Federal learning model training method and high-speed fee evasion behavior identification method

Abstract

The application discloses a federal learning model training method and a high-speed fee-stealing and fee-evading behavior identification method, which comprises the following steps: a plurality of small sample set owners and a label owner respectively complete privacy intersection with a full-amount sample set owner through a circuit with a random mask to obtain intersection information of a corresponding index library; each business party calculates WOE encoding data by using a homomorphic encryption algorithm according to the intersection information and a label value sent by the label owner, and the label value is used to represent whether it is a fee-stealing and fee-evading behavior; each business party trains a logistic regression model locally by using the WOE encoding data and the label value, and performs vertical federal learning with the label owner, and iteratively updates model parameters until a preset training stop condition is met, so that a trained logistic regression model is obtained and deployed locally to score fee-stealing and fee-evading of a vehicle passing through a high-speed area; and the application can realize accurate high-speed fee-stealing and fee-evading auditing and hierarchical classification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data mining, and particularly relates to a federated learning model training method and a high-speed fee evasion behavior identification method. BACKGROUND

[0002] With the nationwide cancellation of provincial boundary toll stations and the comprehensive implementation of ETC construction, the toll mode has changed from the original "closed type" to "open type", which has also brought some new problems and difficulties. The cost of single long-distance transportation is higher, and the temptation of fee evasion is greater, which has led to organized fee evasion behavior, making it more difficult to govern fee evasion.

[0003] In the prior art, machine learning algorithms and data mining technologies such as Logistic regression analysis method, improved genetic algorithm back propagation (BP) neural network, random forest algorithm, random forest and logistic regression combination method and decision tree are used to establish an automatic fee evasion event identification model to actively mine suspicious vehicle information from toll data.

[0004] The patent application document with the publication number CN116433166A introduces a high-speed audit model based on a path algorithm, which uses gantry transaction data, gantry plate recognition data and other internal data to identify long-distance running and short-distance buying high-speed fee evasion behavior. External data such as government and financial data are not used, and gang behavior cannot be associated through external data, making it difficult to identify gang fee evasion behavior. SUMMARY

[0005] The technical problem to be solved by the present application is how to realize accurate high-speed fee evasion audit and hierarchical classification.

[0006] The present application solves the above technical problems by the following technical means:

[0007] On the one hand, the present application proposes a federated learning model training method, which trains a logistic regression model through a label owner and each business party. Each business party includes a plurality of small sample set owners and a full sample set owner. The label owner and the small sample set owner locally deploy a small sample index library, and the full sample set owner locally deploys a full sample index library, comprising:

[0008] The label owner and the plurality of small sample set owners respectively complete private intersection with the full sample set owner through a circuit with a random mask to obtain intersection information of the corresponding index library;

[0009] Each business party calculates WOE encoded data using a homomorphic encryption algorithm according to the intersection information and a label value sent by the label owner, and the label value is used to represent whether it is a fee evasion behavior;

[0010] Each business party trains a logistic regression model locally using the WOE encoding data and the label value, and performs vertical federated learning with the label owner, iteratively updates the model parameters until the preset training stopping condition is met, obtains the trained logistic regression model, and deploys it locally for scoring the toll evasion of vehicles passing through the high-speed area.

[0011] Further, the business parties include financial institutions, industrial and commercial systems, transportation systems, and public security systems, and the label owner is a high-speed operator;

[0012] The small sample index library deployed locally by the high-speed operator stores user features collected by the high-speed operator;

[0013] The small sample index library deployed locally by the financial institution stores user features collected by the financial institution, the small sample index library deployed locally by the industrial and commercial system stores user features collected by the industrial and commercial department, the small sample index library deployed locally by the transportation system stores user features collected by the transportation department, and the full sample index library deployed locally by the public security system stores user features collected by the public security department;

[0014] The user features include basic features and frequency features derived based on time series.

[0015] Further, before the several small sample set owners and the label owner respectively complete the private intersection with the full sample set owner through a circuit with a random mask to obtain intersection information of the corresponding index library, the method further comprises:

[0016] The business party converts non-numeric features contained in the locally deployed index library into numeric features.

[0017] Further, the several small sample set owners and the label owner respectively complete the private intersection with the full sample set owner through a circuit with a random mask to obtain intersection information of the corresponding index library, comprising:

[0018] Partitioning the full sample index library according to the number of partitions, and dividing the elements in each partition into a plurality of first element segments of the same bit length according to the element segmentation number;

[0019] Dividing the elements in the small sample index library into a plurality of second element segments of the same bit length according to the element segmentation number;

[0020] encrypt each of the second element segments to obtain a first encryption result and send to the full sample set owner, so that the full sample set owner calculates a polynomial of the first encryption result, and performs a masking operation on the polynomial after encryption to obtain a second encryption result and returns to the corresponding small sample index library owner;

[0021] The full sample set owner and the corresponding small sample index library owner perform two-party computation using the PEqT algorithm, based on the first encryption result and the second encryption result, to calculate the intersection information of the two-party locally deployed index library.

[0022] Further, the full sample index library is partitioned according to the number of partitions, and the elements in each partition are divided into a plurality of first element segments of the same bit length according to the element segmentation number, comprising:

[0023] The full sample index library is initially hashed on the set X, which is mapped to a fixed-length set X', and X represents the unique identifier set in the full sample index library;

[0024] A simple hash table S is constructed using the set X', and each hash bucket in the simple hash table S is partitioned according to the number of partitions, and the elements in each partition are divided into a plurality of first element segments of the same bit length according to the element segmentation number.

[0025] Further, the elements in the small sample index library are divided into a plurality of second element segments of the same bit length according to the element segmentation number, comprising:

[0026] The small sample index library is initially hashed on the set Y, which is mapped to a fixed-length set Y', and Y represents the unique identifier set in the small sample index library;

[0027] A cuckoo hash table C is constructed using the set Y', and the elements in each hash bucket in the cuckoo hash table C are divided into a plurality of second element segments of the same bit length according to the element segmentation number.

[0028] Further, the second element segments are encrypted to obtain a first encryption result and sent to the full sample set owner, so that the full sample set owner calculates a polynomial of the first encryption result, and performs a masking operation on the polynomial after encryption to obtain a second encryption result and returns to the corresponding small sample index library owner, comprising:

[0029] Each of the second element segments is homomorphically encrypted to obtain a first encryption result and sent to the full sample set owner;

[0030] The full sample set owner calculates a polynomial of the first encryption result , wherein, , belonging to the set of the second element segments;

[0031] The full sample set owner homomorphically encrypts the polynomial and performs a masking operation on the homomorphic encryption result by using a random mask to obtain a second encryption result and returns to the corresponding small sample index library owner.

[0032] Further, the business owner calculates WOE encoding data by using a homomorphic encryption algorithm according to the intersection information and the label value sent by the label owner, including:

[0033] The business owner receives the homomorphic encryption label value sent by the label owner;

[0034] The business owner bins the feature elements in the intersection information and calculates the positive and negative sample numbers according to the homomorphic encryption label value;

[0035] The business owner sends the positive and negative sample numbers to the label owner, so that the label owner calculates the WOE encoding data according to the positive and negative sample numbers;

[0036] The business owner obtains the WOE encoding data returned by the label owner.

[0037] Further, the business owner trains a logistic regression model locally by using the WOE encoding data and the label value, and performs vertical federated learning with the label owner, and iteratively updates the model parameters until a preset training stop condition is met, to obtain a trained logistic regression model, including:

[0038] The business owner trains a logistic regression model locally by using the WOE encoding data and the label value, and sends the model parameters encrypted by using a homomorphic encryption algorithm to the label owner, so that the label owner aggregates the model parameters sent by each business owner, generates global model parameters and encrypts the global model parameters by using a homomorphic encryption algorithm, and returns to each business owner;

[0039] The business owner updates the local logistic regression model by using the global model parameters sent by the label owner, and iterates the model parameters until a preset training stop condition is met, to determine the optimal model parameters of the local logistic regression model of each business owner, and obtain a trained logistic regression model, wherein the model parameters include the weight W and the threshold value b of the model input features.

[0040] In addition, the present application also proposes a high-speed fee evasion behavior recognition method, and each business owner locally deploys a trained logistic regression model obtained by using the federated learning model training method as described above, including:

[0041] convert the output value of the locally deployed logistic regression model of each business party in the current region into an integer score as a score value;

[0042] sum the score values obtained by each logistic regression model to obtain a global score;

[0043] Based on the global score, the probability of vehicle evasion of the current region is classified.

[0044] Further, the output value of the locally deployed logistic regression model of each business party in the current region is converted into an integer score as a score value, which is expressed by the formula:

[0045]

[0046] In the formula, is the score value, is the offset, is the scaling factor, is the feature WOE conversion data, is the model threshold, .

[0047] Further, the score value includes a basic score and a feature score , the basic score is obtained by experts scoring according to the model input initial variable binning level.

[0048] Further, the classification of the probability of vehicle evasion in the current region based on the global score includes:

[0049] Compare the global score with the score threshold to classify the probability of vehicle evasion in the current region, wherein the score threshold is obtained by scoring the modeling samples.

[0050] The advantages of the present application are:

[0051] (1) When the model is trained, the data sources provided by different business parties are used, and the privacy intersection with the full sample index library is completed through the circuit with random mask to obtain the intersection information of the two parties. The data outside the intersection cannot be obtained by the other party. Under the premise of ensuring the privacy data security of the label owning party and the business party, the logistic regression model of user features is trained jointly by the label owning party and each business party, so as to realize the accurate auditing and classification of high-speed evasion of toll fees.

[0052] (2) Based on homomorphic encryption, the weight, gradient and other model parameter information in the forward calculation and back propagation of the longitudinal federal learning model training are updated to ensure the security of parameter information transmission.

[0053] (3) The output of the federated learning model is converted by training to score conversion to form a high-speed fee evasion behavior scoring linear formula, meeting the interpretability requirements of the model.

[0054] Additional aspects and advantages of the present application will be set forth in part in the description that follows, and in part will become apparent to those having ordinary skill in the art upon examination of the following or can be learned from practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 is a flowchart of a federated learning model training method according to an embodiment of the present application;

[0056] Figure 2 is a flowchart of a high-speed fee evasion behavior identification method according to an embodiment of the present application. DETAILED DESCRIPTION

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in a clear and complete manner. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0058] As shown in Figure 1 , the first embodiment of the present application proposes a federated learning model training method. A label owner and each business party perform federated learning training on a logistic regression model. Each business party includes a plurality of small sample set owners and a full sample set owner. The label owner and the small sample set owner locally deploy a small sample index library, and the full sample set owner locally deploys a full sample index library. The method includes the following steps:

[0059] S10, the plurality of small sample set owners and the label owner respectively complete private intersection with the full sample set owner through a circuit with a random mask, to obtain intersection information of the corresponding index library;

[0060] It should be noted that the business parties in the present embodiment include one or more financial institutions, industry and commerce systems, transportation systems, public security systems, and high-speed operation parties. The data related to high-speed fee evasion auditing of each business party is collected, and the corresponding index library is constructed based on the data sources provided by each business party and deployed locally.

[0061] The label owner in the present embodiment is a high-speed operation party, and the financial institutions, industry and commerce systems, transportation systems, and public security systems are business parties.

[0062] S20, each business party adopts a homomorphic encryption algorithm to calculate WOE encoding data according to the intersection information and the label value sent by the label owner, and the label value is used to represent whether it is a fee evasion behavior;

[0063] S30, each business party trains a logistic regression model locally using the WOE encoding data and the label value, and performs vertical federated learning with the label owner, iteratively updates the model parameters until the preset training stop condition is met, obtains the trained logistic regression model, and deploys it locally to score the fee evasion of vehicles passing through the high-speed area.

[0064] When training the model, the embodiment uses data sources provided by different business parties, and completes privacy intersection with the full-sample index library through a circuit with random masking to obtain intersection information of the two parties. The data outside the intersection cannot be obtained by the other party. Under the premise of ensuring the privacy data security of the label owner and the business parties, the logistic regression model of user features is trained jointly by the label owner and each business party, so as to realize accurate auditing and hierarchical classification of high-speed fee evasion behaviors.

[0065] As a further preferred technical solution, each of the business parties includes a financial institution, an industry and commerce system, a transportation system, and a public security system, and the label owner is a high-speed operator;

[0066] The small-sample index library deployed locally by the high-speed operator stores user features collected by the high-speed operator;

[0067] The small-sample index library deployed locally by the financial institution stores user features collected by the financial institution, the small-sample index library deployed locally by the industry and commerce system stores user features collected by the industry and commerce department, the small-sample index library deployed locally by the transportation system stores user features collected by the transportation department, and the full-sample index library deployed locally by the public security system stores user features collected by the public security department;

[0068] The user features include basic features and frequency features derived based on time series.

[0069] Specifically, the present embodiment is directed to the multi-source characteristics of high-speed tax evasion behavior auditing data, and the classification design is carried out according to the source, type and format, and the anti-fraud index library construction of each business party is completed. First, collect the data sources related to high-speed auditing of each business party, such as highway operation party, public security, traffic, industry and commerce, finance and other related data of high-speed toll evasion auditing, including table data, log file, image, video, text and the like. Specifically, the high-speed operation party provides high-speed flow data, including gantry flow, ETC entrance and exit, license plate flow data, etc., the public security system provides driver's household information, social relationship, behavior record, etc., the traffic department provides freight company registration data, vehicle information, transportation punishment information, etc., the industry and commerce provides freight company legal person data, related person data, registration address, etc., and the financial institutions provide personal identity information, asset situation, transaction amount, transaction time, transaction medium, etc.

[0070] Then, the collected data is cleaned to remove duplicates, errors and missing values, and the accuracy and integrity of the data are ensured; feature extraction is performed to select features related to the high-speed toll evasion model, including some basic features (such as ETC number, OBU number, license plate number, vehicle color, axle, toll time, toll distance, driver age, asset situation, industry and commerce registration situation, etc.) and frequency features derived based on time series (such as the number of times of toll evasion, the number of times of tolling on the highway per week, the number of times of tolling on the highway per month, the number of times of driver changing vehicles within a month, etc.); the high-speed auditing index library constructed based on the data sources provided by each business party is deployed locally.

[0071] As a further preferred technical solution, before the step S10: the several small sample set owners and the label owner respectively complete the privacy intersection with the full sample set owner through the circuit with random mask to obtain the intersection information of the corresponding index library, the method further comprises:

[0072] The business party converts the non-numeric features contained in the locally deployed second index library into numeric features.

[0073] It should be noted that since the name and other features in the index library are non-numeric features, the present embodiment encodes the features in the index library of each business party, and converts the non-numeric features into numeric features by using label encoding or one-hot encoding.

[0074] As a further preferred technical solution, the step S10: the several small sample set owners and the label owner respectively complete the privacy intersection with the full sample set owner through the circuit with random mask to obtain the intersection information of the corresponding index library, comprising the following steps:

[0075] S11, partitioning the full-sample index library according to the number of partitions, and dividing the elements in each partition into a plurality of first element segments of the same bit length according to the element segmentation number;

[0076] S12, dividing the elements in the small-sample index library into a plurality of second element segments of the same bit length according to the element segmentation number;

[0077] S13, encrypting each of the second element segments to obtain a first encryption result and sending it to the full-sample set owner, so that the full-sample set owner calculates the polynomial of the first encryption result, and performs a masking operation on the polynomial after encryption to obtain a second encryption result and returns it to the corresponding small-sample index library owner;

[0078] S14, the full-sample set owner and the corresponding small-sample index library owner perform two-party computation using the PEqT algorithm, based on the first encryption result and the second encryption result, to calculate the intersection information of the two-party locally deployed index libraries.

[0079] Specifically, the personnel ID is used as the data identifier in each business party index library. In order to further improve the sample alignment security, the data leakage risk is eliminated by using the circuit privacy intersection algorithm with random masking. In sample alignment, the public security system has relatively complete personnel information, so the full-sample index library is locally deployed. The high-speed operation party, the traffic, the business, and the financial information are less than the public security system, so the small-sample index library is locally deployed.

[0080] As a further preferred technical solution, the step S11: partitioning the full-sample index library according to the number of partitions, and dividing the elements in each partition into a plurality of first element segments of the same bit length according to the element segmentation number, specifically includes:

[0081] S111, performing initial hashing on the full-sample index library set X and mapping it to a fixed-length set X';

[0082] It should be noted that X represents the unique identifier set in the full-sample index library, and X' is the set obtained after the initial hashing of X. Since the hashed X' has a fixed length, the number of buckets and element distribution can be better controlled when building the hash table, and the algorithm efficiency is improved.

[0083] S112, constructing a simple hash table S using the set X', and partitioning each hash bucket in the simple hash table S according to the number of partitions, and dividing the elements in each partition into a plurality of first element segments of the same bit length according to the element segmentation number.

[0084] Specifically, in the offline stage, the present embodiment uses the set X' to construct a simple hash table S, the size of the simple hash table S being O(n y ), where n y is the size of the small sample index library, and such a setting helps subsequent fast lookup and processing of set elements.

[0085] The number of partitions a and the number of element partitions d are selected, and then the following operations are performed on each hash bucket S[k]=B k (where S is the overall hash table, S[k] is a bucket element in S, and B k is the element set corresponding to the kth bucket in the simple hash table S):

[0086] B k is divided into a partitions , and the partitioning operation helps to divide a large set into smaller subsets for subsequent separate processing, reducing computational complexity. The elements in each partition are further divided into d first element segments of the same bit length , and element partitioning can optimize the calculation process to some extent, especially when using homomorphic encryption algorithms.

[0087] Then define the polynomial , .

[0088] As a further preferred technical solution, the step S12: dividing the elements in the small sample index library into a plurality of second element segments of the same bit length according to the element partitioning number, specifically includes:

[0089] S121, performing initial hashing on the small sample index library set Y to map it to a fixed-length set Y';

[0090] It should be noted that the set Y refers to the unique identifier set in the small sample index library, and Y' is the set obtained after the initial hashing of Y. Since the hashed Y' has a fixed length, the number of buckets and element distribution can be better controlled when constructing the hash table, improving the efficiency of the algorithm.

[0091] S122, using the set Y' to construct a cuckoo hash table C, and dividing the elements in each hash bucket of the cuckoo hash table C into a plurality of second element segments of the same bit length according to the element partitioning number.

[0092] Specifically, the size of the cuckoo hash table C is O(n y ), and the cuckoo hash table has the characteristics of fast positioning and processing of elements, especially when determining whether an element belongs to a large set. Then for each cuckoo hash table bucket element C[k]= (where C is the overall hash table, C[k] is a bucket element in C, is the set of elements in the kth bucket in the cuckoo hash table C) performs the following operations:

[0093] Splitting into d second element fragments of the same bit length , so as to match the processing mode of the elements in the full sample index library in the offline stage, facilitating subsequent polynomial calculation.

[0094] As a further preferred technical solution, the step S13: encrypting each of the second element fragments to obtain a first encryption result and sending it to the full sample set owner, so that the full sample set owner calculates the polynomial of the first encryption result, and performs a masking operation on the encrypted polynomial to obtain a second encryption result and returns it to the corresponding small sample index library owner, specifically comprising:

[0095] S131, homomorphically encrypting each of the second element fragments to obtain a first encryption result and sending it to the full sample set owner;

[0096] S132, the full sample set owner calculates the polynomial of the first encryption result , wherein , ;

[0097] S133, the full sample set owner homomorphically encrypts the polynomial to obtain a homomorphic encryption result , and uses a random mask to perform a masking operation on the homomorphic encryption result to obtain a second encryption result and returns it to the corresponding small sample index library owner.

[0098] It should be noted that the present embodiment performs a masking operation on the above homomorphic encryption result by using a random mask and adding noise to ensure the security of the data.

[0099] Subsequently, the full sample set owner sends the second encryption result to the small sample index library owner, and the small sample index library owner decrypts the received ciphertext to obtain p'i .

[0100] It should be noted that the ultimate goal of this algorithm of circuit privacy intersection with a random mask is to obtain the secret sharing form of the set intersection, rather than to restore the original plaintext set elements, so it is not necessary to remove the noise and the mask separately, but to use them as a means to protect privacy and calculate the set intersection, and to perform subsequent Boolean XOR calculation, and finally to obtain the privacy protection result of the set intersection through the calculation of both parties.

[0101] As a further preferred technical solution, the step S14: the full sample set owner and the corresponding small sample index library owner perform two-party computation by using PEqT algorithm, based on the first encryption result and the second encryption result, calculate the intersection information of the two-party locally deployed index library, specifically including:

[0102] Both parties perform two-party computation based on PEqT algorithm, calculate p'i The circuit Boolean value of (if the result is 1, equivalent judgment x Belongs to the intersection X∩Y), and then XOR operation is performed on all partition quantities The result of XOR operation is the secret sharing form of set intersection, and both parties cannot obtain the specific element content, finally output the intersection information.

[0103] It should be noted that PEqT algorithm is to perform XOR operation on two integers with the same bit, and determine whether the two integers are equal by judging whether the value at each bit is equal.

[0104] As a further preferred technical solution, the step S20: each business party calculates WOE encoding data by using homomorphic encryption algorithm according to the intersection information and the label value sent by the label owner, including the following steps:

[0105] S21, the business party receives the homomorphic encryption label value sent by the label owner;

[0106] It should be noted that the label owner homomorphically encrypts the label value y, 1-y to obtain the homomorphic encryption label value

[0107]

[0108] In the formula, is the homomorphic encryption result of , and is the homomorphic encryption result of .

[0109] S22, the business party bins the feature elements in the intersection information, and calculates the positive and negative sample numbers according to the homomorphic encryption label value;

[0110] It should be noted that the business party bins the features in the local index library, and calculates the positive and negative sample numbers according to the encrypted label value, and returns the result to the label owner:

[0111]

[0112] In the formula, is the binning of set, ​to sum the encrypted positive sample labels, to sum the encrypted negative sample labels.

[0113] S23, the business party sends the positive and negative sample numbers to the label owner, so that the label owner calculates the WOE encoding data according to the positive and negative sample numbers;

[0114] Specifically, the label owner calculates the WOE encoding data according to the positive and negative sample numbers as follows:

[0115]

[0116] In the formula, denotes the sample label value of the label owner, E denotes the homomorphic encryption algorithm, denotes the homomorphic encryption, denotes the binning algorithm, denotes the feature variable set of small sample set of financial institutions, transportation systems, etc., which usually adopts equal frequency, equal distance and chi-square binning algorithm, D denotes the decryption of the result, is the sum of the encrypted positive sample labels based on binning calculation, and the number of positive samples in each bin, is the sum of the encrypted negative sample labels based on binning calculation, and the number of negative samples in each bin, is the proportion of the number of positive samples in each bin to the total number of positive samples, is the proportion of the number of negative samples in each bin to the total number of negative samples, is the evidence weight of each bin, is the total number of positive samples in all bins, is the total number of negative samples in all bins.

[0117] S24, the business party obtains the WOE encoding data returned by the label owner.

[0118] It should be noted that the present embodiment encodes discrete variables into numerical variables by WOE encoding, so as to facilitate processing by the logistic regression algorithm. Each business party takes the existing tax evasion behavior label of the high-speed operator as the target Y value, calculates the respective WOE value based on the homomorphic encryption feature engineering, and the high-speed operator that owns the target Y value calculates the WOE value conventionally, while the label owner calculates the WOE value according to the label data and the binning result of the feature data of the party.

[0119] ​As a further preferred technical solution, the step S30: the business party trains a logistic regression model locally using the WOE encoding data and the label value, and performs vertical federated learning with the label owner, iteratively updates the model parameters until the preset training stopping condition is met, to obtain a trained logistic regression model, comprising the following steps:

[0120] S31, the business party trains a logistic regression model locally using the WOE encoding data and the label value, and sends the model parameters encrypted by a homomorphic encryption algorithm to the label owner, so that the label owner aggregates the model parameters sent by each business party, generates global model parameters and encrypts them by a homomorphic encryption algorithm, and returns them to each business party;

[0121] S32, the business party updates the local logistic regression model using the global model parameters sent by the label owner, and determines the optimal model parameters of the local logistic regression model of each business party when the model parameter iteration loop meets the preset training stopping condition, to obtain a trained logistic regression model, wherein the model parameters include the weight W and the threshold value b of the model input features.

[0122] It should be noted that each business party obtains its own WOE result as the input of the model for federated model training. Each business party trains a logistic regression model based on the WOE data on the local data set to obtain preliminary model parameters and gradient information; the updated gradient or model parameters are encrypted based on the homomorphic encryption algorithm and sent to the label owner side node server. The server receives and aggregates the encrypted parameter information from each participant, generates global model parameters, and calculates the loss on the public security side; the label owner side server sends the global model parameters and the loss back to each business party, and each participant updates the local model using the new global model parameters; repeat the above steps until the model reaches the expected performance standard or converges. The weight W and the threshold value b of each model input feature are obtained.

[0123] It should be noted that the present embodiment updates the weight, gradient and other model parameter information in the forward calculation and back propagation of the vertical federated learning model training based on homomorphic encryption, ensuring the security of parameter information transmission.

[0124] In addition, as Figure 2 shown, the second embodiment of the present application proposes a high-speed fee evasion behavior recognition method, each business party locally deploys a trained logistic regression model obtained by using the federated learning model training method as described in the above embodiment, comprising the following steps:

[0125] S100, converting the output value of the logistic regression model locally deployed by each business party in the current area into an integer score as a score value;

[0126] S200, summing up the score values obtained by each logistic regression model to obtain a global score;

[0127] S300, based on the global score, classifying the vehicle fee evasion probability in the current area.

[0128] The embodiment trains a federated learning model by processing public security data and operator and financial data, converts the score, forms a vehicle fee evasion behavior scoring linear formula, meets the interpretability requirement of the model, and outputs early warning information or list information.

[0129] As a further preferred technical solution, the output value of the logistic regression model locally deployed by each business party in the current area is converted into an integer score as a score value, and the formula is represented as:

[0130]

[0131] In the formula, is the score value, is an offset, which ensures that the score system meets the business scenario requirements, is an offset, which ensures that the score system meets the business scenario requirements, is the feature WOE coded data, is the model threshold, .

[0132] It should be noted that the score conversion is to convert the output value of the trained federated learning model into an integer score, not directly using the probability conversion score, but using the linear part to convert into a score. Thus, the score of each feature X is the total score, and if the probability is converted into a score, the total score does not increase the corresponding score value.

[0133] As a further preferred technical solution, the score value includes a basic score and a feature score , the basic score is obtained by scoring the initial variable binning level of the model input by an expert.

[0134] Specifically, the basic score is determined by the initial variable expert scoring of the input model feature variable. For example, there are 30 input model variables for each business party such as the high-speed operation party side, the business side, etc. Each variable is scored by a business expert (high-speed audit personnel, data developer, expert scholar) according to the binning level of the variable. For example, there are 7 bins for the case of missing gantry plate recognition, and the gantry plate recognition variable is scored according to [-30, -20, -10, 0, 10, 20, 30]. The higher the score, the more likely the fee evasion behavior occurs. Finally, the scoring results of each business expert are obtained, and the average value of each business expert is obtained to obtain the vehicle fee evasion basic score value, and the basic score value is distributed between [500, 650].

[0135] For the feature score, the factor value is determined. Based on the current data statistical sample, the vehicle evasion probability is p, and the vehicle normal probability is 1-p. The score conversion requires that the ratio of the vehicle evasion and the vehicle normal probability decreases by 2 times, and the feature score increases by 30 points (the feature score increase value is also calculated according to the actual data of the statistical sample and the experience of the comprehensive business expert). The result is:

[0136]

[0137] Therefore, the factor is -28.85.

[0138] It can be obtained that:

[0139]

[0140] As a further preferred technical solution, the step S300: classifying the vehicle evasion probability in the current region based on the global score, comprises:

[0141] Comparing the global score with a score threshold value to classify the vehicle evasion behavior in the current highway region, wherein the score threshold value is obtained by statistical scoring of the modeling sample.

[0142] Specifically, the trained global score card model is deployed to each online environment of the business party, and the evasion score of the vehicle passing through a certain highway region is updated regularly in batches. The vehicle evasion is classified according to the score value. The score threshold value is obtained by statistical scoring of the modeling sample, which shows the effect of different score thresholds on the business. The threshold value statistics need to calculate the score of each sample first, and then classify the score according to the segmentation grouping.

[0143] According to the proportion of vehicle evasion in the grouping, the high, medium and low risk levels are defined. For example, in the grouping of 750 points or more, the proportion of vehicles evading fees is higher than 80%, which is defined as a high risk level, indicating that the recognized vehicle has a very high evasion behavior characteristic. The score result is between 620 and 750, the proportion of vehicles evading fees is between 60% and 80%, which is defined as a medium risk level, indicating that the recognized vehicle has a high risk and needs to be further checked by the auditing personnel. The score result is less than 620, and the proportion of vehicles evading fees is less than 60%. The classification result can still be refined according to the actual needs of the business.

[0144] The present embodiment adopts different ways of disposal according to the risk level, arranges auditors to directly intercept and audit at the toll entrance for high risk level, needs high-speed operators to further verify in combination with the passing flow data, entrance data and license data for medium risk level, and labels special situations in the system for low risk level, pays attention to whether the subsequent score increases, and improves the probability of fee evasion.

[0145] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily mean the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0146] In addition, the terms "first", "second" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "a plurality of" is at least two, for example, two, three, etc., unless otherwise specifically limited.

[0147] Although the embodiments of the present application have been shown and described above, it can be understood that the above-described embodiments are exemplary and cannot be understood as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above-described embodiments within the scope of the present application.

Claims

1. A federated learning model training method, characterized in that, A logistic regression model is trained using federated learning by the label owner and various business parties. Each business party includes several small sample set owners and one full sample set owner. The label owners and small sample set owners have locally deployed small sample indicator libraries, while the full sample set owner has locally deployed a full sample indicator library, including: Several small sample set owners and the label owners respectively complete privacy intersection with the full sample set owner through circuits with random masks to obtain the intersection information of the corresponding indicator libraries. This includes partitioning the full sample indicator library according to the number of partitions, and dividing the elements in each partition into multiple first element fragments of the same bit length according to the number of element segments; dividing the elements in the small sample indicator library into multiple second element fragments of the same bit length according to the number of element segments; encrypting each second element fragment to obtain a first encryption result and sending it to the full sample set owner, so that the full sample set owner can calculate the polynomial of the first encryption result, encrypt the polynomial, perform a masking operation to obtain a second encryption result, and return it to the corresponding small sample indicator library owner; the full sample set owner and the corresponding small sample indicator library owner use the PEqT algorithm to perform two-party calculations, and calculate the intersection information of the locally deployed indicator libraries of the two parties based on the first encryption result and the second encryption result; Each business unit calculates WOE-encoded data using a homomorphic encryption algorithm based on the intersection information and the tag value sent by the tag owner. The tag value is used to characterize whether it is a fee evasion behavior. Each business unit trains the logistic regression model locally using the WOE encoded data and the label values, and performs longitudinal federated learning with the label owner to iteratively update the model parameters until the preset training stopping condition is met. The trained logistic regression model is then deployed locally to score vehicles evading tolls in the highway area.

2. The federated learning model training method as described in claim 1, characterized in that, The aforementioned business entities include financial institutions, industrial and commercial systems, transportation systems, and public security systems, and the label owner is the highway operator; The local small-sample indicator library deployed by the highway operator stores user characteristics collected by the highway operator. The local small-sample indicator library deployed by the financial institution stores user characteristics collected by the financial institution; the local small-sample indicator library deployed by the industrial and commercial system stores user characteristics collected by the industrial and commercial department; the local small-sample indicator library deployed by the transportation system stores user characteristics collected by the transportation department; and the local full-sample indicator library deployed by the public security system stores user characteristics collected by the public security department. The user features include basic features and frequency features derived from time series data.

3. The federated learning model training method as described in claim 1, characterized in that, Before the owners of the several small sample sets and the owners of the labels respectively complete the privacy intersection with the owner of the full sample set through a circuit with a random mask to obtain the intersection information of the corresponding indicator database, the method further includes: The business team converts the non-numerical features contained in the locally deployed indicator library into numerical features.

4. The federated learning model training method as described in claim 1, characterized in that, The step of partitioning the full sample index database according to the number of partitions, and dividing the elements in each partition into multiple first element fragments of the same bit length according to the number of element segments, includes: The full sample index database is initially hashed to set X, and then mapped to a fixed-length set X', where X represents the set of unique identifiers in the full sample index database. Construct a simple hash table S using a set X', partition each hash bucket in the simple hash table S according to the number of partitions, and divide the elements in each partition into multiple first element fragments of the same bit length according to the number of element splits.

5. The federated learning model training method as described in claim 1, characterized in that, The step of dividing elements in a small sample index library into multiple second element segments of the same bit length according to the number of element segments includes: The small sample index library set Y is initially hashed and mapped to a fixed-length set Y', where Y represents the set of unique identifiers in the small sample index library. Construct a cuckoo hash table C using set Y', and divide the elements in each hash bucket of the cuckoo hash table C into multiple second element fragments of the same bit length according to the number of element splits.

6. The federated learning model training method as described in claim 1, characterized in that, The step of encrypting each second element fragment to obtain a first encryption result and sending it to the owner of the full sample set, so that the owner of the full sample set can calculate the polynomial of the first encryption result, encrypt the polynomial, perform a masking operation to obtain a second encryption result, and return it to the owner of the corresponding small sample index database, includes: Homomorphic encryption is performed on each of the second element fragments to obtain the first encryption result, which is then sent to the owner of the full sample set. The owner of the full sample set computes the polynomial of the first encryption result. ,in, , The set consisting of the fragments of the second element. This is a fragment of the second element; The owner of the full sample set performs homomorphic encryption on the polynomial, and uses a random mask to mask the homomorphic encryption result to obtain a second encryption result, which is then returned to the owner of the corresponding small sample index library.

7. The federated learning model training method as described in claim 1, characterized in that, Each business party calculates WOE-encoded data using a homomorphic encryption algorithm based on the intersection information and the tag value sent by the tag owner, including: The business party receives the homomorphically encrypted tag value sent by the tag owner; The business unit bins the feature elements in the intersection information and calculates the number of positive and negative samples based on the homomorphic encryption tag value; The business party sends the number of positive and negative samples to the tag owner, so that the tag owner can calculate the WOE encoded data based on the number of positive and negative samples; The business party obtains the WOE-encoded data returned by the label owner.

8. The federated learning model training method as described in claim 1, characterized in that, The business unit trains the logistic regression model locally using the WOE-encoded data and the label values, and performs longitudinal federated learning with the label owner, iteratively updating the model parameters until a preset training stopping condition is met, thus obtaining the trained logistic regression model, including: The business unit trains the logistic regression model locally using the WOE encoded data and the label value, and sends the model parameters to the label owner after encrypting them with a homomorphic encryption algorithm. This allows the label owner to aggregate the model parameters sent by each business unit, generate global model parameters, encrypt them with a homomorphic encryption algorithm, and return them to each business unit. The business unit updates its local logistic regression model using the global model parameters sent by the tag owner. When the model parameters iterate and loop until the preset training stopping condition is met, the optimal model parameters of each business unit's local logistic regression model are determined, and the trained logistic regression model is obtained. The model parameters include the model weights W and model thresholds b of the model input features.

9. A method for identifying highway toll evasion, characterized in that, Each business unit has a locally deployed logistic regression model trained using the federated learning model training method described in any one of claims 1-8, including: Convert the output values ​​of the logistic regression models deployed locally by each business unit in the current region into integer scores as the scoring values; The scores obtained from each logistic regression model are summed to obtain the global score; Based on the global score, the probability of vehicles evading tolls in the current area is classified.

10. The method for identifying highway toll evasion as described in claim 9, characterized in that, The step of converting the output values ​​of the logistic regression models deployed locally by each business unit in the current region into integer scores as the scoring values ​​is expressed by the following formula: In the formula, As a rating value, This is the offset. Scaling factor Transform the data for the feature WOE. For the model threshold, .

11. The method for identifying highway toll evasion as described in claim 10, characterized in that, The scoring value includes a base score. and feature scores The basis is obtained by experts scoring the initial variables input into the model according to the binning level.

12. The method for identifying highway toll evasion as described in claim 10, characterized in that, The process of classifying the probability of vehicles evading tolls in the current area based on the global score includes: The global score is compared with a scoring threshold to classify the probability of vehicles evading tolls in the current area, wherein the scoring threshold is obtained from the scoring statistics of the modeled samples.

Images