Federal learning-driven flexible employment compliance decision-making system and method thereof

Abstract

The invention relates to the technical field of artificial intelligence, data privacy protection and compliance management, in particular to a flexible employment compliance decision-making system and method based on a federated learning technology. The system comprises a privacy protection module for performing homomorphic encryption, differential privacy and randomization processing on employee information; the federated learning module is used for carrying out federated learning on the child nodes and generating an employee salary level prediction model; the decision module is used for calculating an individual tax pre-deduction scheme and a risk index; the biological characteristic verification module is used for identifying the identity of the employee and the authenticity of the contract; according to the system, a five-layer progressive privacy protection system is constructed, so that the balance between local data and global intelligence is realized, and the contradiction between data privacy and analysis efficiency is effectively solved; meanwhile, the innovatively designed hierarchical heterogeneous federated learning framework supports different platforms to use different data structures and computing capabilities to participate in training, so that the model adaptability is remarkably improved; on the whole, the method provides efficient, safe and compliant decision support for a flexible employment platform.

Description

Technical Field

[0001] The present invention relates to the technical fields of artificial intelligence, data privacy protection, and compliance management, and particularly relates to a flexible employment compliance decision-making system and method based on federated learning technology. Background Art

[0002] With the rise of the sharing economy and the flexible employment model, more and more enterprises and individuals have participated in the flexible employment market. As a bridge connecting enterprises and flexible employees, flexible employment platforms need to process a large amount of sensitive information such as personal information, work records, and salary data. In this process, data privacy protection, compliance risk management, and efficient matching have become key issues that the industry urgently needs to solve.

[0003] Traditional flexible employment management systems usually adopt a centralized architecture, storing and processing all user data centrally. This method has a serious risk of data privacy leakage. At the same time, due to the lack of an effective prediction mechanism, traditional systems are difficult to identify potential compliance risks in advance and often can only respond passively after problems occur, resulting in enterprises facing legal disputes and economic losses. In addition, traditional personnel matching algorithms are usually based on simple rule matching and cannot fully consider multi-dimensional factors, resulting in low matching efficiency.

[0004] In the prior art, there are solutions that use blockchain technology to ensure data transparency, and there are also methods that use machine learning for risk prediction. However, these technologies are often applied independently and lack systematic integration. Especially in achieving high-quality model training while protecting data privacy, the prior art still has obvious deficiencies.

[0005] Federated learning, as an emerging distributed machine learning technology, allows multiple parties to jointly train a model without sharing raw data, providing a new idea for solving the above problems. However, the application of federated learning in the field of flexible employment compliance management is still in its infancy, lacking a mature system architecture and implementation plan.

[0006] Therefore, there is an urgent need for a flexible employment management system that can achieve efficient compliance decision-making and optimized matching while protecting data privacy. Summary of the Invention

[0007] The object of the present invention is to provide a federated learning-driven flexible employment compliance decision-making system and method, aiming to solve the problems of insufficient data privacy protection, weak compliance risk prediction ability, and low matching efficiency in existing flexible employment management systems.

[0008] The present invention proposes a federated learning-driven flexible employment compliance decision-making system, including:

[0009] A privacy protection module for homomorphic encryption, differential privacy, and randomization processing of employee information on the flexible employment platform;

[0010] A federated learning module, connected to the privacy protection module, for performing federated learning on the processed information on sub-nodes to generate an employee salary level prediction model;

[0011] A decision-making module, connected to the federated learning module, for calculating the individual income tax withholding plan and risk index based on the prediction model;

[0012] A biometric verification module for authenticating the authenticity of employee identities and contract texts.

[0013] Preferably, the privacy protection module includes:

[0014] A data desensitization unit for desensitizing employees' basic registration information, historical salary information, and attendance performance information;

[0015] An encryption unit for performing homomorphic encryption on the desensitized data;

[0016] A differential privacy unit for adding accurately calculated noise to the encrypted data;

[0017] A randomization unit for introducing random factors during data transmission and processing.

[0018] Preferably, the federated learning module includes:

[0019] A local model unit for locally training an employee working hours prediction model on each flexible employment platform;

[0020] A model aggregation unit, connected to the local model unit, for aggregating model parameters of each platform on the federated server;

[0021] A global model unit, connected to the model aggregation unit, for generating a global employee working hours prediction model and a salary level prediction model;

[0022] A model distribution unit, connected to the global model unit, for distributing the global model to each platform.

[0023] Preferably, the decision-making module includes:

[0024] An individual income tax calculation unit for calculating the individual income tax withholding plan based on the salary level prediction model;

[0025] A risk assessment unit for calculating the labor law risk index, and the risk index includes a comprehensive score of legal dimension, tax dimension, contract dimension, and operation dimension;

[0026] A salary adjustment unit for calculating and adjusting the next payment amount based on the current paid salary and the budgeted expenses;

[0027] A visualization unit for visually displaying the individual income tax withholding plan and the risk index.

[0028] Preferably, the biometric verification module includes:

[0029] A fingerprint recognition unit for judging the authenticity of an employee's identity through a distributed fingerprint recognition algorithm;

[0030] An OCR recognition unit for verifying the authenticity of contract texts through a document structure consistency verification algorithm;

[0031] A multi-modal fusion unit for combining multiple biometric features to improve verification accuracy;

[0032] A federal verification unit for comprehensively calculating the verification results of each platform on a federal server.

[0033] Preferably, it further includes:

[0034] A blockchain module connected to the federated learning module and the decision-making module for recording the model update process and the compliance decision results;

[0035] A matching optimization module connected to the decision-making module for optimizing the matching of employees and positions based on a reinforcement learning algorithm.

[0036] Preferably, the blockchain module includes:

[0037] An intelligent contract unit for defining and executing federated learning rules and data access permissions;

[0038] A distributed storage unit for storing model update records and compliance decision results;

[0039] A consensus mechanism unit for combining federated learning weight aggregation with a blockchain consensus algorithm.

[0040] Preferably, the matching optimization module includes:

[0041] A feature extraction unit for extracting the spatio-temporal features of employees and positions;

[0042] A multi-agent unit for establishing matching relationships based on work location, time, position, and skills;

[0043] A reinforcement learning unit for optimizing matching strategies through a reward mechanism;

[0044] A matching result unit for generating and outputting an optimal matching plan.

[0045] Preferably, the local model unit incorporates an LSTM deep learning network to capture the temporal patterns of work and compensation; the global model unit adopts knowledge distillation technology to support the joint training of different structure models.

[0046] A flexible employment compliance decision-making method driven by federated learning, comprising:

[0047] Performing privacy protection processing on the basic registration information, historical salary payment information, and attendance performance information of employees collected on each flexible employment platform by using homomorphic encryption, differential privacy, and randomization techniques;

[0048] Performing calculations on the processed information on each sub-node to implement federated learning of the model and obtain a prediction model for employees' future salary levels;

[0049] Based on the salary level prediction model, calculating the individual income tax withholding plan and multi-dimensional risk index, and calculating and adjusting the next payment amount according to the current paid salary combined with the budgeted expenses;

[0050] Constructing a biometric verification model for employees' working hours information and contract text information to verify employees' identities and identify the authenticity of contracts;

[0051] Recording the model update and decision-making process through blockchain to ensure the transparency and traceability of the system;

[0052] Based on the reinforcement learning algorithm, optimizing the matching relationship between employees and positions.

[0053] By innovatively integrating federated learning, multi-level privacy protection, predictive compliance decision-making, and reinforcement learning-driven matching optimization technologies, the present invention realizes efficient and accurate compliance risk prediction and resource optimization allocation on the premise of ensuring data privacy and security.

[0054] The present invention has the following beneficial effects:

[0055] 1. Constructed a five-layer progressive privacy protection system, realizing global intelligence while ensuring that data does not leave the local area, effectively solving the contradiction between data privacy and analysis efficiency;

[0056] 2. Innovatively designed a hierarchical heterogeneous federated learning framework, supporting different platforms to participate in model training using different data structures and computing capabilities, greatly improving the adaptability of the model;

[0057] 3. Developed a multi-model collaborative prediction architecture, with the prediction accuracy increased by 35% compared with traditional statistical models, changing from passive compliance inspection to active prediction and prevention;

[0058] 4. Realized a multi-dimensional risk assessment system, identifying potential compliance risks 7 days in advance, greatly reducing the potential losses of enterprises;

[0059] 5. The organic combination of blockchain and federated learning ensures the transparency and traceability of model training while protecting the privacy of original data;

[0060] 6. The matching optimization based on multi-agent reinforcement learning improves the matching efficiency by more than 40% and achieves the optimal allocation of resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is the overall architecture diagram of the flexible employment compliance decision-making system driven by federated learning of the present invention;

[0062] Figure 2 It is the structural schematic diagram of the privacy protection module in the present invention;

[0063] Figure 3 It is the working flow chart of the federated learning module in the present invention;

[0064] Figure 4 It is the structural schematic diagram of the decision-making module in the present invention;

[0065] Figure 5 It is the working flow chart of the biometric verification module in the present invention;

[0066] Figure 6 It is the schematic diagram of the combination of the blockchain module and federated learning in the present invention;

[0067] Figure 7 It is the structural schematic diagram of the matching optimization module in the present invention;

[0068] Figure 8 It is the flow chart of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0069] Please refer to the attached Figure 1-8 , and the present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0070] As Figure 1 shown, the flexible employment compliance decision-making system driven by federated learning provided by the present invention includes: a privacy protection module 1, a federated learning module 2, a decision-making module 3, a biometric verification module 4, a blockchain module 5, and a matching optimization module 6.

[0071] As Figure 2 shown, the privacy protection module 1 includes a data desensitization unit 11, an encryption unit 12, a differential privacy unit 13, and a randomization unit 14.

[0072] The data desensitization unit 11 is used to perform preliminary desensitization processing on the basic registration information, historical salary payment information, and attendance performance information of employees collected on each flexible employment platform, and remove direct identifiers such as names and ID numbers. Preferably, the present invention adopts a rule-based desensitization strategy, combines regular expression recognition and replacement of sensitive information, reduces the direct recognition risk while retaining the data analysis value.

[0073] The encryption unit 12 is used to process the desensitized data using homomorphic encryption technology. In an embodiment of the present invention, the Paillier homomorphic encryption algorithm is adopted. This algorithm supports direct addition operations and scalar multiplications on encrypted data. The specific encryption process is as follows:

[0074] 1. Key generation: Select two large prime numbers $ and , calculate and ;

[0075] 2. Calculate ;

[0076] 3. The public key is , and the private key is ;

[0077] For the plaintext , the encryption process is:

[0078] ,

[0079] where is a random number less than , and the encryption result supports the following homomorphic properties:

[0080] ,

[0081] The differential privacy unit 13 is used to add accurately calculated noise during the data aggregation process to protect individual information. In an embodiment of the present invention, the Laplace mechanism is adopted to implement -differential privacy. The specific formula is as follows:

[0082] ,

[0083] where, is the original query function, is the sensitivity, indicating the maximum value that the result of the function may change due to the change of an individual, is the privacy budget, which controls the intensity of privacy protection, and Lap represents the Laplace distribution. Preferably, is set for salary data, is set for attendance data, and , to achieve differential privacy protection.

[0084] The randomization unit 14 is used to introduce random factors during data transmission and processing, further enhancing security. The present invention adopts the randomized response technique. For boolean features , the following transformation is used:

[0085] ,

[0086] where is the probability of retaining the original value, usually set to 0.75.

[0087] Through the synergistic effect of the above four-layer privacy protection mechanism, the present invention can retain the analysis value of the data while ensuring data privacy, providing a secure data foundation for subsequent federated learning.

[0088] As Figure 3 shown, the federated learning module 2 includes a local model unit 21, a model aggregation unit 22, a global model unit 23, and a model distribution unit 24.

[0089] The local model unit 21 is used to locally train an employee working hours prediction model on each flexible employment platform. Preferably, the present invention has an LSTM deep learning network built into the local model unit 21 for capturing the temporal pattern between work and compensation. The LSTM network structure is as follows:

[0090] ,

[0091] ,

[0092] ,

[0093] ,

[0094] ,

[0095] .

[0096] where, is the forget gate, is the input gate, is the output gate, is the cell state, is the hidden state, is the input vector, and are the model parameters, is the sigmoid activation function, is the hyperbolic tangent activation function.

[0097] Input vector It includes features such as the historical working hours, work type, and working period of employees. The model is trained using the backpropagation algorithm, and the loss function is set to the mean absolute error (MAE):

[0098] ,

[0099] where is the actual working hours, is the predicted working hours.

[0100] The model aggregation unit 22 is connected to the local model unit 21 and is used to aggregate the model parameters of each platform on the federated server. The present invention uses the weighted federated average algorithm (FedAvg) for model aggregation, and the specific formula is as follows:

[0101] ,

[0102] where, is the global model parameter, is the local model parameter of the th platform, is the data volume of the th platform, is the total data volume of all platforms, is the total number of platforms participating in federated learning. The present invention further adopts a contribution-based weighting mechanism to improve the above formula to:

[0103] ,

[0104] where, is the data quality coefficient of the th platform, is the weight factor for balancing the data volume and data quality, usually set to 0.7. The data quality coefficient is obtained through the performance evaluation of the model on the validation set.

[0105] The global model unit 23 is connected to the model aggregation unit 22 and is used to generate the global employee working hours prediction model and the salary level prediction model. The present invention uses the knowledge distillation technology to support the joint training of models with different structures. Specifically, let the output of the teacher model (global model) be , and the output of the student model (local model) be , and the distillation temperature be , then the distillation loss function is:

[0106] ,

[0107] where, is the KL divergence, is the softmax function.

[0108] In addition, the global salary level prediction model comprehensively considers factors such as working hours, job type, and regional economic level, and adopts a multi-layer perceptron structure:

[0109] ,

[0110] Among them, is the input feature vector, , , , are model parameters, and ReLU is the rectified linear unit activation function.

[0111] The model distribution unit 24 is connected to the global model unit 23 and is used to distribute the global model to each platform. To reduce communication overhead, the present invention adopts parameter compression technologies, including parameter quantization and sparsification. Specifically, for the model parameter , 8-bit quantization is performed:

[0112] ,

[0113] At the same time, the top 50% of the parameters with the largest absolute values are retained, and the rest are set to zero to achieve sparsification.

[0114] As Figure 4 shown, the decision module 3 includes an individual income tax calculation unit 31, a risk assessment unit 32, a salary adjustment unit 33, and a visualization unit 34.

[0115] The individual income tax calculation unit 31 is used to calculate the individual income tax withholding plan based on the salary level prediction model. According to the provisions of the current Individual Income Tax Law of China and in combination with the predicted salary level, the present invention automatically generates an optimal individual income tax withholding plan. Specifically, the formula for the amount of withholding and prepayment of tax is as follows:

[0116] ,

[0117] Among them, is the amount of withholding and prepayment of tax, is the predicted monthly income, is the withholding rate, is the quick calculation deduction. The withholding rate and the quick calculation deduction are determined according to the cumulative amount of taxable income for withholding and prepayment.

[0118] The risk assessment unit 32 is used to calculate the labor law risk index, which includes a comprehensive score of legal dimension, tax dimension, contract dimension, and operation dimension. The present invention designs a multi-dimensional risk assessment system, and the calculation formulas for the risk indexes of each dimension are as follows:

[0119] ,

[0120] ,

[0121] ,

[0122] ,

[0123] Among them, represents the risk index of each dimension, represents the risk factor weight, represents the risk factor scoring function, represents the number of risk factors in each dimension.

[0124] The comprehensive risk index is calculated through weighted combination:

[0125] ,

[0126] Among them, represents the weight of each dimension, preferably set to , , .

[0127] The risk score is divided into five levels: extremely low risk (0 - 20), low risk (21 - 40), medium risk (41 - 60), high risk (61 - 80), and extremely high risk (81 - 100). When the risk index exceeds 60, the system automatically triggers the warning mechanism.

[0128] The salary adjustment unit 33 is used to calculate and adjust the next payment amount based on the current paid salary and the budgeted expenses. The present invention adopts a dynamic adjustment strategy to dynamically optimize the salary plan according to the risk index, budget constraint, and employee performance. The adjustment formula is as follows:

[0129] ,

[0130] Among them, is the next payment amount, is the current payment amount, is the adjustment coefficient, usually between -0.1 and 0.2, is the deduction amount. The adjustment coefficient The calculation formula is:

[0131] ,

[0132] Among them, represents the employee performance score, represents the comprehensive risk index, represents the budget surplus, Indicates the weights of various factors, preferably set as = 0.6, = 0.3, = 0.1.

[0133] The visualization unit 34 is used to visually display the individual income tax withholding plan and the risk index. The present invention adopts visualization methods such as dashboards, heat maps, and trend charts to intuitively present information related to compliance decisions. At the same time, for high-risk projects, the system automatically generates a detailed risk analysis report and improvement suggestions.

[0134] As Figure 5 shown, the biometric verification module 4 includes a fingerprint recognition unit 41, an OCR recognition unit 42, a multimodal fusion unit 43, and a federated verification unit 44.

[0135] The fingerprint recognition unit 41 is used to judge the authenticity of the employee's identity through a distributed fingerprint recognition algorithm. The present invention adopts a distributed fingerprint recognition technology based on local features, without sharing the complete fingerprint image. The specific steps are as follows:

[0136] 1. Extract local fingerprint features, including the positions of feature points and directions ;

[0137] 2. Perform privacy protection processing on the feature vector:

[0138] ,

[0139] wherein, is the original feature vector, is an irreversible hash function, is a noise vector;

[0140] 3. Send the processed feature vector to the federated server;

[0141] 4. The server calculates the feature similarity and generates a matching score;

[0142] The OCR recognition unit 42 is used to verify the authenticity of the contract text through a document structure consistency verification algorithm. The present invention combines optical character recognition (OCR) and document structure analysis technology to achieve the authenticity recognition of contract texts. The key steps include:

[0143] 1. Text extraction: Use OCR technology to extract the content of the contract text

[0144] 2. Structure analysis: Extract the chapter structure, key terms, and signature positions

[0145] 3. Consistency check: Check the logical relationship between clauses and the consistency of text formats

[0146] 4. Tampering Detection: Identifying Potential Tampering Traces through Image Quality Analysis

[0147] The system sets multiple inspection indicators, including text consistency score, structural integrity score, and tampering trace score, to comprehensively evaluate the authenticity of the contract. When any indicator is lower than the threshold of 0.7, the system determines that there is a risk of authenticity of the contract.

[0148] The multimodal fusion unit 43 is used to combine multiple biometric features to improve the verification accuracy. The present invention adopts a weighted fusion strategy to perform identity verification by integrating fingerprint, face, and behavior features:

[0149] ,

[0150] where, is the final verification score, is the verification score of the th biometric feature, is the corresponding weight, is the number of biometric feature types. Preferably, the fingerprint feature weight is set to 0.5, the face feature weight is set to 0.3, and the behavior feature weight is set to 0.2.

[0151] The federated verification unit 44 is used to comprehensively calculate the verification results of each platform on the federated server. The present invention designs a federated verification protocol to achieve cross-platform identity verification while protecting the privacy of the original biometric data. The specific process includes:

[0152] 1. Each platform locally generates verification results and confidence levels

[0153] 2. Upload the results and confidence levels to the federated server

[0154] 3. The server calculates the final verification result by weighted averaging according to the confidence levels

[0155] 4. Feed back the final result to each platform

[0156] When the final verification score exceeds the threshold of 0.85, the verification is determined to pass; when it is lower than 0.6, the verification is determined to fail; when it is between 0.6 - 0.85, a secondary verification mechanism is triggered.

[0157] As Figure 6 shown, the blockchain module 5 includes a smart contract unit 51, a distributed storage unit 52, and a consensus mechanism unit 53.

[0158] The smart contract unit 51 is used to define and execute federated learning rules and data access permissions. The present invention designs a dedicated smart contract based on the Ethereum platform to achieve automated management of the federated learning process. The key contract functions include:

[0159] 1. Participant registration and permission management;

[0160] 2. Model update submission and verification;

[0161] 3. Incentive mechanism implementation;

[0162] 4. Compliance decision record;

[0163] The sample code structure of the smart contract is as follows:

[0164] solidity

[0165] contract FederatedLearning{

[0166] / / Participant structure

[0167] struct Participant{

[0168] address addr;

[0169] uint256 dataAmount;

[0170] uint256 contributionScore;

[0171] bool isActive;

[0172] }

[0173] / / Model update structure

[0174] struct ModelUpdate{

[0175] address participant;

[0176] string modelHash;

[0177] uint256 timestamp;

[0178] bool isVerified;

[0179] }

[0180] / / State variables

[0181] mapping(address => Participant) public participants;

[0182] ModelUpdate[] public modelUpdates;

[0183] / / Event definitions

[0184] event ParticipantRegistered(address indexed participant);

[0185] event ModelUpdated(address indexed participant, string modelHash);

[0186] / / Function implementation

[0187] function registerParticipant() public {...}

[0188] function submitModelUpdate(string memory modelHash) public {...}

[0189] function verifyModelUpdate(uint256 updateId) public {...}

[0190] function distributeRewards() public {...}

[0191] }

[0192] The distributed storage unit 52 is used to store model update records and compliance decision results. The present invention adopts IPFS (InterPlanetary File System) as the distributed storage solution, and stores model parameters, update history and decision records in a distributed network. The specific storage process is as follows:

[0193] 1. Chunk and encrypt the data;

[0194] 2. Calculate the content-addressable hash value;

[0195] 3. Distribute and store the data in the IPFS network;

[0196] 4. Record the data hash value and access permissions on the blockchain;

[0197] The consensus mechanism unit 53 is used to combine federated learning weight aggregation with the blockchain consensus algorithm. The present invention innovatively designs a Proof of Weight (PoW) consensus mechanism, which combines the model contribution degree in federated learning with the blockchain consensus. The key steps include:

[0198] 1. Calculate the base weight according to the data volume and quality of the participants;

[0199] 2. Evaluate the contribution degree based on the model update quality;

[0200] 3. The comprehensive weight determines the block generation right;

[0201] 4. The verification node confirms the effectiveness of the model update;

[0202] This mechanism encourages the participating parties to provide high-quality data and model updates, while ensuring the security of the system.

[0203] Such as Figure 7 As shown, the matching optimization module 6 includes a feature extraction unit 61, a multi-agent unit 62, a reinforcement learning unit 63, and a matching result unit 64.

[0204] The feature extraction unit 61 is used to extract the spatio-temporal features of employees and positions. The present invention designs a comprehensive feature extraction framework, including:

[0205] 1. Time features: work period preference, available time distribution, historical work pattern;

[0206] 2. Space features: geographical location, commuting distance, regional preference;

[0207] 3. Skill features: professional skills, experience level, evaluation score;

[0208] 4. Position features: job requirements, salary level, working environment;

[0209] The feature extraction adopts an autoencoder structure to compress high-dimensional features into low-dimensional representations:

[0210] ,

[0211] ,

[0212] Among them, is the original feature, is the encoded feature, is the reconstructed feature, and are activation functions, , , , are model parameters.

[0213] The multi-agent unit 62 is used to establish a matching relationship based on the work location, time, position, and skills. The present invention uses a multi-agent system to simulate the interaction behavior between employees and enterprises. Each agent represents a decision-making entity and makes decisions according to the utility function. The utility function of the employee agent is defined as:

[0214] ,

[0215] Among them, Indicates the utility of an employee for a position of the position, Indicates salary satisfaction, Indicates commuting distance, Indicates skill matching degree, Indicates work flexibility, Indicates the weights of various factors.

[0216] The utility function of the enterprise agent is defined as:

[0217] ,

[0218] Among them, Indicates the utility of the enterprise for an employee of the employee, Indicates productivity expectation, Indicates experience level, Indicates availability, Indicates cost, Indicates the weights of various factors.

[0219] The reinforcement learning unit 63 is used to optimize the matching strategy through a reward mechanism. The present invention implements reinforcement learning using a deep Q-network (DQN). The state space includes the set of currently unmatched employees, the set of unassigned positions, and historical matching results; the action space is the possible employee-position assignment schemes; the reward function is defined as:

[0220] ,

[0221] Among them, Indicates the reward for executing action from state to reach state , Indicates the matching set generated by action of action Indicates matching completeness, Indicates matching time, Indicates the weights of various factors.

[0222] The Q-learning update formula is:

[0223] ,

[0224] Among them, Indicates the state-action value function, Indicates the learning rate, Indicates the discount factor, Indicates the immediate reward.

[0225] The matching result unit 64 is used to generate and output the optimal matching solution. The present invention adopts the Hungarian algorithm to achieve global optimal matching, and the steps are as follows:

[0226] 1. Construct a utility matrix , where represents the comprehensive utility of employee and position ;

[0227] 2. Apply the Hungarian algorithm to the matrix to obtain the optimal matching solution;

[0228] 3. Evaluate the matching result and calculate the global utility and satisfaction;

[0229] 4. Generate a detailed matching report, including the reasons for matching and the recommended priorities;

[0230] The following indicators are used for the matching satisfaction evaluation:

[0231] 1. Global utility: the sum of the utilities of all matching pairs;

[0232] 2. Average satisfaction: the average of the employee satisfaction and the enterprise satisfaction;

[0233] 3. Matching completeness: the proportion of successful matches;

[0234] 4. Stability: the degree to which no employee and enterprise tend to change the current match;

[0235] As Figure 8 shown, the flexible employment compliance decision-making method driven by federated learning provided by the present invention includes the following steps:

[0236] Step 1: Perform privacy protection processing on the basic registration information, historical salary payment information, and attendance performance information of employees collected on each flexible employment platform by using homomorphic encryption, differential privacy, and randomization techniques.

[0237] Specifically, first desensitize the original data to remove direct identifiers; then encrypt the data by using the Paillier homomorphic encryption algorithm; then apply the Laplace mechanism during the data aggregation process to achieve differential privacy protection; finally, further improve the data security through the randomized response technique. This multi-level privacy protection mechanism ensures that the original data does not leave the local area while retaining the data analysis value.

[0238] Step 2: Calculate the processed information on each sub-node to implement the federated learning of the model and obtain the employee future salary level prediction model.

[0239] Specifically, each flexible employment platform first trains the LSTM deep learning network locally to capture the temporal characteristics of working hours; then uploads the model parameters (instead of the original data) to the federal server; the federal server aggregates the model parameters using an improved weighted federated averaging algorithm to generate a global model; the global model is then sent down to each platform and combined with local data for further optimization; finally, a high-precision prediction model for employees' working hours and a prediction model for salary levels are formed.

[0240] Step 3: Based on the salary level prediction model, calculate the individual income tax withholding plan and multi-dimensional risk index, and calculate and adjust the next payment amount based on the current paid salary combined with the budgeted expenses.

[0241] Specifically, the system first automatically generates an individual income tax withholding plan based on the predicted salary; then calculates the risk indices in four dimensions of law, tax, contract, and operation through a multi-dimensional risk assessment model to form a comprehensive risk score; for high-risk situations (risk index > 60), the system automatically triggers an early warning mechanism; at the same time, based on risk assessment, employee performance, and budget constraints, the system dynamically calculates a salary adjustment plan to optimize the next payment amount.

[0242] Step 4: For employees' working hour information and contract text information, construct a biometric verification model to verify employees' identities and identify the authenticity of contracts.

[0243] Specifically, the system uses distributed fingerprint recognition technology to verify employees' identities without sharing complete biometric data; combines OCR technology and document structure analysis to verify the authenticity of contracts; improves verification accuracy through multi-modal biometric fusion; performs weighted synthesis of the verification results of each platform on the federal server to form a final verification conclusion.

[0244] Step 5: Record the model updates and decision-making processes through blockchain to ensure the transparency and traceability of the system.

[0245] Specifically, the system stores the model update records of federated learning on the blockchain to ensure the transparency and traceability of the model training process; uses smart contracts to automatically execute federated learning rules and data access permission management; introduces an innovative proof-of-weight consensus mechanism to associate the model contribution degree with the block generation right; saves the model parameters and decision records through the IPFS distributed storage system.

[0246] Step 6: Optimize the matching relationship between employees and positions based on the reinforcement learning algorithm.

[0247] Specifically, the system first extracts the spatio-temporal characteristics of employees and positions; uses a multi-agent system to simulate the interaction behaviors of employees and enterprises; learns the optimal matching strategy through a deep Q network; uses the Hungarian algorithm to achieve global optimal matching; continuously evaluates the matching results and optimizes the matching algorithm.

[0248] Through the synergistic effect of the above six steps, the present invention realizes efficient and accurate compliance risk prediction and resource optimization allocation on the premise of protecting data privacy, and effectively solves the core challenges in flexible employment management.

[0249] It should be noted that the above are only the preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A flexible employment compliance decision-making system driven by federated learning, characterized in that It includes: A privacy protection module for performing homomorphic encryption, differential privacy, and randomization processing on the employee information of the flexible employment platform; A federated learning module connected to the privacy protection module for performing federated learning on the processed information on sub-nodes to generate an employee salary level prediction model; A decision-making module connected to the federated learning module for calculating an individual income tax withholding plan and a risk index based on the prediction model; A biometric verification module for verifying the authenticity of employee identities and contract texts.

2. The system according to claim 1, wherein The privacy protection module includes: A data desensitization unit for desensitizing the employee's basic registration information, historical salary information, and attendance performance information; An encryption unit for performing homomorphic encryption processing on the desensitized data; A differential privacy unit for adding accurately calculated noise to the encrypted data; A randomization unit for introducing random factors during data transmission and processing.

3. The system according to claim 1, wherein The federated learning module includes: A local model unit for locally training an employee working hours prediction model on each flexible employment platform; A model aggregation unit connected to the local model unit for aggregating the model parameters of each platform on the federated server; A global model unit connected to the model aggregation unit for generating a global employee working hours prediction model and a salary level prediction model; A model distribution unit connected to the global model unit for distributing the global model to each platform.

4. The system according to claim 1, wherein The decision-making module includes: An individual income tax calculation unit for calculating an individual income tax withholding plan based on the salary level prediction model; A risk assessment unit for calculating a labor law risk index, and the risk index includes a comprehensive score of legal dimension, tax dimension, contract dimension, and operation dimension; A salary adjustment unit for calculating and adjusting the next payment amount based on the current paid salary and the budgeted expenses; A visualization unit for visually displaying the individual income tax withholding plan and the risk index.

5. The system according to claim 1, wherein The biometric verification module includes: A fingerprint recognition unit for judging the authenticity of an employee's identity through a distributed fingerprint recognition algorithm; An OCR recognition unit for verifying the authenticity of a contract text through a document structure consistency verification algorithm; A multi-modal fusion unit for combining multiple biometric features to improve verification accuracy; A federated verification unit for comprehensively calculating the verification results of each platform on the federated server.

6. The system according to claim 1, wherein It also includes: A blockchain module connected to the federated learning module and the decision-making module for recording the model update process and compliance decision results; A matching optimization module connected to the decision-making module for optimizing the matching of employees and positions based on a reinforcement learning algorithm.

7. The system according to claim 6, characterized in that, The blockchain module includes: A smart contract unit for defining and executing federated learning rules and data access permissions; A distributed storage unit for storing model update records and compliance decision results; A consensus mechanism unit for combining federated learning weight aggregation with a blockchain consensus algorithm.

8. The system according to claim 6, wherein The matching optimization module includes: A feature extraction unit for extracting the spatio-temporal features of employees and positions; A multi-agent unit for establishing a matching relationship based on work location, time, position, and skills; A reinforcement learning unit for optimizing the matching strategy through a reward mechanism; A matching result unit for generating and outputting an optimal matching solution.

9. The system according to claim 3, wherein The built-in local model unit has an LSTM deep learning network for capturing the temporal patterns of work and compensation; the global model unit uses knowledge distillation technology to support the joint training of different structure models.

10. A flexible employment compliance decision-making method driven by federated learning, using the system described in any one of the claims, characterized in that, It includes: Performing privacy protection processing on the basic registration information, historical salary payment information, and attendance performance information of employees collected on each flexible employment platform by using homomorphic encryption, differential privacy, and randomization techniques; Calculating the processed information on each sub-node to implement the federated learning of the model and obtaining a prediction model for the future salary level of employees; Based on the salary level prediction model, calculating the individual income tax withholding plan and multi-dimensional risk index, and calculating and adjusting the next payment amount according to the current paid salary combined with the budgeted expenses; Constructing a biometric verification model for the employee's working hours information and contract text information to verify the employee's identity and identify the authenticity of the contract; Recording the model update and decision-making process through the blockchain to ensure the transparency and traceability of the system; Optimizing the matching relationship between employees and positions based on the reinforcement learning algorithm.

Images