Trans-department data collaborative modeling method and system based on federal learning

Abstract

The invention provides a cross-department data collaborative modeling method and system based on federated learning in the cross technical field of federated learning and privacy computing, and the method comprises the steps: S1, carrying out desensitization operation on department data by each edge computing node, and coding the desensitization data to obtain a feature coding vector; s2, constructing a layered federated learning framework based on a department private feature extractor, a cross-department parameter collaborative aggregation layer and a global task header; s3, extracting data features from the feature coding vectors through a department private feature extractor so as to train the model and extract model parameters; s4, performing hierarchical quantization coding, encryption and compression on the model parameters to obtain compression parameters, and sharing the compression parameters by a cross-department parameter collaborative aggregation layer; and S5, the global task head reads the shared compression parameters to carry out collaborative modeling on the model. The cross-department data collaborative modeling method has the advantages that the precision and efficiency of cross-department data collaborative modeling are greatly improved on the premise that the safety is guaranteed.

Description

Technical Field

[0001] The present invention relates to the interdisciplinary field of federated learning and privacy-preserving computing, and in particular to a method and system for cross-departmental data collaborative modeling based on federated learning. Background Art

[0002] With the rapid development of big data and artificial intelligence technologies, cross-departmental data sharing has become a critical requirement for improving business intelligence in sectors such as healthcare, finance, and government. The core goal of data sharing is to unlock the combined value of distributed data through collaborative modeling of multi-source data. However, traditional data sharing models typically require centralized storage and joint modeling of raw data from different departments. This model faces the following significant challenges: 1. Data privacy and compliance risks: Data from various departments may contain sensitive information (such as personal identity and medical records). Direct sharing is not only non-compliant but also poses a risk of leakage. 2. Data heterogeneity: Cross-departmental data often has different formats (such as structured and unstructured data coexisting), making it difficult for traditional centralized training to efficiently integrate multi-source features. 3. Inefficient collaboration: Failure to optimize for issues such as uneven data volume and limited communication bandwidth between departments affects collaboration efficiency.

[0003] In recent years, federated learning (FL) technology has transformed data sharing into collaborative optimization of model parameters through a "data remains static, model moves" mechanism, providing new insights for distributed modeling under privacy protection. However, existing FL technology still has the following shortcomings in cross-departmental scenarios: 1. The contradiction between privacy and utility: Although traditional differential noise or homomorphic encryption schemes can alleviate privacy leakage, they will significantly reduce model accuracy and cannot meet the accuracy requirements of cross-departmental high-value data collaborative modeling; 2. Bottleneck of heterogeneous data collaboration: Most studies only focus on single scenarios of horizontal federation (same feature space) or vertical federation (same sample space), while cross-departmental data often has cross-heterogeneity of features and samples (such as imaging data of medical departments and text records of medical insurance departments). Existing methods lack the ability to jointly model multimodal and cross-domain features, which in turn affects modeling accuracy; 3. Limited system efficiency: Frequent exchange of high-dimensional model parameters between departments can easily lead to communication bandwidth overload, and the heterogeneous computing resources of edge devices (such as GPU / CPU hybrid deployment) are difficult to support distributed training of complex models.

[0004] This demonstrates that existing technologies have yet to truly achieve the deep coupling of "secure and controllable data sharing" and "efficient and accurate collaborative modeling." For example, in cross-departmental government approval scenarios, departments are unable to share raw data due to privacy concerns, while the inefficient model interaction of traditional federated learning makes it difficult to support real-time risk assessment. This leads to a persistent conflict between data barriers and business needs.

[0005] Therefore, how to provide a cross-departmental data collaborative modeling method and system based on federated learning to improve the accuracy and efficiency of cross-departmental data collaborative modeling while ensuring security has become an urgent technical problem to be solved. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a cross-departmental data collaborative modeling method and system based on federated learning to improve the accuracy and efficiency of cross-departmental data collaborative modeling while ensuring security.

[0007] In a first aspect, the present invention provides a cross-departmental data collaborative modeling method based on federated learning, including the following steps: Step S1: Edge computing nodes deployed in each department respectively preset a neural network model, and set the data sensitivity and department priority of the department data stored in each edge computing node; the department data includes structured data, unstructured data, and time-series data; Step S2: Each edge computing node performs differential desensitization operations on the local department data based on the data sensitivity to obtain desensitized data, performs feature encoding on the desensitized data to obtain a feature encoding vector, and performs feature alignment operations on the feature encoding vectors of each edge computing node through a cross-departmental knowledge graph; Step S3: Each edge computing node respectively constructs a hierarchical federated learning framework based on a department private feature extractor, a cross-departmental parameter collaborative aggregation layer, and a global task head; Step S4: Each edge computing node extracts data features from the feature encoding vector through a department private feature extractor, dynamically allocates computing resources through a heterogeneous computing resource adapter, uses each data feature to train the neural network model, and extracts the model parameters of the trained neural network model; Step S5: Each edge computing node performs hierarchical quantization encoding on the model parameters to obtain an encoding result, encrypts the encoding result into encrypted parameters through a homomorphic encryption algorithm, compresses each encrypted parameter through a bandwidth-aware compression strategy to obtain compressed parameters, and the cross-departmental parameter collaborative aggregation layer shares the compressed parameters to the cross-departmental parameter collaborative aggregation layers of the remaining edge computing nodes through a dynamic gating mechanism; Step S6: Each edge computing node reads the compressed parameters shared by the remaining edge computing nodes from the cross-departmental parameter collaborative aggregation layer through the global task head, and performs collaborative modeling on the local neural network model through the shared compressed parameters and department priorities; Step S7: Each edge computing node records the collaborative modeling log in real time, and stores the collaborative modeling log through a blockchain.

[0008] In a second aspect, the present invention provides a cross-departmental data collaborative modeling system based on federated learning, including the following modules: An initialization module, configured to pre-set a neural network model on edge computing nodes deployed in each department, and set the data sensitivity and department priority of the department data stored in each edge computing node; the department data includes structured data, unstructured data, and time-series data; A feature alignment module, configured to perform differential desensitization operations on local department data based on data sensitivity by each edge computing node to obtain desensitized data, perform feature encoding on the desensitized data to obtain feature encoding vectors, and perform feature alignment operations on the feature encoding vectors of each edge computing node through a cross-department knowledge graph; A hierarchical federated learning framework construction module, configured to construct a hierarchical federated learning framework on each edge computing node based on a department private feature extractor, a cross-department parameter collaborative aggregation layer, and a global task head respectively; A model training module, configured to extract data features from the feature encoding vectors by each edge computing node through a department private feature extractor, dynamically allocate computing resources through a heterogeneous computing resource adapter, so as to train the neural network model using each of the data features, and extract the model parameters of the trained neural network model; A parameter collaboration module, configured to perform hierarchical quantization encoding on the model parameters by each edge computing node to obtain an encoding result, encrypt the encoding result into encrypted parameters through a homomorphic encryption algorithm, compress each of the encrypted parameters through a bandwidth-aware compression strategy to obtain compressed parameters, and the cross-department parameter collaborative aggregation layer shares the compressed parameters to the cross-department parameter collaborative aggregation layers of the remaining edge computing nodes through a dynamic gating mechanism; A collaborative modeling module, configured to read the compressed parameters shared by the remaining edge computing nodes from the cross-department parameter collaborative aggregation layer by each edge computing node through the global task head, and perform collaborative modeling on the local neural network model through each of the shared compressed parameters and department priorities; A collaborative modeling log management module, configured to record the collaborative modeling log in real time by each edge computing node, and store the collaborative modeling log through a blockchain.

[0009] The advantages of the present invention are as follows: 1. Pre-set a neural network model on each edge computing node deployed in each department, set the data sensitivity and department priority of the department data stored in each edge computing node; each edge computing node performs differentiated desensitization operations on the local department data based on data sensitivity to obtain desensitized data, performs feature encoding on the desensitized data to obtain feature encoding vectors, and performs feature alignment operations on the feature encoding vectors of each edge computing node through a cross-departmental knowledge graph; each edge computing node builds a hierarchical federated learning framework based on a department-private feature extractor, a cross-departmental parameter collaborative aggregation layer, and a global task head, extracts data features from the feature encoding vector through a department-private feature extractor, and dynamically allocates computing resources through a heterogeneous computing resource adapter. The neural network model is trained by utilizing the features of each data, and the model parameters of the trained neural network model are extracted; each edge computing node performs hierarchical quantization encoding on the model parameters to obtain the encoding results, encrypts the encoding results into encryption parameters through the homomorphic encryption algorithm, and compresses each encryption parameter through the bandwidth-aware compression strategy to obtain compression parameters. The cross-department parameter collaborative aggregation layer shares the compression parameters to the cross-department parameter collaborative aggregation layer of the remaining edge computing nodes through the dynamic gating mechanism; each edge computing node reads the compression parameters shared by the remaining edge computing nodes from the cross-department parameter collaborative aggregation layer through the global task header, and collaboratively models the local neural network model through the shared compression parameters and department priorities, and records the collaborative modeling in real time. Logs, use blockchain to store collaborative modeling logs; that is, through differentiated desensitization and feature encoding of data sensitivity, combined with layered quantization encoding and homomorphic encryption, dual optimization of privacy protection and model accuracy is achieved - differentiated desensitization reduces information loss, feature encoding retains key information, and homomorphic encryption and layered quantization are linked to ensure the security of parameter transmission while reducing accuracy loss, breaking the contradiction between privacy and utility in traditional solutions; through cross-departmental knowledge graph feature alignment and layered federated learning framework, multi-modal data collaborative links are opened, cross-departmental knowledge graphs align features of different modal data, departmental private feature extractors adapt to local data structures, and cross-departmental parameter collaborative aggregation layers fuse multi-dimensional features of different departments, thereby solving horizontal / vertical The "feature-sample heterogeneous intersection" problem that is difficult for federation to handle is improved, and modeling accuracy is improved; the computing load is optimized by dynamically allocating GPU / CPU resources through heterogeneous computing resource adapters, and the parameter transmission volume is significantly reduced by combining bandwidth-aware compression strategies with layered quantization coding; at the same time, the dynamic gating mechanism filters redundant parameter transmissions. The three work together to reduce the pressure on communication bandwidth while taking into account the adaptation of edge device computing power, breaking the bottleneck of high-dimensional parameter transmission and heterogeneous resources, and improving distributed training efficiency; that is, through multi-dimensional technical innovations such as differentiated desensitization, layered quantization coding, homomorphic encryption, knowledge graph feature alignment, layered federated architecture, and dynamic resource adaptation, the accuracy and efficiency of cross-departmental data collaborative modeling are greatly improved while ensuring security.

[0010] 2. Through the deep integration of federated learning and edge computing, on the basis of ensuring data privacy (differential desensitization, homomorphic encryption) and security and trustworthiness (blockchain-based evidence storage), efficient collaborative modeling of cross-department heterogeneous data (structured / unstructured / temporal data) is achieved. Its dynamic resource adaptation mechanism (heterogeneous computing resource allocation, bandwidth-aware compression) significantly reduces communication and computing costs. The hierarchical federated architecture and cross-department knowledge graph effectively alleviate the data distribution shift problem and improve the model generalization ability through feature alignment and priority scheduling. At the same time, blockchain-based evidence storage and dynamic gating mechanism enhance the transparency of multi-party collaboration and the ability to resist "free-riding", and the support for the autonomy and ecological compatibility of edge nodes makes it highly implementable in complex scenarios in multiple fields, providing a systematic solution for breaking data islands and achieving secure and efficient cross-domain intelligent collaboration.

[0011] 3. The dynamic orchestration of computing resources is achieved through the hybrid container architecture of Docker + Kubernetes. The containerized pre-deployment of neural network models supports gray release and A / B testing, ensuring business continuity during model updates and effectively improving deployment efficiency.

[0012] 4. By introducing an exponential function with a data age decay factor λ = 0.5, the dynamic optimization of storage policies is achieved. Through a third-order weight adjustment mechanism (w c , w t , w s ), dynamic adjustment is supported according to compliance requirements. Based on the introduction of a Tmax threshold control in the calculation of data urgency, an alarm escalation is automatically triggered when Δt2 exceeds the maximum processing cycle, ensuring the timeliness requirements of critical business data.

[0013] 5. Through a data sensitivity quantification and evaluation system of a dynamic decay model (Ds = (C*w c + T*w t + S*w s ) / Σw), real-time optimization of data security classification and storage policies is achieved. Through a multi-factor coupling algorithm for department priorities (P = I*w i + U*w u + D*w d ), combined with data urgency threshold control and cross-node federated learning mechanism, intelligent scheduling and load balancing of computing resources are achieved.

[0014] 6. By refining data sensitivity into four levels (public / internal / confidential / top secret), and dynamically matching cryptographic algorithms (such as HMAC-SHA256, Paillier) and privacy protection technologies (k-anonymity, differential privacy) for different levels, through the positive correlation design of algorithm strength and sensitivity, while ensuring the security of high-level data, the waste of computing resources caused by over-encryption of low-level data is avoided.

[0015] 7. For structured data, hash replacement (public level) is adopted to balance reversibility and anonymity, k-anonymity (internal level) is used to balance data availability and privacy risks, and the Paillier algorithm is used at the top-secret level to implement homomorphic encryption to support data analysis under encrypted states; for unstructured data, semantic-preserving transformation (confidential level) is introduced for text data to avoid semantic breaks caused by traditional replacement, and fine-grained protection is achieved by combining paragraph-level AES encryption (top-secret level); for images / videos, reversible pixel scrambling technology is used to balance visual privacy protection and data recoverability, breaking through the limitations of the irreversibility of traditional mosaics; for audio data, a combination strategy of segment replacement and voiceprint transformation is used to effectively eliminate voiceprint features and prevent identity re-identification attacks; that is, a complete desensitization technology matrix covering multi-modal data is established, filling the gap in traditional solutions that only target single data types.

[0016] 8. By adopting a three-stage processing architecture (modal encoding layer → cross-modal fusion layer → feature compression layer): dual-modal interaction channels achieve deep association of structured / unstructured / temporal features (such as causal analysis of text semantics and time series); the joint learning strategy dynamically adjusts cross-modal weights through the attention mechanism to solve the problem of feature redundancy in traditional multi-modal fusion; low-dimensional dense vectors are generated through regularization dimensionality reduction, which not only retains cross-modal association features but also meets the high-efficiency computing requirements of subsequent graph neural networks.

[0017] 9. By limiting high-computation-complexity operations (such as Paillier encryption and Fourier transform) to the top-secret data processing link and using lightweight algorithms (HMAC-SHA256) for low-secrecy-level data, the average power consumption of edge nodes is significantly reduced; through the decoupled design of the sensitivity grading standard and the desensitization algorithm library, it supports dynamic expansion of new algorithms through configuration files (such as replacing Paillier with an LWE-based solution), meeting the requirements of NIST's post-quantum cryptography migration.

[0018] 10. By implementing double protection of "encryption + desensitization" for top-secret data (such as performing shape-preserving transformation on Paillier-encrypted data), it resists the coordinated attacks of model inversion attacks and password cracking, breaks through the traditional single-layer protection thinking, and increases the attack cost through algorithm nesting; the shape-preserving transformation technology for time-series data (internal / confidential level) destroys the correlation between time series and external data while retaining statistical features, preventing the inference of user identity through time-series association.

[0019] 11. By building a dynamic desensitization system driven by sensitivity grading, it innovatively integrates cryptographic algorithms, privacy protection technologies and multimodal data processing capabilities to achieve differentiated and precise protection of structured, unstructured and time-series data (such as lightweight hash replacement of public-level data and Paillier homomorphic encryption of top-secret data). It also introduces breakthrough new desensitization methods such as reversible pixel scrambling and semantic-preserving deformation to maintain business availability while protecting data privacy. Through the collaboration of multimodal encoders and graph attention networks, it enables cross-modal feature alignment and knowledge sharing at the edge nodes, solves the data silo problem and improves the efficiency of federated learning. Its modular architecture is compatible with dynamic compliance configuration and intelligent resource scheduling, supporting low-power and high-reliability deployment in edge computing scenarios.

[0020] 12. By using a private feature extractor to process feature coding vectors only locally, there is no need to share original data, thus meeting the requirement for local processing of sensitive data and avoiding the risk of privacy leakage caused by cross-departmental transmission. By abstracting feature coding vectors, it is easier to implement privacy protection measures such as noise addition than with original data, providing a compliance basis for subsequent parameter sharing.

[0021] 13. By building a three-layer architecture of "local feature extraction - cross-departmental parameter collaboration - global model fine-tuning", it innovatively solves the problems of data privacy, communication efficiency and model generalization collaboration in edge computing scenarios. Its core advantages are: ① The adoption of local feature encoding and compression parameter sharing mechanism effectively reduces communication overhead while ensuring data local storage requirements; ② Through the federated transfer learning architecture of the global task head, dynamic feature alignment and personalization-generalization dual-objective optimization under Non-IID data distribution are achieved, effectively improving model accuracy.

[0022] 14. By abstracting different types of computing resources into a unified resource model, global perception and intelligent scheduling of diversified hardware resources such as CPU and GPU in edge computing nodes are achieved, significantly improving resource utilization and reducing task response delay. Combined with dynamic energy efficiency optimization algorithms, power consumption is reduced while ensuring service quality, forming a complete technical barrier covering resource modeling, task scheduling and energy efficiency management, with outstanding compatibility and scalability.

[0023] 15. Divide the core layer and the auxiliary layer based on the sensitivity of model parameters to inference accuracy. The core layer uses 8-bit mixed-precision quantization (balancing accuracy and computational efficiency), and the auxiliary layer uses 4-bit sparse coding (significantly reducing storage and transmission overhead). On the premise of ensuring the model inference quality, achieve the optimal allocation of computing resources; by recording the quantization error distribution and generating a compensation vector, effectively offset the accuracy loss caused by low-bit quantization and improve the model robustness; through the error distribution recording and compensation vector generation, form a quantization-compensation dynamic closed-loop to reduce the risk of model performance degradation caused by low-precision quantization.

[0024] 16. By setting the core parameters to adopt the CKKS scheme (supporting floating-point homomorphic operations and suitable for high-precision inference), and the auxiliary parameters and compensation vector to adopt the BFV scheme (higher integer operation efficiency), balance security and computational efficiency and avoid the performance bottleneck caused by a single encryption scheme; through the combination of encryption and quantization coding, prevent the parameters from being reverse-cracked during transmission and collaboration.

[0025] 17. Predict the future transmission window state by real-time monitoring of network metrics (such as link bandwidth, delay jitter, and packet loss rate), and solve the problem of dynamic fluctuations in the edge network environment; adaptively adjust the compression algorithm according to the prediction results (such as selecting entropy coding or differential coding) to maximize the compression ratio and transmission efficiency and reduce the communication delay between edge nodes.

[0026] 18. Adjust the parameter transmission granularity through a gating mechanism (such as transmitting by block or by layer), effectively balance the transmission load and the timeliness of model aggregation; from quantization, encryption, compression to transmission, the full-process technical closed-loop comprehensively reduces the computing, storage, and communication resource consumption of edge nodes.

[0027] 19. Through hierarchical quantization coding and error compensation mechanism, significantly reduce the computing and transmission overhead while ensuring the inference accuracy of the neural network; adopt a mixed encryption strategy of CKKS and BFV to achieve hierarchical security protection of core parameters and auxiliary parameters, taking into account privacy and operation efficiency; combine LSTM network prediction and dynamic compression mode selection to adaptively cope with network fluctuations and optimize transmission efficiency; through dynamic gating granularity adjustment and lightweight encapsulation, achieve efficient collaboration and resource load balancing across edge nodes.

[0028] 20. Through the collaborative sharing mechanism of compressed and encrypted parameters between edge computing nodes, the efficiency of distributed modeling is significantly improved while ensuring data privacy and secure transmission: parameter compression is used to reduce communication overhead, and encryption and decompression are used to achieve end-to-end security. Combined with priority-driven dynamic queue scheduling, it not only optimizes the collaborative utilization of cross-departmental resources (enhancing model generalization through weighted aggregation), but also supports the rapid adaptation of local data (based on fine-tuning to retain personalized features). At the same time, through asynchronous queue processing and distributed storage architecture, it takes into account system scalability, fault tolerance and computing load balancing, providing a solution that takes into account both global optimization and local agility for the safe and efficient collaboration of multiple subjects in edge computing scenarios.

[0029] 21. A layered encryption structure is constructed by adopting hybrid encryption algorithms such as SM3 / SM4 (national secret algorithm), 3DES, and Blowfish. Different algorithms protect the header, body, and tail of the log respectively, forming algorithm-level security isolation. The key management system uses SM2 asymmetric encryption to protect symmetric keys (K1 / K2 / K3), realizing security reinforcement of the key transmission link. The AES-GCM algorithm is used to ensure the confidentiality and integrity of the final encrypted log, and dynamic parameters are combined to enhance timeliness protection.

[0030] 22. By inserting random strings into / dev / urandom (5 bytes of noise are inserted every 10 bytes), the entropy value of the data stream is significantly increased; the dual obfuscation technology of Base64 encoding and Caesar cipher is used to effectively resist pattern recognition attacks; the segmented encryption strategy destroys the continuity of the data structure and increases the difficulty of data parsing.

[0031] 23. Generate dynamic parameters by binding the system timestamp to the MAC address to establish dual verification of device fingerprint and time dimension; ensure the authenticity of data source and prevent replay attacks through the AES-GCM authentication tag mechanism; form an unalterable chain of evidence through HMAC data fingerprint and blockchain evidence storage.

[0032] 24. By building a multi-level security protection system (integrating national secret algorithms, hybrid encryption and dynamic noise injection) and a dynamic defense mechanism (timestamp-MAC address binding and quantum-resistant design), while ensuring data confidentiality, integrity and tamper-resistance, the coordinated optimization of encryption efficiency and system reliability in edge computing scenarios is achieved, forming an independent and controllable technical solution covering the entire link of "encryption-storage-verification-traceability", which has both high security and strong compatibility.

[0033] 25. A dynamic data masking strategy based on data sensitivity classification (public / internal / confidential / top secret) that uses different technologies such as HMAC-SHA256, k-anonymity, differential privacy, and Paillier encryption for structured / unstructured / temporal data respectively to achieve precise privacy protection while ensuring data availability; reversible pixel scrambling and audio voiceprint transformation technologies for unstructured data effectively balance the privacy and feature retention requirements of multimedia data.

[0034] 26. Through a differential classification and masking mechanism (such as using k-anonymity, pixel scrambling, and Fourier encryption for structured / unstructured / temporal data respectively) and a hierarchical federated learning architecture (departmental private feature extractor + cross-departmental parameter collaborative aggregation layer), efficient collaborative modeling of cross-departmental multi-modal data (structured / unstructured / temporal) is achieved while ensuring data privacy (supporting homomorphic encryption, differential privacy, and blockchain-based evidence storage); combined with dynamic resource scheduling (heterogeneous computing resource adapter to optimize GPU / FPGA utilization) and bandwidth-aware compression strategies (LSTM prediction network quality + hierarchical quantization coding), communication overhead and computing latency are significantly reduced. At the same time, through knowledge graph-driven feature alignment (graph attention network) and quantization error compensation mechanism (error feedback network), the loss of model accuracy is effectively reduced; its modular design (Docker / K8s containerized deployment) and dynamic priority scheduling (departmental weight allocation based on data urgency / dependency) further support flexible expansion and prioritized processing of critical tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The present invention will be further described below with reference to the accompanying drawings in conjunction with embodiments.

[0036] Figure 1 It is a flowchart of a cross-departmental data collaborative modeling method based on federated learning according to the present invention.

[0037] Figure 2 It is a schematic structural diagram of a cross-departmental data collaborative modeling system based on federated learning according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0038] The technical solution in the embodiment of the present application has the following overall idea: through differentiated desensitization and feature encoding of data sensitivity, combined with layered quantization encoding and homomorphic encryption, dual optimization of privacy protection and model accuracy is achieved - differentiated desensitization reduces information loss, feature encoding retains key information, and homomorphic encryption and layered quantization are linked to ensure the security of parameter transmission while reducing precision loss, breaking the contradiction between privacy and utility in traditional solutions; through cross-departmental knowledge graph feature alignment and layered federated learning framework, multi-modal data collaborative links are opened up, cross-departmental knowledge graphs perform feature alignment on different modal data, departmental private feature extractors adapt to local data structures, and cross-departmental parameter collaborative aggregation layers fuse multi-dimensional features of different departments, thereby solving horizontal / vertical The "feature-sample heterogeneous intersection" problem that is difficult for federation to handle is improved, and modeling accuracy is improved; the computing load is optimized by dynamically allocating GPU / CPU resources through heterogeneous computing resource adapters, and the parameter transmission volume is significantly reduced by combining bandwidth-aware compression strategies with layered quantization coding; at the same time, the dynamic gating mechanism filters redundant parameter transmissions. The three work together to reduce the pressure on communication bandwidth while taking into account the adaptation of edge device computing power, breaking the bottleneck of high-dimensional parameter transmission and heterogeneous resources, and improving distributed training efficiency; that is, through multi-dimensional technical innovations such as differentiated desensitization, layered quantization coding, homomorphic encryption, knowledge graph feature alignment, layered federated architecture, and dynamic resource adaptation, the accuracy and efficiency of cross-departmental data collaborative modeling can be improved while ensuring security.

[0039] Please refer to Figures 1 to 2 As shown, a preferred embodiment of the cross-departmental data collaborative modeling method based on federated learning of the present invention includes the following steps: Step S1: Preset a neural network model on each edge computing node deployed in each department, and set the data sensitivity and department priority of the department data stored in each edge computing node; the department data includes structured data, unstructured data, and time series data; Step S2: Each edge computing node performs differentiated desensitization operations on local department data based on data sensitivity to obtain desensitized data, performs feature encoding on the desensitized data to obtain feature encoding vectors, and performs feature alignment operations on the feature encoding vectors of each edge computing node through a cross-departmental knowledge graph; the cross-departmental knowledge graph is dynamically updated based on a preset update trigger mechanism; Step S3: Each edge computing node builds a hierarchical federated learning framework based on the department's private feature extractor, the cross-department parameter collaborative aggregation layer, and the global task head; Step S4: Each edge computing node extracts data features from the feature encoding vector using a department-specific feature extractor, dynamically allocates computing resources through a heterogeneous computing resource adapter, trains a neural network model using the data features, and extracts model parameters of the trained neural network model; The model parameters are the "knowledge" learned by the neural network model during the training process. These model parameters determine how the neural network model processes the input data and generates the output results. Simply put, these model parameters are like the "memory" of the neural network model, recording the patterns learned by the neural network model from the training data. By extracting the model parameters, they can be used to save the model, analyze the model, and perform transfer learning.

[0040] Step S5: Each edge computing node performs hierarchical quantization encoding on the model parameters to obtain an encoding result, encrypts the encoding result into encrypted parameters through a homomorphic encryption algorithm, compresses each encrypted parameter through a bandwidth-aware compression strategy to obtain compressed parameters, and the cross-department parameter collaborative aggregation layer shares the compressed parameters to the cross-department parameter collaborative aggregation layers of the remaining edge computing nodes through a dynamic gating mechanism; Step S6: Each edge computing node reads the compressed parameters shared by the remaining edge computing nodes from the cross-department parameter collaborative aggregation layer through the global task header, and performs collaborative modeling on the local neural network model through the shared compressed parameters and department priorities; Step S7: Each edge computing node records the collaborative modeling log in real time and stores the collaborative modeling log through the blockchain; Through the deep integration of federated learning and edge computing, on the basis of ensuring data privacy (differential desensitization, homomorphic encryption) and security and trustworthiness (blockchain-based evidence storage), it realizes the efficient collaborative modeling of cross-department heterogeneous data (structured / unstructured / temporal data); its dynamic resource adaptation mechanism (heterogeneous computing resource allocation, bandwidth-aware compression) significantly reduces communication and computing costs, and the hierarchical federated architecture and cross-department knowledge graph effectively alleviate the data distribution drift problem and improve the model generalization ability through feature alignment and priority scheduling; at the same time, blockchain-based evidence storage and dynamic gating mechanism enhance the transparency of multi-party collaboration and the ability to resist "free-riding", and the support for the autonomy and ecological compatibility of edge nodes makes it highly implementable in multi-domain complex scenarios, providing a systematic solution for breaking data islands and achieving secure and efficient cross-domain intelligent collaboration.

[0041] Specifically, step S1 is as follows: Edge computing nodes deployed in each department respectively pre-set a neural network model through a containerized operating environment built by Docker and Kubernetes; The dynamic orchestration of computing resources is achieved through the hybrid container architecture of Docker+Kubernetes; the containerized pre-setting of the neural network model supports gray release and A / B testing, ensuring business continuity during model updates and effectively improving the deployment efficiency.

[0042] Calculate the data sensitivity of the department data stored in each edge computing node based on the data content, data timeliness, and data source: D s = (C * w c + T * w t + S * w s ) / (w c + w t + w s ); Among them, D s represents the data sensitivity; C represents the data content sensitivity, and its value range is [0, 1]. When the data attributes of the department data are public data, restricted data, and core sensitive data respectively, the value ranges of the data content sensitivity are [0, 0.4], (0.4, 0.7], (0.7, 1]; T represents the data timeliness sensitivity, and its value range is [0, 1]. T = e^-λ*Δt1, where e represents the natural constant, λ represents the decay coefficient, λ = 0.5, and Δt1 represents the data age; S represents the data source sensitivity. When the data sources of the department data are internal core database, third-party partners, and public web crawling respectively, the values of the data source sensitivity are 1, 0.6, and 0.2; w c , w t , w s all represent weight coefficients; Calculate the department priority of the department data stored in each edge computing node based on the data importance, data urgency, and data dependence: P = I * w i + U * w u + D * w d ; Among them, P represents the department priority; I represents the data importance, and its value range is [0, 1]. When the data uses of the department data are auxiliary data, key business data, and strategic-level data respectively, the value ranges of the data importance are [0, 0.5], (0.5, 0.8], (0.8, 1]; U represents the data urgency, U = max(1 - Δt2 / T max , 0), where Δt2 represents the difference between the current time and the deadline, and T max represents the maximum allowed processing cycle; D represents the data dependence, and its value range is [0, 1]. For example, for a core patent document that is relied on by multiple subsequent processes (such as examination response, international stage submission), the value of the data dependence is 1; the value of the data dependence of intermediate data is 0.5; the value of the data dependence of independent data is 0.2; w i , w u , w d all represent weight coefficients; The department data includes structured data, unstructured data, and time-series data; the structured data includes at least personal information, financial data, approval data, statistical data, medical data, and educational data; the unstructured data includes at least document files, image data, audio data, video data, social media data, log data, and emails; the time-series data includes at least economic data, meteorological data, traffic data, environmental monitoring data, medical monitoring data, and industrial production data.

[0043] By introducing an exponential function with a data age decay factor λ = 0.5, the dynamic optimization of the storage strategy is achieved; through a third-order weight adjustment mechanism (w c , w t , w s ), it supports dynamic adjustment according to compliance requirements; based on the calculation of data urgency, a Tmax threshold control is introduced. When Δt2 exceeds the maximum processing cycle, an alarm upgrade is automatically triggered to ensure the timeliness requirements of critical business data.

[0044] Through a data sensitivity quantification and evaluation system of a dynamic decay model (Ds = (C*w c +T*w t +S*w s ) / Σw), the real-time optimization of data security classification and storage strategy is achieved; through a multi-factor coupling algorithm for department priorities (P = I*w i +U*w u +D*w d ), combined with data urgency threshold control and a cross-node federated learning mechanism, intelligent scheduling and load balancing of computing resources are achieved.

[0045] The specific steps of S2 are as follows: Each edge computing node maps the data sensitivity to public level, internal level, confidential level, or top-secret level based on a preset sensitivity classification standard; For the structured data in the local department data: when the data sensitivity is at the public level, the user identifier is calculated using HMAC-SHA256 to obtain a hash value H1, and the user identifier is replaced based on the hash value H1 to perform a desensitization operation to obtain desensitized data; when the data sensitivity is at the internal level, a k-anonymity algorithm (ensuring that each record has at least k - 1 identical records on the quasi-identifier) is used to perform a desensitization operation to obtain desensitized data; when the data sensitivity is at the confidential level, differential privacy noise addition is performed to perform a desensitization operation to obtain desensitized data; when the data sensitivity is at the top-secret level, it is encrypted using the Paillier algorithm to perform a desensitization operation to obtain desensitized data; For unstructured data in local department data: For text in document files, social media data, log data, and emails, when the data sensitivity is public or internal, perform entity replacement (replace the patient name with the [PATIENT] tag, retaining gender pronouns) to perform the desensitization operation to obtain desensitized data; when the data sensitivity is confidential, perform semantic-preserving transformation (rewrite the clinical description, maintaining medical logic but changing surface features) to perform the desensitization operation to obtain desensitized data; when the data sensitivity is top secret, use the AES-256-CTR algorithm for paragraph encryption; for image data, video data, social media data, and images in emails, perform reversible pixel scrambling on the privacy areas in the images to perform the desensitization operation to obtain desensitized data; for audio data, perform segment replacement, add noise, or voiceprint transformation to perform the desensitization operation to obtain desensitized data; For time series data in local department data: When the data sensitivity is public, perform timestamp obfuscation (convert the time accurate to seconds to an hour interval) to perform the desensitization operation to obtain desensitized data; when the data sensitivity is internal or confidential, perform shape-preserving transformation (perform dynamic time warping on the ECG signal, maintaining waveform characteristics but changing the time correspondence) to perform the desensitization operation to obtain desensitized data; when the data sensitivity is top secret, perform segmented Fourier transform encryption (decompose the physiological signal into frequency domain components and selectively encrypt key frequency bands) to perform the desensitization operation to obtain desensitized data; Use a multi-modal encoder to perform feature encoding on the desensitized data to obtain a feature encoding vector, and through a graph attention network, use a pre-created and desensitized cross-departmental knowledge graph to perform a feature alignment operation on the feature encoding vectors of each edge computing node; The multi-modal encoder is constructed based on a modal encoding layer, a cross-modal fusion layer, and a feature compression layer; the modal encoding layer is used to perform feature encoding on structured data, unstructured data, and time series data in the desensitized data to obtain a structured feature vector, an unstructured feature vector, and a time series feature vector; the cross-modal fusion layer is used to perform cross-modal feature fusion on the structured feature vector, unstructured feature vector, and time series feature vector through a bimodal interaction channel and a joint learning strategy to obtain a joint feature vector; the feature compression layer is used to perform dimensionality reduction and regularization processing on the joint feature vector to generate a feature encoding vector.

[0046] The modal encoding layer is constructed based on a structured encoding module, an unstructured module, and a temporal encoding module. The structured encoding module uses a residual fully-connected network to extract high-order interaction features of structured data through a multi-layer perceptron, that is, to extract structured feature vectors. The unstructured data uses a bidirectional Transformer structure to capture the context semantic features of the unstructured data through a self-attention mechanism, that is, unstructured feature vectors. The temporal encoding module uses a hybrid architecture of a temporal convolutional network and LSTM to expand the receptive field of temporal features through dilated convolution and extract temporal feature vectors. The cross-modal fusion layer is constructed based on a first interaction channel and a second interaction channel. The first interaction channel uses a multi-head cross-attention mechanism to align features with temporal features as Query and structured features as Key-Value. The second interaction channel uses a feature concatenation-gating fusion architecture to dynamically weight-fuse unstructured features and structured features. The outputs of the first interaction channel and the second interaction channel are combined through an adaptive weighting strategy to obtain a joint feature vector. The feature compression layer is constructed based on an orthogonal projection module and a variance normalization module. The orthogonal projection module reduces the dimensionality of the joint feature vector through a fully-connected layer with orthogonal constraints, and the variance normalization module uses a moving average algorithm to perform variance stabilization processing on the joint feature vector, thereby generating a feature encoding vector.

[0047] By refining data sensitivity into four levels (public / internal / confidential / top secret) and dynamically matching cryptographic algorithms (such as HMAC-SHA256, Paillier) and privacy protection technologies (k-anonymity, differential privacy) for different levels, through the positive correlation design of algorithm strength and sensitivity, while ensuring the security of high-level data, it avoids the waste of computing resources caused by excessive encryption of low-level data.

[0048] By adopting a three-stage processing architecture (modal encoding layer → cross-modal fusion layer → feature compression layer): the dual-modal interaction channel realizes the deep association of structured / unstructured / temporal features (such as causal analysis of text semantics and time series); the joint learning strategy dynamically adjusts cross-modal weights through the attention mechanism to solve the feature redundancy problem in traditional multi-modal fusion; through regularization dimensionality reduction, low-dimensional dense vectors are generated, which not only retain cross-modal association features but also meet the efficient computing requirements of subsequent graph neural networks.

[0049] By limiting operations with high computational complexity (such as Paillier encryption and Fourier transform) to top-secret data processing, and using a lightweight algorithm (HMAC-SHA256) for low-level data, the average power consumption of edge nodes is significantly reduced. By decoupling the sensitivity grading standard from the desensitizing algorithm library, it supports dynamic expansion of new algorithms through configuration files (such as replacing Paillier with an LWE-based scheme), meeting NIST post-quantum cryptography migration requirements.

[0050] By implementing a dual "encryption + desensitization" protection approach for top-secret data (such as performing shape-preserving deformation on data after Paillier encryption), we can resist the coordinated attacks of model inversion attacks and password cracking, break through the traditional single-layer protection thinking, and increase the attack cost through algorithm nesting; the shape-preserving deformation technology for time series data (internal / confidential level) preserves statistical features while destroying the correlation between time series and external data, preventing the inference of user identity through time series correlation.

[0051] In step S3, the department-private feature extractor is used to extract data features from the local feature encoding vector; the cross-department parameter collaborative aggregation layer is used for different edge computing nodes to share compression parameters; the global task head is used to read the compression parameters shared by the remaining edge computing nodes from the cross-department parameter collaborative aggregation layer to jointly fine-tune the local neural network model; By using a private feature extractor to process feature coding vectors only locally, there is no need to share original data, which meets the requirement for local processing of sensitive data and avoids the risk of privacy leakage caused by cross-departmental transmission. By abstracting feature coding vectors, it is easier to implement privacy protection measures such as noise addition than with original data, providing a compliance basis for subsequent parameter sharing.

[0052] In step S4, the heterogeneous computing resource adapter is used to abstract different types of computing resources of the edge computing node into a unified resource model, and allocate computing resources, schedule tasks, and optimize performance of the computing tasks of the edge computing node through the resource model.

[0053] Resource abstraction abstracts different types of computing resources into a unified resource model, allowing upper-level applications to use these resources transparently. Allocating computing resources is resource mapping, which maps application requirements to specific computing resources and selects the most appropriate resources for allocation based on the characteristics of the task. Task scheduling dynamically schedules computing tasks to different computing resources based on the usage of computing resources and the priority of tasks to achieve load balancing and efficient utilization. Performance optimization improves the overall performance of the system by optimizing the execution path of computing tasks and the use of computing resources.

[0054] By abstracting different types of computing resources into a unified resource model, global awareness and intelligent scheduling of diverse hardware resources such as CPUs and GPUs in edge computing nodes are achieved, significantly improving resource utilization and reducing task response latency. Combining dynamic energy efficiency optimization algorithms reduces power consumption while ensuring service quality, forming a complete technical barrier covering resource modeling, task scheduling, and energy efficiency management, with outstanding compatibility and scalability.

[0055] The specific steps of step S5 are as follows: Each edge computing node divides each of the model parameters into a core layer (such as classifier weights) and an auxiliary layer (such as intermediate feature transformation matrices) based on the precision sensitivity of the neural network model inference with respect to the model parameters; the model parameters of the core layer are quantized and encoded using 8-bit mixed precision quantization encoding (retaining the sign bit + dynamic exponent bit) to obtain a first quantization result, and the model parameters of the auxiliary layer are quantized and encoded using 4-bit sparse coding quantization encoding (constructing a sparse index table based on a Huffman coding tree) to obtain a second quantization result. During the quantization encoding process, the quantization error distribution of the model parameters of each layer is recorded, and a compensation vector is generated by an error feedback network based on the quantization error distribution. The first quantization result, the second quantization result, and the compensation vector are jointly encapsulated as an encoding result; The first quantization result in the encoding result is homomorphically encrypted through the CKKS scheme, and the second quantization result and the compensation vector in the encoding result are homomorphically encrypted through the BFV scheme to obtain encrypted parameters; Edge computing nodes continuously monitor network metrics including at least link bandwidth, delay jitter, and packet loss rate, input the network metrics into a pre-trained LSTM network, predict the bandwidth quality for the next 3 transmission windows, and dynamically select a compression mode based on the bandwidth quality to compress each encrypted parameter to obtain compressed parameters; the bandwidth quality is high bandwidth, medium bandwidth, or low bandwidth. Lossless compression (Zstandard dictionary encoding + entropy splitting and recombination) is enabled during high bandwidth periods, lossy-lossless hybrid compression (retaining lossless for the core layer and using FPGA-accelerated KL transform compression for the auxiliary layer) is used during medium bandwidth periods, and an emergency mode (quadratic sampling of core layer parameters + Taylor polynomial parameter fitting) is activated during low bandwidth periods; The cross-departmental parameter collaborative aggregation layer adjusts the parameter transmission granularity through a dynamic gating mechanism and shares the compressed parameters to the cross-departmental parameter collaborative aggregation layers of the remaining edge computing nodes; the dynamic gating mechanism (Dynamic Gating Mechanism) is a technology that dynamically controls information flow or resource allocation in a system. It can intelligently decide whether to allow certain operations or data transmissions based on current inputs, states, or environmental conditions.

[0056] By setting core parameters to adopt the CKKS scheme (supporting floating-point homomorphic operations, suitable for high-precision inference), and using the BFV scheme for auxiliary parameters and compensation vectors (higher integer operation efficiency), both security and computational efficiency are considered, avoiding performance bottlenecks caused by a single encryption scheme; through the combination of encryption and quantization coding, reverse cracking of parameters during transmission and collaboration is prevented.

[0057] By real-time monitoring of network metrics (such as link bandwidth, delay jitter, and packet loss rate) to predict the future transmission window state, the problem of dynamic fluctuations in the edge network environment is solved; according to the prediction results, the compression algorithm is adaptively adjusted (such as selecting entropy coding or differential coding) to maximize the compression ratio and transmission efficiency, and reduce the communication delay between edge nodes.

[0058] Through hierarchical quantization coding and error compensation mechanism, while ensuring the inference accuracy of the neural network, the computational and transmission overheads are significantly reduced; the CKKS and BFV hybrid encryption strategy is adopted to achieve hierarchical security protection of core parameters and auxiliary parameters, taking into account privacy and operation efficiency; combined with LSTM network prediction and dynamic compression mode selection, it adaptively copes with network fluctuations and optimizes transmission efficiency; through dynamic gating granularity adjustment and lightweight encapsulation, efficient collaboration and resource load balancing across edge nodes are achieved.

[0059] The specific content of step S6 is as follows: Each edge computing node reads the compressed parameters shared by the other edge computing nodes from the cross-department parameter collaboration and aggregation layer through the global task header, decompresses each of the shared compressed parameters to obtain encrypted parameters, adds each encrypted parameter to a preset parameter queue based on the department priority, and sequentially selects encrypted parameters from the parameter queue to perform weighted aggregation training and local fine-tuning on the local neural network model for collaborative modeling.

[0060] The specific content of step S7 is as follows: Each edge computing node records in real time the collaborative modeling log including at least collaborative modeling time, encrypted parameters, model parameters, knowledge graph alignment record, model convergence curve, quantization error compensation record, dynamic gating trigger record, packet loss retransmission event, and error code. Use the SM3 algorithm to calculate the hash value of the collaborative modeling log to generate a 256-bit hash value H2. Divide the collaborative modeling log into a header, a body, and a tail according to preset rules. Encrypt the header using the SM4 algorithm with the key K1 to obtain header encrypted data, encrypt the body using the 3DES algorithm with the key K2 to obtain body encrypted data, encrypt the tail using the Blowfish algorithm with the key K3 to obtain tail encrypted data, and splice the header encrypted data, body encrypted data, and tail encrypted data in sequence with the hash value H2 to obtain spliced data; In the splicing data, a 5-byte random string generated by / dev / urandom is inserted every 10 bytes to form obfuscated data; Perform Base64 encoding on the obfuscated data, and then perform cyclic displacement on the letters in the encoding result through the Caesar cipher (displacement amount = 7) to generate ciphertext data; Encrypt the key K1, key K2, and key K3 into a key package through the SM2 algorithm; the key K1 is preset by the administrator; the key K2 is dynamically generated through HMAC-SHA256; the key K3 is derived based on the current system timestamp; Obtain the current system time and the MAC address of the local machine, splice the system time and the MAC address into dynamic parameters, encrypt the ciphertext data, the key package, and the dynamic parameters into an encrypted log through the AES-GCM algorithm, and perform distributed storage on the encrypted log; Calculate the data fingerprint of the encrypted log through the HMAC algorithm, and store the data fingerprint on the blockchain.

[0061] Construct a hierarchical encryption structure by adopting hybrid encryption algorithms such as SM3 / SM4 (national cryptography algorithm), 3DES, and Blowfish. Different algorithms respectively protect the header, body, and tail of the log to form algorithm-level security isolation; the key management system uses SM2 asymmetric encryption to protect the symmetric keys (K1 / K2 / K3) to achieve security reinforcement in the key transmission link; ensure the confidentiality and integrity of the final encrypted log through the AES-GCM algorithm, and enhance the timeliness protection by combining dynamic parameters.

[0062] Significantly increase the entropy value of the data stream by inserting / dev / urandom random strings (inserting 5 bytes of noise every 10 bytes); effectively resist pattern recognition attacks through the dual obfuscation techniques of Base64 encoding and Caesar cipher; disrupt the data structure continuity through the segmented encryption strategy and increase the difficulty of data parsing.

[0063] Generate dynamic parameters by binding the system timestamp and the MAC address to establish double verification of the device fingerprint and the time dimension; ensure the authenticity of the data source through the authentication tag mechanism of AES-GCM to prevent replay attacks; form an immutable evidence chain through the HMAC data fingerprint and blockchain storage.

[0064] A preferred embodiment of a cross-departmental data collaborative modeling system based on federated learning according to the present invention includes the following modules: An initialization module is used to preset a neural network model for each edge computing node deployed in each department, and set the data sensitivity and department priority of the department data stored in each edge computing node; the department data includes structured data, unstructured data, and time-series data; A feature alignment module, which is used for each edge computing node to perform differential desensitization operations on local department data based on data sensitivity to obtain desensitized data, perform feature encoding on the desensitized data to obtain feature encoding vectors, and perform feature alignment operations on the feature encoding vectors of each edge computing node through a cross-department knowledge graph; the cross-department knowledge graph is dynamically updated based on a preset update trigger mechanism; A hierarchical federated learning framework construction module, which is used for each edge computing node to construct a hierarchical federated learning framework respectively based on a department private feature extractor, a cross-department parameter collaborative aggregation layer, and a global task head; a model training module, which is used for each edge computing node to extract data features from the feature encoding vectors through the department private feature extractor, dynamically allocate computing resources through a heterogeneous computing resource adapter, so as to train a neural network model using each of the data features, and extract the model parameters of the trained neural network model; The model parameters are the "knowledge" learned by the neural network model during the training process. These model parameters determine how the neural network model processes input data and generates output results. Simply put, these model parameters are like the "memory" of the neural network model, recording the rules learned by the neural network model from the training data. By extracting the model parameters, they can be used for model saving, model analysis, and transfer learning.

[0065] A parameter collaboration module, which is used for each edge computing node to perform hierarchical quantization encoding on the model parameters to obtain an encoding result, encrypt the encoding result into encrypted parameters through a homomorphic encryption algorithm, compress each of the encrypted parameters through a bandwidth-aware compression strategy to obtain compressed parameters, and the cross-department parameter collaborative aggregation layer shares the compressed parameters to the cross-department parameter collaborative aggregation layers of the remaining edge computing nodes through a dynamic gating mechanism; A collaborative modeling module, which is used for each edge computing node to read the compressed parameters shared by the remaining edge computing nodes from the cross-department parameter collaborative aggregation layer through the global task head, and perform collaborative modeling on the local neural network model through each of the shared compressed parameters and department priorities; A collaborative modeling log management module, which is used for each edge computing node to record the collaborative modeling log in real time, and store the collaborative modeling log through a blockchain; Through the deep integration of federated learning and edge computing, on the basis of ensuring data privacy (differential desensitization, homomorphic encryption) and security and trustworthiness (blockchain evidence storage), efficient collaborative modeling of cross-department heterogeneous data (structured / unstructured / temporal data) is achieved; its dynamic resource adaptation mechanism (heterogeneous computing resource allocation, bandwidth-aware compression) significantly reduces communication and computing costs, and the hierarchical federated architecture and cross-department knowledge graph effectively alleviate the data distribution shift problem and improve the model generalization ability through feature alignment and priority scheduling; at the same time, blockchain evidence storage and dynamic gating mechanism enhance the transparency of multi-party collaboration and the ability to resist "free-riding", and the support for the autonomy and ecological compatibility of edge nodes enables it to have strong implementability in complex scenarios in multiple fields, providing a systematic solution for breaking data islands and achieving secure and efficient cross-domain intelligent collaboration.

[0066] The initialization module is specifically used for: Edge computing nodes deployed in each department respectively pre-set a neural network model through a containerized operating environment built by Docker and Kubernetes; The dynamic orchestration of computing resources is achieved through the hybrid container architecture of Docker+Kubernetes; the containerized pre-setting of the neural network model supports gray release and A / B testing, ensuring business continuity during model updates and effectively improving the deployment efficiency.

[0067] Calculate the data sensitivity of the department data stored in each edge computing node based on the data content, data timeliness, and data source: D s = (C * w c + T * w t + S * w s ) / (w c + w t + w s ); Among them, D s represents the data sensitivity; C represents the data content sensitivity, and its value range is [0,1]. When the data attributes of the department data are public data, restricted data, and core sensitive data respectively, the value ranges of the data content sensitivity are [0,0.4], (0.4,0.7], (0.7,1]; T represents the data timeliness sensitivity, and its value range is [0,1], T = e^-λ*Δt1, e represents the natural constant, λ represents the decay coefficient, λ = 0.5, Δt1 represents the data age; S represents the data source sensitivity. When the data sources of the department data are the internal core database, third-party partners, and public network crawling respectively, the values of the data source sensitivity are 1, 0.6, and 0.2; w c 、w t 、w s all represent weight coefficients; Calculate the department priority of the department data stored in each edge computing node based on data importance, data urgency, and data dependence: P = I * w i + U * w u + D * w d ; Where, P represents the department priority; I represents data importance, and its value range is [0, 1]. When the data usage of the department data is auxiliary data, key business data, and strategic-level data respectively, the value ranges of data importance are [0, 0.5], (0.5, 0.8], (0.8, 1]; U represents data urgency, U = max(1 - Δt2 / T max , 0), Δt2 represents the difference between the current time and the deadline, T max represents the maximum allowed processing cycle; D represents data dependence, and its value range is [0, 1]. For example, for core patent documents that are relied on by multiple subsequent processes (such as examination responses, international stage submissions), the value of data dependence is 1; the value of data dependence for intermediate data is 0.5; the value of data dependence for independent data is 0.2; w i , w u , w d all represent weight coefficients; The department data includes structured data, unstructured data, and time-series data; the structured data includes at least personal information, financial data, approval data, statistical data, medical data, and educational data; the unstructured data includes at least document files, image data, audio data, video data, social media data, log data, and e-mails; the time-series data includes at least economic data, meteorological data, traffic data, environmental monitoring data, medical monitoring data, and industrial production data.

[0068] By introducing an exponential function with a data age decay factor λ = 0.5, dynamic optimization of the storage strategy is achieved; through a third-order weight adjustment mechanism (w c , w t , w s ), support for dynamic adjustment according to compliance requirements is provided; based on the calculation of data urgency, a Tmax threshold control is introduced. When Δt2 exceeds the maximum processing cycle, an alarm upgrade is automatically triggered to ensure the timeliness requirements of key business data.

[0069] Through a data sensitivity quantification evaluation system of a dynamic decay model (Ds = (C * w c + T * w t + S * w s ) / Σw), real-time optimization of data security classification and storage strategy is achieved; through a department priority multi-factor coupling algorithm (P = I * w i+U*w u +D*w d ), combined with data urgency threshold control and cross-node federated learning mechanism, to achieve intelligent scheduling of computing resources and load balancing.

[0070] The feature alignment module is specifically used for: Each edge computing node maps the data sensitivity to public level, internal level, confidential level or top secret level based on a preset sensitivity grading standard; For structured data in local department data: When the data sensitivity is public level, calculate the hash value H1 of the user identifier using HMAC-SHA256, and replace the user identifier based on the hash value H1 to perform the desensitization operation to obtain desensitized data; When the data sensitivity is internal level, perform the desensitization operation through the k-anonymity algorithm (ensuring that each record has at least k-1 identical records on the quasi-identifier) to obtain desensitized data; When the data sensitivity is confidential level, perform differential privacy noise addition to perform the desensitization operation to obtain desensitized data; When the data sensitivity is top secret level, encrypt through the Paillier algorithm to perform the desensitization operation to obtain desensitized data; For unstructured data in local department data: For the text in document files, social media data, log data, and emails, when the data sensitivity is public level or internal level, perform entity replacement (replace the patient name with the [PATIENT] tag and retain gender pronouns) to perform the desensitization operation to obtain desensitized data; When the data sensitivity is confidential level, perform semantic-preserving transformation (rewrite the clinical description, maintaining medical logic but changing surface features) to perform the desensitization operation to obtain desensitized data; When the data sensitivity is top secret level, use the AES-256-CTR algorithm to encrypt paragraphs; For image data, video data, social media data, and images in emails, perform reversible pixel scrambling on the privacy area in the image to perform the desensitization operation to obtain desensitized data; For audio data, perform segment replacement, add noise, or voiceprint transformation to perform the desensitization operation to obtain desensitized data; For time series data in local department data: When the data sensitivity is public level, perform timestamp obfuscation (convert the time accurate to seconds to an hour interval) to perform the desensitization operation to obtain desensitized data; When the data sensitivity is internal level or confidential level, perform shape-preserving transformation (perform dynamic time warping on the ECG signal, maintaining waveform features but changing the time correspondence) to perform the desensitization operation to obtain desensitized data; When the data sensitivity is top secret level, perform segmented Fourier transform encryption (decompose the physiological signal into frequency domain components and selectively encrypt key frequency bands) to perform the desensitization operation to obtain desensitized data; The desensitized data is subjected to feature encoding by a multi-modal encoder to obtain a feature encoding vector. Through a graph attention network, a feature alignment operation is performed on the feature encoding vectors of each edge computing node by using a pre-created and desensitized cross-department knowledge graph. The multi-modal encoder is constructed based on a modal encoding layer, a cross-modal fusion layer, and a feature compression layer. The modal encoding layer is used to perform feature encoding on structured data, unstructured data, and time-series data in the desensitized data to obtain a structured feature vector, an unstructured feature vector, and a time-series feature vector. The cross-modal fusion layer is used to perform cross-modal feature fusion on the structured feature vector, the unstructured feature vector, and the time-series feature vector through a bimodal interaction channel and a joint learning strategy to obtain a joint feature vector. The feature compression layer is used to perform dimensionality reduction and regularization processing on the joint feature vector to generate a feature encoding vector.

[0071] The modal encoding layer is constructed based on a structured encoding module, an unstructured module, and a time-series encoding module. The structured encoding module uses a residual fully connected network to extract high-order interaction features of structured data through a multi-layer perceptron, that is, to extract a structured feature vector. The unstructured data uses a bidirectional Transformer structure to capture the context semantic features of the unstructured data through a self-attention mechanism, that is, an unstructured feature vector. The time-series encoding module uses a hybrid architecture of a temporal convolutional network and an LSTM, expands the receptive field of the time-series features through dilated convolution, and extracts a time-series feature vector. The cross-modal fusion layer is constructed based on a first interaction channel and a second interaction channel. The first interaction channel uses a multi-head cross-attention mechanism to perform feature alignment with the time-series feature as the Query and the structured feature as the Key-Value. The second interaction channel uses a feature concatenation-gating fusion architecture to perform dynamic weighted fusion of the unstructured feature and the structured feature. The outputs of the first interaction channel and the second interaction channel are combined through an adaptive weighting strategy to obtain a joint feature vector. The feature compression layer is constructed based on an orthogonal projection module and a variance normalization module. The orthogonal projection module performs feature dimensionality reduction on the joint feature vector through a fully connected layer with orthogonal constraints. The variance normalization module uses a moving average algorithm to perform variance stabilization processing on the joint feature vector, and then generates a feature encoding vector.

[0072] By refining the data sensitivity into four levels (public / internal / confidential / top secret), and dynamically matching cryptographic algorithms (such as HMAC-SHA256, Paillier) and privacy protection technologies (k-anonymity, differential privacy) for different levels, through the positive correlation design of the algorithm strength and sensitivity, while ensuring the security of high-level data, it avoids the waste of computing resources caused by excessive encryption of low-level data.

[0073] By adopting a three-stage processing architecture (modal encoding layer → cross-modal fusion layer → feature compression layer): the dual-modal interaction channel realizes the deep association of structured / unstructured / temporal features (such as the causal analysis of text semantics and time series); the joint learning strategy dynamically adjusts the cross-modal weights through the attention mechanism to solve the feature redundancy problem in traditional multi-modal fusion; through regularization dimensionality reduction, low-dimensional dense vectors are generated, which not only retain the cross-modal associated features but also meet the high-efficiency computing requirements of subsequent graph neural networks.

[0074] By limiting high-computation-complexity operations (such as Paillier encryption and Fourier transform) to the top-secret data processing link and using lightweight algorithms (HMAC-SHA256) for low-classification data, the average power consumption of edge nodes is significantly reduced; through the decoupled design of the sensitivity grading standard and the desensitization algorithm library, it supports dynamically expanding new algorithms through configuration files (such as replacing Paillier with an LWE-based scheme), meeting the requirements of NIST's post-quantum cryptography migration.

[0075] By implementing "encryption + desensitization" dual protection for top-secret data (such as performing shape-preserving transformation on Paillier-encrypted data), it resists the coordinated attacks of model inversion attacks and password cracking, breaks through the traditional single-layer protection thinking, and increases the attack cost through algorithm nesting; the shape-preserving transformation technology for time-series data (internal / confidential level) destroys the correlation between the time series and external data while retaining statistical features, preventing the inference of user identity through time-series correlation.

[0076] In the hierarchical federated learning framework construction module, the department private feature extractor is used to extract data features from the local feature encoding vector; the cross-department parameter collaborative aggregation layer is used for sharing compressed parameters among different edge computing nodes; the global task head is used to read the compressed parameters shared by the remaining edge computing nodes from the cross-department parameter collaborative aggregation layer to jointly fine-tune the local neural network model. By only processing the feature encoding vector locally through the private feature extractor without sharing the original data, it meets the requirement of localizing the processing of sensitive data and avoids the privacy leakage risk caused by cross-department transmission; through the abstract processing of the feature encoding vector, it is easier to implement privacy protection means such as adding noise compared to the original data, providing a compliance basis for subsequent parameter sharing.

[0077] In the model training module, the heterogeneous computing resource adapter is used to abstract different types of computing resources of the edge computing node into a unified resource model, and through the resource model, computing resources are allocated, tasks are scheduled, and performance is optimized for the computing tasks of the edge computing node.

[0078] Resource abstraction abstracts different types of computing resources into a unified resource model, enabling upper-layer applications to transparently use these resources; allocating computing resources, i.e., resource mapping, maps the application's requirements to specific computing resources and selects the most suitable resources for allocation according to the characteristics of the tasks; task scheduling dynamically schedules computing tasks to different computing resources based on the usage of computing resources and the priorities of the tasks to achieve load balancing and efficient utilization; performance optimization improves the overall performance of the system by optimizing the execution path of computing tasks and the usage mode of computing resources.

[0079] By abstracting different types of computing resources into a unified resource model, global perception and intelligent scheduling of diverse hardware resources such as CPUs and GPUs in edge computing nodes are achieved, significantly improving resource utilization and reducing task response latency. Combining with a dynamic energy efficiency optimization algorithm reduces power consumption while ensuring service quality, forming a complete technical barrier covering resource modeling, task scheduling, and energy efficiency management, with outstanding compatibility and scalability.

[0080] The parameter coordination module is specifically used for: Each edge computing node divides each of the model parameters into a core layer (such as classifier weights) and an auxiliary layer (such as an intermediate feature transformation matrix) based on the accuracy sensitivity of the neural network model inference to the model parameters; the model parameters of the core layer are quantized and encoded using 8-bit mixed precision quantization (retaining the sign bit + dynamic exponent bit) to obtain a first quantization result, and the model parameters of the auxiliary layer are quantized and encoded using 4-bit sparse coding quantization (constructing a sparse index table based on a Huffman coding tree) to obtain a second quantization result. During the quantization encoding process, the quantization error distribution of the model parameters of each layer is recorded, and a compensation vector is generated based on the quantization error distribution through an error feedback network. The first quantization result, the second quantization result, and the compensation vector are jointly encapsulated into an encoding result; The first quantization result in the encoding result is homomorphically encrypted through the CKKS scheme, and the second quantization result and the compensation vector in the encoding result are homomorphically encrypted through the BFV scheme to obtain encrypted parameters; Edge computing nodes continuously monitor network metrics including at least link bandwidth, delay jitter, and packet loss rate, input the network metrics into a pre-trained LSTM network to predict the bandwidth quality in the next 3 transmission windows, and dynamically select a compression mode based on the bandwidth quality to compress each encrypted parameter to obtain compressed parameters; the bandwidth quality is high bandwidth, medium bandwidth, or low bandwidth. Lossless compression (Zstandard dictionary coding + entropy segmentation and recombination) is enabled during high bandwidth periods, lossy-lossless hybrid compression (retaining lossless for the core layer and using FPGA-accelerated KL transform compression for the auxiliary layer) is used during medium bandwidth periods, and an emergency mode (sub-sampling of core layer parameters + Taylor polynomial parameter fitting) is activated during low bandwidth periods; The cross - department parameter collaborative aggregation layer adjusts the parameter transmission granularity through a dynamic gating mechanism and shares the compressed parameters to the cross - department parameter collaborative aggregation layers of other edge computing nodes; the dynamic gating mechanism (Dynamic Gating Mechanism) is a technology that dynamically controls information flow or resource allocation in a system. It can intelligently decide whether to allow certain operations or data transmissions according to the current input, state, or environmental conditions.

[0081] By setting the core parameters to adopt the CKKS scheme (supporting floating - point homomorphic operations and suitable for high - precision inference), and the auxiliary parameters and compensation vectors to adopt the BFV scheme (with higher integer operation efficiency), it balances security and computing efficiency and avoids performance bottlenecks caused by a single encryption scheme; through the combination of encryption and quantization coding, it prevents the parameters from being reverse - cracked during transmission and collaboration.

[0082] By real - time monitoring network metrics (such as link bandwidth, delay jitter, and packet loss rate, etc.) to predict the future transmission window state, it solves the problem of dynamic fluctuations in the edge network environment; according to the prediction results, it adaptively adjusts the compression algorithm (such as selecting entropy coding or differential coding) to maximize the compression rate and transmission efficiency and reduce the communication delay between edge nodes.

[0083] Through hierarchical quantization coding and error compensation mechanism, while ensuring the inference accuracy of the neural network, it significantly reduces the computing and transmission overhead; it adopts a hybrid encryption strategy of CKKS and BFV to achieve hierarchical security protection of core parameters and auxiliary parameters, balancing privacy and operation efficiency; it combines LSTM network prediction and dynamic compression mode selection to adaptively cope with network fluctuations and optimize transmission efficiency; through dynamic gating granularity adjustment and lightweight encapsulation, it realizes efficient collaboration and resource load balancing across edge nodes.

[0084] The collaborative modeling module is specifically used for: Each edge computing node reads the compressed parameters shared by other edge computing nodes from the cross - department parameter collaborative aggregation layer through the global task header, decompresses each of the shared compressed parameters to obtain encrypted parameters, adds each encrypted parameter to a preset parameter queue based on the department priority, and sequentially selects encrypted parameters from the parameter queue to perform weighted aggregation training and local fine - tuning on the local neural network model for collaborative modeling.

[0085] The collaborative modeling log management module is specifically used for: Each edge computing node records in real time a collaborative modeling log that at least includes collaborative modeling time, encryption parameters, model parameters, knowledge graph alignment records, model convergence curves, quantization error compensation records, dynamic gating trigger records, packet loss retransmission events, and error codes. Use the SM3 algorithm to calculate the hash of the collaborative modeling log to generate a 256-bit hash value H2. Divide the collaborative modeling log into a header, a body, and a tail according to preset rules. Encrypt the header using the SM4 algorithm with the key K1 to obtain header encrypted data, encrypt the body using the 3DES algorithm with the key K2 to obtain body encrypted data, and encrypt the tail using the Blowfish algorithm with the key K3 to obtain tail encrypted data. Concatenate the header encrypted data, body encrypted data, and tail encrypted data in sequence with the hash value H2 to obtain concatenated data; In the concatenated data, insert a 5-byte random string generated by / dev / urandom every 10 bytes to form obfuscated data; Perform Base64 encoding on the obfuscated data, and then cyclically shift the letters in the encoding result by the Caesar cipher (shift amount = 7) to generate ciphertext data; Encrypt the key K1, key K2, and key K3 into a key package using the SM2 algorithm; the key K1 is preset by the administrator; the key K2 is dynamically generated by HMAC-SHA256; the key K3 is derived based on the current system timestamp; Obtain the current system time and the MAC address of the local machine, concatenate the system time and the MAC address into dynamic parameters, and encrypt the ciphertext data, key package, and dynamic parameters into an encrypted log using the AES-GCM algorithm, and store the encrypted log distributively; Calculate the data fingerprint of the encrypted log using the HMAC algorithm, and store the data fingerprint on the blockchain.

[0086] Construct a hierarchical encryption structure by using hybrid encryption algorithms such as SM3 / SM4 (national cryptographic algorithm), 3DES, and Blowfish. Different algorithms respectively protect the header, body, and tail of the log to form algorithm-level security isolation; the key management system uses SM2 asymmetric encryption to protect the symmetric keys (K1 / K2 / K3) to achieve security reinforcement in the key transmission link; ensure the confidentiality and integrity of the final encrypted log through the AES-GCM algorithm, and enhance the timeliness protection by combining dynamic parameters.

[0087] By inserting random strings from / dev / urandom (inserting 5 bytes of noise every 10 bytes), the entropy value of the data stream is significantly increased; through the dual obfuscation techniques of Base64 encoding and Caesar cipher, pattern recognition attacks can be effectively resisted; by segmenting the encryption strategy to disrupt the continuity of the data structure, the difficulty of data parsing is increased.

[0088] Through the generation of dynamic parameters by binding the system timestamp and the MAC address, a dual verification of the device fingerprint and the time dimension is established; through the authentication tag mechanism of AES-GCM, the authenticity of the data source is ensured to prevent replay attacks; through the HMAC data fingerprint and blockchain evidence storage, an immutable evidence chain is formed.

[0089] Although the specific implementation manners of the present invention have been described above, those skilled in the art of this technology should understand that the specific embodiments we described are illustrative rather than used to limit the scope of the present invention. Equivalent modifications and changes made by those skilled in the art in accordance with the spirit of the present invention should be covered by the scope protected by the claims of the present invention.

Claims

1. A cross-departmental data collaborative modeling method based on federated learning, characterized in that: It includes the following steps: Step S1: Edge computing nodes deployed in each department respectively preset a neural network model, and set the data sensitivity and department priority of the department data stored in each edge computing node; the department data includes structured data, unstructured data, and time-series data; Step S2: Each edge computing node performs differential desensitization operations on the local department data based on the data sensitivity to obtain desensitized data, performs feature encoding on the desensitized data to obtain a feature encoding vector, and performs feature alignment operations on the feature encoding vectors of each edge computing node through a cross-department knowledge graph; Step S3: Each edge computing node respectively constructs a hierarchical federated learning framework based on a department-private feature extractor, a cross-department parameter collaborative aggregation layer, and a global task head; Step S4: Each edge computing node extracts data features from the feature encoding vector through a department-private feature extractor, dynamically allocates computing resources through a heterogeneous computing resource adapter, and uses each of the data features to train the neural network model, and extracts the model parameters of the trained neural network model; Step S5: Each edge computing node performs hierarchical quantization encoding on the model parameters to obtain an encoding result, encrypts the encoding result into encrypted parameters through a homomorphic encryption algorithm, compresses each of the encrypted parameters through a bandwidth-aware compression strategy to obtain compressed parameters, and the cross-department parameter collaborative aggregation layer shares the compressed parameters to the cross-department parameter collaborative aggregation layers of the remaining edge computing nodes through a dynamic gating mechanism; Step S6: Each edge computing node reads the compressed parameters shared by the remaining edge computing nodes from the cross-department parameter collaborative aggregation layer through the global task head, and performs collaborative modeling on the local neural network model through each of the shared compressed parameters and the department priority; Step S7: Each edge computing node records the collaborative modeling log in real time, and stores the collaborative modeling log through a blockchain.

2. The cross-departmental data collaborative modeling method based on federated learning according to claim 1, wherein: The specific content of step S1 is: Edge computing nodes deployed in each department respectively preset a neural network model through a containerized operating environment built by Docker and Kubernetes; Calculate the data sensitivity of the department data stored in each edge computing node based on the data content, data timeliness, and data source: D s = (C * w c + T * w t + S * w s ) / (w c + w t + w s ); Among them, D s represents the data sensitivity; C represents the data content sensitivity, and its value range is [0, 1]. When the data attributes of department data are public data, restricted data, and core sensitive data respectively, the value ranges of the data content sensitivity are [0, 0.4], (0.4, 0.7], and (0.7, 1]; T represents the data timeliness sensitivity, and its value range is [0, 1]. T = e^-λ*Δt1, where e represents the natural constant, λ represents the decay coefficient, λ = 0.5, and Δt1 represents the data age; S represents the data source sensitivity. When the data sources of department data are internal core databases, third-party partners, and public web crawls respectively, the data source sensitivities are 1, 0.6, and 0.2; w c 、w t 、w s all represent weight coefficients; Calculate the department priority of the department data stored in each edge computing node based on the data importance, data urgency, and data dependence: P = I * w i + U * w u + D * w d ; Among them, P represents the department priority; I represents the data importance, and its value range is [0, 1]. When the data usage of department data is auxiliary data, key business data, and strategic data respectively, the value ranges of data importance are [0, 0.5], (0.5, 0.8], and (0.8, 1]; U represents the data urgency, U = max(1 - Δt2 / T max , 0), where Δt2 represents the difference between the current time and the deadline, and T max represents the maximum allowed processing period; D represents the data dependency, and its value range is [0, 1]; w i , w u , and w d all represent weight coefficients; The department data includes structured data, unstructured data, and time-series data; the structured data at least includes personal information, financial data, approval data, statistical data, medical data, and educational data; the unstructured data at least includes document files, image data, audio data, video data, social media data, log data, and e-mails; the time-series data at least includes economic data, meteorological data, traffic data, environmental monitoring data, medical monitoring data, and industrial production data; The specific content of step S2 is: Each edge computing node maps the data sensitivity to public level, internal level, confidential level, or top-secret level based on a preset sensitivity grading standard; For structured data in local department data: When the data sensitivity is at the public level, calculate the hash value H1 of the user identifier using HMAC-SHA256, and replace the user identifier based on the hash value H1 to perform the desensitization operation to obtain desensitized data; when the data sensitivity is at the internal level, perform the desensitization operation using the k-anonymity algorithm to obtain desensitized data; when the data sensitivity is at the confidential level, perform differential privacy adding noise to perform the desensitization operation to obtain desensitized data; when the data sensitivity is at the top-secret level, encrypt using the Paillier algorithm to perform the desensitization operation to obtain desensitized data; For unstructured data in local department data: For the text in document files, social media data, log data, and emails, when the data sensitivity is at the public level or internal level, perform entity replacement to perform the desensitization operation to obtain desensitized data; when the data sensitivity is at the confidential level, perform semantic-preserving transformation to perform the desensitization operation to obtain desensitized data; when the data sensitivity is at the top-secret level, use the AES-256-CTR algorithm for paragraph encryption; for image data, video data, social media data, and images in emails, perform reversible pixel scrambling on the privacy areas in the images to perform the desensitization operation to obtain desensitized data; for audio data, perform segment replacement, add noise, or voiceprint transformation to perform the desensitization operation to obtain desensitized data; For time-series data in local department data: When the data sensitivity is at the public level, perform timestamp fuzzification to perform the desensitization operation to obtain desensitized data; when the data sensitivity is at the internal level or confidential level, perform shape-preserving transformation to perform the desensitization operation to obtain desensitized data; when the data sensitivity is at the top-secret level, perform segmented Fourier transform encryption to perform the desensitization operation to obtain desensitized data; Use a multi-modal encoder to perform feature encoding on the desensitized data to obtain a feature encoding vector, and through a graph attention network, use a pre-created and desensitized cross-department knowledge graph to perform a feature alignment operation on the feature encoding vectors of each edge computing node; The multi-modal encoder is constructed based on a modal encoding layer, a cross-modal fusion layer, and a feature compression layer; the modal encoding layer is used to perform feature encoding on the structured data, unstructured data, and time-series data in the desensitized data to obtain a structured feature vector, an unstructured feature vector, and a time-series feature vector; the cross-modal fusion layer is used to perform cross-modal feature fusion on the structured feature vector, unstructured feature vector, and time-series feature vector through a bimodal interaction channel and a joint learning strategy to obtain a joint feature vector; the feature compression layer is used to perform dimensionality reduction and regularization processing on the joint feature vector to generate a feature encoding vector.

3. A cross-departmental data collaborative modeling method based on federated learning according to claim 1, characterized in that: In the step S3, the department private feature extractor is used to extract data features from the local feature encoding vector; the cross-department parameter collaborative aggregation layer is used for sharing compressed parameters among different edge computing nodes; the global task header is used to read the compressed parameters shared by the remaining edge computing nodes from the cross-department parameter collaborative aggregation layer to jointly fine-tune the local neural network model; In the step S4, the heterogeneous computing resource adapter is used to abstract different types of computing resources of the edge computing node into a unified resource model, and allocate computing resources, task scheduling, and performance optimization for the computing tasks of the edge computing node through the resource model.

4. The cross-departmental data collaborative modeling method based on federated learning according to claim 1, characterized in that: The step S5 is specifically as follows: Each edge computing node divides each model parameter into a core layer and an auxiliary layer based on the accuracy sensitivity of the neural network model inference with respect to the model parameters; the model parameters of the core layer are quantized and encoded using 8-bit mixed precision to obtain a first quantization result, and the model parameters of the auxiliary layer are quantized and encoded using 4-bit sparse coding to obtain a second quantization result. During the quantization encoding process, the quantization error distribution of the model parameters of each layer is recorded, and a compensation vector is generated by an error feedback network based on the quantization error distribution. Each of the first quantization result, the second quantization result, and the compensation vector are jointly encapsulated into an encoding result; The first quantization result in the encoding result is homomorphically encrypted through the CKKS scheme, and the second quantization result and the compensation vector in the encoding result are homomorphically encrypted through the BFV scheme to obtain encrypted parameters; The edge computing node real-time monitors network metrics including at least link bandwidth, delay jitter, and packet loss rate, inputs the network metrics into a pre-trained LSTM network, predicts the bandwidth quality of the next 3 transmission windows, and dynamically selects a compression mode based on the bandwidth quality to compress each encrypted parameter to obtain compressed parameters; The cross-department parameter collaborative aggregation layer adjusts the parameter transmission granularity through a dynamic gating mechanism and shares the compressed parameters to the cross-department parameter collaborative aggregation layer of the remaining edge computing nodes; The step S6 is specifically as follows: Each edge computing node reads the compressed parameters shared by the remaining edge computing nodes from the cross-department parameter collaborative aggregation layer through the global task header, decompresses each of the shared compressed parameters to obtain encrypted parameters, adds each encrypted parameter to a preset parameter queue based on the department priority, and sequentially selects encrypted parameters from the parameter queue to perform weighted aggregation training and local fine-tuning on the local neural network model for collaborative modeling.

5. The cross-departmental data collaborative modeling method based on federated learning according to claim 1, characterized in that: The step S7 is specifically as follows: Each edge computing node records in real time a collaborative modeling log that at least includes collaborative modeling time, encryption parameters, model parameters, knowledge graph alignment records, model convergence curves, quantization error compensation records, dynamic gating trigger records, packet loss retransmission events, and error codes. Use the SM3 algorithm to calculate the hash of the collaborative modeling log to generate a 256-bit hash value H2. Divide the collaborative modeling log into a header, a body, and a tail according to a preset rule. Encrypt the header using the SM4 algorithm with the key K1 to obtain header encrypted data, encrypt the body using the 3DES algorithm with the key K2 to obtain body encrypted data, and encrypt the tail using the Blowfish algorithm with the key K3 to obtain tail encrypted data. Concatenate the header encrypted data, body encrypted data, and tail encrypted data in sequence with the hash value H2 to obtain concatenated data; In the concatenated data, insert a 5-byte random string generated by / dev / urandom every 10 bytes to form obfuscated data; Perform Base64 encoding on the obfuscated data, and then perform a cyclic shift on the letters in the encoding result using the Caesar cipher to generate ciphertext data; Encrypt the key K1, key K2, and key K3 into a key package using the SM2 algorithm; Obtain the current system time and the MAC address of the local machine, concatenate the system time and the MAC address into dynamic parameters, and encrypt the ciphertext data, key package, and dynamic parameters into an encrypted log using the AES-GCM algorithm, and perform distributed storage on the encrypted log; Calculate the data fingerprint of the encrypted log using the HMAC algorithm, and store the data fingerprint on the blockchain.

6. A cross-departmental data collaborative modeling system based on federated learning, characterized in that: Includes the following modules: Initialization module, used to pre-set a neural network model for each edge computing node deployed in each department, and set the data sensitivity and department priority of the department data stored in each edge computing node; the department data includes structured data, unstructured data, and time series data; Feature alignment module, used for each edge computing node to perform differential desensitization operations on the local department data based on data sensitivity to obtain desensitized data, perform feature encoding on the desensitized data to obtain feature encoding vectors, and perform feature alignment operations on the feature encoding vectors of each edge computing node through a cross-department knowledge graph; Hierarchical federated learning framework construction module, used for each edge computing node to construct a hierarchical federated learning framework based on a department private feature extractor, a cross-department parameter collaborative aggregation layer, and a global task header respectively; Model training module, used for each edge computing node to extract data features from the feature encoding vectors through a department private feature extractor, dynamically allocate computing resources through a heterogeneous computing resource adapter, and use each of the data features to train a neural network model, and extract the model parameters of the trained neural network model; The parameter collaboration module is used for each edge computing node to perform hierarchical quantization encoding on the model parameters to obtain an encoding result, encrypt the encoding result into encrypted parameters through a homomorphic encryption algorithm, compress each of the encrypted parameters through a bandwidth-aware compression strategy to obtain compressed parameters, and the cross-department parameter collaborative aggregation layer shares the compressed parameters to the cross-department parameter collaborative aggregation layers of the remaining edge computing nodes through a dynamic gating mechanism; The collaborative modeling module is used for each edge computing node to read the compressed parameters shared by the remaining edge computing nodes from the cross-department parameter collaborative aggregation layer through the global task header, and perform collaborative modeling on the local neural network model through the shared compressed parameters and department priorities; The collaborative modeling log management module is used for each edge computing node to record the collaborative modeling log in real time and store the collaborative modeling log through the blockchain.

7. The cross-departmental data collaborative modeling system based on federated learning according to claim 6, characterized in that: The initialization module is specifically used for: Edge computing nodes deployed in each department respectively preset a neural network model through a containerized operating environment built by Docker and Kubernetes; Calculating the data sensitivity of the department data stored in each edge computing node based on the data content, data timeliness, and data source; D s = (C * w c + T * w t + S * w s ) / (w c + w t + w s ); Among them, D s represents data sensitivity; C represents data content sensitivity, and its value range is [0, 1]. When the data attributes of department data are public data, restricted data, and core sensitive data respectively, the value ranges of data content sensitivity are [0, 0.4], (0.4, 0.7], and (0.7, 1]; T represents data timeliness sensitivity, and its value range is [0, 1]. T = e^-λ*Δt1, where e represents the natural constant, λ represents the decay coefficient, λ = 0.5, and Δt1 represents data age; S represents data source sensitivity. When the data sources of department data are internal core database, third-party partners, and public web crawling respectively, the values of data source sensitivity are 1, 0.6, and 0.2; w c 、w t 、w s all represent weight coefficients; Calculating the department priority of the department data stored in each edge computing node based on the data importance, data urgency, and data dependency; P = I * w i + U * w u + D * w d ; Among them, P represents the department priority; I represents the data importance, and its value range is [0, 1]. When the data usage of department data is auxiliary data, key business data, and strategic-level data respectively, the value ranges of data importance are [0, 0.5], (0.5, 0.8], and (0.8, 1]; U represents the data urgency, U = max(1 - Δt2 / T max , 0), where Δt2 represents the difference between the current time and the deadline, and T max represents the maximum allowed processing cycle; D represents the data dependency, and its value range is [0, 1]; w i , w u , and w d all represent weight coefficients; The department data includes structured data, unstructured data, and time-series data; the structured data at least includes personal information, financial data, approval data, statistical data, medical data, and educational data; the unstructured data at least includes document files, image data, audio data, video data, social media data, log data, and e-mails; the time-series data at least includes economic data, meteorological data, traffic data, environmental monitoring data, medical monitoring data, and industrial production data; The feature alignment module is specifically used for: Each edge computing node maps the data sensitivity to public level, internal level, confidential level, or top secret level based on a preset sensitivity grading standard; For the structured data in the local department data: when the data sensitivity is public level, calculate the hash value H1 of the user identifier using HMAC-SHA256, and replace the user identifier based on the hash value H1 to perform a desensitization operation to obtain desensitized data; when the data sensitivity is internal level, perform a desensitization operation through the k-anonymity algorithm to obtain desensitized data; when the data sensitivity is confidential level, perform differential privacy adding noise to perform a desensitization operation to obtain desensitized data; when the data sensitivity is top secret level, encrypt through the Paillier algorithm to perform a desensitization operation to obtain desensitized data; For unstructured data in local department data: For the text in document files, social media data, log data, and emails, when the data sensitivity is public or internal, entity replacement is performed to execute the desensitization operation to obtain desensitized data; when the data sensitivity is confidential, semantic-preserving transformation is performed to execute the desensitization operation to obtain desensitized data; when the data sensitivity is top-secret, the AES-256-CTR algorithm is used for paragraph encryption; for image data, video data, social media data, and images in emails, reversible pixel scrambling is performed on the privacy areas in the images to execute the desensitization operation to obtain desensitized data; for audio data, segment replacement, noise addition, or voiceprint transformation is performed to execute the desensitization operation to obtain desensitized data; For time-series data in local department data: When the data sensitivity is public, timestamp obfuscation is performed to execute the desensitization operation to obtain desensitized data; when the data sensitivity is internal or confidential, shape-preserving transformation is performed to execute the desensitization operation to obtain desensitized data; when the data sensitivity is top-secret, segmented Fourier transform encryption is performed to execute the desensitization operation to obtain desensitized data; The desensitized data is feature-encoded by a multi-modal encoder to obtain a feature-encoding vector. Through a graph attention network, the feature-encoding vectors of each edge computing node are subjected to a feature alignment operation using a pre-created and desensitized cross-department knowledge graph; The multi-modal encoder is constructed based on a modal encoding layer, a cross-modal fusion layer, and a feature compression layer; the modal encoding layer is used to perform feature encoding on structured data, unstructured data, and time-series data in the desensitized data to obtain structured feature vectors, unstructured feature vectors, and time-series feature vectors; the cross-modal fusion layer is used to perform cross-modal feature fusion on the structured feature vectors, unstructured feature vectors, and time-series feature vectors through a bimodal interaction channel and a joint learning strategy to obtain a joint feature vector; the feature compression layer is used to perform dimensionality reduction and regularization processing on the joint feature vector to generate a feature-encoding vector.

8. The cross-departmental data collaborative modeling system based on federated learning according to claim 6, wherein: In the hierarchical federated learning framework construction module, the department private feature extractor is used to extract data features from the local feature-encoding vector; the cross-department parameter collaborative aggregation layer is used for sharing compressed parameters among different edge computing nodes; the global task head is used to read the compressed parameters shared by the other edge computing nodes from the cross-department parameter collaborative aggregation layer to jointly fine-tune the local neural network model; In the model training module, the heterogeneous computing resource adapter is used to abstract different types of computing resources of the edge computing node into a unified resource model, and the resource model is used to allocate computing resources, task scheduling, and performance optimization for the computing tasks of the edge computing node.

9. The cross-departmental data collaborative modeling system based on federated learning according to claim 6, characterized in that: The parameter collaboration module is specifically used for: Each edge computing node divides each of the model parameters into a core layer and an auxiliary layer based on the precision sensitivity of the neural network model inference with respect to the model parameters; the model parameters of the core layer are quantized and encoded using 8-bit mixed precision to obtain a first quantization result, and the model parameters of the auxiliary layer are quantized and encoded using 4-bit sparse coding to obtain a second quantization result. During the quantization encoding process, the quantization error distribution of the model parameters of each layer is recorded, and a compensation vector is generated by an error feedback network based on the quantization error distribution. Each of the first quantization result, the second quantization result, and the compensation vector are jointly encapsulated into an encoding result; The first quantization result in the encoding result is homomorphically encrypted through the CKKS scheme, and the second quantization result and the compensation vector in the encoding result are homomorphically encrypted through the BFV scheme to obtain encrypted parameters; The edge computing node real-time monitors network metrics including at least link bandwidth, delay jitter, and packet loss rate, inputs the network metrics into a pre-trained LSTM network, predicts the bandwidth quality of the next 3 transmission windows, and dynamically selects a compression mode based on the bandwidth quality to compress each of the encrypted parameters to obtain compressed parameters; The cross-department parameter collaborative aggregation layer adjusts the parameter transmission granularity through a dynamic gating mechanism and shares the compressed parameters to the cross-department parameter collaborative aggregation layers of the remaining edge computing nodes; The collaborative modeling module is specifically used for: Each edge computing node reads the compressed parameters shared by the remaining edge computing nodes from the cross-department parameter collaborative aggregation layer through the global task header, decompresses each of the shared compressed parameters to obtain encrypted parameters, adds each of the encrypted parameters to a preset parameter queue based on the department priority, and sequentially selects encrypted parameters from the parameter queue to perform weighted aggregation training and local fine-tuning on the local neural network model for collaborative modeling.

10. The cross-departmental data collaborative modeling system based on federated learning according to claim 6, characterized in that: The collaborative modeling log management module is specifically used for: Each edge computing node real-time records a collaborative modeling log including at least collaborative modeling time, encrypted parameters, model parameters, knowledge graph alignment records, model convergence curves, quantization error compensation records, dynamic gating trigger records, packet loss retransmission events, and error codes, calculates a hash value H2 of 256 bits for the collaborative modeling log using the SM3 algorithm, divides the collaborative modeling log into a header, a body, and a tail according to a preset rule, encrypts the header using the SM4 algorithm by invoking a key K1 to obtain header encrypted data, encrypts the body using the 3DES algorithm by invoking a key K2 to obtain body encrypted data, encrypts the tail using the Blowfish algorithm by invoking a key K3 to obtain tail encrypted data, and splices the header encrypted data, the body encrypted data, and the tail encrypted data in sequence with the hash value H2 to obtain spliced data; In the spliced data, a group of 5-byte random strings generated by / dev / urandom are inserted every 10 bytes to form obfuscated data; Base64 encode the obfuscated data, and then perform a cyclic shift on the letters in the encoding result through the Caesar cipher to generate ciphertext data; Encrypt the key K1, key K2, and key K3 into a key package through the SM2 algorithm; Obtain the current system time and the MAC address of the local machine, concatenate the system time and the MAC address into dynamic parameters, encrypt the ciphertext data, the key package, and the dynamic parameters into an encrypted log through the AES-GCM algorithm, and perform distributed storage on the encrypted log; Calculate the data fingerprint of the encrypted log through the HMAC algorithm, and store the data fingerprint on the blockchain.

Images