Business data protection system and method based on federal learning

Abstract

The invention relates to a federated learning-based commercial data protection system, and the system comprises a federated protection module which is used for obtaining multi-modal behavior feature vectors, forming encryption features, carrying out aggregation operation, forming an encryption view, constructing a relation graph, and carrying out federated risk scoring; the self-adaptive control module is used for selecting a group of response action candidate sets, performing cost-benefit simulation evaluation on candidate actions in the response action candidate sets, generating a response action sequence, issuing a control measure command, verifying whether risk misjudgment occurs or not, and feeding back a verification result to the federal protection module; and the compliance guarantee module is used for acquiring data, carrying out noise addition on the data, locally completing encryption, constructing a verifiable trust and audit chain, hiding technical operation to corresponding legal requirements and dynamically generating a compliance state report. According to the technical scheme, the technical problems that in the prior art, misjudgment and early warning are prone to occurring, effective blocking cannot be conducted before IP leakage occurs, and privacy regulation requirements are not met are solved.

Description

Technical Field

[0001] This invention relates to the field of enterprise information security technology, and in particular to a business data protection system and method based on federated learning. Background Technology

[0002] Traditional technical solutions require centralized collection of employee behavior data. Based on rules and relying on pre-defined, explicit conditions, they identify IP leakage behavior, such as triggering an alert if an employee copies a file larger than 10MB to a USB device. Because employees have different work habits, such rules cannot be applied to every employee, leading to incorrect judgments and warnings. Furthermore, existing solutions are mostly reactive, unable to effectively prevent IP leakage before it occurs. Summary of the Invention

[0003] In view of the above problems, the present invention provides a business data protection system and method based on federated learning to solve the technical problems of existing technologies, such as easy misjudgment and early warning, inability to effectively block IP leakage before it occurs, and non-compliance with privacy regulations.

[0004] This invention provides a business data protection system based on federated learning. The system includes: a federated protection module for collecting code development behavior, network access behavior, document operation behavior, and physical access behavior; obtaining multimodal behavioral feature vectors; performing privacy protection processing on these vectors to form encrypted features that cannot be directly deciphered; performing aggregation operations on several encrypted features to form an encrypted view; integrating the encrypted view with entity relationship information to construct a relationship graph; and performing a federated risk score based on the relationship graph: RiskScore = 0.35×BehaviorAnomaly + 0.30×IPExposure + 0.20×ExternalActivity + 0.15×TemporalFactors: Update and distribute the global risk assessment model, where BehaviorAnomaly is the behavioral anomaly index, IPExposure is the data exposure index, ExternalActivity is the external dangerous behavior index, and TemporalFactors is the time factor index. The adaptive control module, connected to the federated protection module, combines federated risk scoring and contextual diagnosis to select a set of response action candidates from a pre-set policy library when data security risks occur. It then performs cost-benefit simulations and evaluations on these candidate actions, generating a response action sequence accordingly. The system issues control measures commands to various devices. During the execution of these control measures, it verifies whether any risk misjudgment occurs and feeds the verification results back to the Federal Protection Module to optimize the global risk assessment model. The Compliance Assurance Module is connected to both the Federal Protection Module and the Adaptive Control Module. It acquires data generated by these modules, adds noise to the data using a differential privacy protection layer, and encrypts it locally using homomorphic encryption. Through zero-knowledge proof technology and blockchain technology, it constructs a verifiable trust and audit chain, mapping technical operations to relevant legal requirements and dynamically generating compliance status reports, thereby achieving both privacy protection and compliance assurance.

[0005] Furthermore, the federated protection module includes: a behavior collector for collecting raw behavior data and extracting preliminary features; a federated learning security framework connected to the behavior collector for providing encrypted transmission of the collected data and extracted preliminary features, performing aggregation operations, and forming an encrypted view; a graph analyzer connected to the federated learning security framework for constructing and analyzing the relationship network between employees, projects, and data assets; and a federated risk scoring engine connected to the behavior collector and the graph analyzer, respectively, for calculating federated risk scores based on graphs and federated learning.

[0006] Furthermore, the adaptive control module includes: a policy management console, which provides a user interface for policy configuration and selection, and selects a set of response action candidates from a preset policy library when a data security risk occurs by combining federal risk scoring and contextual diagnostics; an adaptive response controller, connected to the policy management console, which performs cost-benefit simulation and evaluation of the candidate actions in the candidate set, generates a response action sequence, and issues control measure commands to various devices accordingly, and performs automatic rollback when the response fails; a real-time alarm center, connected to the adaptive response controller, which verifies whether a risk misjudgment occurs during the execution of control measures, and feeds back the verification results to the federal protection module to optimize the global risk assessment model; and an external interaction integration module, which is connected to the adaptive response controller, the real-time alarm center, and the adaptive controller, respectively, and interacts with external systems to execute specific control measures according to the control measure commands.

[0007] Furthermore, the compliance assurance module includes: a differential privacy protector for adding privacy noise during data processing; a homomorphic encryption processor connected to the differential privacy protector for performing encryption locally using homomorphic encryption technology; a secure communication gateway connected to the homomorphic encryption processor for providing authentication and communication security based on zero-knowledge proof technology; an audit verifier connected to the secure communication gateway for providing tamper-proof audit records and compliance verification based on blockchain technology; and an audit report generator connected to the audit verifier for generating audit reports that comply with regulatory requirements.

[0008] This invention also provides a method for a business data protection system based on federated learning. The method includes: Step 1, collecting code development behavior, network access behavior, document operation behavior, and physical access behavior to obtain multimodal behavior feature vectors; Step 2, adding noise to the data through a differential privacy protection layer and encrypting it locally using homomorphic encryption technology to form encrypted features that cannot be directly interpreted; Step 3, performing aggregation operations on several encrypted features to form an encrypted view, integrating the encrypted view with entity relationship information to construct a relationship graph; Step 4, based on the relationship graph, performing a federated risk score: RiskScore = 0.35×BehaviorAnomaly + 0.30×IPExposure + 0.20×ExternalActivity + Step 5: When data security risks occur, combine the federal risk score and context diagnosis to select a set of response action candidates from the preset policy library, conduct cost-benefit simulation evaluation of the candidate actions, generate a response action sequence, and issue control measures commands to each device accordingly. Step 6: During the execution of control measures, verify whether there is a risk misjudgment, and feed the verification results back to the federal protection module to optimize the global risk assessment model. Step 7: Through zero-knowledge proof technology and blockchain technology, construct a verifiable trust and audit chain, map technical operations to corresponding legal requirements, and dynamically generate compliance status reports to achieve privacy protection and compliance assurance.

[0009] Furthermore, step 1 includes: step 11, the code behavior collector parses the version control system logs and extracts code submission frequency, time, and sensitive file access sequence patterns; step 12, the network behavior collector analyzes proxy server and firewall logs to identify abnormal external connections and large-scale data upload behaviors; step 13, the document and physical behavior collectors respectively track access and printing records of sensitive files, as well as the usage of access control and USB devices.

[0010] Furthermore, step 4 includes: Step 41, updating and distributing the global risk assessment model according to the following function for subsequent risk analysis, thereby forming an optimization loop, W_{t+1} = W_t - η· W_global, W_global = ∑( W_i + Noise_i) / n, where, W_global represents the global gradient, W_{t+1} represents the updated global risk assessment model, W_t represents the original global risk assessment model, and η represents the learning rate. W_i + Noise_i is the gradient after adding noise, and n is the number of participants; Step 42, according to the function Risk(v) = α·InitialRisk(v) + β·∑ u∈N(v) w uv Risk(u) is used to simulate risk on the relationship graph, where Risk(v) is the exposure risk score, α is the individual risk weight, InitialRisk(v) is the initial individual risk score, β is the relationship path weight, and w uv For risk propagation capability, Risk(u) is the risk score of the employee to which the relationship path points; Step 43, based on the relationship graph, perform risk scoring through the global risk assessment model: RiskScore = 0.35×BehaviorAnomaly + 0.30×IPExposure +0.20×ExternalActivity + 0.15×TemporalFactors, where BehaviorAnomaly is the behavioral anomaly index, IPExposure is the data contact index, ExternalActivity is the external dangerous behavior index, and TemporalFactors is the time factor index.

[0011] Furthermore, step 5 includes: Step 51, when a data security risk occurs, performing contextual diagnosis to analyze whether the current time is a peak business period, whether the employee is handling critical tasks, and their historical behavior patterns, and determining the risk level based on the federal risk score; Step 52, combining the risk level and contextual diagnosis results, selecting a preliminary set of response action candidates from a predefined strategy library; Step 53, performing cost-benefit simulation evaluation on the candidate actions in the response action candidate set to predict the business interruption cost that may be caused by executing the action, and how much risk can be reduced; Step 54, generating a priority-ordered response action sequence based on the prediction results to ensure that the maximum security effect is achieved with the minimum business cost; Step 55, the multi-level response executor, based on the response action sequence, issues control measure commands to various infrastructures within the enterprise through an integration interface.

[0012] Furthermore, step 6 includes: step 61, after the control measure command is issued, continuously monitor relevant feedback and logs to confirm that the control measure has taken effect; step 62, during the execution of the control measure, verify whether a risk misjudgment has occurred. If a risk misjudgment is found, the control measure is revoked through an automatic rollback mechanism to restore business; step 63, use the final effect of each control measure as feedback data to optimize the global risk assessment model.

[0013] Furthermore, step 7 includes: step 71, storing the encrypted information to be audited; step 72, using zero-knowledge proof technology to write key operation instructions, risk alarms, and response logs into the blockchain audit chain in real time; step 73, mapping the technical operations to the corresponding legal requirements and dynamically generating a compliance status report, thereby completing privacy protection and compliance assurance.

[0014] This invention provides a business data protection system and method based on federated learning. It aims to resolve the contradiction between privacy protection and IP protection in the prior art, and to provide a precise, real-time, and adaptive enterprise intellectual property protection solution that can protect employee privacy. It is mainly used to solve the technical problems of data silos and privacy contradictions, insufficient detection accuracy, and delayed response in the prior art. Attached Figure Description

[0015] Figure 1 A schematic diagram of a business data protection system based on federated learning provided by the present invention; Figure 2 A flowchart of a business data protection method based on federated learning provided by the present invention; Figure 3 This is a flowchart of the method for obtaining feature vectors provided by the present invention; Figure 4 This is a flowchart of the method for performing federal risk scoring and updating the global risk assessment model provided by the present invention; Figure 5 This is a flowchart of the method for issuing control measures commands provided by the present invention; Figure 6 This is a flowchart of the method for optimizing the feedback verification result model provided by the present invention; Figure 7 This is a flowchart of the method for generating a compliance status report provided by the present invention. Detailed Implementation

[0016] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0017] Device Implementation Example: This invention provides a commercial data protection system based on federated learning, such as... Figure 1 As shown, the system is divided into five layers: infrastructure layer, data acquisition layer, federated computing layer, core business layer, and user interaction layer, in addition to external system integration. The system includes a federated protection module, an adaptive control module, and a compliance assurance module, with each module component distributed across its respective layer, as detailed below.

[0018] The Federal Protection module collects code development behavior, network access behavior, document manipulation behavior, and physical access behavior to obtain multimodal behavioral feature vectors. These vectors undergo privacy protection processing to form encrypted features, making them unreadable. Several encrypted features are aggregated to form an encrypted view. This encrypted view is then integrated with entity relationship information to construct a relationship graph. Based on this graph, a federal risk score is calculated: RiskScore = 0.35×BehaviorAnomaly + 0.30×IPExposure + 0.20×ExternalActivity + 0.15×TemporalFactors. The global risk assessment model is then updated and distributed, where BehaviorAnomaly is the behavioral anomaly index, IPExposure is the data access index, ExternalActivity is the external dangerous behavior index, and TemporalFactors is the time factor index. Figure 1As shown, the federated protection module includes a behavior collector, a federated learning security framework, a graph analyzer, and a federated risk scoring engine, located at the data acquisition layer, the federated computing layer, and the core business side, respectively. Specifically, it includes the following: The behavior collector, belonging to the data acquisition layer, is used for collecting raw behavioral data and extracting preliminary features. The system starts multiple acquisition engines, i.e., behavior collectors, in parallel at each edge node (such as each R&D center). The behavior collectors include code behavior collectors, network behavior collectors, document behavior collectors, and physical behavior collectors. The code behavior collector parses the logs of version control systems (such as Git) to extract patterns such as code commit frequency, time, and sensitive file access sequences. The network behavior collector analyzes proxy server and firewall logs to identify behaviors such as abnormal external connections and large-scale data uploads. The document and physical behavior collectors track access and printing records of sensitive files, as well as the usage of access control systems and USB devices, respectively. All this raw data never leaves the local node but is transformed into structured "behavioral feature vectors" in real time locally. The Federated Learning Security Framework, deployed at the Federated Computing Layer and connected to the Behavior Collector, provides encrypted transmission of collected data and extracted preliminary features, performs aggregation operations, and forms an encrypted view. The generated feature vectors immediately enter the privacy-preserving pipeline for encrypted transmission and aggregation operations. After local privacy processing and encryption, the data packets prepared for upload are encrypted and cannot be directly deciphered. Federated machine learning (Federated machine learning), also known as federated machine learning, consortium learning, or alliance learning, is a machine learning framework that effectively helps multiple organizations use data and perform machine learning modeling while meeting user privacy protection, data security, and government regulatory requirements. The Graph Analyzer, deployed at the core business layer and connected to the Federated Learning Security Framework, is used to build and analyze the relationship network of employees, projects, and data assets. Each node device uploads its locally encrypted feature data packets to the central node through a secure channel. Without decryption, the central node utilizes the properties of homomorphic encryption to perform secure aggregation operations (such as federated averaging) on ​​the encrypted features from multiple nodes, forming a global, encrypted feature view. Meanwhile, the knowledge graph fusion unit at the central node integrates the anonymized entity relationship information uploaded by each node (such as "Employee A participated in Project X"), dynamically constructing and updating a global "Employee-Project-Data Asset" relationship graph. The federated risk scoring engine, deployed at the core business layer, connects to the behavior collector and graph analyzer, respectively, and is used to calculate federated risk scores based on the graph and federated learning. Based on encrypted features and a dynamic relationship graph, the federated risk scoring engine begins operation. It updates the global risk assessment model on one hand, and simulates risks on the knowledge graph on the other (e.g., the risk impact of a high-risk employee on colleagues within their project).This process calculates not a single score, but a comprehensive assessment encompassing multiple dimensions, including individual behavioral anomalies, IP exposure, and network association risks. Ultimately, the updated, more accurate global risk assessment model is encrypted and distributed to all edge nodes for the next round of more efficient local analysis, thus forming a continuously evolving intelligent cycle. The federal risk score formula is: RiskScore = 0.35×BehaviorAnomaly + 0.30×IPExposure + 0.20×ExternalActivity + 0.15×TemporalFactors, that is: Federal Risk Score = 0.35*Behavioral Anomaly Index + 0.3*Data Exposure Index + 0.2*External Risk Behavior Index (e.g., accessing recruitment websites) + 0.15*Time Factor Index (e.g., whether it's a career milestone - performance evaluation, resignation application, etc.).

[0019] The adaptive control module, connected to the federated protection module, combines federated risk scoring and contextual diagnostics to select a set of response action candidates from a pre-set policy library when data security risks arise. It then performs cost-benefit simulations and evaluations of these candidate actions, generates a response action sequence, and issues control measures commands to various devices accordingly. During the execution of these control measures, it verifies whether any risk misjudgments have occurred and feeds the verification results back to the federated protection module to optimize the global risk assessment model. Figure 1As shown, the adaptive control module includes a policy management console, an adaptive response controller, a real-time alarm center, and an external interaction integration module, located in the user interaction layer, the core business layer, and the external system integration layer, respectively, as detailed below. The policy management console, deployed in the user interaction layer, provides a user interface for policy configuration and selection. It combines federal risk scoring and contextual diagnostics to select a set of response action candidates from a pre-defined policy library when data security risks arise. When the core business layer generates a high-risk alarm (e.g., a risk score for an employee), the intelligent policy selection engine immediately activates. It does not mechanically apply rules but first performs "contextual diagnostics": analyzing whether the current time is a peak business period, whether the employee is handling critical tasks, and their historical behavior patterns. Combining the risk level (e.g., "high risk") and the context, the engine filters a preliminary set of response action candidates (e.g., "restrict code submission permissions," "trigger HR review") from the predefined policy library. It predicts the business interruption cost that executing a certain action (e.g., "freeze account") may cause and estimates how much risk it can reduce. Through this trade-off calculation, the system automatically generates a prioritized, optimized sequence of response actions to ensure maximum security with minimal business costs. The adaptive response controller, deployed at the core business layer and connected to the policy management console, performs cost-benefit simulations and evaluations of candidate actions in the candidate set, generates a response action sequence, and issues control measures commands to various devices accordingly. Automatic rollback is executed when the response fails. The cost-benefit simulation evaluation involves the system predicting the potential business interruption cost of executing a certain action (such as "freezing an account") and estimating how much risk it can reduce. Through this trade-off calculation, the system automatically generates a prioritized, optimized sequence of response actions to ensure maximum security with minimal business costs. The real-time alarm center, deployed at the user exchange layer and connected to the adaptive response controller, verifies whether there are any misjudgments of risk during the execution of control measures and feeds the verification results back to the federated protection module to optimize the global risk assessment model. Multi-level response executors receive instructions and issue commands to various internal infrastructures through the system-integrated API. This might include: notifying the version control system (such as GitLab) to freeze the employee's write permissions to the code repository; instructing the network firewall to block their upload channels to external cloud drives; and simultaneously sending a pending review task to the HR system's workflow platform. After the command is executed, the real-time effect verification module continuously monitors the feedback and logs of the relevant systems to confirm whether the control measures have taken effect. The entire execution process is recorded in detail. If the verification finds an overreaction (a false alarm) or underreaction, the automatic rollback mechanism can quickly undo the operation and restore business operations.More importantly, the final effect of each response—whether it truly prevented the threat and what its impact on business—is fed back as feedback data to the strategy engine and global risk assessment model to optimize future decisions, making the system increasingly "intelligent" and accurate. The external interaction integration module, deployed in the external system integration section, connects to the adaptive response controller, real-time alarm center, and adaptive controller. It interacts with external systems to execute specific control measures based on control action commands, including version control systems, network devices, endpoint security systems, and HR management systems.

[0020] The compliance assurance module, connected to both the federal protection and adaptive control modules, acquires data generated by these modules. It adds noise to the data using a differential privacy protection layer and encrypts it locally using homomorphic encryption. Through zero-knowledge proof and blockchain technology, it constructs a verifiable trust and audit chain, mapping technical operations to relevant legal requirements and dynamically generating compliance status reports, thus achieving both privacy and compliance assurance. This module is not an independent component but rather acts as a "capillary" and "cornerstone," deeply embedded in every step of the first two modules, providing verifiable privacy protection throughout the entire lifecycle. Figure 1As shown, the compliance assurance module includes: a differential privacy protector, a homomorphic encryption processor, a secure communication gateway, an audit verifier, and an audit report generator, deployed at the federated computing layer, infrastructure layer, and user interaction layer, respectively, as detailed below. The differential privacy protector adds privacy noise during data processing. From the moment data is generated, the principle of data minimization is applied, collecting only necessary feature information. Before transmission, the data is noise-added via the differential privacy protector. The homomorphic encryption processor, connected to the differential privacy protector, performs encryption locally using homomorphic encryption technology. Local encryption using homomorphic encryption technology achieves "data usable but not visible." In the core federated computing process, all aggregation operations are performed on ciphertext, and the central node has no way of knowing any of the original information. Finally, any audit information that needs to be stored is anonymized. The secure communication gateway, connected to the homomorphic encryption processor, provides authentication and communication security based on zero-knowledge proof technology. The system uses zero-knowledge proof technology to allow edge nodes to prove to the central node, "I have indeed processed the real data locally according to the prescribed algorithm," without revealing the data itself. The audit verifier, connected to the secure communication gateway, provides tamper-proof audit records and compliance verification based on blockchain technology. The aforementioned verification process, along with all critical operational instructions, risk alerts, and response logs, is written to the blockchain audit chain in real time. The immutability of the blockchain ensures that the entire system's operational records are complete, reliable, and traceable. The audit report generator, connected to the audit verifier, generates audit reports that comply with regulatory requirements. The system's built-in compliance verifier maintains a machine-readable regulatory knowledge base (such as key provisions of GDPR and the Personal Information Protection Act). It automatically maps the system's technical operations (e.g., "using differential privacy with ε=0.1") to corresponding legal requirements (e.g., "taking technical measures to ensure data processing security") and dynamically generates compliance status reports. The audit report generator can generate structured evidence documents that meet internal and external audit requirements with a single click, transforming complex technical processes into clear legal compliance statements, significantly reducing compliance costs and risks for enterprises.

[0021] This system provides a brand-new enterprise intellectual property protection solution. It not only achieves a balance between privacy protection and security performance in terms of technology, but also builds an intelligent, adaptive, and trustworthy secure collaborative network for enterprises at the business level, providing a solid guarantee for the core assets of enterprises in digital competition.

[0022] Method Example: This invention provides a method for a business data protection system based on federated learning, such as... Figure 2 As shown, the method includes: Step 1: Collect code development behavior, network access behavior, document operation behavior, and physical access behavior to obtain multimodal behavior feature vectors; like Figure 3 As shown, step 1 includes: Step 11: The code behavior collector parses the version control system logs and extracts code commit frequency, time, and sensitive file access sequence patterns. Step 12: The network behavior collector analyzes the proxy server and firewall logs to identify abnormal outbound connections and large-scale data uploads. Step 13: The document and physical behavior collectors track access and printing records of sensitive documents, as well as the usage of access control and USB devices.

[0023] All of this raw data never leaves the local node; instead, it is transformed into structured "behavioral feature vectors" in real time on the local machine.

[0024] Step 2: Add noise to the data using a differential privacy protection layer and encrypt it locally using homomorphic encryption technology to form encrypted features that make it impossible to be directly deciphered; The generated feature vector immediately enters the privacy protection pipeline. First, the differential privacy protector adds random noise conforming to a strict mathematical definition to the feature, ensuring that the original data of any single employee cannot be inferred from the result. Next, the homomorphic encryption processor uses an encryption algorithm (such as CKKS) to convert the noisy feature vector into ciphertext. At this point, local processing is complete, and the encrypted data packet, which cannot be directly deciphered, is ready for upload. Differential privacy is a technique in cryptography designed to maximize the accuracy of data queries when querying statistical databases, while minimizing the chance of identifying records. Homomorphic encryption, a cryptographic technique based on the computational complexity theory of mathematical problems, is a class of encryption methods with special natural properties. Compared to general encryption algorithms, homomorphic encryption, in addition to performing basic encryption operations, can also perform various computational functions between ciphertexts; that is, computation before decryption is equivalent to decryption before computation. This characteristic is of great significance for protecting information security. Homomorphic encryption technology can be used to perform calculations on multiple ciphertexts before decryption, avoiding the high computational cost of decrypting each ciphertext individually. Homomorphic encryption technology can also enable keyless computation of ciphertexts, reducing communication costs and shifting computational tasks, thus balancing the computational costs for all parties. Furthermore, homomorphic encryption technology ensures that the decrypting party can only know the final result and not the individual ciphertext messages, thereby improving information security.

[0025] Step 3: Perform aggregation operations on several encryption features to form an encryption view, and integrate the encryption view with entity relationship information to construct a relationship graph; Step 4: Based on the relationship graph, calculate the federal risk score: RiskScore = 0.35 × BehaviorAnomaly + 0.30 × IPExposure + 0.20 × ExternalActivity + 0.15 × TemporalFactors. Update and distribute the global risk assessment model, where BehaviorAnomaly is the behavioral anomaly index, IPExposure is the data exposure index, ExternalActivity is the external dangerous behavior index, and TemporalFactors is the time factor index. like Figure 4 As shown, step 4 includes: Step 41: Update and distribute the global risk assessment model according to the following function for subsequent risk analysis, thus forming an optimization loop. W_{t+1} = W_t - η· W_global, W_global = ∑( W_i + Noise_i) / n, in, W_global represents the global gradient, W_{t+1} represents the updated global risk assessment model, W_t represents the original global risk assessment model, and η represents the learning rate. W_i+Noise_i is the gradient after adding noise, and n is the number of participants; The noise added here is Laplacian noise: Noise ~ Lap(Δf / ε), randomly sampled from a Laplacian distribution. It is used to irreversibly blur the gradient. After this step, even if an attacker obtains the gradient with added noise, the noise will remain obfuscated. Even with W_i+Noise_i, the original gradient cannot be inferred with high confidence. W_i, thus making it impossible to trace back to the original individual data that generated the gradient. This provides a mathematically provable privacy guarantee. Δf is the sensitivity, which quantifies "the maximum impact that a single participant's data can have on the computational outcome." In federated learning, it is typically the maximum norm (e.g., L1 or L2 norm) of the local gradient vector. We can explicitly control Δf by gradient clipping (setting an upper limit C), making it a known constant (e.g., Δf=C). Δf is the "anchor" of the noise scale. Higher sensitivity means a greater influence of a single person's data, and more noise needs to be added to mask that influence. ε is the privacy budget, the core parameter of differential privacy, also known as the "privacy loss parameter." The smaller the ε value, the stronger the privacy protection provided. ε directly controls the scale of the noise. The smaller ε is, the smaller the denominator in the formula, resulting in a larger noise scale Δf / ε, more intense added noise, and more thorough privacy protection, but also a greater impairment to the model's usefulness. Lap(b) is the Laplace distribution. Safely aggregating noisy gradients: W_global = ∑( W_i + Noise_i) / n. The central server (coordinating node) collects all noisy local gradient updates uploaded by participants and averages them to obtain the global gradient update. This is used to leverage collective intelligence while protecting individual privacy. Although individual noisy gradients are distorted, the average of a large number of distorted gradients amplifies the signal (pointing to the true optimization direction) while relatively weakening the noise, resulting in a global direction that is still useful and safe for model updates. The gradients aggregated by the server are not pure gradients. W_i, but the version contaminated with noise. W_i + Noise_i. The server never has the opportunity to access a pure gradient. The noise_i added by each participant is independent and identically distributed, with a mean of 0. When averaging a large number of such noise gradients, according to the law of large numbers, these independent noises will partially cancel each other out. The final global gradient... The expected value of the "net noise" contained in W_global is zero, and its variance is inversely proportional to the number of participants, n. Because the noise is averaged during aggregation, to achieve the same global model accuracy, the amount of noise each participant actually needs to add is less than the amount required to directly publish a local model. This is known as the "amplification effect" in privacy analysis and is an advantage of federated learning architectures. Update the global risk assessment model: W_{t+1} = W_t - η· W_global. This is the standard (stochastic) gradient descent update step, using the computed, privacy-preserving global gradient. W_global is used to update the shared global risk assessment model parameters W. This drives the model to evolve in a direction that is effective on group data and independent of any individual data. After multiple iterations, the final model is a "privacy-safe model" that both learns collective knowledge and satisfies the definition of differential privacy. η is the learning rate, and the choice of η is particularly important. Because... W_global contains residual noise. If η is too large, the noise will be amplified, potentially leading to unstable model updates, oscillations, or even divergence. If η is too small, the convergence speed will be very slow, requiring more training epochs, each epoch consuming additional privacy budget ε. Therefore, η needs careful tuning to strike a balance between convergence speed and model stability. The "privacy-utility" tradeoff is reflected here: Utility (model accuracy): W_global should be as close as possible to the ideal global gradient without noise, ensuring accurate model update direction and good final model performance. Privacy (protection strength): Adding sufficiently large noise (i.e., using a smaller ε) provides strong protection; ε and η jointly control this tradeoff. Stronger privacy (smaller ε) means... The W_global model has more noise, which may require smaller η and more training epochs to compensate for, ultimately affecting the model's utility.

[0026] Step 42, according to the function Risk(v) = α·InitialRisk(v) + β·∑ u∈N(v) w uv Risk(u) is used to simulate risk on the relationship graph, where Risk(v) is the exposure risk score, α is the individual risk weight, InitialRisk(v) is the initial individual risk score, β is the relationship path weight, and w uv For risk propagation capability, Risk(u) is the risk score of the employee to whom the relationship path points; The function provided in this step is the core iterative equation of a "risk propagation model." It describes how risk values ​​are transmitted and calculated between nodes (knowledge graphs) through edges (relationships) in a graph consisting of nodes such as employees, projects, and knowledge assets. The IP exposure risk score is obtained through the employee-project-knowledge asset association graph G(V,E). First, the risk values ​​of each node are initialized. Then, the risk propagation weights are calculated based on the relationship strength. Next, risk propagation is iteratively executed until convergence: Risk(v) = α·InitialRisk(v) + β·∑ u∈N(v) w uv• Risk(u), i.e., IP exposure risk score = individual risk weight * initial individual risk score (i.e., the RiskScore result above) + risk weight of social network "infection" * summation (relationship path weight * risk score of the employee (colleague) to which the relationship path points). Where w uv The "pipeline thickness," representing risk propagation, is a value between 0 and 1, quantifying the closeness, importance, and risk transmission capability of the relationship between two entities. For employee-project relationships: the weight may be based on the employee's participation in the project and the importance of their role (e.g., core developer vs. peripheral tester). For employee-employee relationships: the weight may be based on the frequency of code collaboration (e.g., joint Git commits), communication frequency, and distance within the organizational structure. α and β represent the "risk balance" between the individual and the environment. These can be customized according to the enterprise environment; for example, highly independent R&D teams have high α and low β. Strongly collaborative, project-based teams have low α and high β.

[0027] Step 43: Based on the relationship graph, perform a federal risk score using a global risk assessment model: RiskScore = 0.35 × BehaviorAnomaly + 0.30 × IPExposure + 0.20 × ExternalActivity + 0.15 × TemporalFactors, where BehaviorAnomaly is the behavioral anomaly index, IPExposure is the data exposure index, ExternalActivity is the external dangerous behavior index, and TemporalFactors is the time factor index.

[0028] Step 5: When a data security risk occurs, combine the federal risk score and context diagnosis to select a set of response action candidates from the preset policy library, conduct cost-benefit simulation evaluation of the candidate actions, generate a response action sequence, and issue control action commands to each device accordingly. like Figure 5 As shown, step 5 includes: Step 51: When a data security risk occurs, perform contextual diagnosis to analyze whether the current time is a peak business period, whether the employee is handling critical tasks, their historical behavior patterns, and determine the risk level based on the federal risk score results. The federal risk score and its corresponding risk level are as follows: 0-0.3 is low risk Level 1, 0.3-0.6 is medium risk Level 2, 0.6-0.8 is high risk Level 3, and 0.8-1.0 is severe risk Level 4.

[0029] Step 52: Combining the risk level and contextual diagnostic results, select a preliminary set of candidate response actions from the predefined strategy library; The risk levels and corresponding candidate response actions are as follows: Level 1 response actions: logging, notification management; Level 2 response actions: enhanced monitoring, multi-factor authentication, network restriction; Level 3 response actions: code access restriction, USB device disabling, download freezing; Level 4 response actions: immediate access blocking, legal proceedings initiation, professional agreement activation.

[0030] Step 53: Perform a cost-benefit simulation evaluation on the candidate actions in the response action candidate set to predict the business interruption costs that may result from executing the action and how much risk can be reduced. Step 54: Based on the prediction results, generate a priority-ordered sequence of response actions to ensure maximum security effect with minimal business cost; Step 55: Based on the response action sequence, the multi-level response actuator issues control action commands to various infrastructures within the enterprise through the integration interface.

[0031] Step 6: During the implementation of control measures, verify whether there is any misjudgment of risk, and feed the verification results back to the federal protection module in order to optimize the global risk assessment model; like Figure 6 As shown, step 6 includes: Step 61: After the control measures order is issued, continuously monitor the relevant feedback and logs to confirm that the control measures have taken effect; Step 62: During the execution of control measures, verify whether there is a misjudgment of risk. If a misjudgment of risk is found, the control measures are revoked through the automatic rollback mechanism and the business is restored. Step 63: Use the final effect of each control measure as feedback data to optimize the global risk assessment model.

[0032] Step 7: Using zero-knowledge proof technology and blockchain technology, a verifiable trust and audit chain is constructed, mapping technical operations to corresponding legal requirements and dynamically generating compliance status reports, thereby achieving privacy protection and compliance assurance.

[0033] like Figure 7 As shown, step 7 includes: Step 71: Store the encrypted information to be audited; Step 72: Using zero-knowledge proof technology, key operation instructions, risk alarms, and response logs are written to the blockchain audit chain in real time; Step 73 maps the technical operations to the corresponding legal requirements and dynamically generates a compliance status report, thereby achieving privacy protection and compliance assurance.

[0034] This invention provides a business data protection system based on federated learning. It aims to resolve the contradiction between privacy protection and IP protection in existing technologies, and to provide a precise, real-time, and adaptive enterprise intellectual property protection solution that can protect employee privacy. It is mainly used to solve technical problems such as data silos, privacy contradictions, insufficient detection accuracy, and delayed response in existing technologies.

[0035] In summary, this invention provides a business data protection system and method based on federated learning. This technical solution resolves the long-standing contradiction between privacy protection and security effectiveness in the field of enterprise intellectual property protection, and introduces a new adaptive and trustworthy protection paradigm. Its beneficial effects are mainly reflected in the following four aspects: First, it achieves precise protection under strict privacy protection. Traditional solutions, in pursuit of detection effects, centrally process raw data, facing significant legal and ethical risks. This system, through a knowledge federation technology framework, fundamentally reshapes the protection model: Data remains stationary, model moves: The raw behavioral data of each business unit (such as algorithm and hardware teams) does not need to leave the local machine; feature extraction and encryption are performed locally, eliminating the risk of privacy leakage at the source. Second, it is usable but invisible: Through homomorphic encryption and differential privacy technology, the central node can aggregate and analyze the encrypted features and train the model to obtain global risk insights without accessing any plaintext sensitive information. Third, it provides self-verification of compliance: The system's built-in zero-knowledge proof and blockchain audit and evidence storage mechanism can provide verifiable privacy protection proofs to regulatory agencies, ensuring that all operations comply with domestic and international data privacy regulations. The system enhances the depth and foresight of risk perception, breaking through the limitations of traditional shallow detection based on static rules. It achieves deep insights into complex and hidden threats through: Multi-dimensional behavioral fusion analysis: The system comprehensively analyzes multi-modal data such as employee code development, network access, document operations, and even physical behavior, enabling it to identify collaborative anomalies that cannot be detected by a single dimension. Knowledge graph-based association reasoning: By constructing an employee-project-knowledge asset association graph, the system can understand the flow of enterprise knowledge assets, accurately assess the IP exposure risk of specific employees, and predict the potential propagation path of risks in the collaborative network. Dynamic risk baseline: The system establishes a dynamic behavioral baseline for each employee and role, enabling it to keenly perceive "behavioral drift" that deviates from its normal pattern, rather than mechanically triggering static thresholds, thus discovering potential risk intentions earlier. A closed-loop, intelligent, and controllable response system has been constructed, transforming the traditional lagging "alarm-manual handling" model and achieving automation and intelligence in the response process: Adaptive hierarchical response: The system can automatically select and execute a tiered response strategy, from "logging" to "immediate blocking," based on risk assessment results and business context (such as whether it is a critical project period), avoiding insufficient protection or overreaction. Cost-effectiveness optimization: Before executing a response action, the system estimates the business impact and risk reduction benefits of the action, prioritizing the most cost-effective measures to minimize interference with normal business operations. Self-healing and continuous optimization: The system has an automatic rollback mechanism, enabling rapid recovery when a response is confirmed to be a false alarm; simultaneously, the response results serve as feedback data to optimize subsequent strategy selection, forming a closed loop that becomes increasingly intelligent with use.This system lays the technological foundation for cross-domain collaboration and secure trust, providing enterprises, especially large, geographically dispersed R&D companies, with a unified IP protection infrastructure: Breaking down data silos: Under strict compliance with data localization regulations, it enables risk information sharing and collaborative protection among various R&D centers and departments, forming a globally unified security view. Establishing a trust mechanism: Through federated learning and privacy computing technologies, the system establishes a verifiable trust and cooperation mechanism among multiple untrusted parties (such as different subsidiaries), making cross-organizational collaboration possible. Improving security operation efficiency: It frees security teams from massive false alarms and tedious manual audits, allowing them to focus on higher-value strategic analysis and incident response, reshaping the enterprise security operation model.

[0036] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A commercial data protection system based on federated learning, characterized in that, The system includes: The Federation Protection Module collects code development behavior, network access behavior, document operation behavior, and physical access behavior to obtain multimodal behavioral feature vectors. These vectors undergo privacy protection processing to form encrypted features that cannot be directly deciphered. Several encrypted features are aggregated to form an encrypted view. This encrypted view is then integrated with entity relationship information to construct a relationship graph. Based on this graph, a federated risk score is calculated: RiskScore = 0.35×BehaviorAnomaly + 0.30×IPExposure + 0.20×ExternalActivity + 0.15×TemporalFactors. The global risk assessment model is then updated and saved, where BehaviorAnomaly is the behavioral anomaly index, IPExposure is the data access index, ExternalActivity is the external dangerous behavior index, and TemporalFactors is the time factor index. The adaptive control module, connected to the federal protection module, combines federal risk scoring and contextual diagnosis to select a set of response action candidates from a preset policy library when data security risks occur. It performs cost-benefit simulation and evaluation on the candidate actions, generates a response action sequence, and issues control measures commands to various devices accordingly. During the execution of control measures, it verifies whether there is a risk misjudgment and feeds the verification results back to the federal protection module to optimize the global risk assessment model. The compliance assurance module is connected to the federal protection module and the adaptive control module respectively. It is used to obtain the data generated in the federal protection module and the adaptive control module, add noise to the data through a differential privacy protection layer, and encrypt it locally using homomorphic encryption technology. Through zero-knowledge proof technology and blockchain technology, a verifiable trust and audit chain is built, mapping the technical operations to the corresponding legal requirements, and dynamically generating compliance status reports, thereby completing privacy protection and compliance assurance.

2. The commercial data protection system based on federated learning according to claim 1, characterized in that, The federal protection module includes: A behavior collector is used to collect raw behavior data and extract preliminary features. The federated learning security framework, connected to the behavior collector, is used to provide encrypted transmission of the collected data and extracted preliminary features, perform aggregation operations, and form an encrypted view; The graph analyzer, connected to the federated learning security framework, is used to build and analyze the relationship network of employees, projects, and data assets. The federated risk scoring engine is connected to the behavior collector and the graph analyzer, respectively, and is used to compute a comprehensive risk score based on the graph and federated learning.

3. The commercial data protection system based on federated learning according to claim 1, characterized in that, The adaptive control module includes: The policy management console provides a user interface for policy configuration and selection. It combines federal risk scoring and contextual diagnostics to select a set of response action candidates from a pre-defined policy library when data security risks occur. The adaptive response controller, connected to the policy management console, is used to perform cost-benefit simulation and evaluation of candidate actions in the candidate set, generate a sequence of response actions, and issue control measures commands to various devices accordingly. When the response fails, it performs automatic rollback. The real-time alarm center, connected to the adaptive response controller, is used to verify whether a risk misjudgment has occurred during the execution of control measures, and to feed the verification results back to the federal protection module in order to optimize the global risk assessment model. The external interaction integration module is connected to the adaptive response controller, the real-time alarm center, and the adaptive controller, respectively, and is used to interact with external systems to execute specific control measures according to control measure commands.

4. The commercial data protection system based on federated learning according to claim 1, characterized in that, The compliance assurance module includes: Differential privacy protectors are used to add privacy noise during data processing; Homomorphic encryption processor, connected to differential privacy protector, is used to perform encryption locally using homomorphic encryption technology; A secure communication gateway, connected to a homomorphic encryption processor, is used to provide authentication and communication security based on zero-knowledge proof technology; An audit verifier, connected to a secure communication gateway, is used to provide tamper-proof audit records and compliance verification based on blockchain technology. An audit report generator, connected to an audit validator, is used to generate audit reports that comply with regulatory requirements.

5. A method for using a business data protection system based on federated learning as described in any one of claims 1-4, characterized in that, The method includes: Step 1: Collect code development behavior, network access behavior, document operation behavior, and physical access behavior to obtain multimodal behavior feature vectors; Step 2: Add noise to the data using a differential privacy protection layer and encrypt it locally using homomorphic encryption technology to form encrypted features that make it impossible to be directly deciphered; Step 3: Perform aggregation operations on several encryption features to form an encryption view, and integrate the encryption view with entity relationship information to construct a relationship graph; Step 4: Based on the relationship graph, perform a federal risk score: RiskScore = 0.35×BehaviorAnomaly +0.30×IPExposure + 0.20×ExternalActivity + 0.15×TemporalFactors. Update the global risk assessment model and distribute and save it. Here, BehaviorAnomaly is the behavioral anomaly index, IPExposure is the data exposure index, ExternalActivity is the external dangerous behavior index, and TemporalFactors is the time factor index. Step 5: When a data security risk occurs, combine the federal risk score and context diagnosis to select a set of response action candidates from the preset policy library, conduct cost-benefit simulation evaluation of the candidate actions, generate a response action sequence, and issue control action commands to each device accordingly. Step 6: During the implementation of control measures, verify whether there is any misjudgment of risk, and feed the verification results back to the federal protection module in order to optimize the global risk assessment model; Step 7: Using zero-knowledge proof technology and blockchain technology, a verifiable trust and audit chain is constructed, mapping technical operations to corresponding legal requirements and dynamically generating compliance status reports, thereby achieving privacy protection and compliance assurance.

6. The commercial data protection method based on federated learning according to claim 5, characterized in that, Step 1 includes: Step 11: The code behavior collector parses the version control system logs and extracts code commit frequency, time, and sensitive file access sequence patterns. Step 12: The network behavior collector analyzes the proxy server and firewall logs to identify abnormal outbound connections and large-scale data uploads. Step 13: The document and physical behavior collectors track access and printing records of sensitive documents, as well as the usage of access control and USB devices.

7. The commercial data protection method based on federated learning according to claim 5, characterized in that, Step 4 includes: Step 41: Update and distribute the global risk assessment model according to the following function for subsequent risk analysis, thus forming an optimization loop. W_{t+1} = W_t - η· W_global, W_global = ∑( W_i + Noise_i) / n, in, W_global represents the global gradient, W_{t+1} represents the updated global risk assessment model, W_t represents the original global risk assessment model, and η represents the learning rate. W_i + Noise_i is the gradient after adding noise, and n is the number of participants; Step 42, according to the function Risk(v) = α·InitialRisk(v) + β·∑ u∈N(v) w uv Risk(u) is used to simulate risk on the relationship graph, where Risk(v) is the exposure risk score, α is the individual risk weight, InitialRisk(v) is the initial individual risk score, β is the relationship path weight, and w uv For risk propagation capability, Risk(u) is the risk score of the employee to whom the relationship path points; Step 43: Based on the relationship graph, perform a federal risk score using the global risk assessment model: RiskScore = 0.35×BehaviorAnomaly + 0.30×IPExposure + 0.20×ExternalActivity + 0.15×TemporalFactors, where BehaviorAnomaly is the behavioral anomaly index, IPExposure is the data exposure index, ExternalActivity is the external dangerous behavior index, and TemporalFactors is the time factor index.

8. The commercial data protection method based on federated learning according to claim 5, characterized in that, Step 5 includes: Step 51: When a data security risk occurs, perform contextual diagnosis to analyze whether the current time is a peak business period, whether the employee is handling critical tasks, their historical behavior patterns, and determine the risk level based on the federal risk score results. Step 52: Combining the risk level and contextual diagnostic results, select a preliminary set of candidate response actions from the predefined strategy library; Step 53: Perform a cost-benefit simulation evaluation on the candidate actions in the response action candidate set to predict the business interruption costs that may result from executing the action and how much risk can be reduced. Step 54: Based on the prediction results, generate a priority-ordered sequence of response actions to ensure maximum security effect with minimal business cost; Step 55: Based on the response action sequence, the multi-level response actuator issues control action commands to various infrastructures within the enterprise through the integration interface.

9. The commercial data protection method based on federated learning according to claim 5, characterized in that, Step 6 includes: Step 61: After the control measures order is issued, continuously monitor the relevant feedback and logs to confirm that the control measures have taken effect; Step 62: During the execution of control measures, verify whether there is a misjudgment of risk. If a misjudgment of risk is found, the control measures are revoked through the automatic rollback mechanism and the business is restored. Step 63: Use the final effect of each control measure as feedback data to optimize the global risk assessment model.

10. The commercial data protection method based on federated learning according to claim 5, characterized in that, Step 7 includes: Step 71: Store the encrypted information to be audited; Step 72: Using zero-knowledge proof technology, key operation instructions, risk alarms, and response logs are written to the blockchain audit chain in real time; Step 73 maps the technical operations to the corresponding legal requirements and dynamically generates a compliance status report, thereby achieving privacy protection and compliance assurance.

Images