IoT机具中存储的所述机具隐私数据,并且,对所述机具隐私数据进行累计处理或向量表征处理,得到所述机具特征;或,从所述第一IoT机具接收所述机具特征,所述机具特征由所述第一IoT机具对其自身存储的所述机具隐私数据进行累计处理或向量表征处理而得到。
在一个实施例中,所述机具隐私数据包括以下中的一种或多种:所述第一IoT机具的开机时间、关机时间和位置信息,对所述第一IoT机具进行解绑、换绑产生的操作数据。
在一个实施例中,所述MPC技术包括同态加密技术,所述第一计算节点和第二计算节点中还存储基于所述同态加密技术生成的第一公钥,所述第一支付平台中还存储所述第一公钥和对应的第一私钥;其中利用安全多方计算MPC技术,提供所述第一中间结果,包括:利用所述第一公钥对所述第一中间结果进行加密,得到第一加密结果;将所述第一加密结果发送至所述第二计算节点,以使所述第二计算节点对所述第一加密结果和其对第二中间结果加密得到的第二加密结果进行第一同态加操作,进而使所述第一支付平台对所述第一同态加操作得到的第一操作结果和其对第三中间结果加密得到的第三加密结果进行第二同态加操作,并利用所述私钥对所述第二同态加操作得到的第二操作结果进行解密,得到所述训练损失;其中获取所述训练损失,包括:从所述第一支付平台接收所述训练损失。
根据第二方面,提供另一种多方联合训练针对IoT机具的风险评估模型的方法,其中多方包括第一计算节点、第二计算节点和第一支付平台,各自维护风险评估模型中的部分参数;所述第一计算节点与第一IoT机具相关联,存储与所述第一IoT机具相关的机具特征,所述第二计算节点与绑定所述第一IoT机具的第一商户相关联,存储所述第一商户的商户隐私数据,所述第一支付平台存储与所述第一商户相关的支付隐私数据和指示所述第一IoT机具风险情况的风险标签;所述方法应用于所述第二计算节点。该方法包括:基于所述商户隐私数据及所述第二计算节点维护的第二参数,确定第二中间结果;利用安全多方计算MPC技术,提供所述第二中间结果,用于结合所述第一计算节 点基于所述机具特征及其维护的第一参数确定的第一中间结果,所述第一支付平台基于所述支付隐私数据及其维护的第三参数和所述风险标签确定的第三中间结果,确定针对所述第一IoT机具的训练损失;获取所述训练损失,并且,利用所述训练损失调整所述第二参数。
在一个实施例中,在基于所述商户隐私数据及所述第二计算节点维护的第二参数,确定第二中间结果之前,所述方法还包括:获取所述第一商户中存储的所述商户隐私数据。
在一个实施例中,所述第二计算节点为可信计算节点,在基于所述商户隐私数据及所述第二计算节点维护的第二参数,确定第二中间结果之前,所述方法还包括:生成第二公钥和第二私钥,并且,将所述第二公钥发送至多个商户,所述多个商户中包括所述第一商户;从所述第一商户接收加密隐私数据,所述加密隐私数据由所述第一商户利用所述第二公钥对所述商户隐私数据进行加密而得到;利用所述第二私钥对所述加密隐私数据进行解密,得到所述商户隐私数据。
在一个实施例中,所述商户隐私数据包括所述第一商户在所述第一支付平台以外的其他支付平台中产生的交易信息,具体包括以下中的一种或多种:交易金额、交易地点、交易时间、商品种类、风险事件。
在一个实施例中,所述MPC技术包括同态加密技术,所述第一计算节点和第二计算节点中还存储基于所述同态加密技术生成的第一公钥,所述第一支付平台中还存储所述第一公钥和对应的第一私钥;其中利用安全多方计算MPC技术,提供所述第二中间结果,包括:利用第一公钥对所述第二中间结果进行加密,得到第二加密结果;从所述第一计算节点接收其利用所述第一公钥对所述第一中间结果加密得到的第一加密结果;对所述第一加密结果和第二加密结果进行同态加操作,得到第一操作结果;将所述第一操作结果发送至所述第一支付平台,以使所述第一支付平台对所述第一操作结果和其对第三中间结果加密得到的第三加密结果进行第二同态加操作,并利用所述私钥对所述第二同态加操作得到的第二操作结果进行解密,得到所述训练损失;其中获取所述训练损失,包括:从所述第一支付平台接收所述训练损失。
根据第三方面,提供又一种多方联合训练针对IoT机具的风险评估模型的方法,其中多方包括第一计算节点、第二计算节点和第一支付平台,各自维护风险评估模型中的部分参数;所述第一计算节点与第一IoT机具相关联,存储基于所述第一IoT机具的机具隐私数据而确定的机具特征,所述第二计算节点与绑定所述第一IoT机具的第一商户 相关联,存储所述第一商户的商户隐私数据,所述第一支付平台存储与所述第一商户相关的支付隐私数据和指示所述第一IoT机具风险情况的风险标签;所述方法应用于所述第一支付平台。该方法包括:
基于所述支付隐私数据及所述第一支付平台维护的第三参数和所述风险标签,确定第三中间结果;利用安全多方计算MPC技术,提供所述第三中间结果,用于结合所述第一计算节点基于所述机具特征及其维护的第一参数确定的第一中间结果,所述第二计算节点基于所述商户隐私数据及其维护的第二参数确定的第二中间结果,确定针对所述第一IoT机具的训练损失;获取所述训练损失,并且,利用所述训练损失调整所述第三参数。
在一个实施例中,所述支付隐私特征包括所述第一商户与所述第一支付平台的签约信息,在所述第一支付平台中产生的交易信息,具体包括以下中的一种或多种:交易用户的用户信息、交易金额、交易地点、交易时间、商品种类、风险事件。
在一个实施例中,所述MPC技术包括同态加密技术,所述第一计算节点和第二计算节点中还存储基于所述同态加密技术生成的第一公钥,所述第一支付平台中还存储所述第一公钥和对应的第一私钥;其中利用安全多方计算MPC技术,提供所述第三中间结果,包括:从所述第二计算节点接收第一操作结果,所述第一操作结果通过对第二加密结果和从第一计算节点接收的第一加密结果进行第一同态加操作而得到,所述第二加密结果是利用所述第一公钥对所述第二中间结果进行加密而得到,所述第一加密结果是利用所述第一公钥对所述第一中间结果进行加密而得到;利用所述第一公钥对所述第三中间结果进行加密,得到第三加密结果;对所述第一操作结果和所述第三加密结果进行第二同态加操作,得到第二操作结果;其中获取所述训练损失,包括:利用所述私钥对所述第二操作结果进行解密,得到所述训练损失。
根据第四方面,提供一种多方联合训练针对IoT机具的风险评估模型的装置,其中多方包括第一计算节点、第二计算节点和第一支付平台,各自维护风险评估模型中的部分参数;所述第一计算节点与第一IoT机具相关联,存储基于所述第一IoT机具的机具隐私数据而确定的机具特征,所述第二计算节点与绑定所述第一IoT机具的第一商户相关联,存储所述第一商户的商户隐私数据,所述第一支付平台存储与所述第一商户相关的支付隐私数据和指示所述第一IoT机具风险情况的风险标签。所述装置集成于所述第一计算节点,所述装置包括:中间结果确定单元,配置为基于所述机具特征及第一计算节点维护的第一参数,确定第一中间结果;中间结果提供单元,配置为利用安全多方计 算MPC技术,提供所述第一中间结果,用于结合所述第二计算节点基于所述商户隐私数据及其维护的第二参数确定的第二中间结果,所述第一支付平台基于所述支付隐私数据及其维护的第三参数和所述风险标签确定的第三中间结果,确定针对所述第一IoT机具的训练损失;损失获取单元,配置为获取所述训练损失;调参单元,配置为利用所述训练损失调整所述第一参数。
根据第五方面,提供另一种多方联合训练针对IoT机具的风险评估模型的装置,其中多方包括第一计算节点、第二计算节点和第一支付平台,各自维护风险评估模型中的部分参数;所述第一计算节点与第一IoT机具相关联,存储与所述第一IoT机具相关的机具特征,所述第二计算节点与绑定所述第一IoT机具的第一商户相关联,存储所述第一商户的商户隐私数据,所述第一支付平台存储与所述第一商户相关的支付隐私数据和指示所述第一IoT机具风险情况的风险标签;所述装置集成于所述第二计算节点,所述装置包括:中间结果确定单元,配置为基于所述商户隐私数据及第二计算节点维护的第二参数,确定第二中间结果;中间结果提供单元,配置为利用安全多方计算MPC技术,提供所述第二中间结果,用于结合所述第一计算节点基于所述机具特征及其维护的第一参数确定的第一中间结果,所述第一支付平台基于所述支付隐私数据及其维护的第三参数和所述风险标签确定的第三中间结果,确定针对所述第一IoT机具的训练损失;损失获取单元,配置为获取所述训练损失;调参单元,配置为利用所述训练损失调整所述第二参数。
根据第六方面,提供又一种多方联合训练针对IoT机具的风险评估模型的装置,其中多方包括第一计算节点、第二计算节点和第一支付平台,各自维护风险评估模型中的部分参数;所述第一计算节点与第一IoT机具相关联,存储基于所述第一IoT机具的机具隐私数据而确定的机具特征,所述第二计算节点与绑定所述第一IoT机具的第一商户相关联,存储所述第一商户的商户隐私数据,所述第一支付平台存储与所述第一商户相关的支付隐私数据和指示所述第一IoT机具风险情况的风险标签。所述装置集成于所述第一支付平台,所述装置包括:中间结果确定单元,配置为基于所述支付隐私数据及第一支付平台维护的第三参数和所述风险标签,确定第三中间结果;中间结果提供单元,配置为利用安全多方计算MPC技术,提供所述第三中间结果,用于结合所述第一计算节点基于所述机具特征及其维护的第一参数确定的第一中间结果,所述第二计算节点基于所述商户隐私数据及其维护的第二参数确定的第二中间结果,确定针对所述第一IoT机具的训练损失;损失获取单元,配置为获取所述训练损失;调参单元,配置为利用所述训练损失调整所述第三参数。
根据第七方面,提供了一种计算机可读存储介质,其上存储有计算机程序,当所述计算机程序在计算机中执行时,令计算机执行第一方面或第二方面或第三方面的方法。
根据第八方面,提供了一种计算设备,包括存储器和处理器,所述存储器中存储有可执行代码,所述处理器执行所述可执行代码时,实现第一方面或第二方面或第三方面的方法。
综上,采用本说明书实施例提供的方法及装置,可以在保障多方数据安全的同时,实现充分、全面利用各方有效数据,训练得到性能优良的风险评估模型,进而通过使用该风险评估模型,对IoT机具的进行全面、准确的风险评估。
附图说明
为了更清楚地说明本发明实施例的技术方案,下面将对实施例描述中所需要使用的附图作简单地介绍,显而易见地,下面描述中的附图仅仅是本发明的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动的前提下,还可以根据这些附图获得其它的附图。
图1示出根据一个实施例的多方联合对IoT机具进行风控的框架图。
图2示出根据一个实施例的可信计算节点的搭建流程示意图。
图3示出根据一个实施例的多方联合训练风险评估模型的框架图。
图4示出根据一个实施例的基于同态加密的多方交互图。
图5示出根据一个实施例的多方联合训练风险评估模型的架构示意图。
图6示出根据一个实施例的多方联合训练针对IoT机具的风险评估模型的装置结构图。
图7示出根据另一个实施例的多方联合训练针对IoT机具的风险评估模型的装置结构图。
图8示出根据又一个实施例的多方联合训练针对IoT机具的风险评估模型的装置结构图。
具体实施方式
下面结合附图,对本说明书提供的方案进行描述。
如前所述,需要对IoT机具进行风险识别和管控。在一个方案中,考虑到IoT机具背后的收款方通常是商户,由此可以基于某个支付平台(如支付宝)中采集的商户数据(如商户营业执照、经营状况等)对该商户进行风险识别和管控,从而实现对该商户使用的IoT机具进行风险管控。然而,此种基于商户的风控方案受商户数据质量的影响大,对于大部分的中长尾商户(如小型商户、营业额较少的商户等)和新增商户,经常存在商户数据的数据量少,关键信息缺失等问题,导致风险识别的准确度较低。
在另一种方案中,可以通过对IoT机具中的单笔交易进行风险评估,实现对IoT机具的风控。具体地,对于IoT机具中通过某个支付平台进行的交易,可以从该某个支付平台中获取该交易的交易信息(包括买家ID,交易的时间、位置和金额等),并从该某个支付平台中获取交易买家的买家信息(包括历史交易记录等),实现对该笔交易的风险识别和管控。然而,此种基于用户的风控方案,是从单个用户、单笔交易的角度进行风险识别,得到的识别结果直接应用于IoT机具,会存在识别准确率低和误打扰的问题。比如说,在识别出一笔交易高风险的情况下,可能是买家存在问题,并非是商户存在问题,此时根据该高风险的识别结果对IoT机具和IoT机具背后的商户进行管控,显然是不妥当的,将存在误打扰的问题。
此外发明人还发现,上述两种方案均未使用IoT场景的特殊数据,如IoT机具的开机时间,商家账号解绑换绑频率,IoT机具的位置信息等。
基于以上观察和统计,发明人提出一种融合多方数据对IoT机具进行风控的方案,可以在保障多方数据隐私安全的技术上,实现数据的融合共享。在一个实施例中,图1示出根据一个实施例的多方联合对IoT机具进行风控的框架图,如图1所示,其中融合IoT机具端(图1中的收款机具)产生的机具隐私数据、与IoT机具相绑定商户的商户隐私数据、支付平台中与该商户相关的支付隐私数据(包括交易信息、买家信息和商家信息),利用MPC(Secure Multi-Party Computation,安全多方计算)技术,实现对IoT机具进行全面、准确的风险评估。
融合多方数据进行风险评估是基于风险评估模型实现的,具体地,本说明书实施例披露一种多方联合训练风险评估模型的方法。下面,先对其中执行联合训练的多方进行介绍,再对训练方法的实施流程进行介绍。
具体地,上述多方包括第一计算节点、第二计算节点和第一支付平台。首先需要说明的是,其中的“第一”、“第二”以及文中它处的类似用语,仅用于区分同类事务,不具有其他限定作用。
上述第一计算节点与第一IoT机具相关联,存储基于第一IoT机具的机具隐私数据而确定的机具特征。可以理解,实际存在多个IoT机具,其中第一IoT机具可以为多个IoT机具中的任意一个。
在一个实施例中,计算节点与IoT机具存在一一对应的关系,也就是说,针对每个IoT机具都构建有与之对应的计算节点。在一个具体的实施例中,第一计算节点与第一IoT机具集成在一起。在另一个具体的实施例中,可以直接将第一IoT机具作为第一计算节点。此时,第一计算节点可以从第一IoT机具中获取存储的机具隐私数据进行处理,以确定机具特征。在另一个实施例中,多个IoT机具可以共用一个计算节点,这意味着,第一计算节点与多个IoT机具相关联。此时,因考虑到数据安全合规等要求,IoT机具中的隐私数据不允许外泄,可以由IoT机具对其中的机具隐私数据进行特征聚合处理,并将聚合得到的机具特征发送至第一计算节点,需要理解,根据机具特征通常是很难还原出原始隐私数据的,如此可以有效防止机具隐私数据的泄漏。
在一个实施例中,其中机具隐私数据可以包括:第一IoT机具的开机时间、关机时间和位置信息,对所述第一IoT机具进行解绑、换绑产生的操作数据。在一个具体的实施例中,其中开机时间和关机时间可以包括在历史时段内的多个时刻,如开机时间包括周一上午6:00,周二上午7:00等。在一个具体的实施例中,其中位置信息可以包括利用LBS(Location Based Services,基于位置的服务)采集的位置信息,如经纬度信息等。在一个具体的实施例中,其中解绑、换绑产生的操作数据可以包括解绑、换绑的操作时刻、操作频率和涉及到的商户(或商户账号)的数量。
在一个实施例中,上述机具特征可以包括累计特征或特征向量。在一个具体的实施例中,上述对机具隐私数据的特征聚合处理可以包括累计处理和向量表征处理等。在一个例子中,其中累计处理可以包括,确定机具隐私数据对应的累计特征,作为上述机具特征。在一个具体的例子中,其中累计特征可以包括IoT机具每天开关机的平均次数。在一个例子中,其中向量表征处理可以包括利用表征学习算法,如神经网络等,计算出机具隐私数据对应的特征向量,作为上述机具特征。
以上,主要对第一计算节点与第一IoT机具之间的关系,以及机具隐私数据和机具特征进行介绍。
上述第二计算节点与绑定第一IoT机具的第一商户(或第一商户的商户账号)相关联,存储第一商户的商户隐私数据。需要理解,第一商户参与的多笔交易通常涉及到多个支付平台,比如目前比较主流的支付宝支付平台和微信支付平台等,而对于第一商户 在第一支付平台以外的其他支付平台产生的交易数据,第一支付平台通常是无法获取的,对于这部分交易数据,可以由第一商户提供。由此,其中商户隐私数据可以包括:第一商户在第一支付平台以外的其他支付平台中产生的交易信息,具体可以包括:交易金额、交易地点、交易时间、商品种类、风险事件。
在一个实施例中,第二计算节点可以是第一商户自己搭建的节点,如此,第二计算节点可以直接获取第一商户中的商户隐私数据,进行模型训练。在另一个实施例中,考虑到大部分中长尾商户,没有能力去搭建一个计算节点。因此,发明人提出可以引入TEE(Trusted execution environment,可信计算环境),帮助多个商户(包括第一商户)在保障自身数据安全的情况下,将各自的商户隐私数据加密至TEE环境,以实现可信计算节点(作为第二计算节点)的构建。
需要说明的是,TEE环境可以采用Intel SGX,HyperVisor等技术实现。下面以Intel SGX技术实现TEE环境为例,说明构建可信计算节点的过程。在一个具体的实施例中,图2示出根据一个实施例的可信计算节点的搭建流程示意图。如图2所示,其中可信计算节点是采用SGX技术实现的可信计算围圈Enclave,具体地,通过提供一系列CPU指令码,允许用户代码创建具有高访问权限的私有内存区域而形成计算围圈Enclave。任何商户都无法访问围圈Enclave中的数据,因此,存储在Enclave中的隐私数据无法被窃取或篡改。图2示出的搭建流程包括以下步骤:首先,在步骤S21,商户向可信计算围圈(Enclave)请求软件运行报告。此时,Enclave中的Intel CPU会根据后续用于对商户隐私数据进行计算的算法C++代码生成一个公钥,私钥和算法代码签名。接着,在步骤S22,Enclave将生成的公钥,算法代码签名作为软件运行报告中的内容返回给商户,而私钥则存在Intel CPU中,以保证任何商户都无法对其他商户利用公钥加密后的商户隐私数据进行解密。然后,在步骤S23,商户将接收到的软件运行报告发送至Intel公司的认证接口做一个第三方认证。再接着,在步骤S24,Intel公司可以告知商户,软件运行报告是可信的,这意味着,其中包括的公钥和算法代码签名确实是Intel CPU生成的,中间没有串改,是可信的。再然后,在步骤S25,商户在接收到认证成功的结果后,利用公钥对自己的商户隐私数据加密,并且,在步骤S26,将加密数据发送至Enclave,由此,Enclave可以利用私钥对加密数据进行解密,得到原始的商户隐私数据,用于模型训练。如此,可以实现各商户在保证自身数据安全的情况下,提供其隐私数据以用于模型训练。
以上,主要对第二计算节点和第一商户之间的关系,以及商户隐私数据进行介绍。
上述第一支付平台存储与所述第一商户相关的支付隐私数据和指示所述第一IoT机具风险情况的风险标签。在一个实施例中,其中支付隐私数据可以包括第一商户与第一支付平台的签约信息,在所述第一支付平台中产生的交易信息,且此交易信息具体包括以下中的一种或多种:交易用户的用户信息、交易金额、交易地点、交易时间、商品种类、风险事件。在一个具体的实施例中,其中签约信息可以包括第一商户的营业执照、签约时刻、签约时长等信息。在一个具体的实施例中,其中用户信息可以包括用户的基本属性信息、交易偏好和历史交易记录。在一个例子中,其中基本属性信息可以包括性别、年龄、职业、常驻地、兴趣爱好等。在一个例子中,其中交易偏好可以包括最常购买的商品种类(如电子商品)、最常购物的时段(如晚上21:00-22:00)。在一个具体的实施例中,其中风险事件可以包括第一商户曾经发生的高风险事件,如售卖非法产品(如赌博产品)。
在一个实施例中,上述风险标签可以包括有风险和无风险。在另一个实施例中,上述风险标签还可以包括多个风险等级。在一个具体的实施例中,可以包括高风险、中风险和低风险。
以上,主要对第一支付平台中存储的支付隐私数据和风险标签进行介绍。
此外,上述第一计算节点、第二计算节点和第一支付平台中,还各自维护风险评估模型中的部分参数。通常,此三方维护的部分参数是互不相同的。在一个实施例中,每方所维护的部分参数与其所提供数据对应的样本特征相关联。在一个实施例中,每方具体维护风险评估模型中参数的哪一部分,可以通过MPC技术而确定。另一方面,在一个实施例中,风险评估模型可以采用逻辑回归算法、决策树算法、神经网络等实现。
进一步地,图3示出根据一个实施例的多方联合训练风险评估模型的框架图。如图3所示,在训练过程中,上述第一计算节点、第二计算节点和第一支付平台各自利用自身存储的隐私数据和模型参数进行计算,得到各自的中间计算结果,再利用MPC技术提供各自的中间计算结果,实现数据的融合共享,并完成对风险评估模型的训练。
针对训练过程,下面先对各方各自计算出中间结果的过程进行介绍,再对各方利用MPC技术进行数据融合共享,并各自对其维护的模型参数进行调整的过程进行介绍。
如图3所示,其中第一计算节点基于机具特征及其维护的第一参数,确定第一中间结果。在一个实施例中,利用第一参数对机具特征进行计算,可以得到第一中间结果。在一个具体的实施例中,可以用θ 1表示第一参数,用z 1表示基于机具隐私数据x 1确定的 机具特征,由此可以确定出第一中间结果
Figure PCTCN2020124289-appb-000001
在另一个实施例中,多方中还各自维护针对风险评估模型的损失函数的部分计算式。相应地,第一计算节点可以利用其维护的第一计算式和第一参数对机具特征进行计算,得到第一中间结果。在一个具体的实施例中,可以用θ 1表示第一参数,用z 1表示机具特征,l 1()表示第一计算式,由此可以确定出第一中间结果l 1(z 1;θ 1)。
在一个实施例中,在确定第一中间结果之前,所述训练方法还可以包括:第一计算节点获取第一IoT机具中存储的机具隐私数据,并且,对所述机具隐私数据进行累计处理或向量表征处理,得到所述机具特征。在另一个实施例中,在确定第一中间结果之前,所述训练方法还可以包括:从所述第一IoT机具接收所述机具特征,所述机具特征由所述第一IoT机具对其自身存储的所述机具隐私数据进行累计处理或向量表征处理而得到。需要说明的是,对于机具特征、机具隐私数据、累计处理和向量表征处理的描述,可以参见前述实施例中的相关描述,不作赘述。
如此,第一计算节点可以确定出第一计算结果。
图3中示出的第二计算节点可以基于商户隐私数据及其维护的第二参数,确定第二中间结果。在一个实施例中,利用第二参数对商户隐私数据进行计算,可以得到第二中间结果。在一个具体的实施例中,可以利用θ 2表示第二参数,用x 2表示商户隐私数据,由此可以确定出第二中间结果
Figure PCTCN2020124289-appb-000002
在另一个实施例中,多方中还各自维护针对风险评估模型的损失函数的部分计算式。相应地,第二计算节点可以利用其维护的第二计算式和第二参数对商户隐私数据进行计算,得到第二中间结果。在一个具体的实施例中,可以用θ 1表示第二参数,用x 2表示商户隐私数据,用l 2()表示第二计算式,由此可以确定出第二中间结果l 2(x 2;θ 2)。
在一个实施例中,在确定第二中间结果之前,所述训练方法还可以包括:第二计算节点获取第一商户中存储的所述商户隐私数据。在另一个实施例中,第二计算节点为可信计算节点,在确定第二中间结果之前,所述训练方法还可以包括:第二计算节点生成第二公钥和第二私钥,并且,将所述第二公钥发送至多个商户,所述多个商户中包括所述第一商户;第二计算节点从所述第一商户接收加密隐私数据,所述加密隐私数据由所述第一商户利用所述第二公钥对所述商户隐私数据进行加密而得到;第二计算节点利用所述第二私钥对所述加密隐私数据进行解密,得到所述商户隐私数据。需要说明的是,对其中可信计算节点、第二公钥、第二私钥等的描述,可以参见前述实施例中的相关描述,在此不作赘述。
如此,第二计算节点可以确定出第二计算结果。
图3中示出的第一支付平台可以基于支付隐私数据及其维护的第三参数和风险标签,确定第三中间结果。在一个实施例中,利用第三参数对支付隐私特征进行计算,并将此计算得到的结果与标签进行比对,可以得到第三中间结果。在一个具体的实施例中,可以用θ 3表示第三参数,用x 3表示支付隐私数据,用y表示样本标签,由此可以确定出第三中间结果
Figure PCTCN2020124289-appb-000003
在另一个实施例中,多方中还各自维护针对风险评估模型的损失函数的部分计算式。相应地,第一支付平台可以利用其维护的第三计算式和第三参数对机具特征和风险标签进行计算,得到第三中间结果。在一个具体的实施例中,用θ 3表示第三参数,用x 3表示支付隐私数据,用y表示样本标签,用l 3()表示第三计算式,由此可以确定出第三中间结果l 3(x 3;y;θ 3)。
需要说明的是,对于其中支付隐私数据和风险标签等的介绍,可以参见前述实施例中的相关描述,不作赘述。
如此,第一支付平台可以确定出第三计算结果。
以上,第一计算节点、第二计算节点和第一支付平台可以各自计算出第一中间结果、第二中间结果和第三中间结果。基于此,可以利用MPC技术,实现数据的融合共享。在一个实施例中,利用的MPC技术可以包括同态加密技术、秘密分享技术和混淆电路技术等。
在一种实施方式下,可以采用同态加密技术,此时,第一计算节点和第二计算节点中还存储基于同态加密技术生成的第一公钥,第一支付平台中还存储所述第一公钥和对应的第一私钥。在一个实施例中,第一公钥和第一私钥可以是第一支付平台生成的。在另一个实施例中,其中第一公钥和第一私钥可以是第三方可信机构生成的。图4示出根据一个实施例的基于同态加密的多方交互图。如图4所示,多方交互过程可以包括以下步骤:
步骤S401,第一计算节点利用第一公钥对第一中间结果进行加密,得到第一加密结果。在一个实施例中,对第一中间结果
Figure PCTCN2020124289-appb-000004
加密,可以得到第一加密结果
Figure PCTCN2020124289-appb-000005
在另一个实施例中,对第一中间结果l 1(z 1;θ 1)加密,可以得到第一加密结果E pk(l 1(z 1;θ 1))。
步骤S402,第一计算节点将第一加密结果发送至第二计算节点。
步骤S403,第二计算节点利用第一公钥对第二中间结果进行加密,得到第二加密 结果。在一个实施例中,对第二中间结果
Figure PCTCN2020124289-appb-000006
加密,可以得到第二加密结果
Figure PCTCN2020124289-appb-000007
在另一个实施例中,对第二中间结果l 2(x 2;θ 2)加密,可以得到第二加密结果E pk(l 2(x 2;θ 2))。
步骤S404,第二计算节点对第一加密结果和第二加密结果进行第一同态加操作,得到第一操作结果。在一个实施例中,其中第一同态加操作为对第一加密结果和第二加密结果的相乘操作。在一个具体的实施例中,得到的第一操作结果可以为
Figure PCTCN2020124289-appb-000008
Figure PCTCN2020124289-appb-000009
在另一个具体的实施例中,得到的第一操作结果可以为E pk(l 1(z 1;θ 1))*E pk(l 2(x 2;θ 2))。
步骤S405,第二计算节点将第一操作结果发送至第一支付平台。
步骤S406,第一支付平台利用第一公钥对第三中间结果进行加密,得到第三加密结果。在一个实施例中,对第三中间结果
Figure PCTCN2020124289-appb-000010
加密,可以得到第三加密结果
Figure PCTCN2020124289-appb-000011
在另一个实施例中,对第三中间结果l 3(x 3;θ 3)加密,可以得到第三加密结果E pk(l 3(x 3;y;θ 3))。
步骤S407,第一支付平台对第一操作结果和第三加密结果进行第二同态加操作,得到第二操作结果。在一个实施例中,其中第二同态加操作为对第一操作结果和第三加密结果的相乘操作。在一个实施例中,得到的第二操作结果为
Figure PCTCN2020124289-appb-000012
在另一个实施例中,得到的第二操作结果可以为E pk(l 1(z 1;θ 1))*E pk(l 2(x 2;θ 2))*E pk(l 3(x 3;y;θ 3))。
步骤S408,第一支付平台利用第一私钥对第二操作结果进行解密,得到针对第一IoT机具的训练损失。在一个实施例中,解密得到的训练损失可以为:
Figure PCTCN2020124289-appb-000013
Figure PCTCN2020124289-appb-000014
在另一个实施例中,解密得到的训练损失可以为:L=l 1(z 1;θ 1)+l 2(x 2;θ 2)+l 3(x 3;y;θ 3)。
步骤S409,第一支付平台利用训练损失调整其维护的第三参数。
步骤S410,第一支付平台将训练损失发送至第一计算节点。
步骤S411,第一计算节点利用训练损失调整其维护的第一参数。
步骤S412,第一支付平台将训练损失发送至第二计算节点。
步骤S413,第二计算节点利用训练损失调整其维护的第二参数。
需要说明的是,图4中步骤的标号,并不构成对步骤顺序的限定。此外,图4中 是第二计算节点进行第一同态加操作,需要理解,还可以是第二计算节点将第二加密结果发送给第一计算节点,由第一计算节点进行第一同态加操作,再第一操作结果发送给第一支付平台。如此,第一计算节点、第二计算节点和第一支付平台可以根据基于MPC技术实现数据融合而得到的训练损失,各自调节自身维护的风险评估模型的部分参数。
在另一种实施方式下,还可以采用秘密分享的方式,实现数据融合共享。具体可以参考现有技术进行实施,在此不作赘述。
综上,采用本说明书实施例提供的训练方法,通过构建安全计算节点,各自维护自身隐私数据、部分模型参数,并且各自计算中间结果,再结合MPC技术进行数据的融合共享,可以实现在保障各方数据隐私安全的情况下,共同训练风险评估模型。在多次执行上述训练过程后,可以得到最终训练好的风险评估模型,用于对IoT机具的风险识别和管控。
下面再结合一个具体的例子,对本说明书实施例披露的多方联合训练针对IoT机具的风险评估模型的方法进行介绍。图5示出根据一个实施例的多方联合训练风险评估模型的架构示意图,如图5所示,在机具端计算节点(参见上述第一计算节点)中,多个IoT机具中的各机具对自身的机具隐私数据进行特征聚合,得到各自的机具特征;在商户端计算节点(参见上述第二计算节点)中,多个商户中的各商户将自身的商户隐私数据加密到TEE中;在支付平台计算节点(参见上述第一支付平台)中,支付平台数据库存储与多个商户相关的支付隐私数据。
基于此,根据训练时设定的批量样本数量(Batch Size),如一批5个样本或20个样本等,机具端计算节点将对应数量的一批机具(如5或20个机具)的机具特征输入子模型1(其中包括上述第一参数)中,得到第一中间结果;商户端计算节点对绑定该批机具的一批商户的商户加密数据进行解密,得到对应的商户解密数据,并输入子模型2(其中包括上述第二参数)中,得到第二中间结果;支付平台计算节点获取该批商户对应的支付隐私数据,输入子模型3(包括上述第三参数)中,并结合子模型输出结果和获取的该批机具对应的风险标签确定第三中间结果。
进一步地,机具端计算节点、商户端计算节点和支付平台计算节点利用同态加密或秘密分享等MPC技术,分别提供第一中间结果、第二中间结果和第三中间结果,并对这三个中间结果进行融合确定出训练损失,进而调整各自维护的子模型1、子模型2和子模型3中的参数。可以理解,子模型1、子模型2和子模型3共同构成风险评估模型。如此,经过多次训练直到收敛,可以得到最终训练好的风险评估模型,用于针对IoT 机具的风险评估。
以上对风险评估模型的训练过程进行介绍。下面,对训练好的风险评估模型的使用方法进行简单介绍。针对待评估的目标IoT机具(以下简称目标机具),与目标机具关联的目标机具端计算节点,可以基于该目标机具的机具特征和调整好的一部分模型参数,给出第一风险评分;与绑定该目标IoT机具的目标商户相关联的目标商户端计算节点,可以基于该目标商户的商户隐私数据和调整好的另一部分模型参数,给出第二风险评分;第一支付平台可以基于与该目标商户相关的支付隐私数据和调整好的再一部分模型参数,给出第三风险评分。
进一步地,在一个实施例中,这三方可以利用MPC技术,分别提供第一风险评分、第二风险评分和第三风险评分,以得到最终的综合风险评分。在另一个实施例中,这三方可以将各自确定出的风险评分发送给第三方可信结构,再由第三方可信机构进行汇总,并将汇总得到的综合评分分别返回给三方中的每一方。
如此,可以全面利用各方提供的有效数据,通过该风险评估模型,得到准确的、可用性高的风险评估结果,进而实现对IoT机具或IoT机具的绑定商户或发生交易的精准管控。例如,在风险评估结果指示风险低的情况下,允许通过该IoT机具完成交易。又例如,在风险评估结果指示风险极高的情况下,对该IoT机具进行禁用干预,如显示交易失败,甚至对该IoT机具的绑定商户进行账号冻结,或者对使用该IoT机具进行支付的用户账户进行冻结。
综上,采用本说明书实施例披露的多方联合训练针对IoT机具的风险评估模型的方法,可以在保障多方数据安全的同时,实现充分、全面利用各方有效数据,训练得到性能优良的风险评估模型,进而通过使用该风险评估模型,对IoT机具的进行全面、准确的风险评估。
与上述训练方法相对应的,本说明书实施例还披露一种训练装置。具体如下:
图6示出根据一个实施例的多方联合训练针对IoT机具的风险评估模型的装置结构图,其中多方包括第一计算节点、第二计算节点和第一支付平台,各自维护风险评估模型中的部分参数;所述第一计算节点与第一IoT机具相关联,存储基于所述第一IoT机具的机具隐私数据而确定的机具特征,所述第二计算节点与绑定所述第一IoT机具的第一商户相关联,存储所述第一商户的商户隐私数据,所述第一支付平台存储与所述第一商户相关的支付隐私数据和指示所述第一IoT机具风险情况的风险标签。所述装置600 集成于所述第一计算节点,如图6所示,所述装置600包括:中间结果确定单元610,配置为基于所述机具特征及所述第一计算节点维护的第一参数,确定第一中间结果。中间结果提供单元620,配置为利用安全多方计算MPC技术,提供所述第一中间结果,用于结合所述第二计算节点基于所述商户隐私数据及其维护的第二参数确定的第二中间结果,所述第一支付平台基于所述支付隐私数据及其维护的第三参数和所述风险标签确定的第三中间结果,确定针对所述第一IoT机具的训练损失。损失获取单元630,配置为获取所述训练损失;调参单元640,配置为利用所述训练损失调整所述第一参数。
在一个实施例中,所述装置600还包括:特征获取单元650,配置为获取所述第一IoT机具中存储的所述机具隐私数据,并且,对所述机具隐私数据进行累计处理或向量表征处理,得到所述机具特征;或配置为,从所述第一IoT机具接收所述机具特征,所述机具特征由所述第一IoT机具对其自身存储的所述机具隐私数据进行累计处理或向量表征处理而得到。
在一个实施例中,所述机具隐私数据包括以下中的一种或多种:所述第一IoT机具的开机时间、关机时间和位置信息,对所述第一IoT机具进行解绑、换绑产生的操作数据。
在一个实施例中,所述MPC技术包括同态加密技术,所述第一计算节点和第二计算节点中还存储基于所述同态加密技术生成的第一公钥,所述第一支付平台中还存储所述第一公钥和对应的第一私钥。其中中间结果提供单元620具体配置为:利用所述第一公钥对所述第一中间结果进行加密,得到第一加密结果;将所述第一加密结果发送至所述第二计算节点,以使所述第二计算节点对所述第一加密结果和其对第二中间结果加密得到的第二加密结果进行第一同态加操作,进而使所述第一支付平台对所述第一同态加操作得到的第一操作结果和其对第三中间结果加密得到的第三加密结果进行第二同态加操作,并利用所述私钥对所述第二同态加操作得到的第二操作结果进行解密,得到所述训练损失。其中损失获取单元630具体配置为:从所述第一支付平台接收所述训练损失。
图7示出根据另一个实施例的多方联合训练针对IoT机具的风险评估模型的装置结构图,其中多方包括第一计算节点、第二计算节点和第一支付平台,各自维护风险评估模型中的部分参数;所述第一计算节点与第一IoT机具相关联,存储与所述第一IoT机具相关的机具特征,所述第二计算节点与绑定所述第一IoT机具的第一商户相关联,存储所述第一商户的商户隐私数据,所述第一支付平台存储与所述第一商户相关的支付隐 私数据和指示所述第一IoT机具风险情况的风险标签。所述装置700集成于所述第二计算节点,所述装置700包括:中间结果确定单元710,配置为基于所述商户隐私数据及所述第二计算节点维护的第二参数,确定第二中间结果。中间结果提供单元720,配置为利用安全多方计算MPC技术,提供所述第二中间结果,用于结合所述第一计算节点基于所述机具特征及其维护的第一参数确定的第一中间结果,所述第一支付平台基于所述支付隐私数据及其维护的第三参数和所述风险标签确定的第三中间结果,确定针对所述第一IoT机具的训练损失。损失获取单元730,配置为获取所述训练损失。调参单元740,配置为利用所述训练损失调整所述第二参数。
在一个实施例中,所述装置700还包括隐私数据获取单元750,配置为:获取所述第一商户中存储的所述商户隐私数据。
在一个实施例中,所述第二计算节点为可信计算节点,所述装置700还包括隐私数据获取单元750,配置为:生成第二公钥和第二私钥,并且,将所述第二公钥发送至多个商户,所述多个商户中包括所述第一商户;从所述第一商户接收加密隐私数据,所述加密隐私数据由所述第一商户利用所述第二公钥对所述商户隐私数据进行加密而得到;利用所述第二私钥对所述加密隐私数据进行解密,得到所述商户隐私数据。
在一个实施例中,所述商户隐私数据包括所述第一商户在所述第一支付平台以外的其他支付平台中产生的交易信息,具体包括以下中的一种或多种:交易金额、交易地点、交易时间、商品种类、风险事件。
在一个实施例中,所述MPC技术包括同态加密技术,所述第一计算节点和第二计算节点中还存储基于所述同态加密技术生成的第一公钥,所述第一支付平台中还存储所述第一公钥和对应的第一私钥。其中中间结果提供单元720具体配置为:利用所述第一公钥对所述第二中间结果进行加密,得到第二加密结果;从所述第一计算节点接收其利用所述第一公钥对所述第一中间结果加密得到的第一加密结果;对所述第一加密结果和第二加密结果进行同态加操作,得到第一操作结果;将所述第一操作结果发送至所述第一支付平台,以使所述第一支付平台对所述第一操作结果和其对第三中间结果加密得到的第三加密结果进行第二同态加操作,并利用所述私钥对所述第二同态加操作得到的第二操作结果进行解密,得到所述训练损失。其中损失获取单元730具体配置为:从所述第一支付平台接收所述训练损失。
图8示出根据又一个实施例的多方联合训练针对IoT机具的风险评估模型的装置结构图,其中多方包括第一计算节点、第二计算节点和第一支付平台,各自维护风险评估 模型中的部分参数;所述第一计算节点与第一IoT机具相关联,存储基于所述第一IoT机具的机具隐私数据而确定的机具特征,所述第二计算节点与绑定所述第一IoT机具的第一商户相关联,存储所述第一商户的商户隐私数据,所述第一支付平台存储与所述第一商户相关的支付隐私数据和指示所述第一IoT机具风险情况的风险标签。所述装置800集成于所述第一支付平台,所述装置800包括:中间结果确定单元810,配置为基于所述支付隐私数据及所述第一支付平台维护的第三参数和所述风险标签,确定第三中间结果。中间结果提供单元820,配置为利用安全多方计算MPC技术,提供所述第三中间结果,用于结合所述第一计算节点基于所述机具特征及其维护的第一参数确定的第一中间结果,所述第二计算节点基于所述商户隐私数据及其维护的第二参数确定的第二中间结果,确定针对所述第一IoT机具的训练损失。损失获取单元830,配置为获取所述训练损失。调参单元840,配置为利用所述训练损失调整所述第三参数。
在一个实施例中,所述支付隐私特征包括所述第一商户与所述第一支付平台的签约信息,在所述第一支付平台中产生的交易信息,具体包括以下中的一种或多种:交易用户的用户信息、交易金额、交易地点、交易时间、商品种类、风险事件。
在一个实施例中,所述MPC技术包括同态加密技术,所述第一计算节点和第二计算节点中还存储基于所述同态加密技术生成的第一公钥,所述第一支付平台中还存储所述第一公钥和对应的第一私钥。
其中中间结果提供单元820,具体配置为:从所述第二计算节点接收第一操作结果,所述第一操作结果通过对第二加密结果和从第一计算节点接收的第一加密结果进行第一同态加操作而得到,所述第二加密结果是利用所述第一公钥对所述第二中间结果进行加密而得到,所述第一加密结果是利用所述第一公钥对所述第一中间结果进行加密而得到;利用所述第一公钥对所述第三中间结果进行加密,得到第三加密结果;对所述第一操作结果和所述第三加密结果进行第二同态加操作,得到第二操作结果;其中损失获取单元830具体配置为:利用所述私钥对第二操作结果进行解密,得到所述训练损失。
综上,采用本说明书实施例披露的多方联合训练针对IoT机具的风险评估模型的装置,可以在保障多方数据安全的同时,实现充分、全面利用各方有效数据,训练得到性能优良的风险评估模型,进而通过使用该风险评估模型,对IoT机具的进行全面、准确的风险评估。
根据另一方面的实施例,还提供一种计算机可读存储介质,其上存储有计算机程序,当所述计算机程序在计算机中执行时,令计算机执行结合图3或图4或图5所描述 的方法。
根据再一方面的实施例,还提供一种计算设备,包括存储器和处理器,所述存储器中存储有可执行代码,所述处理器执行所述可执行代码时,实现结合图3或图4或图5所述的方法。
本领域技术人员应该可以意识到,在上述一个或多个示例中,本发明所描述的功能可以用硬件、软件、固件或它们的任意组合来实现。当使用软件实现时,可以将这些功能存储在计算机可读介质中或者作为计算机可读介质上的一个或多个指令或代码进行传输。
以上所述的具体实施方式,对本发明的目的、技术方案和有益效果进行了进一步详细说明,所应理解的是,以上所述仅为本发明的具体实施方式而已,并不用于限定本发明的保护范围,凡在本发明的技术方案的基础之上,所做的任何修改、等同替换、改进等,均应包括在本发明的保护范围之内。

Claims (17)

  1. 一种多方联合训练针对IoT机具的风险评估模型的方法,其中,所述多方包括第一计算节点、第二计算节点和第一支付平台,各自维护风险评估模型中的部分参数;所述第一计算节点与第一IoT机具相关联,存储基于所述第一IoT机具的机具隐私数据而确定的机具特征,所述第二计算节点与绑定所述第一IoT机具的第一商户相关联,存储所述第一商户的商户隐私数据,所述第一支付平台存储与所述第一商户相关的支付隐私数据和指示所述第一IoT机具风险情况的风险标签;所述方法应用于所述第一计算节点,所述方法包括:
    基于所述机具特征及所述第一计算节点维护的第一参数,确定第一中间结果;
    利用安全多方计算MPC技术,提供所述第一中间结果,用于结合所述第二计算节点基于所述商户隐私数据及其维护的第二参数确定的第二中间结果,所述第一支付平台基于所述支付隐私数据及其维护的第三参数和所述风险标签确定的第三中间结果,确定针对所述第一IoT机具的训练损失;
    获取所述训练损失,并且,利用所述训练损失调整所述第一参数。
  2. 根据权利要求1所述的方法,其中,在基于所述机具特征及所述第一计算节点维护的第一参数和第一计算式,确定第一中间结果之前,所述方法还包括:
    获取所述第一IoT机具中存储的所述机具隐私数据,并且,对所述机具隐私数据进行累计处理或向量表征处理,得到所述机具特征;或,
    从所述第一IoT机具接收所述机具特征,所述机具特征由所述第一IoT机具对其自身存储的所述机具隐私数据进行累计处理或向量表征处理而得到。
  3. 根据权利要求1或2所述的方法,其中,所述机具隐私数据包括以下中的一种或多种:所述第一IoT机具的开机时间、关机时间和位置信息,对所述第一IoT机具进行解绑、换绑产生的操作数据。
  4. 根据权利要求1所述的方法,其中,所述MPC技术包括同态加密技术,所述第一计算节点和第二计算节点中还存储基于所述同态加密技术生成的第一公钥,所述第一支付平台中还存储所述第一公钥和对应的第一私钥;
    其中利用安全多方计算MPC技术,提供所述第一中间结果,包括:
    利用所述第一公钥对所述第一中间结果进行加密,得到第一加密结果;
    将所述第一加密结果发送至所述第二计算节点,以使所述第二计算节点对所述第一加密结果和其对第二中间结果加密得到的第二加密结果进行第一同态加操作,进而使所述第一支付平台对所述第一同态加操作得到的第一操作结果和其对第三中间结果加密 得到的第三加密结果进行第二同态加操作,并利用所述私钥对所述第二同态加操作得到的第二操作结果进行解密,得到所述训练损失;
    其中获取所述训练损失,包括:
    从所述第一支付平台接收所述训练损失。
  5. 一种多方联合训练针对IoT机具的风险评估模型的方法,其中,所述多方包括第一计算节点、第二计算节点和第一支付平台,各自维护风险评估模型中的部分参数;所述第一计算节点与第一IoT机具相关联,存储与所述第一IoT机具相关的机具特征,所述第二计算节点与绑定所述第一IoT机具的第一商户相关联,存储所述第一商户的商户隐私数据,所述第一支付平台存储与所述第一商户相关的支付隐私数据和指示所述第一IoT机具风险情况的风险标签;所述方法应用于所述第二计算节点,所述方法包括:
    基于所述商户隐私数据及所述第二计算节点维护的第二参数,确定第二中间结果;
    利用安全多方计算MPC技术,提供所述第二中间结果,用于结合所述第一计算节点基于所述机具特征及其维护的第一参数确定的第一中间结果,所述第一支付平台基于所述支付隐私数据及其维护的第三参数和所述风险标签确定的第三中间结果,确定针对所述第一IoT机具的训练损失;
    获取所述训练损失,并且,利用所述训练损失调整所述第二参数。
  6. 根据权利要求5所述的方法,其中,在基于所述商户隐私数据及所述第二计算节点维护的第二参数,确定第二中间结果之前,所述方法还包括:
    获取所述第一商户中存储的所述商户隐私数据。
  7. 根据权利要求5所述的方法,其中,所述第二计算节点为可信计算节点,在基于所述商户隐私数据及所述第二计算节点维护的第二参数,确定第二中间结果之前,所述方法还包括:
    生成第二公钥和第二私钥,并且,将所述第二公钥发送至多个商户,所述多个商户中包括所述第一商户;
    从所述第一商户接收加密隐私数据,所述加密隐私数据由所述第一商户利用所述第二公钥对所述商户隐私数据进行加密而得到;
    利用所述第二私钥对所述加密隐私数据进行解密,得到所述商户隐私数据。
  8. 根据权利要求5-7中任一项所述的方法,其中,所述商户隐私数据包括所述第一商户在所述第一支付平台以外的其他支付平台中产生的交易信息,具体包括以下中的一种或多种:交易金额、交易地点、交易时间、商品种类、风险事件。
  9. 权利要求5所述的方法,其中,所述MPC技术包括同态加密技术,所述第一计 算节点和第二计算节点中还存储基于所述同态加密技术生成的第一公钥,所述第一支付平台中还存储所述第一公钥和对应的第一私钥;
    其中利用安全多方计算MPC技术,提供所述第二中间结果,包括:
    利用所述第一公钥对所述第二中间结果进行加密,得到第二加密结果;
    从所述第一计算节点接收其利用所述第一公钥对所述第一中间结果加密得到的第一加密结果;
    对所述第一加密结果和第二加密结果进行同态加操作,得到第一操作结果;
    将所述第一操作结果发送至所述第一支付平台,以使所述第一支付平台对所述第一操作结果和其对第三中间结果加密得到的第三加密结果进行第二同态加操作,并利用所述私钥对所述第二同态加操作得到的第二操作结果进行解密,得到所述训练损失;
    其中获取所述训练损失,包括:
    从所述第一支付平台接收所述训练损失。
  10. 一种多方联合训练针对IoT机具的风险评估模型的方法,其中,所述多方包括第一计算节点、第二计算节点和第一支付平台,各自维护风险评估模型中的部分参数;所述第一计算节点与第一IoT机具相关联,存储基于所述第一IoT机具的机具隐私数据而确定的机具特征,所述第二计算节点与绑定所述第一IoT机具的第一商户相关联,存储所述第一商户的商户隐私数据,所述第一支付平台存储与所述第一商户相关的支付隐私数据和指示所述第一IoT机具风险情况的风险标签;所述方法应用于所述第一支付平台,所述方法包括:
    基于所述支付隐私数据及所述第一支付平台维护的第三参数和所述风险标签,确定第三中间结果;
    利用安全多方计算MPC技术,提供所述第三中间结果,用于结合所述第一计算节点基于所述机具特征及其维护的第一参数确定的第一中间结果,所述第二计算节点基于所述商户隐私数据及其维护的第二参数确定的第二中间结果,确定针对所述第一IoT机具的训练损失;
    获取所述训练损失,并且,利用所述训练损失调整所述第三参数。
  11. 根据权利要求10所述的方法,其中,所述支付隐私特征包括所述第一商户与所述第一支付平台的签约信息,在所述第一支付平台中产生的交易信息,具体包括以下中的一种或多种:交易用户的用户信息、交易金额、交易地点、交易时间、商品种类、风险事件。
  12. 权利要求10所述的方法,其中,所述MPC技术包括同态加密技术,所述第一 计算节点和第二计算节点中还存储基于所述同态加密技术生成的第一公钥,所述第一支付平台中还存储所述第一公钥和对应的第一私钥;
    其中利用安全多方计算MPC技术,提供所述第三中间结果,包括:
    从所述第二计算节点接收第一操作结果,所述第一操作结果通过对第二加密结果和从第一计算节点接收的第一加密结果进行第一同态加操作而得到,所述第二加密结果是利用所述第一公钥对所述第二中间结果进行加密而得到,所述第一加密结果是利用所述第一公钥对所述第一中间结果进行加密而得到;
    利用所述第一公钥对所述第三中间结果进行加密,得到第三加密结果;
    对所述第一操作结果和所述第三加密结果进行第二同态加操作,得到第二操作结果;
    其中获取所述训练损失,包括:
    利用所述私钥对所述第二操作结果进行解密,得到所述训练损失。
  13. 一种多方联合训练针对IoT机具的风险评估模型的装置,其中,所述多方包括第一计算节点、第二计算节点和第一支付平台,各自维护风险评估模型中的部分参数;所述第一计算节点与第一IoT机具相关联,存储基于所述第一IoT机具的机具隐私数据而确定的机具特征,所述第二计算节点与绑定所述第一IoT机具的第一商户相关联,存储所述第一商户的商户隐私数据,所述第一支付平台存储与所述第一商户相关的支付隐私数据和指示所述第一IoT机具风险情况的风险标签;所述装置集成于所述第一计算节点,所述装置包括:
    中间结果确定单元,配置为基于所述机具特征及所述第一计算节点维护的第一参数,确定第一中间结果;
    中间结果提供单元,配置为利用安全多方计算MPC技术,提供所述第一中间结果,用于结合所述第二计算节点基于所述商户隐私数据及其维护的第二参数确定的第二中间结果,所述第一支付平台基于所述支付隐私数据及其维护的第三参数和所述风险标签确定的第三中间结果,确定针对所述第一IoT机具的训练损失;
    损失获取单元,配置为获取所述训练损失;
    调参单元,配置为利用所述训练损失调整所述第一参数。
  14. 一种多方联合训练针对IoT机具的风险评估模型的装置,其中,所述多方包括第一计算节点、第二计算节点和第一支付平台,各自维护风险评估模型中的部分参数;所述第一计算节点与第一IoT机具相关联,存储与所述第一IoT机具相关的机具特征,所述第二计算节点与绑定所述第一IoT机具的第一商户相关联,存储所述第一商户的商户隐私数据,所述第一支付平台存储与所述第一商户相关的支付隐私数据和指示所述第 一IoT机具风险情况的风险标签;所述装置集成于所述第二计算节点,所述装置包括:
    中间结果确定单元,配置为基于所述商户隐私数据及所述第二计算节点维护的第二参数,确定第二中间结果;
    中间结果提供单元,配置为利用安全多方计算MPC技术,提供所述第二中间结果,用于结合所述第一计算节点基于所述机具特征及其维护的第一参数确定的第一中间结果,所述第一支付平台基于所述支付隐私数据及其维护的第三参数和所述风险标签确定的第三中间结果,确定针对所述第一IoT机具的训练损失;
    损失获取单元,配置为获取所述训练损失;
    调参单元,配置为利用所述训练损失调整所述第二参数。
  15. 一种多方联合训练针对IoT机具的风险评估模型的装置,其中,所述多方包括第一计算节点、第二计算节点和第一支付平台,各自维护风险评估模型中的部分参数;所述第一计算节点与第一IoT机具相关联,存储基于所述第一IoT机具的机具隐私数据而确定的机具特征,所述第二计算节点与绑定所述第一IoT机具的第一商户相关联,存储所述第一商户的商户隐私数据,所述第一支付平台存储与所述第一商户相关的支付隐私数据和指示所述第一IoT机具风险情况的风险标签;所述装置集成于所述第一支付平台,所述装置包括:
    中间结果确定单元,配置为基于所述支付隐私数据及所述第一支付平台维护的第三参数和所述风险标签,确定第三中间结果;
    中间结果提供单元,配置为利用安全多方计算MPC技术,提供所述第三中间结果,用于结合所述第一计算节点基于所述机具特征及其维护的第一参数确定的第一中间结果,所述第二计算节点基于所述商户隐私数据及其维护的第二参数确定的第二中间结果,确定针对所述第一IoT机具的训练损失;
    损失获取单元,配置为获取所述训练损失;
    调参单元,配置为利用所述训练损失调整所述第三参数。
  16. 一种计算机可读存储介质,其上存储有计算机程序,其中,当所述计算机程序在计算机中执行时,令计算机执行权利要求1-12中任一项的所述的方法。
  17. 一种计算设备,包括存储器和处理器,其中,所述存储器中存储有可执行代码,所述处理器执行所述可执行代码时,实现权利要求1-12中任一项所述的方法。
PCT/CN2020/124289 2019-12-13 2020-10-28 多方联合训练针对IoT机具的风险评估模型的方法及装置 WO2021114922A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201911284459.8 2019-12-13
CN201911284459.8A CN111027981B (zh) 2019-12-13 2019-12-13 多方联合训练针对IoT机具的风险评估模型的方法及装置

Publications (1)

Publication Number Publication Date
WO2021114922A1 true WO2021114922A1 (zh) 2021-06-17

Family

ID=70208989

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/124289 WO2021114922A1 (zh) 2019-12-13 2020-10-28 多方联合训练针对IoT机具的风险评估模型的方法及装置

Country Status (3)

Country Link
CN (1) CN111027981B (zh)
TW (1) TWI720888B (zh)
WO (1) WO2021114922A1 (zh)

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111027981B (zh) * 2019-12-13 2021-04-27 支付宝(杭州)信息技术有限公司 多方联合训练针对IoT机具的风险评估模型的方法及装置
CN113034153B (zh) * 2021-03-26 2022-08-02 支付宝(杭州)信息技术有限公司 一种基于共享学习的被欺诈风险识别方法、装置及设备
CN113569263A (zh) * 2021-07-30 2021-10-29 拉扎斯网络科技(上海)有限公司 跨私域数据的安全处理方法、装置及电子设备
CN114205155B (zh) * 2021-12-07 2023-09-15 四川启睿克科技有限公司 一种基于安全多方计算的供应商风险评估系统及方法
CN114817943A (zh) * 2022-02-28 2022-07-29 中国银联股份有限公司 一种数据匹配方法、装置、设备及介质

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109241016A (zh) * 2018-08-14 2019-01-18 阿里巴巴集团控股有限公司 多方安全计算方法及装置、电子设备
CN109359470A (zh) * 2018-08-14 2019-02-19 阿里巴巴集团控股有限公司 多方安全计算方法及装置、电子设备
WO2019051116A1 (en) * 2017-09-06 2019-03-14 Hodo Patrick G SECURING PRIVATE INFORMATION IN COMPUTER DEVICE TRANSACTIONS HOSTED BY SEVERAL PARTS
CN109543453A (zh) * 2018-11-30 2019-03-29 北京八分量信息科技有限公司 一种基于安全多方计算的订单信息处理方法、装置及系统
CN110335037A (zh) * 2019-04-19 2019-10-15 矩阵元技术(深圳)有限公司 基于区块链和多方安全计算的交易方法、装置及存储介质
CN111027981A (zh) * 2019-12-13 2020-04-17 支付宝(杭州)信息技术有限公司 多方联合训练针对IoT机具的风险评估模型的方法及装置

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9819650B2 (en) * 2014-07-22 2017-11-14 Nanthealth, Inc. Homomorphic encryption in a healthcare network environment, system and methods
US9860221B2 (en) * 2015-03-10 2018-01-02 Intel Corporation Internet of things group formation using a key-based join protocol
CN106504445B (zh) * 2016-11-07 2021-08-31 中国银联股份有限公司 基于物联网的pos终端监控方法以及监控系统
US11599883B2 (en) * 2017-03-08 2023-03-07 Visa International Service Association System and method for fraud risk analysis in IoT
US20180260815A1 (en) * 2017-03-09 2018-09-13 International Business Machines Corporation Internet of things recognition of questionable activity
CN109003075A (zh) * 2017-06-07 2018-12-14 阿里巴巴集团控股有限公司 一种风险识别方法及装置
CN107835247B (zh) * 2017-11-08 2020-08-28 中国科学技术大学 一种信用认证、保障系统及方法
CN110377590B (zh) * 2018-04-11 2021-01-19 深圳富桂精密工业有限公司 基于物联网的数据采集系统、方法及计算机可读存储介质
CN110365703B (zh) * 2019-07-30 2022-01-21 国家电网有限公司 物联网终端异常状态检测方法、装置及终端设备

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2019051116A1 (en) * 2017-09-06 2019-03-14 Hodo Patrick G SECURING PRIVATE INFORMATION IN COMPUTER DEVICE TRANSACTIONS HOSTED BY SEVERAL PARTS
CN109241016A (zh) * 2018-08-14 2019-01-18 阿里巴巴集团控股有限公司 多方安全计算方法及装置、电子设备
CN109359470A (zh) * 2018-08-14 2019-02-19 阿里巴巴集团控股有限公司 多方安全计算方法及装置、电子设备
CN109543453A (zh) * 2018-11-30 2019-03-29 北京八分量信息科技有限公司 一种基于安全多方计算的订单信息处理方法、装置及系统
CN110335037A (zh) * 2019-04-19 2019-10-15 矩阵元技术(深圳)有限公司 基于区块链和多方安全计算的交易方法、装置及存储介质
CN111027981A (zh) * 2019-12-13 2020-04-17 支付宝(杭州)信息技术有限公司 多方联合训练针对IoT机具的风险评估模型的方法及装置

Also Published As

Publication number Publication date
TWI720888B (zh) 2021-03-01
CN111027981B (zh) 2021-04-27
CN111027981A (zh) 2020-04-17
TW202123123A (zh) 2021-06-16

Similar Documents

Publication Publication Date Title
WO2021114922A1 (zh) 多方联合训练针对IoT机具的风险评估模型的方法及装置
CN110245510B (zh) 用于预测信息的方法和装置
CN111770201B (zh) 一种数据验证方法、装置及设备
US10547444B2 (en) Cloud encryption key broker apparatuses, methods and systems
US20180349968A1 (en) Systems and methods for product review management with distributed database
JP2024054307A (ja) 暗号論的に署名された記録のセキュアな交換
CN111770200B (zh) 一种信息共享方法和系统
US20200084051A1 (en) Systems and methods for preserving privacy and incentivizing third-party data sharing
US20200019865A1 (en) System and method for processing data and managing information
CN111291401B (zh) 基于隐私保护的业务预测模型训练方法和装置
US20120304273A1 (en) Tokenizing Sensitive Data
US11811933B2 (en) System and method for fair, secure n-party computation using at least one blockchain
CN109034796A (zh) 基于联盟链的交易监管方法、电子装置及可读存储介质
US11068881B2 (en) System for resource distribution within an offline environment
EP3867849A1 (en) Secure digital wallet processing system
CN112788001A (zh) 一种基于数据加密的数据处理业务处理方法、装置及设备
CN112199697A (zh) 基于共享根密钥的信息处理方法、装置、设备及介质
CN112000978B (zh) 隐私数据的输出方法、数据处理系统及存储介质
CN111861462A (zh) 基于区块链的金融产品交易方法及装置
CN115599959A (zh) 数据共享方法、装置、设备及存储介质
JP6889073B2 (ja) 仮想通貨管理装置、仮想通貨管理方法、およびプログラム
WO2019123676A1 (ja) 仮想通貨管理システム、プログラム
CN113378198B (zh) 一种保护用户标识的模型的联邦训练系统、方法及装置
CN117077816B (zh) 联邦模型的训练方法及系统
CN112822152B (zh) 定向信息展示处理方法及相关设备

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20897761

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20897761

Country of ref document: EP

Kind code of ref document: A1