A model training method and device based on privacy protection

Abstract

Embodiments of the present specification provide a privacy protection based model training method and device, obtaining an initial event sequence set used for training a target business model; for each initial event sequence in the initial event sequence set, according to policy information of a first enhancement strategy, based on a plurality of data enhancement units used for data perturbation modification, performing a first number of data enhancement operations on the initial event sequence to obtain a corresponding enhanced event sequence, wherein one data enhancement operation is performed using one data enhancement unit; the policy information at least indicates the first number; using a first enhanced sequence set formed by each enhanced event sequence to perform model training to obtain a privacy protected first business model.

Description

Technical Field

[0001] This specification relates to the field of privacy and security technology, and in particular to a privacy-preserving model training method and apparatus. Background Technology

[0002] Black-box attacks targeting member inference refer to attacks that, given access to specific data and the API interface of a target business model (i.e., black-box access to the target business model), determine whether that specific data belongs to the training dataset of the target business model. Training data in the training dataset typically includes real user data, which may contain user privacy data. If specific data is determined to belong to the training dataset, then the user's privacy data is obtained, potentially leading to a breach of user privacy.

[0003] For example, if the target business model is an anti-fraud model in an anti-fraud scenario, the training data of its training dataset includes users' real transfer and consumption information. If the black box of the target business model is successfully attacked, it can easily lead to the leakage of user privacy, such as users' real transfer and consumption information. Summary of the Invention

[0004] This specification provides one or more embodiments of a privacy-preserving model training method and apparatus to train a privacy-preserving business model.

[0005] Based on the first aspect, a privacy-preserving model training method is provided, including:

[0006] Obtain an initial set of event sequences for training the target business model;

[0007] For each initial event sequence in the initial event sequence set, based on the strategy information of the first enhancement strategy and a preset number of data enhancement units for data perturbation modification, a first number of data enhancement operations are performed on the initial event sequence to obtain the corresponding enhanced event sequence. Each data enhancement operation is performed using one data enhancement unit. The strategy information indicates at least the first number.

[0008] The model is trained using the first set of augmented sequences formed by the various augmented event sequences to obtain the first business model for privacy protection.

[0009] In one alternative implementation, the data augmentation unit is used to perturb and modify one or more of the following: the order, number, and attributes of events in an event sequence.

[0010] In one alternative implementation, it further includes:

[0011] Obtain the first performance result corresponding to the first business model;

[0012] Based on the first performance result and the first enhancement strategy, a second enhancement strategy is determined for the next round.

[0013] In one alternative implementation, it further includes:

[0014] Obtain the first performance result corresponding to the first business model;

[0015] Based at least on the first performance result and the first business model, determine the target privacy protection model corresponding to the target business model, or its alternative model.

[0016] In one optional implementation, determining the target privacy protection model corresponding to the target business model, or its alternative model, includes:

[0017] A plurality of second enhanced sequence sets corresponding to the initial event sequence set are obtained, wherein each second enhanced sequence set is obtained by performing data enhancement operations on the initial event sequence set based on the policy information of the corresponding enhancement strategy and the plurality of data enhancement units;

[0018] Each set of second enhancement sequences is used to train the model, resulting in several corresponding second business models;

[0019] Obtain several second performance results corresponding to several second business models;

[0020] Based on the first performance result and the second performance result, the target privacy protection model or the alternative model is determined from the first business model and several second business models.

[0021] In one alternative implementation, it further includes:

[0022] Obtain the target performance result corresponding to the target business model, which is trained using the initial event sequence set;

[0023] The step of determining the target privacy protection model, or the alternative model, from the first business model and several second business models includes:

[0024] Determine the performance changes of the first performance result and several second performance results relative to the target performance result;

[0025] Based on the changes in performance, the business model with better performance in terms of performance changes is determined from the first business model and the plurality of second business models, and is used as the target privacy protection model.

[0026] In one alternative implementation, the first performance result includes at least one of the following: business forecast accuracy and the probability of a successful black-box attack.

[0027] In one alternative implementation, the strategy information further includes: a disturbance intensity value, wherein the first number of data augmentation operations includes a first augmentation operation, the first augmentation operation utilizing a first augmentation operation unit to perform data perturbation conforming to the disturbance intensity value.

[0028] In one alternative implementation, the data perturbation modification includes perturbation modification of the order and / or number of events in the event sequence; the plurality of data augmentation units include data augmentation units for performing at least one of the following operations:

[0029] Delete a portion of the events from the event sequence;

[0030] Extract several first events from the event sequence, and then sample and rearrange the several first events without replacement and put them back into the event sequence;

[0031] Several second events are extracted from the event sequence, and the several second events are rearranged with replacement and then returned to the event sequence.

[0032] Modify the third event in the event sequence to be the fourth event;

[0033] Maintain the event sequence.

[0034] In one alternative implementation, the data perturbation modification includes perturbation modification of attributes of events in the event sequence; the plurality of data augmentation units include data augmentation units for performing at least one of the following operations:

[0035] For the first numerical attribute of the fifth event in the event sequence, the attribute value of the first attribute is modified by perturbation using the Gaussian distribution corresponding to each attribute value of the first attribute.

[0036] For the second attribute of the category of the sixth event in the event sequence, the attribute value of the second attribute is perturbed and modified by using the prior distribution corresponding to each attribute value of the second attribute.

[0037] In one optional implementation, the strategy information further includes a perturbation strength value; the step of modifying the attribute value of the first attribute by perturbation using the Gaussian distribution corresponding to each attribute value of the first attribute includes:

[0038] Based on the Gaussian distribution corresponding to the values ​​of each attribute of the first attribute, the perturbation value corresponding to the first attribute of the fifth event is generated.

[0039] Based on the perturbation value and the perturbation intensity value, the attribute value of the first attribute is modified by perturbation.

[0040] In one alternative implementation, the data augmentation unit used in a single data augmentation operation is randomly selected from the plurality of data augmentation units in an equally probabilistic manner.

[0041] According to the second aspect, a privacy-preserving model training device is provided, comprising:

[0042] The first acquisition module is configured to acquire an initial set of event sequences for training the target business model;

[0043] The data augmentation module is configured to perform a first number of data augmentation operations on each initial event sequence in the initial event sequence set, based on the strategy information of a first augmentation strategy and a preset number of data augmentation units for data perturbation modification, to obtain a corresponding augmented event sequence. Each data augmentation operation is performed using one data augmentation unit. The strategy information indicates at least the first number.

[0044] The model training module is configured to train the model using a first set of augmented sequences formed by various augmented event sequences, thereby obtaining a first business model with privacy protection.

[0045] In one alternative implementation, the data augmentation unit is used to perturb and modify one or more of the following: the order, number, and attributes of events in an event sequence.

[0046] In one alternative embodiment, the device further includes:

[0047] The second acquisition module is configured to acquire the first performance result corresponding to the first business model;

[0048] The first determining module is configured to determine a second enhancement strategy for the next round based on the first performance result and the first enhancement strategy.

[0049] In one alternative embodiment, the device further includes:

[0050] The third acquisition module is configured to acquire the first performance result corresponding to the first business model;

[0051] The second determining module is configured to determine, at least based on the first performance result and the first business model, the target privacy protection model corresponding to the target business model, or its alternative model.

[0052] In one alternative implementation, the second determining module includes:

[0053] The first obtaining submodule is configured to obtain a plurality of second enhanced sequence sets corresponding to the initial event sequence set, wherein each second enhanced sequence set is obtained by performing data enhancement operations on the initial event sequence set based on the policy information of the corresponding enhancement strategy and the plurality of data enhancement units;

[0054] The model training submodule is configured to train the model using each set of second augmentation sequences to obtain several corresponding second business models.

[0055] The second acquisition submodule is configured to acquire several second performance results corresponding to several second business models;

[0056] The determination submodule is configured to determine the target privacy protection model or the alternative model from the first business model and several second business models based on the first performance result and the second performance result.

[0057] In one optional implementation, the second determining module further includes:

[0058] The third acquisition submodule is configured to obtain the target performance result corresponding to the target business model, wherein the target business model is trained using the initial event sequence set;

[0059] The determining submodule is specifically configured to determine the performance changes of the first performance result and several second performance results relative to the target performance result.

[0060] Based on the changes in performance, the business model with better performance in terms of performance changes is determined from the first business model and the plurality of second business models, and is used as the target privacy protection model.

[0061] In one alternative implementation, the first performance result includes at least one of the following: business forecast accuracy and the probability of a successful black-box attack.

[0062] In one alternative implementation, the strategy information further includes: a disturbance intensity value, wherein the first number of data augmentation operations includes a first augmentation operation, the first augmentation operation utilizing a first augmentation operation unit to perform data perturbation conforming to the disturbance intensity value.

[0063] In one alternative implementation, the data perturbation modification includes perturbation modification of the order and / or number of events in the event sequence; the plurality of data augmentation units include data augmentation units for performing at least one of the following operations:

[0064] Delete a portion of the events from the event sequence;

[0065] Extract several first events from the event sequence, and then sample and rearrange the several first events without replacement and put them back into the event sequence;

[0066] Several second events are extracted from the event sequence, and the several second events are rearranged with replacement and then returned to the event sequence.

[0067] Modify the third event in the event sequence to be the fourth event;

[0068] Maintain the event sequence.

[0069] In one alternative implementation, the data perturbation modification includes perturbation modification of attributes of events in the event sequence; the plurality of data augmentation units include data augmentation units for performing at least one of the following operations:

[0070] For the first numerical attribute of the fifth event in the event sequence, the attribute value of the first attribute is modified by perturbation using the Gaussian distribution corresponding to each attribute value of the first attribute.

[0071] For the second attribute of the category of the sixth event in the event sequence, the attribute value of the second attribute is perturbed and modified by using the prior distribution corresponding to each attribute value of the second attribute.

[0072] In one optional implementation, the strategy information further includes a perturbation intensity value; a data augmentation unit is specifically configured to generate a perturbation value corresponding to the first attribute of the fifth event based on the Gaussian distribution corresponding to each attribute value of the first attribute;

[0073] Based on the perturbation value and the perturbation intensity value, the attribute value of the first attribute is modified by perturbation.

[0074] In one alternative implementation, the data augmentation unit used in a single data augmentation operation is randomly selected from the plurality of data augmentation units in an equally probabilistic manner.

[0075] According to a third aspect, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method described in the first aspect.

[0076] According to a fourth aspect, a computing device is provided, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, it implements the method described in the first aspect.

[0077] According to the method and apparatus provided in the embodiments of this specification, based on the strategy information of the first enhancement strategy, and based on a plurality of preset data enhancement units for data perturbation modification, data enhancement operations are performed on the initial event sequence to mask the real event information in the initial event sequence, thereby obtaining the corresponding enhanced event sequence. Then, the enhanced event sequence that masks the real event information is used for model training to obtain a privacy-protected first business model, which can improve the model's defense capability to a certain extent and reduce the risk of leakage of real event information in the model training data. Attached Figure Description

[0078] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0079] Figure 1 This is a schematic diagram illustrating the implementation framework of one embodiment disclosed in this specification;

[0080] Figure 2 A schematic flowchart of a privacy-preserving model training method provided for an embodiment;

[0081] Figure 3 A schematic diagram of a data augmentation process provided for an embodiment;

[0082] Figure 4 This is a schematic block diagram of a privacy-preserving model training device provided in an embodiment. Detailed Implementation

[0083] The technical solutions of the embodiments of this specification will now be described in detail with reference to the accompanying drawings.

[0084] This specification discloses a privacy-preserving model training method and apparatus. The application scenarios and technical concepts of the method are first introduced below:

[0085] Currently, training data used to train business models typically includes collected real user data and its corresponding label data. However, black-box attacks targeting member inference have a certain probability of inferring whether the specific data used is the training data for the business model, which can easily lead to the leakage of real user data from the training data.

[0086] In view of this, the inventors propose a privacy-preserving model training method. Figure 1This diagram illustrates an implementation framework of an embodiment disclosed in this specification. In this embodiment, a system implementing a privacy-preserving model training method is associated with multiple preset data augmentation units, such as data augmentation unit 1, data augmentation unit 2, ..., data augmentation unit N. These multiple data augmentation units are used to modify the data perturbation of event sequences to mask real event information in the actual event sequences. Specifically, the system may include: a first acquisition module acquiring multiple event sequences for training a target business model to form an initial event sequence set, wherein each event sequence is a real event sequence collected for a sample object, and each event sequence includes several events arranged in chronological order of the events triggered by the sample object. Different event sequences may correspond to different sample objects or the same sample object.

[0087] For example, if the sample objects are users of an electronic payment platform, the events in each event sequence could be a series of events triggered by each user within a certain period of time based on the electronic payment platform. Similarly, if the sample objects are merchants on an e-commerce platform, the events in each event sequence could be a series of events triggered by each merchant within a certain period of time based on the e-commerce platform. It is understood that the above are merely illustrative examples of sample objects and do not constitute a limitation on the specific type of sample objects. Sample objects can be any object that can trigger a series of events to form an event sequence, such as users of social media platforms or financial management platforms.

[0088] The data augmentation module, for each initial event sequence in the initial event sequence set, performs a certain number of data augmentation operations on the initial event sequence based on the policy information of the current augmentation strategy and multiple data augmentation units. This masks the real event information (e.g., the order of events, event attributes, and the number of events) in the initial event sequence, resulting in an augmented event sequence corresponding to the initial event sequence. Each data augmentation operation is performed using one data augmentation unit, and the policy information indicates at least the number of such augmentation operations. Furthermore, the model training module uses the first augmented sequence set formed by the augmented event sequences to train a model, obtaining a first privacy-preserving business model.

[0089] The target business model can be obtained by training the initialization model based on the initial event sequence set and the label data corresponding to each initial event sequence.

[0090] In one implementation, the model training module can use the first set of augmented sequences and the original label data corresponding to each augmented event sequence therein to further train the target business model, thereby obtaining a privacy-preserving first business model. This ensures the accuracy of the business prediction of the first business model to a certain extent, while improving the defense capability and generalization of the first business model, increasing the attack difficulty of the first business model, and achieving the purpose of privacy protection.

[0091] In another implementation, the model training module can use the first set of augmented sequences and the original label data corresponding to each augmented event sequence in it to train the initial model and obtain the first business model with privacy protection. The first business model can be obtained by directly using the augmented event sequences that mask the real event sequences, thus better achieving privacy protection.

[0092] The original label data corresponding to each augmented event sequence is the label data of the initial event sequence corresponding to that augmented event sequence.

[0093] In one exemplary scenario, the target business model can be a fraud detection model. Correspondingly, each initial event sequence in the initial event sequence set can be an event sequence collected from each user of the electronic payment platform. Each event in this sequence can include, but is not limited to, user transfer events, consumption events, account opening events, and bank card binding events. Tag data can include tags characterizing whether the user corresponding to the event sequence has been defrauded. The electronic payment platform can be any platform that supports asset consumption management.

[0094] In another exemplary scenario, the target business model can be a merchant classification model. Correspondingly, each initial event sequence in the initial event sequence set can be an event sequence collected from each merchant on the e-commerce platform. Each event in the event sequence can include, but is not limited to, merchant transfer events, payment collection events, and consumption events. Tag data can include category tags representing the merchants corresponding to the event sequences.

[0095] In this embodiment, data augmentation is performed on the initial event sequence to mask the real event information in the initial event sequence, resulting in a corresponding augmented event sequence. The augmented event sequence, which masks the real event information, is then used to train the model, resulting in a privacy-preserving first business model. This can improve the model's defense capabilities to a certain extent and reduce the risk of leakage of real event information (privacy data) in the model training data.

[0096] The privacy-preserving model training method provided in this specification will be described in detail below with reference to specific embodiments.

[0097] Figure 2A flowchart of a privacy-preserving model training method according to one embodiment of this specification is shown. This method can be implemented using any device, equipment, platform, device cluster, etc., with computing and processing capabilities. The method includes the following steps S210-S230:

[0098] S210: Obtain an initial set of event sequences for training the target business model. This initial set of event sequences includes multiple real event sequences collected for the sample objects. Each event sequence includes several events, which are events triggered sequentially by a certain sample object within a certain time period. The events included in the event sequence can be ordered according to the order of their occurrence time. Each event corresponds to several attributes, which can describe the corresponding event from various dimensions.

[0099] Each event's attributes may include, but are not limited to: basic event attributes, such as: the sample object identifier corresponding to the event, the event's own identifier and category (e.g., transfer event, consumption event, etc.), the event occurrence time, the event's geographical location, the device type where the event occurred, and the corresponding operation path; category-related attributes, such as: when the event category is a transfer event, category-related attributes include: the amount of funds transferred, the identifier of the fund transferee, the identifier of the fund transferor, and transfer remarks information. When the event category is a consumption event, category-related attributes include: the amount of funds consumed, the identifier of the fund transferee and related information, the identifier of the fund transferor and related information, and information on the items purchased (or sold) during the consumption. The corresponding operation path can refer to the set of sequential operations executed by the sample object when triggering the event. For example, if the event is a transfer event, the corresponding operation path may include: clicking the fund transferee icon, clicking to enter the corresponding transfer operation page, selecting the transfer method (e.g., bank card, account balance), entering the transfer amount, triggering the transfer function, and entering the transfer password.

[0100] The target business model can be obtained by training the initialization model based on the initial event sequence set and the corresponding label data of each initial event sequence. The target business model can be any business prediction model based on event sequences, such as a fraud detection model or a merchant classification model.

[0101] S220: For each initial event sequence in the initial event sequence set, based on the strategy information of the first enhancement strategy and using a preset number of data enhancement units for data perturbation modification, a first number of data enhancement operations are performed on the initial event sequence to obtain the corresponding enhanced event sequence. Each data enhancement operation is performed using one data enhancement unit; the strategy information indicates at least the first number.

[0102] First, multiple preset data augmentation units for data perturbation modification can be obtained. These data augmentation units are used to perturb and modify one or more of the following: the order, quantity, and attributes of events in the event sequence, in order to mask the real event information in the original event sequence. The current augmentation strategy, i.e., the first augmentation strategy, is obtained. This strategy information indicates at least a first number, representing the number of data augmentation operations required for the initial event sequence.

[0103] Specifically, for each initial event sequence in the initial event sequence set, the following steps are performed to obtain the enhanced event sequence corresponding to each initial event sequence. Specifically, in each data augmentation operation, a data augmentation unit is selected from the plurality of data augmentation units as the target data augmentation unit. Data augmentation operations are performed on the event sequence based on the target data augmentation unit to obtain the output sequence, which is then used as the input sequence for the next round of data augmentation operations. This process continues until a first number of data augmentation operations are performed on the event sequence to obtain the enhanced event sequence corresponding to the initial event sequence. The input sequence for the first data augmentation operation is the initial event sequence; the input sequence for subsequent data augmentation operations is the output sequence of the previous round of data augmentation operations.

[0104] In one implementation, the data augmentation unit used in each data augmentation operation is randomly selected from a set of preset data augmentation units in an equally probabilistic manner.

[0105] S230: A first set of augmented sequences, formed by various augmented event sequences, is used to train the model, resulting in a privacy-preserving first business model. In one implementation, the first set of augmented sequences and the original label data corresponding to each augmented event sequence can be used to further train the target business model trained based on the initial event sequences, resulting in a privacy-preserving first business model. This approach can, to a certain extent, ensure the accuracy of the first business model's business predictions while increasing the attack difficulty of the first business model, thus achieving the goal of privacy protection. In another implementation, the first set of augmented sequences and the original label data corresponding to each augmented event sequence can be used to train the initial model, resulting in a privacy-preserving first business model. This directly utilizes augmented event sequences that mask the real event sequences to train the first business model, further enhancing privacy protection.

[0106] The original label data corresponding to each augmented event sequence is the label data of the initial event sequence corresponding to that augmented event sequence.

[0107] In this embodiment, data augmentation is performed on the initial event sequence to mask the real event information within it, resulting in an augmented event sequence. This augmented sequence, now masking the real event information, is then used for model training to obtain a privacy-preserving first business model. This approach can improve the model's defense capabilities to some extent and reduce the risk of real event information leakage from the training data. Furthermore, this embodiment provides a method for data augmentation of event sequences, which can effectively increase the amount of training data available for model training, thus facilitating model development.

[0108] Furthermore, when the first business model is obtained by further training the target business model using the first enhanced sequence set, the training data of the first business model increases, and its generalization ability is higher than that of the target business model to a certain extent.

[0109] In one embodiment of this specification, the strategy information may further include: a disturbance intensity value, and a first number of data augmentation operations including a first augmentation operation, wherein the first augmentation operation utilizes a first augmentation operation unit to perform data disturbance conforming to the disturbance intensity value. The first augmentation operation may be one or more. The first augmentation operation utilizes any one of the multiple data augmentation units. In one case, the disturbance intensity value may range from [0, 1].

[0110] In one embodiment of this specification, the data perturbation modification includes perturbation modification of the order and / or number of events in an event sequence; the plurality of data augmentation units include data augmentation units for performing at least one of the following operations. It is understood that the event sequence may refer to an initial event sequence or an event sequence obtained after performing several data augmentation operations on the initial event sequence. The operations include:

[0111] Delete a portion of the events in the event sequence; specifically, delete a portion of the events in the event sequence in a random manner or based on the aforementioned perturbation strength value. The aforementioned perturbation strength value can characterize the probability of each event in the event sequence being deleted.

[0112] Several first events are extracted from the event sequence, and the first events are sampled and rearranged without replacement and then returned to the event sequence with replacement. The first events can be extracted randomly or based on the aforementioned perturbation strength value, and then the first events are sampled and rearranged without replacement and then returned to the event sequence with replacement. The aforementioned perturbation strength value can characterize the probability of each event in the event sequence being extracted. For example, given the event sequence A, B, C, D, E, and F, events A, C, and D are extracted as the first event. Then, events A, C, and D are rearranged without replacement. Specifically, one event is sampled from A, C, and D, for example, event C, which becomes the first event in the rearranged sequence. Next, one event is sampled from A and D, for example, event A, which becomes the second event in the rearranged sequence. Finally, only one event remains, D, which becomes the third event in the rearranged sequence. The rearranged events are then placed back into the event sequence, resulting in the event sequence C, B, A, D, E, and F, which is the output sequence of this data augmentation operation.

[0113] Several second events are extracted from the event sequence, and the second events are sampled and rearranged with replacement and then returned to the event sequence; wherein, the second events can be extracted randomly or based on the aforementioned perturbation strength value, and the second events are sampled and rearranged with replacement and then returned to the event sequence, and the aforementioned perturbation strength value can characterize the probability that each event in the event sequence is extracted (as a second event). For example, given the event sequence B, A, D, E, and F, events B, A, E, and F are extracted as the second event. Then, events B, A, E, and F are rearranged using sampling with replacement. First, an event is sampled from B, A, E, and F, for example, event A, as the first event in the rearranged sequence. Next, another event is sampled from B, A, E, and F, for example, event E, as the second event in the rearranged sequence. Then, another event is sampled from B, A, E, and F, for example, event E, as the third event in the rearranged sequence. Finally, an event is sampled from B, A, E, and F, for example, event B, as the fourth event in the rearranged sequence. The rearranged events are then placed back into the event sequence, resulting in the event sequence A, E, D, E, and B, which is the output sequence of this data augmentation operation.

[0114] The third event in the event sequence is modified into a fourth event. This modification can be done randomly or based on the aforementioned perturbation strength value, which characterizes the probability of each event in the event sequence being modified. In one scenario, both the third and fourth events can be events in the event sequence. In another scenario, the fourth event can be an event outside the event sequence, generated based on the attributes of events in the event sequence.

[0115] Preserve the event sequence, that is, do not perturb the event sequence.

[0116] In one embodiment of this specification, the data perturbation modification further includes perturbation modification of the attributes of events in the event sequence; the plurality of data augmentation units include a data augmentation unit for performing at least one of the following operations: an operation for perturbating and modifying the attribute value of a numerical first attribute of a fifth event in the event sequence using a Gaussian distribution corresponding to each attribute value of the first attribute. Specifically, a corresponding Gaussian distribution can first be constructed using each attribute value of the first attribute in the event sequence (the attribute value of the first attribute of each event), and then the attribute value of the first attribute of the fifth event can be perturbated and modified using the Gaussian distribution, wherein the fifth event can be some or all of the events in the event sequence.

[0117] In one implementation, a pre-defined Gaussian perturbation function is used to generate a perturbation value for the first attribute based on the mean and variance of the Gaussian distribution. This perturbation value is then superimposed on the attribute values ​​of the first attribute to modify the attribute value. In another implementation, a pre-defined Gaussian perturbation function is used to generate a perturbation value for the first attribute based on the mean and variance of the Gaussian distribution. Based on this perturbation value and the aforementioned perturbation intensity value, the attribute value of the first attribute is modified by perturbation. Specifically, the product of the perturbation value and the aforementioned perturbation intensity value is superimposed on the attribute values ​​of the first attribute to modify the attribute value, thereby obtaining a perturbed event sequence that conforms to reality.

[0118] The numerical first attribute can be one or more. When there are multiple first attributes, for each numerical first attribute of the fifth event in the event sequence, the attribute value of the first attribute is modified by perturbation using the Gaussian distribution corresponding to each attribute value.

[0119] Another operation involves perturbating categorical attributes. Categorical attributes, also known as discrete attributes, typically include categories indicated by several discrete values, such as occupation, city, and equipment type. Therefore, the operation in the data augmentation unit could include perturbing the value of the categorical second attribute of the sixth event in the event sequence, using the prior distribution corresponding to the attribute values ​​of that second attribute. The sixth event can be some or all of the events in the event sequence.

[0120] In one implementation, several events can be selected from the event sequence based on the aforementioned perturbation strength value, as the sixth event. The perturbation strength value represents the probability of each event being selected from the event sequence. Subsequently, all (or part) of the categorical attributes of the sixth event are used as the second attribute. For each second attribute, a corresponding prior distribution is constructed using the attribute values ​​of the second attribute (the attribute values ​​of the second attribute for each event in the event sequence). The attribute value of the second attribute of the sixth event is then perturbed and modified using this prior distribution. For example, if the second attribute is the event occurrence device type, and its corresponding prior distribution represents that the event occurrence device types of the second attribute for each event in the event sequence include device type 1, device type 2, and device type 3, then the perturbation and modification of the attribute value of the second attribute of the sixth event could involve changing the event occurrence device type 1 of event A to device type 3, keeping the device type 2 of event B, changing the device type 1 of event C to device type 2, and so on. For example, the prior distribution can also represent that the probability of device type 3 occurring is greater than the probability of device type 2 occurring, and the probability of device type 2 occurring is equal to the probability of device type 1 occurring. Accordingly, the attribute value of the second attribute of the sixth event can be perturbed and modified according to the probability of each device type occurring.

[0121] In another implementation, all (or some) events in the event sequence can be considered as the sixth event. Then, based on the aforementioned perturbation strength value, several attributes are selected from all categorical attributes corresponding to each event in the event sequence as the second attributes of the sixth event. The aforementioned perturbation strength value represents the probability of each categorical attribute being selected in the event sequence. For each second attribute, a corresponding prior distribution is constructed using the attribute values ​​of that second attribute (the attribute values ​​of that second attribute for each event in the event sequence). This prior distribution is then used to perturb and modify the attribute value of that second attribute for the sixth event.

[0122] The specific operations for data perturbation modification described in the above embodiments are illustrative examples of data perturbation modification and do not constitute a limitation on the specific operations for data perturbation modification in this specification. In another embodiment, perturbation modification of the order and / or quantity of events in an event sequence may also include operations such as randomly swapping the order of two events in the event sequence, or randomly replacing one event with another in the event sequence. Perturbation modification of the attributes of events in an event sequence may also include operations such as random positive fluctuation or random negative fluctuation for numerical attributes.

[0123] The following specific example illustrates the data augmentation process for the initial event sequence, such as... Figure 3As shown, the initial event sequence 1 includes events A, B, C, D, E, F and G. It is assumed that the policy information of the first enhancement strategy includes a first number of 4, and the multiple data enhancement units are data enhancement units 1-7.

[0124] First, in the first round of data augmentation, data augmentation unit 2 is selected from multiple data augmentation units. Specifically, this involves extracting several events from the event sequence, rearranging the extracted events without replacement, and then replacing them back into the event sequence. Figure 3 As shown, the initial event sequence 1 is input into data augmentation unit 2 to obtain the corresponding output sequence 1, which serves as the input sequence for the second round of data augmentation. In the second round of data augmentation, data augmentation unit 7 is selected from multiple data augmentation units; for example, the specific operation is to preserve the event sequence, such as... Figure 3 As shown, the output sequence 1 is input into data augmentation unit 7 to obtain the corresponding output sequence 2, which serves as the input sequence for the third round of data augmentation. The third round of data augmentation involves selecting data augmentation unit 4 from multiple data augmentation units. For example, the specific operation involves perturbing and modifying the attribute values ​​of several numerical attributes of several events in the event sequence, such as... Figure 3 As shown, the output sequence 2 is input into data augmentation unit 4 to obtain the corresponding output sequence 3, which serves as the input sequence for the fourth round of data augmentation. In the fourth round of data augmentation, data augmentation unit 6 is selected from multiple data augmentation units. For example, the specific operation involves perturbing and modifying the attribute values ​​of several categorical attributes of several events in the event sequence, such as... Figure 3 As shown, the output sequence 3 is input into the data augmentation unit 6 to obtain the corresponding output sequence 4, which serves as the augmented event sequence corresponding to the initial event sequence 1.

[0125] In the above embodiments, a first set of augmented sequences is generated using the aforementioned multiple data augmentation units through the current first augmentation strategy, and a first business model is trained using this set. It is understood that different augmentation strategies can produce different sets of augmented sequences, thereby training business models with different performance levels. To ensure the performance of the finally trained business model, the augmentation strategy can be optimized to determine the set of augmented sequences that yields good data augmentation results, thus obtaining a more efficient business model. Accordingly, in some embodiments, the method further includes the following steps 11-12:

[0126] Step 11: Obtain the first performance result corresponding to the first business model; in one implementation, the first performance result includes at least one of the following results: business prediction accuracy and the probability of successful black-box attack. The higher the business prediction accuracy and the lower the probability of successful black-box attack, the better the performance of the first business model, and correspondingly, the better the data augmentation effect of the augmented event sequence used to train the first business model.

[0127] Step 12: Based on the first performance result and the first enhancement strategy, determine the second enhancement strategy for the next round.

[0128] In one implementation, test data for testing a first business model can be obtained. This test data includes several test event sequences and their corresponding label data. For clarity, the label data corresponding to the test event sequences is referred to as test label data. Then, the several test event sequences are input into the first business model, causing the first business model to perform business predictions on each of the test event sequences, obtaining business prediction data corresponding to each test event sequence. The test label data and business prediction data corresponding to each test event sequence are compared to determine the accuracy of the business predictions of the first business model. Furthermore, a black-box testing tool targeting member inference is used to perform a black-box attack on the first business model to obtain the probability of a successful black-box attack, thereby obtaining the performance result corresponding to the first business model, which is used as the first performance result.

[0129] In another implementation, after obtaining the first business model, the first business model can be sent to a model testing device to test the first business model and obtain the first performance result corresponding to the first business model.

[0130] Subsequently, in one scenario, a pre-defined optimization algorithm can be used to determine a second enhancement strategy for the next round, based on the first performance result and the first enhancement strategy. This pre-defined optimization algorithm may include, but is not limited to, brute-force search optimization algorithms, genetic algorithm-based optimization algorithms, reinforcement learning-based optimization algorithms, and grid search-based optimization algorithms.

[0131] In another scenario, the direction for adjusting the enhancement strategy can be determined based on the first performance result and the first enhancement strategy. Based on this direction, a second enhancement strategy for the next round can be determined. For example, determining the direction for adjusting the enhancement strategy could be as follows: if the first number in the strategy information of the first enhancement strategy is greater than the number in the strategy information of the previous round's enhancement strategy, the disturbance intensity value in the strategy information of the first enhancement strategy is greater than the disturbance intensity value in the strategy information of the previous round's enhancement strategy, and the first performance result indicates that the first business model has better performance, then it can be determined that the number in the strategy information and the disturbance intensity value can be further increased to obtain a second enhancement strategy for the next round.

[0132] Different enhancement strategies can determine different sets of enhancement sequences, and the performance of models trained based on different sets of enhancement sequences varies. In this embodiment, to improve the user experience of the model prediction service, it is necessary to obtain a business model with good performance, i.e., high business prediction accuracy and high attack defense capability (low probability of successful black-box attacks). Accordingly, a high-performing business model can be determined by the performance results of the business model trained based on the set of enhancement sequences. In another embodiment of this specification, the method may further include the following steps 21-22:

[0133] Step 21: Obtain the first performance result corresponding to the first business model.

[0134] Step 22: Based at least on the first performance result and the first business model, determine the target privacy protection model or its alternative model corresponding to the initial event sequence set.

[0135] In one implementation, the expected performance result of the business model can be preset. If the first performance result meets the expected performance result, the first business model can be directly determined as the target privacy protection model corresponding to the initial business model. Alternatively, in another implementation, if the first performance result indicates that the performance of the first business model is improved relative to the performance of the target business model, the first business model can be used as a candidate model for determining the target privacy protection model.

[0136] In another implementation, in order to determine a more efficient business model, step 22 may include the following steps 221-224:

[0137] Step 221: Obtain several second augmentation sequence sets corresponding to the initial event sequence set. Each second augmentation sequence set is obtained by performing data augmentation operations on the initial event sequence set based on multiple data augmentation units, according to the policy information of the corresponding augmentation strategy. The specific number of data augmentation operations and / or the specific numerical value of the perturbation intensity value may differ for the policy information corresponding to different second augmentation sequence sets. The specific determination process for each second augmentation sequence set is the same as the determination process for the first augmentation sequence set, and will not be repeated here.

[0138] Step 222: Train the model using each set of second augmented sequences to obtain several corresponding second business models. Any one of these second business models can be obtained by training the target business model (or the aforementioned initialization model) using its corresponding set of second augmented sequences and the label data corresponding to each second augmented event sequence. The label data corresponding to each second augmented event sequence is the label data of the initial event sequence corresponding to each second augmented event sequence.

[0139] Step 223: Obtain several second performance results corresponding to several second business models. The method for obtaining the second performance results for each second business model is the same as the method for obtaining the first performance results, and will not be repeated here.

[0140] Step 224: Based on the first performance result and the second performance result, determine the target privacy protection model, or its alternative models, from the first business model and several second business models. One approach is to determine the performance result with the best performance (highest business prediction accuracy and highest model attack defense capability, i.e., lowest probability of successful black-box attack) from the first and second performance results, and then select the business model corresponding to this best performance result from the first and several second business models as the target privacy protection model. Alternatively, based on the expected performance result, the first performance result, and the second performance result, determine the business model from the first and several second business models whose performance result meets the expected performance result, and use this as an alternative model for the target privacy protection model. Subsequently, models that meet business requirements can be selected from the alternative models as the target privacy protection model, and so on.

[0141] In one implementation, a target privacy protection model, or its alternative model, can be determined based on the performance changes of the first business model and several second business models relative to the target business model. Accordingly, before step 224, the method further includes step 225:

[0142] Step 225: Obtain the target performance result corresponding to the target business model, which is trained using the initial event sequence set.

[0143] Accordingly, step 224 includes the following steps 2241-2242:

[0144] Step 2241: Determine the performance changes of the first performance result and several second performance results relative to the target performance result. These performance changes may include changes in the accuracy of the business model's prediction results and changes in the probability of a successful black-box attack.

[0145] Step 2242: Based on the changes in performance, determine the business model with better performance in terms of performance changes from the first business model and several second business models, and use it as the target privacy protection model.

[0146] In this implementation, the performance changes of the first performance result and several second performance results relative to the target performance result can characterize the performance improvement, decrease, or maintenance of the first business model and several second business models relative to the target business model. Specifically, based on the performance changes, several business models representing the performance improvement relative to the target business model can be identified from the first business model and several second business models, serving as intermediate business models. Furthermore, based on the performance changes, the business model with the largest improvement can be identified from the intermediate business models, i.e., the business model with superior performance characteristics, serving as the target privacy-preserving model. Therefore, using this target privacy-preserving model for business prediction can ensure both the accuracy of business prediction and the protection of privacy.

[0147] For example, performance improvement may include, but is not limited to, the following situations: First, the business prediction accuracy of the first business model and several second business models is improved relative to the business prediction accuracy of the target business model, and the attack defense capability of the first business model and several second business models is improved relative to the attack defense capability of the target business model (i.e., the probability of the first business model and several second business models being successfully attacked by a black box is reduced relative to the probability of the target business model being successfully attacked by a black box); Second, the business prediction accuracy of the first business model and several second business models is reduced relative to the business prediction accuracy of the target business model, but the reduction is within a preset allowable range, and the attack defense capability of the first business model and several second business models is improved relative to the attack defense capability of the target business model (the improvement reaches the expected level), etc.

[0148] Through the above embodiments, a better augmentation strategy, a set of augmentation sequences with better data augmentation effect, and a target privacy protection model with better performance can be determined for the initial event sequence set.

[0149] The foregoing description describes specific embodiments of this specification; other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than those shown in the embodiments, and the desired result may still be achieved. Furthermore, the processes depicted in the drawings do not necessarily need to follow the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0150] Corresponding to the above method embodiments, this specification provides an apparatus 400, the schematic block diagram of which is shown below. Figure 4 As shown, it includes:

[0151] The first acquisition module 410 is configured to acquire an initial set of event sequences for training the target business model;

[0152] The data augmentation module 420 is configured to perform a first number of data augmentation operations on each initial event sequence in the initial event sequence set, based on the strategy information of a first augmentation strategy and a preset number of data augmentation units for data perturbation modification, to obtain a corresponding augmented event sequence. Each data augmentation operation is performed using one data augmentation unit. The strategy information indicates at least the first number.

[0153] The model training module 430 is configured to train the model using a first set of augmented sequences formed by various augmented event sequences, thereby obtaining a first business model with privacy protection.

[0154] In one alternative implementation, the data augmentation unit is used to perturb and modify one or more of the following: the order, number, and attributes of events in an event sequence.

[0155] In one alternative embodiment, the device further includes:

[0156] The second obtaining module (not shown in the figure) is configured to obtain the first performance result corresponding to the first business model;

[0157] The first determining module (not shown in the figure) is configured to determine a second enhancement strategy for the next round based on the first performance result and the first enhancement strategy.

[0158] In one alternative embodiment, the device further includes:

[0159] The third acquisition module (not shown in the figure) is configured to obtain the first performance result corresponding to the first business model;

[0160] The second determining module (not shown in the figure) is configured to determine, at least based on the first performance result and the first business model, the target privacy protection model corresponding to the initial event sequence set, or its alternative model.

[0161] In one alternative implementation, the second determining module includes:

[0162] The first obtaining submodule (not shown in the figure) is configured to obtain a plurality of second enhanced sequence sets corresponding to the initial event sequence set, wherein each second enhanced sequence set is obtained by performing data enhancement operations on the initial event sequence set based on the policy information of the corresponding enhancement strategy and the plurality of data enhancement units;

[0163] The model training submodule (not shown in the figure) is configured to train the model using each set of second enhancement sequences to obtain several corresponding second business models.

[0164] The second acquisition submodule (not shown in the figure) is configured to acquire several second performance results corresponding to several second business models;

[0165] The determination submodule (not shown in the figure) is configured to determine the target privacy protection model or the alternative model from the first business model and several second business models based on the first performance result and the second performance result.

[0166] In one optional implementation, the second determining module further includes:

[0167] The third acquisition submodule (not shown in the figure) is configured to obtain the target performance result corresponding to the target business model, which is trained using the initial event sequence set;

[0168] The determining submodule is specifically configured to determine the performance changes of the first performance result and several second performance results relative to the target performance result.

[0169] Based on the changes in performance, the business model with better performance in terms of performance changes is determined from the first business model and the plurality of second business models, and is used as the target privacy protection model.

[0170] In one alternative implementation, the first performance result includes at least one of the following: business forecast accuracy and the probability of a successful black-box attack.

[0171] In one alternative implementation, the strategy information further includes: a disturbance intensity value, wherein the first number of data augmentation operations includes a first augmentation operation, the first augmentation operation utilizing a first augmentation operation unit to perform data perturbation conforming to the disturbance intensity value.

[0172] In one alternative implementation, the data perturbation modification includes perturbation modification of the order and / or number of events in the event sequence; the plurality of data augmentation units include data augmentation units for performing at least one of the following operations:

[0173] Delete a portion of the events from the event sequence;

[0174] Extract several first events from the event sequence, and then sample and rearrange the several first events without replacement and put them back into the event sequence;

[0175] Several second events are extracted from the event sequence, and the several second events are rearranged with replacement and then returned to the event sequence.

[0176] Modify the third event in the event sequence to be the fourth event;

[0177] Maintain the event sequence.

[0178] In one alternative implementation, the data perturbation modification includes perturbation modification of attributes of events in the event sequence; the plurality of data augmentation units include data augmentation units for performing at least one of the following operations:

[0179] For the first numerical attribute of the fifth event in the event sequence, the attribute value of the first attribute is modified by perturbation using the Gaussian distribution corresponding to each attribute value of the first attribute.

[0180] For the second attribute of the category of the sixth event in the event sequence, the attribute value of the second attribute is perturbed and modified by using the prior distribution corresponding to each attribute value of the second attribute.

[0181] In one optional implementation, the strategy information further includes a perturbation intensity value and a data augmentation unit, specifically configured to generate a perturbation value corresponding to the first attribute of the fifth event based on a Gaussian distribution corresponding to each attribute value of the first attribute.

[0182] Based on the perturbation value and the perturbation intensity value, the attribute value of the first attribute is modified by perturbation.

[0183] In one alternative implementation, the data augmentation unit used in a single data augmentation operation is randomly selected from the plurality of data augmentation units in an equally probabilistic manner.

[0184] The above-described apparatus embodiments correspond to the method embodiments, and detailed descriptions can be found in the description of the method embodiments section, which will not be repeated here. The apparatus embodiments are derived based on the corresponding method embodiments and have the same technical effects as the corresponding method embodiments; detailed descriptions can be found in the corresponding method embodiments.

[0185] This specification also provides a computer-readable storage medium storing a computer program that, when executed on a computer, causes the computer to perform the asset transfer method in the payment platform described in this specification.

[0186] This specification also provides a computing device, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, it implements the asset transfer method in the payment platform provided in this specification.

[0187] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for storage media and computing devices are basically similar to the method embodiments, so they are described more simply; relevant parts can be referred to the descriptions of the method embodiments.

[0188] Those skilled in the art will recognize that the functions described in the embodiments of the present invention in one or more of the above examples can be implemented using hardware, software, firmware, or any combination thereof. When implemented in software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium.

[0189] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, or improvements made based on the technical solutions of the present invention should be included within the scope of protection of the present invention.

Claims

1. A privacy-preserving model training method, comprising: Multiple event sequences are obtained for training the target business model to form an initial event sequence set. Each event sequence is a real event sequence collected for a sample object. Each event sequence includes several events, which are arranged in the order of the time when the sample object triggers the event. For each event sequence in the initial event sequence set, according to the first number indicated by the strategy information of the first enhancement strategy, a first number of data enhancement operations are performed on the events in the event sequence to obtain the corresponding enhanced event sequence. In each data enhancement operation, a data enhancement unit is selected from a plurality of preset data enhancement units for data perturbation modification, and the event sequence is enhanced based on the data enhancement unit. The data perturbation modification includes: perturbing the order of events in the event sequence. The model is trained using the first set of augmented sequences formed by each augmented event sequence and the original label data corresponding to each augmented event sequence to obtain the first business model for privacy protection. The original label data corresponding to each augmented event sequence is the label data of the initial event sequence corresponding to the augmented event sequence.

2. The method according to claim 1, wherein, The data augmentation unit is used to perturb and modify one or more of the following: the order, number, and attributes of events in an event sequence.

3. The method according to claim 1, further comprising: Obtain the first performance result corresponding to the first business model; Based on the first performance result and the first enhancement strategy, a second enhancement strategy is determined for the next round.

4. The method according to claim 1, further comprising: Obtain the first performance result corresponding to the first business model; Based at least on the first performance result and the first business model, determine the target privacy protection model corresponding to the target business model, or its alternative model.

5. The method according to claim 4, wherein, The step of determining the target privacy protection model corresponding to the target business model, or its alternative model, includes: A plurality of second enhanced sequence sets corresponding to the initial event sequence set are obtained, wherein each second enhanced sequence set is obtained by performing data enhancement operations on the initial event sequence set based on the policy information of the corresponding enhancement strategy and the plurality of data enhancement units; Each set of second enhancement sequences is used to train the model, resulting in several corresponding second business models; Obtain several second performance results corresponding to several second business models; Based on the first performance result and the second performance result, the target privacy protection model or the alternative model is determined from the first business model and several second business models.

6. The method according to claim 5, further comprising: Obtain the target performance result corresponding to the target business model, which is trained using the initial event sequence set; The step of determining the target privacy protection model, or the alternative model, from the first business model and several second business models includes: Determine the performance changes of the first performance result and several second performance results relative to the target performance result; Based on the changes in performance, the business model with better performance in terms of performance changes is determined from the first business model and the plurality of second business models, and is used as the target privacy protection model.

7. The method according to claim 3, wherein, The first performance result includes at least one of the following: business forecast accuracy, probability of being successfully attacked by a black box.

8. The method according to any one of claims 1-7, wherein, The strategy information also includes: a disturbance intensity value, wherein the first number of data augmentation operations includes a first augmentation operation, and the first augmentation operation uses a first augmentation operation unit to perform data disturbance that conforms to the disturbance intensity value.

9. The method according to any one of claims 1-7, wherein, The data perturbation modification includes perturbation modification of the order and / or number of events in the event sequence; the plurality of data augmentation units include data augmentation units for performing at least one of the following operations: Delete a portion of the events from the event sequence; Extract several first events from the event sequence, and then sample and rearrange the several first events without replacement and put them back into the event sequence; Several second events are extracted from the event sequence, and the several second events are rearranged with replacement and then returned to the event sequence. Modify the third event in the event sequence to be the fourth event; Maintain the event sequence.

10. The method according to any one of claims 1-7, wherein, The data perturbation modification includes perturbation modification of the attributes of events in the event sequence; the plurality of data augmentation units include data augmentation units for performing at least one of the following operations: For the first numerical attribute of the fifth event in the event sequence, the attribute value of the first attribute is modified by perturbation using the Gaussian distribution corresponding to each attribute value of the first attribute. For the second attribute of the category of the sixth event in the event sequence, the attribute value of the second attribute is perturbed and modified by using the prior distribution corresponding to each attribute value of the second attribute.

11. The method according to claim 10, wherein, The strategy information also includes a perturbation strength value; the step of using the Gaussian distribution corresponding to each attribute value of the first attribute to perturb and modify the attribute value of the first attribute includes: Based on the Gaussian distribution corresponding to the values ​​of each attribute of the first attribute, the perturbation value corresponding to the first attribute of the fifth event is generated. Based on the perturbation value and the perturbation intensity value, the attribute value of the first attribute is modified by perturbation.

12. The method according to any one of claims 1-7, wherein, The data augmentation unit used in a single data augmentation operation is randomly selected from the plurality of data augmentation units in an equally probabilistic manner.

13. A privacy-preserving model training device, comprising: The first acquisition module is configured to acquire multiple event sequences for training the target business model to form an initial event sequence set, wherein each event sequence is a real event sequence collected for the sample object, and each event sequence includes several events, which are arranged in the order of the time when the sample object triggers the event. The data augmentation module is configured to perform a first number of data augmentation operations on each event sequence in the initial event sequence set, according to a first number indicated by the strategy information of the first augmentation strategy, to obtain a corresponding augmented event sequence. In each data augmentation operation, a data augmentation unit is selected from a set of preset data augmentation units for data perturbation modification, and the event sequence is augmented based on that data augmentation unit. The data perturbation modification includes perturbing the order of events in the event sequence. The model training module is configured to use the first set of augmented sequences formed by each augmented event sequence and the original label data corresponding to each augmented event sequence to train the model and obtain the first business model with privacy protection. The original label data corresponding to each augmented event sequence is the label data of the initial event sequence corresponding to the augmented event sequence.

14. A computing device comprising a memory and a processor, wherein, The memory stores executable code, and when the processor executes the executable code, it implements the method of any one of claims 1-12.

Images