Human resource cross-domain privacy screening system and method based on federated learning
Patent Information
- Application Number
- CN202610679003.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-18
- Publication Date
- 2026-10-13
AI Technical Summary
[0003]现有技术中，大多数人力资源筛选系统依赖于集中式数据汇聚的方式进行候选人评估，要求各参与方将候选人原始数据统一上传至中心服务器进行处理，无法满足不同组织对候选人多源敏感信息的隐私保护需求，难以实现在候选人原始数据不出域前提下的跨域能力评估与筛选，在涉及多个人力资源节点、多岗位类型的大规模跨域招聘场景中，仅使用简单的数据脱敏或静态加密方式进行候选人数据保护，面对跨节点的海量候选人特征数据融合时，容易出现隐私保护粒度不足、敏感维度区分缺失和数据泄露风险不可控的问题
[0014]本发明的有益效果在于：本发明构建分层隐私编码与渐进式跨域筛选体系，将各能力评估维度按隐私暴露风险系数从低到高划分至对应隐私层级并施加递增强度的差分隐私噪声，实现对候选人特征数据的差异化隐私保护，在初轮筛选阶段仅释放低隐私风险层级的隐私编码，对通过筛选的候选人逐步释放更高层级的隐私编码进行精筛，被淘汰的候选人仅暴露最低限度的隐私信息，提高隐私保护的精细化程度和候选人敏感数据的安全性，通过构建跨域身份碰撞检测与多源评估融合机制，结合脱敏身份哈希值的一致性比对确认不同人力资源节点中的同一自然人记录，对跨域候选人在不同节点的各轮次筛选评分进行加权融合，提升筛选结果的完整性和准确性，同时基于各隐私层级包含的各能力评估维度的隐私暴露风险系数计算各轮次的层级可信度权重，通过可信度加权求和得到各目标候选人的综合适配评分，实现评分结果对隐私噪声干扰的自适应补偿，确保最终跨域隐私筛选方案的科学性和合理性。
Smart Images
Figure CN122905613A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of privacy screening technology, specifically to a cross-domain privacy screening system and method for human resources based on federated learning. Background Technology
[0002] Human resource screening, as a core part of talent recruitment, directly impacts recruitment quality and candidate rights protection through its privacy protection capabilities, cross-domain collaboration efficiency, and the scientific rigor of screening results. The entire cross-domain screening process involves multiple key control points for candidates, and any deficiency in privacy protection at any stage may lead to the exposure of sensitive candidate information or deviations in screening results. Therefore, it is essential to build a human resource cross-domain privacy screening system based on federated learning to achieve accurate candidate screening and intelligent cross-domain data fusion under multi-dimensional privacy constraints.
[0003] In existing technologies, most human resource screening systems rely on centralized data aggregation for candidate evaluation, requiring all participating parties to upload candidate raw data to a central server for processing. This fails to meet the privacy protection needs of different organizations for multi-source sensitive information of candidates and makes it difficult to achieve cross-domain capability assessment and screening without leaving the candidate's original data domain. In large-scale cross-domain recruitment scenarios involving multiple human resource nodes and multiple job types, using only simple data desensitization or static encryption methods for candidate data protection is prone to problems such as insufficient privacy protection granularity, lack of differentiation of sensitive dimensions, and uncontrollable data leakage risks when facing the fusion of massive candidate feature data across nodes.
[0004] Meanwhile, existing technologies lack the ability to hierarchically encode and progressively release candidate feature data according to privacy sensitivity levels. They also lack proactive quantification and differentiated protection mechanisms for privacy exposure risks across different capability assessment dimensions. They cannot effectively avoid the problems of loss of assessment accuracy in low-sensitivity dimensions or insufficient privacy protection in high-sensitivity dimensions caused by uniform noise intensity. Furthermore, existing systems often release all candidate feature data at once during cross-domain screening, resulting in excessive privacy exposure, excessive leakage of information on eliminated candidates, and lack of privacy control between screening rounds. In terms of identifying duplicate candidates across domains and fusing multi-source assessment data, they cannot identify candidate records pointing to the same natural person in different human resource nodes through privacy-encoded sequence similarity analysis and desensitized identity hash collision detection, leading to duplicate recommendations or fragmented multi-source assessment information in the final screening results. Summary of the Invention
[0005] The purpose of this invention is to provide a cross-domain privacy screening system and method for human resources based on federated learning, which solves the problems existing in the background technology.
[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The first aspect of the present invention provides a cross-domain privacy screening system for human resources based on federated learning, including: a cross-domain hierarchical privacy coding module, used to establish a secure aggregation channel between the server and each human resources node, wherein each human resources node analyzes the local feature gradient and job requirement coding vector of each candidate locally, and performs hierarchical division and differential privacy perturbation processing, thereby generating privacy codes of each level of each candidate in each human resources node.
[0007] The cross-domain screening and privacy control module is used to upload the job requirement code vector and the first-level privacy code of each candidate in each human resources node to the server through a secure aggregation channel, and perform the initial cross-domain screening to obtain each candidate in the initial screening of each human resources node. It also uploads the second-level privacy code of each candidate in the initial screening of each human resources node, and performs cross-domain screening in the fine screening rounds in a loop to obtain each target candidate in each human resources node and their screening scores in each round.
[0008] The cross-domain identity collision and fusion module is used to perform cross-domain privacy identity collision detection on each target candidate at each human resources node, thereby constructing a cross-domain fusion profile of each target candidate.
[0009] The scheme compliance generation module is used to sort the suitability and verify the privacy compliance of each target candidate based on the cross-domain integrated profile, and generate a cross-domain privacy screening scheme.
[0010] The second aspect of the present invention provides a cross-domain privacy screening method for human resources based on federated learning, which is implemented based on the aforementioned cross-domain privacy screening system for human resources based on federated learning, including: Step 1. Establish a secure aggregation channel between the server and each human resources node. Each human resources node analyzes the local feature gradient and job requirement encoding vector of each candidate locally, and performs hierarchical division and differential privacy perturbation processing to generate privacy codes for each level of each candidate in each human resources node.
[0011] Step 2. Upload the job requirement code vector and the first-level privacy code of each candidate in each human resources node to the server through a secure aggregation channel, and perform the initial cross-domain screening to obtain the initial screening candidates of each human resources node. Upload the second-level privacy code of each initial screening candidate in each human resources node, and perform the fine screening rounds of cross-domain screening in a loop to obtain the target candidates of each human resources node and their screening scores in each round.
[0012] Step 3. Perform cross-domain privacy identity collision detection on each target candidate at each human resources node to construct a cross-domain fusion profile of each target candidate.
[0013] Step 4. Based on the cross-domain integrated profiles of each target candidate, sort them by suitability and verify their privacy compliance to generate a cross-domain privacy screening scheme.
[0014] The beneficial effects of this invention are as follows: This invention constructs a hierarchical privacy coding and progressive cross-domain screening system. Each capability assessment dimension is divided into corresponding privacy levels according to its privacy exposure risk coefficient from low to high, and differential privacy noise with progressively increasing intensity is applied. This achieves differentiated privacy protection for candidate feature data. In the initial screening stage, only privacy codes at low privacy risk levels are released. Higher-level privacy codes are gradually released for further screening of candidates who pass the initial screening. Eliminated candidates are exposed only to the minimum amount of privacy information, improving the refinement of privacy protection and the security of sensitive candidate data. This is achieved by constructing a cross-domain identity... The system integrates collision detection and multi-source evaluation, combining consistency comparison of de-identified identity hash values to confirm the same natural person record in different human resources nodes. It weights and fuses the screening scores of cross-domain candidates in each round at different nodes to improve the completeness and accuracy of the screening results. At the same time, it calculates the level credibility weight of each round based on the privacy exposure risk coefficient of each capability assessment dimension included in each privacy level. The comprehensive fit score of each target candidate is obtained by weighted summation of credibility, realizing adaptive compensation of the scoring results for privacy noise interference, and ensuring the scientificity and rationality of the final cross-domain privacy screening scheme. Attached Figure Description
[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0016] Figure 1 This is a schematic diagram of the system modules of the present invention.
[0017] Figure 2 This is a schematic diagram of the method flow of the present invention. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] Reference Figure 1As shown, the present invention provides a cross-domain privacy screening system and method for human resources based on federated learning, including: a cross-domain hierarchical privacy coding module, a cross-domain screening and privacy management module, a cross-domain identity collision and fusion module, and a scheme compliance generation module.
[0020] It should be noted that the cross-domain layered privacy coding module is connected to the cross-domain filtering and privacy management module, the cross-domain filtering and privacy management module is connected to the cross-domain identity collision and fusion module, and the cross-domain identity collision and fusion module is connected to the solution compliance generation module.
[0021] The cross-domain hierarchical privacy coding module is used to establish a secure aggregation channel between the server and each human resources node. Each human resources node analyzes the local feature gradient and job requirement coding vector of each candidate locally, and performs hierarchical division and differential privacy perturbation processing to generate privacy codes for each candidate at each level in each human resources node.
[0022] In one specific implementation, a secure aggregation channel is established. The specific method is as follows: the server and each human resources node conduct two-way authentication through a key negotiation protocol, negotiate and generate a session key, and encrypt the data transmitted in the channel end-to-end based on the session key, thereby establishing a secure aggregation channel between the server and each human resources node. The privacy codes and job requirement code vectors uploaded by each human resources node at each level are transmitted through this encrypted channel. The server can only receive the encrypted data and cannot obtain the original candidate data stored locally by each human resources node.
[0023] In a specific embodiment of the present invention, the local feature gradient and job requirement encoding vector of each candidate are analyzed. The specific method is as follows: each human resources node collects the structured resume data, performance evaluation records and job experience data of each candidate locally.
[0024] Each human resources node decomposes the structured resume data of each candidate into capability dimensions using a pre-deployed feature coding model, extracts the feature vectors of each candidate in each capability assessment dimension, and performs backpropagation calculations in combination with the performance evaluation records and job resume data of each candidate to generate the comprehensive feature values of each candidate in each capability assessment dimension in each human resources node, and uses them as the local feature gradient of each candidate in each human resources node.
[0025] It should be noted that the pre-deployed feature encoding model is a deep neural network model that is uniformly trained by the server and distributed to each human resources node during the system initialization phase. This model takes the candidate's structured resume data as input and decomposes the candidate's education, skills, work experience, and other information into competency dimensions through a multi-layer encoding network. It outputs feature vectors for each candidate in each competency assessment dimension. Each human resources node runs this model independently on its local machine without uploading the candidate's raw data to the server.
[0026] In one specific implementation, the comprehensive feature values of each candidate in each competency assessment dimension are generated at each human resources node. The specific method is as follows: Each human resources node extracts the initial competency score for each candidate in each competency assessment dimension based on the candidate's feature vector in each competency assessment dimension; based on each candidate's performance evaluation records, it extracts the performance score for each candidate in each competency assessment dimension for each historical assessment period; it calculates the arithmetic mean of the performance scores for each historical assessment period, and then divides it by the maximum performance score to obtain the performance correction coefficient for each candidate in each competency assessment dimension; based on each candidate's job experience... According to the data, the cumulative years of work experience and the highest job level number of each candidate in each competency assessment dimension are extracted. The cumulative years of work experience are divided by the benchmark years of work experience for that competency assessment dimension to obtain the years of work experience ratio. The highest job level number is divided by the benchmark job level number for that competency assessment dimension to obtain the level ratio. The arithmetic mean of the years of work experience ratio and the level ratio is used as the resume weighting coefficient for each candidate in each competency assessment dimension. The initial competency score, performance correction coefficient and resume weighting coefficient of each candidate in each competency assessment dimension are multiplied to calculate the comprehensive characteristic value of each candidate in each competency assessment dimension.
[0027] Each human resources node collects the text of the target job description of each candidate and performs semantic segmentation and vectorization encoding to generate the job requirement encoding vector for each candidate in each human resources node.
[0028] In one specific implementation, semantic segmentation and vectorization encoding are performed to generate job requirement encoding vectors for each candidate in each human resources node. The specific method is as follows: the existing semantic segmentation and vectorization encoding technologies are relatively mature, and job requirement encoding vectors for each candidate in each human resources node can be generated based on the existing technologies.
[0029] In a specific embodiment of the present invention, the privacy codes of each candidate at each level in each human resources node are generated. The specific method is as follows: the privacy-sensitive weights of each ability assessment dimension are obtained from the local database, and the privacy exposure risk coefficients of each candidate in each human resources node in each ability assessment dimension are evaluated based on the comprehensive feature values of each candidate in each human resources node in each ability assessment dimension.
[0030] In a specific implementation, the privacy exposure risk coefficient of each candidate in each ability assessment dimension at each human resources node is evaluated. The specific method is as follows: based on the privacy sensitivity weight of each ability assessment dimension, the comprehensive feature value of each candidate in each ability assessment dimension is multiplied to calculate the privacy exposure risk coefficient of each candidate in each ability assessment dimension. The higher the privacy sensitivity weight and the larger the comprehensive feature value of the ability assessment dimension, the higher its privacy exposure risk coefficient, indicating that the amount of candidate privacy information that can be inferred when the data in that dimension is leaked is greater.
[0031] The system retrieves the capacity and ranking of each privacy level from the local database. It then sorts and groups each capability assessment dimension from low to high based on the privacy exposure risk coefficient. The sorted capability assessment dimensions are then sequentially assigned to the privacy levels until the corresponding privacy level has reached its capacity, thus obtaining the capability assessment dimensions included in each privacy level.
[0032] Based on the comprehensive feature values of each candidate's ability assessment dimensions at each privacy level in each human resources node, differential privacy noise with progressively increasing intensity is applied from low to high privacy level to generate privacy codes for each candidate at each level in each human resources node.
[0033] In one specific implementation, differential privacy noise with increasing intensity is applied to generate privacy codes for each candidate at each level in each human resources node. The specific method is as follows: obtain the basic noise intensity parameter and the level increment coefficient from the local database. The noise intensity of each privacy level is equal to the basic noise intensity parameter multiplied by the level increment coefficient raised to the power of the level number. The higher the level number, the more exponentially the noise intensity increases. For the comprehensive feature values of each candidate's ability assessment dimensions at each privacy level, a pseudo-random noise value with the corresponding noise intensity as a parameter is superimposed. The comprehensive feature values of each ability assessment dimension after superposition of noise are combined to generate the privacy codes for each candidate at each level.
[0034] The cross-domain screening and privacy control module is used to upload the job requirement code vector and the first-level privacy code of each candidate in each human resources node to the server through a secure aggregation channel, and perform the initial cross-domain screening to obtain each candidate in the initial screening of each human resources node. It also uploads the second-level privacy code of each candidate in the initial screening of each human resources node, and performs cross-domain screening in the fine screening rounds in a loop to obtain each target candidate in each human resources node and their screening scores in each round.
[0035] In a specific embodiment of the present invention, the preliminary screening candidates of each human resources node are obtained by means of: based on the job requirement coding vector of each candidate in each human resources node, and combined with the privacy coding of the first level of each candidate in each human resources node, the deviation between the ability and the corresponding requirement of each candidate in each ability assessment dimension in the first level is evaluated, and the initial matching score of each candidate in each human resources node is obtained by weighted fusion according to the privacy sensitive weight of each ability assessment dimension.
[0036] In a specific implementation, the deviation between the abilities of each candidate in each ability assessment dimension at the first level and the corresponding requirements is evaluated at each human resources node. The specific method is as follows: based on the job requirement coding vector of each candidate, the requirement benchmark value of each ability assessment dimension of the target job in the first level is obtained by parsing. Based on the privacy coding of each candidate in the first level, noise is restored to obtain the comprehensive feature value of each candidate in each ability assessment dimension at the first level. The absolute value of the comprehensive feature value of each candidate in each ability assessment dimension is subtracted from the corresponding requirement benchmark value, and then divided by the requirement benchmark value to obtain the deviation of each candidate in each ability assessment dimension at the first level. The smaller the deviation, the closer the ability is to the job requirements.
[0037] In a specific implementation, the initial fit score of each candidate at each human resources node is obtained. The specific method is as follows: based on the deviation of each candidate in each ability assessment dimension at the first level. Privacy sensitivity weights for each capability assessment dimension , where x is the candidate number, y is a positive integer greater than 2, and n is the number of each ability assessment dimension. Where m is a positive integer greater than 2, the initial fit score for each candidate is obtained. , where e represents the natural constant.
[0038] The initial screening threshold is obtained from the local database, and candidates whose initial fit scores are higher than the initial screening threshold are selected as the initial screening candidates for each human resources node.
[0039] In a specific embodiment of the present invention, each target candidate and its screening scores in each round are obtained from each human resources node. The specific method is as follows: each human resources node uploads the second-level privacy code of each initial screening candidate to the server through a secure aggregation channel.
[0040] Based on the privacy codes of each initial screening candidate at the first and second levels of each human resources node, and combined with the job requirement code vector of each initial screening candidate, a second screening is conducted to obtain the second-round fit score of each initial screening candidate at each human resources node, and then each second screening candidate at each human resources node is selected.
[0041] Upload the third-level privacy codes of each candidate in the second screening at each human resources node to perform subsequent fine screening rounds, and repeat the process until all levels of privacy codes have been used for screening or the number of candidates is lower than the preset target number, at which point the cycle terminates.
[0042] The candidates who pass the final screening will be designated as target candidates for each human resources node, and the fit scores from each round will be aggregated as the screening scores for each target candidate in each round.
[0043] The cross-domain identity collision and fusion module is used to perform cross-domain privacy identity collision detection on each target candidate at each human resources node, thereby constructing a cross-domain fusion profile of each target candidate.
[0044] In a specific embodiment of the present invention, a cross-domain fusion profile of each target candidate is constructed. The specific method is as follows: the privacy codes of each level released by each target candidate in each human resources node are spliced according to the level to generate the release privacy code sequence of each target candidate in each human resources node, and the cross-node similarity calculation is performed on the release privacy code sequences of each target candidate in different human resources nodes, so as to obtain the similarity between each target candidate in the same human resources node and other target candidates in different human resources nodes.
[0045] In one specific implementation, the similarity between each target candidate in the same human resources node and other target candidates in different human resources nodes is obtained by the following method: the released privacy codes of each target candidate in each human resources node are concatenated hierarchically to obtain the released privacy code sequence, which is regarded as a high-dimensional vector. The target candidates in different human resources nodes are paired up, and the cosine similarity calculation method is used to calculate the vector dot product of the released privacy code sequence vector of a target candidate in the same human resources node and the released privacy code sequence vector of a target candidate in different human resources nodes. Then, the product of the magnitudes of the two vectors is divided to obtain the similarity between the two.
[0046] The similarity threshold is obtained from the local database. If the similarity between a target candidate in the same human resources node and another target candidate in different human resources nodes is greater than the similarity threshold, then the target candidate pair to be verified is marked as a target candidate pair to be verified. Each target candidate pair to be verified is obtained and then de-identified and verified to confirm each cross-domain candidate.
[0047] For each cross-domain candidate, the average of their screening scores in each round at different human resources nodes is calculated. For non-cross-domain target candidates, the screening scores in each round are retained, thus obtaining the final screening scores of each target candidate in each round, which are used as the cross-domain integrated profile of each target candidate.
[0048] In a specific embodiment of the present invention, de-identification verification is performed to confirm each cross-domain candidate. The specific method is as follows: for each pair of target candidates to be verified, a de-identification verification request is initiated to the human resources node to which the two target candidates belong. A one-way hash function is used to generate the de-identified identity hash value of the two target candidates in each pair of target candidates to be verified. The de-identified identity hash value is then uploaded to the server through a secure aggregation channel.
[0049] In one specific implementation, the method for generating de-identified identity hash values for the two target candidates in each pair of target candidates to be verified is as follows: After receiving the de-identification verification request initiated by the server, each human resources node extracts the preset identity identifier field of the corresponding target candidate locally, uses the identity identifier field as input, and performs an irreversible hash operation through a preset one-way hash function to generate a fixed-length de-identified identity hash value. This hash value cannot be reversed to restore the original identity identifier field.
[0050] The server performs a consistency comparison of the de-identified identity hash values of the two target candidates in each pair of target candidates to be verified. If they match, the server confirms that the pair of target candidates to be verified belongs to the same natural person and marks them as cross-domain candidates, thereby confirming each cross-domain candidate.
[0051] The scheme compliance generation module is used to sort the suitability and verify the privacy compliance of each target candidate based on the cross-domain integrated profile, and generate a cross-domain privacy screening scheme.
[0052] In a specific embodiment of the present invention, a cross-domain privacy screening scheme is generated by performing fit ranking and privacy compliance verification. The specific method is as follows: based on the final screening scores of each target candidate in each round, and combined with the privacy exposure risk coefficients of each capability assessment dimension included in each privacy level corresponding to each round, a weighted fusion is performed to obtain the comprehensive fit score of each target candidate.
[0053] The candidates are sorted from highest to lowest based on their overall fit score, thus obtaining the final ranking sequence of each target candidate.
[0054] The target number of candidates to be screened is obtained from the local database. Candidates for the target number of candidates are selected according to the final sorting sequence, thereby obtaining each selected candidate. The comprehensive fit score of each selected candidate is extracted and combined to form a cross-domain privacy screening scheme.
[0055] In a specific embodiment of the present invention, the comprehensive fit score of each target candidate is obtained by means of: determining each privacy level participating in the evaluation in each round of each target candidate based on the final screening score of each target candidate in each round, and extracting the privacy exposure risk coefficient of each capability evaluation dimension included in each privacy level.
[0056] A credibility assessment is conducted based on the privacy exposure risk coefficients of each capability assessment dimension included in each privacy level of each target candidate in each round of evaluation, resulting in the credibility weight of each target candidate in each round.
[0057] In one specific implementation, the credibility weight of each target candidate in each round is obtained by: extracting the privacy exposure risk coefficients of each capability assessment dimension included in each privacy level based on the privacy level in which each target candidate participates in the evaluation in each round, and calculating the mean to obtain the average privacy exposure risk value of each target candidate in each round. , where p represents the number of each target candidate. q is a positive integer greater than 2, and i represents the round number. Let j be a positive integer greater than 2, to obtain the hierarchical credibility weight of each target candidate in each round. Since the differential privacy noise intensity applied to the higher level of privacy exposure risk coefficient is greater, the credibility of its score is lower, and the higher the average privacy exposure risk value, the smaller the credibility weight of the level.
[0058] The comprehensive suitability score of each target candidate is obtained by weighting and summing the screening scores of each round and the corresponding level credibility weight of each round.
[0059] In one specific implementation, a weighted summation is performed to obtain the comprehensive suitability score for each target candidate. The specific method is as follows: based on the final screening scores of each target candidate in each round... Corresponding to the level credibility weight of the round The comprehensive suitability score of each target candidate is obtained. .
[0060] Reference Figure 2 As shown, the second aspect of the present invention provides a cross-domain privacy screening method for human resources based on federated learning, which is implemented based on the aforementioned cross-domain privacy screening system for human resources based on federated learning, including: Step 1. Establish a secure aggregation channel between the server and each human resources node. Each human resources node analyzes the local feature gradient and job requirement encoding vector of each candidate locally, and performs hierarchical division and differential privacy perturbation processing to generate privacy codes for each level of each candidate in each human resources node.
[0061] Step 2. Upload the job requirement code vector and the first-level privacy code of each candidate in each human resources node to the server through a secure aggregation channel, and perform the initial cross-domain screening to obtain the initial screening candidates of each human resources node. Upload the second-level privacy code of each initial screening candidate in each human resources node, and perform the fine screening rounds of cross-domain screening in a loop to obtain the target candidates of each human resources node and their screening scores in each round.
[0062] Step 3. Perform cross-domain privacy identity collision detection on each target candidate at each human resources node to construct a cross-domain fusion profile of each target candidate.
[0063] Step 4. Based on the cross-domain integrated profiles of each target candidate, sort them by suitability and verify their privacy compliance to generate a cross-domain privacy screening scheme.
[0064] All formulas in this manual are dimensionless and calculated numerically. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.
[0065] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications or additions should fall within the protection scope of the present invention.
Claims
1. A cross-domain privacy screening system for human resources based on federated learning, characterized in that, include: The cross-domain hierarchical privacy coding module is used to establish a secure aggregation channel between the server and each human resources node. Each human resources node analyzes the local feature gradient and job requirement coding vector of each candidate locally, and performs hierarchical division and differential privacy perturbation processing to generate privacy codes for each candidate at each level in each human resources node. The cross-domain screening and privacy control module is used to upload the job requirement code vector and the first-level privacy code of each candidate in each human resources node to the server through a secure aggregation channel, and perform the initial cross-domain screening to obtain each candidate in the initial screening of each human resources node. It also uploads the second-level privacy code of each candidate in the initial screening of each human resources node, and performs cross-domain screening in the fine screening rounds in a loop to obtain each target candidate in each human resources node and their screening scores in each round. The cross-domain identity collision and fusion module is used to perform cross-domain privacy identity collision detection on each target candidate at each human resources node, thereby constructing a cross-domain fusion profile of each target candidate. The scheme compliance generation module is used to sort the suitability and verify the privacy compliance of each target candidate based on the cross-domain integrated profile, and generate a cross-domain privacy screening scheme.
2. The human resource cross-domain privacy screening system based on federated learning according to claim 1, characterized in that, The specific method for analyzing the local feature gradients and job requirement encoding vectors of each candidate is as follows: Each human resources node collects structured resume data, performance evaluation records, and job experience data of each candidate locally; Each human resources node decomposes the structured resume data of each candidate into capability dimensions through a pre-deployed feature coding model, extracts the feature vectors of each candidate in each capability assessment dimension, and performs backpropagation calculations in combination with the performance evaluation records and job resume data of each candidate to generate the comprehensive feature values of each candidate in each capability assessment dimension in each human resources node, and uses them as the local feature gradients of each candidate in each human resources node. Each human resources node collects the text of the target job description of each candidate and performs semantic segmentation and vectorization encoding to generate the job requirement encoding vector for each candidate in each human resources node.
3. The human resource cross-domain privacy screening system based on federated learning according to claim 2, characterized in that, The specific method for generating privacy codes for each candidate at each level in each human resources node is as follows: The privacy-sensitive weights of each ability assessment dimension are obtained from the local database. Based on the comprehensive feature values of each candidate in each ability assessment dimension in each human resources node, the privacy exposure risk coefficient of each candidate in each ability assessment dimension in each human resources node is evaluated. The capacity and ranking of each privacy level are obtained from the local database. Each capability assessment dimension is sorted and grouped from low to high according to the privacy exposure risk coefficient. The sorted capability assessment dimensions are then assigned to the privacy level in sequence until the capacity of the corresponding privacy level is reached, thus obtaining the capability assessment dimensions included in each privacy level. Based on the comprehensive feature values of each candidate's ability assessment dimensions at each privacy level in each human resources node, differential privacy noise with progressively increasing intensity is applied from low to high privacy level to generate privacy codes for each candidate at each level in each human resources node.
4. The human resource cross-domain privacy screening system based on federated learning according to claim 3, characterized in that, The specific method for obtaining the initial screening candidates at each human resources node is as follows: Based on the job requirement coding vector of each candidate in each human resources node, and combined with the privacy coding of the first level of each candidate in each human resources node, the deviation between the ability and the corresponding requirement of each candidate in each ability assessment dimension in the first level is evaluated. The initial matching score of each candidate in each human resources node is obtained by weighted fusion according to the privacy sensitive weight of each ability assessment dimension. The initial screening threshold is obtained from the local database, and candidates whose initial fit scores are higher than the initial screening threshold are selected as the initial screening candidates for each human resources node.
5. The human resource cross-domain privacy screening system based on federated learning according to claim 4, characterized in that, The specific method for obtaining the target candidates and their screening scores for each round at each human resources node is as follows: Each human resources node uploads the second-level privacy code of each candidate in the initial screening to the server through a secure aggregation channel; Based on the privacy codes of each initial screening candidate at the first and second levels of each human resources node, and combined with the job requirement code vector of each initial screening candidate, a second screening is conducted to obtain the second round of matching scores for each initial screening candidate at each human resources node, and then each second screening candidate at each human resources node is selected. Upload the third-level privacy codes of each candidate in the second screening at each human resources node to perform subsequent fine screening rounds, and repeat the process until all levels of privacy codes have been used for screening or the number of candidates is lower than the preset target number, at which point the cycle terminates. The candidates who pass the final screening will be designated as target candidates for each human resources node, and the fit scores from each round will be aggregated as the screening scores for each target candidate in each round.
6. The human resource cross-domain privacy screening system based on federated learning according to claim 3, characterized in that, The specific method for constructing the cross-domain fusion profile of each target candidate is as follows: The privacy codes of each target candidate released at each level in each human resources node are concatenated according to the level to generate the release privacy code sequence of each target candidate in each human resources node. Cross-node similarity calculation is performed on the release privacy code sequences of each target candidate in different human resources nodes to obtain the similarity between each target candidate in the same human resources node and other target candidates in different human resources nodes. The similarity threshold is obtained from the local database. If the similarity between a target candidate in the same human resources node and another target candidate in different human resources nodes is greater than the similarity threshold, then it is marked as a target candidate pair to be verified, thereby obtaining each target candidate pair to be verified, and performing de-identification verification on it to confirm each cross-domain candidate. For each cross-domain candidate, the average of their screening scores in each round at different human resources nodes is calculated. For non-cross-domain target candidates, the screening scores in each round are retained, thus obtaining the final screening scores of each target candidate in each round, which are used as the cross-domain integrated profile of each target candidate.
7. The human resource cross-domain privacy screening system based on federated learning according to claim 6, characterized in that, The specific method for performing de-identification verification to confirm each cross-domain candidate is as follows: For each pair of target candidates to be verified, a de-identification verification request is sent to their respective human resources nodes. A one-way hash function is used to generate the de-identified identity hash values of the two target candidates in each pair of target candidates to be verified. The de-identified identity hash values are then uploaded to the server through a secure aggregation channel. The server performs a consistency comparison of the de-identified identity hash values of the two target candidates in each pair of target candidates to be verified. If they match, the server confirms that the pair of target candidates to be verified belongs to the same natural person and marks them as cross-domain candidates, thereby confirming each cross-domain candidate.
8. The human resource cross-domain privacy screening system based on federated learning according to claim 6, characterized in that, The specific method for performing compatibility ranking and privacy compliance verification to generate a cross-domain privacy screening scheme is as follows: Based on the final screening scores of each target candidate in each round, and combined with the privacy exposure risk coefficients of each capability assessment dimension included in each privacy level in each round, a weighted fusion is performed to obtain the comprehensive suitability score of each target candidate. The candidates are sorted from highest to lowest based on their overall fit score to obtain the final ranking sequence of each target candidate. The target number of candidates to be screened is obtained from the local database. Candidates for the target number of candidates are selected according to the final sorting sequence, thereby obtaining each selected candidate. The comprehensive fit score of each selected candidate is extracted and combined to form a cross-domain privacy screening scheme.
9. The human resource cross-domain privacy screening system based on federated learning according to claim 8, characterized in that, The specific method for obtaining the comprehensive suitability score of each target candidate is as follows: Based on the final screening scores of each target candidate in each round, the privacy levels of each target candidate in each round of evaluation are determined, and the privacy exposure risk coefficients of each capability evaluation dimension included in each privacy level are extracted. Based on the privacy exposure risk coefficients of each capability assessment dimension included in each privacy level of each target candidate in each round of evaluation, a credibility assessment is conducted to obtain the level credibility weight of each target candidate in each round. The comprehensive suitability score of each target candidate is obtained by weighting and summing the screening scores of each round and the corresponding level credibility weight of each round.
10. A method for cross-domain privacy screening of human resources based on federated learning, implemented based on the cross-domain privacy screening system for human resources based on federated learning as described in any one of claims 1-9, characterized in that, include: Step 1. Establish a secure aggregation channel between the server and each human resources node. Each human resources node analyzes the local feature gradient and job requirement encoding vector of each candidate locally, and performs hierarchical division and differential privacy perturbation processing to generate privacy codes for each candidate at each level in each human resources node. Step 2. Upload the job requirement code vector and the first-level privacy code of each candidate in each human resources node to the server through a secure aggregation channel, and perform the initial cross-domain screening to obtain each candidate in the initial screening of each human resources node. Upload the second-level privacy code of each candidate in the initial screening of each human resources node, and perform the fine screening rounds of cross-domain screening in a loop to obtain each target candidate in each human resources node and their screening scores in each round. Step 3. Perform cross-domain privacy identity collision detection on each target candidate at each human resources node to construct a cross-domain fusion profile of each target candidate; Step 4. Based on the cross-domain integrated profiles of each target candidate, sort them by suitability and verify their privacy compliance to generate a cross-domain privacy screening scheme.