Commercial finance big data security processing method and system

Abstract

The invention relates to the technical field of data security processing, in particular to a commercial finance big data security processing method, which comprises the following steps: S1, receiving uploaded data from a plurality of data provider nodes, the uploaded data comprising local feature vectors encrypted by using a joint public key based on a threshold homomorphic encryption mechanism, the commercial financial subject anonymous identifier is encrypted by using the same joint public key and corresponds to the feature vector sample; according to the method, security alignment and cooperative calculation of financial data under full-flow ciphertext are realized, privacy is guaranteed and processing efficiency is greatly improved through decentralized architecture and hardware acceleration, and a reliable and efficient solution is provided for cross-mechanism data fusion.

Description

Technical Field

[0001] This invention relates to the field of data security processing technology, specifically to a method and system for secure processing of commercial and financial big data. Background Technology

[0002] With the full arrival of the digital economy era, the commercial and financial sector is undergoing profound changes. Banks, insurance companies, securities firms, and various financial institutions have accumulated massive amounts of high-value data in their daily operations and services. This data contains profound insights into market trends, customer behavior, and risk characteristics. At the same time, the need for cross-industry and cross-institutional data collaboration and joint modeling is becoming increasingly urgent. For example, in scenarios such as anti-fraud, precision marketing, and corporate credit assessment, integrating data from multiple parties can build more accurate and robust models, releasing enormous commercial value and social benefits. According to CN116776357A, a method and system for secure processing of commercial financial big data is disclosed. This technology discloses the following technical solutions: "It solves the technical problems of the stored financial data itself not being altered and being encrypted. When processing non-public data securely, it is necessary to use encrypted storage. During the storage process, firstly, according to the different segments existing within the corresponding non-public data, each group of different segments is extracted. Then, according to the specific capacity of the segment, the corresponding area of ​​the encrypted segment circle is determined. The area of ​​each encrypted segment circle is different, so it is difficult for external personnel to steal the corresponding data through the encrypted segment circle. Subsequently, the corresponding markers are deleted, and the data stream is divided to improve the encryption effect of the entire non-public data and improve the overall security processing effect." In existing cross-institutional commercial and financial big data collaborative modeling practices, a core issue is how to securely align data samples without exposing the original data of each party. Specifically, when multiple financial institutions (such as banks, insurance companies, and credit reporting agencies) need to conduct collaborative modeling based on their respective customer data (such as transaction records, policy information, and credit records), they must first confirm which records belong to the same customer or business entity, i.e., perform "customer ID alignment." However, due to constraints such as trade secrets, privacy regulations, and competitive relationships, institutions cannot directly exchange or centrally compare customer identifiers (such as ID card numbers and unified social credit codes). If traditional hash-based or encryption-based methods are used to directly exchange identifiers, although the original ID plaintext can be protected, hash collision risks and encryption key management challenges remain. More importantly, participating parties may use the obtained ciphertext or hash value to launch reverse collision attacks, thereby inferring the customer identity information of other institutions. This privacy leakage risk, which exists in the initial stage of data fusion, has become the primary technical obstacle to the secure release of the value of financial data. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method and system for secure processing of commercial and financial big data. It achieves secure alignment and collaborative computation of financial data in encrypted form throughout the entire process. Through a decentralized architecture and hardware acceleration, it significantly improves processing efficiency while ensuring privacy, providing a reliable and efficient solution for cross-institutional data fusion.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for secure processing of commercial financial big data, comprising the following steps: S1, receive uploaded data from multiple data provider nodes. The uploaded data includes a local feature vector encrypted using a joint public key based on a threshold homomorphic encryption mechanism, and an anonymous identifier of a commercial financial entity corresponding to the feature vector sample, encrypted using the same joint public key. S2, without decrypting any encrypted anonymous identifiers, perform a predetermined homomorphic operation on any two encrypted anonymous identifiers Enc(ID_i) and Enc(ID_j) from different nodes to generate an encrypted matching decision factor Enc(F_ij); S3, the encrypted matching decision factor Enc(F_ij) is sent to a designated security adjudication node; the security adjudication node uses its share of the joint private key to perform a partial decryption operation on Enc(F_ij), obtains a result, and generates a binary judgment result of "match" or "not match" based on whether the result is a predetermined value; during this process, the security adjudication node cannot know the specific value of ID_i or ID_j; S4. Based on all matching results, identify the set of encrypted samples representing the same commercial and financial entity. For each set, concatenate the encrypted feature vectors from different data provider nodes in the ciphertext domain to form an encrypted joint feature vector, and summarize the encrypted joint feature vectors of all entities to generate an encrypted joint feature matrix. S5 distributes the encrypted joint feature matrix and encrypted model parameters to one or more secure computing nodes. Each secure computing node executes the specified computational subtask in parallel in the encrypted state, generates encrypted intermediate results and returns them. S6. Homomorphically aggregate the returned intermediate encrypted results to form the final encrypted result; coordinate multiple data provider nodes holding a share of the joint private key to partially decrypt the final encrypted result under the condition of meeting a preset threshold number, and finally synthesize the plaintext aggregated result. S7 securely distributes the plaintext aggregation results or the updated parameters of the model generated based on them to each data provider node.

[0005] Preferably, the predetermined homomorphic operation in step S2 and the judgment logic in step S3 are as follows: In step S2, the central coordination node locally generates a random blinding factor r and a random offset s, where s≠0, and calculates: Enc(F_ij)=Enc(r*(ID_i-ID_j)+s); In step S3, after the security adjudication node decrypts, it obtains F_ij=r*(ID_i-ID_j)+s, and determines: if F_ij=s, then ID_i and ID_j are the same, and "match" is returned; otherwise, "mismatch" is returned.

[0006] Preferably, the threshold homomorphic encryption mechanism is as follows: the joint public key is generated jointly by all data provider nodes; the corresponding joint private key is divided into n shares through secret sharing technology, of which n-1 shares are distributed to each data provider node, and the remaining 1 share is distributed to the security adjudication node; in step S6, at least t shares of data provider nodes and shares of security adjudication node are required to participate in decryption, where t < n.

[0007] Preferably, in step S5, the specified computational subtask is logistic regression or forward propagation computation of a neural network; the secure computing node is equipped with a dedicated hardware acceleration card that supports the CKKS homomorphic encryption scheme and SIMD operations, for performing parallel batch processing operations on the packaged and encoded ciphertext data.

[0008] Preferably, before step S1, each data provider node adds Laplace noise or Gaussian noise that meets differential privacy requirements to the sensitive feature dimensions in its local feature vector that conform to preset rules, and then encrypts and uploads them.

[0009] Preferably, after step S7, the central coordinating node uploads the task identifier of this round of processing, the anonymized information of the participating nodes, and the digital fingerprint of the plaintext aggregation result to the blockchain network for immutable evidence storage via a smart contract.

[0010] This invention also discloses a commercial financial big data security processing system, comprising: Multiple data provider nodes are used to generate and encrypt local signature data and anonymous identifiers; The central coordinating node, which communicates with the data provider nodes, includes: The ciphertext alignment module executes steps S2 and S3 and includes a random number generator, a homomorphic operator, and a secure adjudication communication interface. The task scheduling and matrix assembly module executes step S4 and assembles the encrypted joint feature matrix based on the matching results; The secure computing scheduling module executes step S5 and distributes the encrypted computing task to external computing resources; The secure adjudication node is deployed independently of the data provider node. It includes a joint private key share storage unit and an adjudication logic execution unit, performs the decryption and judgment operations in step S3, and communicates with the secure adjudication communication interface of the ciphertext alignment module. The secure computing cluster, which communicates with the secure computing scheduling module, consists of multiple computing nodes equipped with homomorphic encryption hardware acceleration cards and executes the encrypted computing tasks distributed by the secure computing scheduling module. The distributed key management service communicates with data provider nodes, central coordinating nodes, security adjudication nodes, and secure computing clusters to generate and maintain joint key pairs based on threshold homomorphic encryption mechanisms, and manage the distribution and verification of private key shares.

[0011] Preferably, the secure computing scheduling module allocates computing subtasks using a dynamic load balancing algorithm based on the size of the encrypted joint feature matrix and the real-time load of each computing node in the secure computing cluster; the communication channel between the adjudication logic execution unit of the secure adjudication node and the ciphertext alignment module of the central coordination node is encrypted using a transport layer security protocol.

[0012] This invention provides a method and system for secure processing of commercial financial big data. Compared with existing technologies, it has the following advantages: 1. By deeply integrating homomorphic encryption, threshold secret sharing, and differential privacy technology, a complete privacy computing chain is constructed, from encrypted data upload, ciphertext alignment and matching, secure collaborative computation to threshold decryption. In this process, the original data of each participant always exists in ciphertext form or is securely segmented. The aggregated result can only be decrypted and aggregated after final authorization, which effectively prevents the risk of data leakage in each stage of transmission, alignment, computation, and decryption, and meets the strict requirements of data sovereignty and privacy compliance in the commercial and financial field.

[0013] 2. Through modular design and physical isolation, the system decomposes core functions such as key management, privacy adjudication, and ciphertext computation to independent trusted nodes, enforcing the principles of separation of privileges and least privilege. In particular, the introduction of independent secure adjudication nodes and threshold decryption mechanisms ensures that critical operations must be completed collaboratively by multiple parties. No single entity (including the central coordinating node) can independently decrypt data or tamper with the adjudication logic. This eliminates single points of failure and the risk of collusive attacks at the mechanism level, greatly improving the overall security and anti-attack capabilities of the system.

[0014] 3. To address the bottleneck of high computational overhead in homomorphic encryption, this invention combines a dedicated hardware acceleration card supporting the CKKS scheme and SIMD instructions to achieve parallel batch processing of ciphertext data. Simultaneously, a dynamic load balancing algorithm optimizes task distribution, reducing the ciphertext computation time for large-scale financial data models from hours to minutes. This enables the system to scale linearly to meet the demands of massive data processing, transforming the theoretical advantages of cutting-edge cryptographic technology into a practical, high-performance engineering solution, providing reliable support for scenarios with high real-time requirements such as financial risk control and precision marketing. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system architecture block diagram of the present invention.

[0016] In the diagram: 100, Data Provider Node; 200, Central Coordination Node; 210, Ciphertext Alignment Module; 211, Random Number Generator; 212, Homomorphic Calculator; 213, Secure Adjudication Communication Interface; 220, Task Scheduling and Matrix Assembly Module; 230, Secure Computation Scheduling Module; 300, Secure Adjudication Node; 310, Joint Private Key Share Storage Unit; 320, Adjudication Logic Execution Unit; 400, Secure Computation Cluster; 410, Computation Node; 500, Distributed Key Management Service. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 - Figure 2 This invention provides a technical solution: a method for secure processing of commercial financial big data, comprising the following steps: S1, receive uploaded data from multiple data provider nodes. The uploaded data includes a local feature vector encrypted using a joint public key based on a threshold homomorphic encryption mechanism, and an anonymous identifier of a commercial financial entity corresponding to the feature vector sample, encrypted using the same joint public key. S2, without decrypting any encrypted anonymous identifiers, perform a predetermined homomorphic operation on any two encrypted anonymous identifiers Enc(ID_i) and Enc(ID_j) from different nodes to generate an encrypted matching decision factor Enc(F_ij); S3, the encrypted matching decision factor Enc(F_ij) is sent to a designated security adjudication node; the security adjudication node uses its share of the joint private key to perform a partial decryption operation on Enc(F_ij), obtains a result, and generates a binary judgment result of "match" or "not match" based on whether the result is a predetermined value; during this process, the security adjudication node cannot know the specific value of ID_i or ID_j; S4. Based on all matching results, identify the set of encrypted samples representing the same commercial and financial entity. For each set, concatenate the encrypted feature vectors from different data provider nodes in the ciphertext domain to form an encrypted joint feature vector, and summarize the encrypted joint feature vectors of all entities to generate an encrypted joint feature matrix. S5 distributes the encrypted joint feature matrix and encrypted model parameters to one or more secure computing nodes. Each secure computing node executes the specified computational subtask in parallel in the encrypted state, generates encrypted intermediate results and returns them. S6. Homomorphically aggregate the returned intermediate encrypted results to form the final encrypted result; coordinate multiple data provider nodes holding a share of the joint private key to partially decrypt the final encrypted result under the condition of meeting a preset threshold number, and finally synthesize the plaintext aggregated result. S7 securely distributes the plaintext aggregation results or the updated parameters of the model generated based on them to each data provider node.

[0019] In this implementation plan, a chain-based privacy computing paradigm of "ciphertext alignment—cooperative adjudication—secure computation—threshold decryption" is constructed to ensure that commercial and financial data is "usable but not visible" and "process auditable" throughout the entire process. The logic of this method begins with each data provider node using a distributedly generated joint public key to homomorphically encrypt its local feature vector and the corresponding anonymous identifier of the commercial entity before uploading it, thus achieving data privacy from the source. The key is that after receiving the encrypted data, the central coordinating node does not need to decrypt it, but instead constructs an encrypted matching decision factor through precise homomorphic operations (e.g., calculating Enc(r*(ID_i-ID_j)+s)) and sends this factor to the independently deployed secure adjudication node. The secure adjudication node performs partial decryption using its share of the private key, only obtaining a randomized and blinded intermediate result. It outputs a binary conclusion of "match" or "not match" by judging whether the result is equal to a preset value (such as s), without being able to infer any original identifier information, thereby achieving cross-institutional sample alignment while protecting privacy. Based on the alignment results, the system automatically concatenates features from different data providers in the ciphertext domain to form an encrypted joint feature matrix, preparing input for subsequent modeling. Subsequently, this matrix and encrypted model parameters are distributed to a secure computing node cluster equipped with dedicated hardware. These nodes perform calculations such as logistic regression or neural network forward propagation directly on the ciphertext in a parallel, batch processing manner, generating encrypted intermediate results. Finally, the system homomorphically aggregates these results and coordinates multiple data provider nodes holding private key shares with security adjudication nodes to collaboratively complete threshold decryption under a preset threshold, ultimately obtaining the plaintext aggregation result or model update parameters and securely distributing them. This method features a tightly integrated logic, employing multiple technologies such as homomorphic encryption, threshold secret sharing, differential privacy (optional), and blockchain notarization (optional). It not only achieves high-strength privacy protection at each stage of data processing, alignment, computation, and decryption but also ensures process reliability and anti-attack capabilities through a distributed collaborative mechanism. This provides a systematic and verifiable solution for the secure integration and value mining of cross-institutional commercial and financial big data.

[0020] Specifically, the pre-defined homomorphic operation in step S2 and the judgment logic in step S3 are as follows: In step S2, the central coordination node locally generates a random blinding factor r and a random offset s, where s≠0, and calculates: Enc(F_ij)=Enc(r*(ID_i-ID_j)+s); In step S3, after the security adjudication node decrypts, it obtains F_ij=r*(ID_i-ID_j)+s, and determines: if F_ij=s, then ID_i and ID_j are the same, and "match" is returned; otherwise, "mismatch" is returned.

[0021] In this embodiment, the described ciphertext equivalence comparison algorithm is based on homomorphic encryption randomization masking technology. Its core principle lies in constructing a mathematical structure with information hiding properties. The random blinding factor r generated by the central coordinating node provides unpredictable randomness to the linear transformation, ensuring that even if the secure adjudicating node can decrypt Enc(F_ij), it cannot deduce the original identifier difference (ID_i-ID_j) from the decryption result F_ij. This is because r, as an unknown multiplier, nonlinearly perturbs the difference. At the same time, the introduction of the non-zero random offset s cleverly transforms the judgment logic from "whether the result is zero" to "whether the result is equal to a known random value s". This design allows the secure adjudicating node to perform an equivalence comparison once with the secret value s it holds, which is synchronized from the central coordinating node (or the promised value obtained in advance through the secure channel), after decryption, and output a "match" or "mismatch" judgment, completely isolating it from the possibility of contacting or deducing any original identifiers ID_i and ID_j. This dual protection mechanism of "blinding + offset" achieves perfect privacy separation in cryptography—the central coordinating node has access to the random parameters but cannot see the plaintext results, while the security adjudication node can see the calculation results but cannot associate them with the original identifier, thus securely completing the cross-organizational ciphertext identity alignment in a distributed environment. Functionally, this algorithm achieves precise equivalence determination in encrypted states, providing a reliable foundation for subsequent feature fusion, and the determination results are cryptographically verifiable. In terms of security, the algorithm implements information theory-based privacy protection; even if the security adjudication node is maliciously attacked or colludes with other nodes, it cannot leak any information about ID_i or ID_j through F_ij (unless the difference is exactly zero, but the probability is negligible), greatly enhancing the system's ability to resist internal threats. In terms of efficiency, the scheme requires only one homomorphic multiplication and one homomorphic addition to generate the determination factor, with controllable computational overhead, and avoids complex multi-party secure computation protocols. This makes the privacy alignment process both highly secure and practically feasible, providing an efficient and reliable technical solution for large-scale cross-entity data association in the commercial and financial fields.

[0022] Specifically, the threshold homomorphic encryption mechanism works as follows: the joint public key is generated jointly by all data provider nodes; the corresponding joint private key is divided into n shares through secret sharing technology, of which n-1 shares are distributed to each data provider node, and the remaining 1 share is distributed to the security adjudication node; in step S6, at least t shares of data provider nodes and shares of the security adjudication node are required to participate in decryption, where t < n.

[0023] In this embodiment, in the key generation phase, all data provider nodes jointly participate in the distributed key generation protocol, enabling the generation of a joint public-private key pair through multi-party collaboration, thus ensuring that no single node (including the central coordination node) can obtain the complete private key, eliminating the risk of key leakage or being controlled by a single entity from the source; the private key splitting scheme splits the private key into n shares, distributes n - 1 shares to each data party, and leaves the remaining 1 share with an independent security adjudication node. This distribution strategy cleverly disperses the key control among all business participants (data providers) and a neutral arbitration party (security adjudication node); in the decryption phase, a decryption threshold mechanism is designed, requiring that at least t data provider shares and the security adjudication node share must jointly participate to successfully decrypt. This means that the decryption operation not only requires authorization from the majority of data parties (satisfying the condition of t < n), but also must obtain the collaboration of the neutral adjudication party, and no party can decrypt alone or in combination with a minority. In terms of security, it achieves decentralized trust, and by mathematical means, it enforces multi-party collaboration, effectively preventing systemic risks caused by insiders maliciously decrypting data or external attackers breaking a single node. Even if up to n - t - 1 shares are lost or leaked, the system can still remain secure; in terms of permission control, this design deeply binds the business logic and security logic, making the key decryption operation naturally require the majority consent of data providers (data owners) and the review of the security adjudication node (independent supervision party), providing a two-layer authorization guarantee for data usage, and strengthening data governance and compliance; in terms of availability and fault tolerance, the threshold mechanism allows the system to still be able to complete necessary operations by other nodes that meet the threshold when some nodes are offline or refuse to participate, improving the robustness of the system and the flexibility of actual deployment, providing a reliable security infrastructure for common dynamic member scenarios in commercial financial alliances.

[0024] Specifically, in step S5, the specified computational subtask is the forward propagation calculation of logistic regression or neural network; the secure computing node is equipped with a dedicated hardware acceleration card that supports the CKKS homomorphic encryption scheme and SIMD operations, and is used to perform parallel batch processing operations on the packed and encoded ciphertext data.

[0025] In this embodiment, for typical numerical features (such as amount, interest rate, statistical indicators) and models (such as logistic regression, neural networks) in commercial finance scenarios, the CKKS homomorphic encryption scheme, which supports floating-point approximation operations, is selected. This scheme encodes floating-point vectors into complex polynomials, allowing key operations such as addition, multiplication, and polynomial approximation to be performed directly in the ciphertext state, thus completely preserving the arithmetic logic required for forward propagation. The dedicated hardware acceleration card, through its built-in SIMD (Single Instruction Multiple Data) architecture, can maximize the "packaging encoding" feature of the CKKS scheme—a single instruction can simultaneously perform the same operation on hundreds or thousands of ciphertext data slots, realizing batch and parallel processing of the encrypted joint feature matrix. The computation task scheduler breaks down large matrix calculations into multiple parallelizable subtasks (such as calculations of different batches of samples or calculations of different network layers), which are executed simultaneously by multiple acceleration card nodes, and finally complete the global calculation through homomorphic aggregation. In terms of computational efficiency, hardware-level parallelization reduces the computational overhead of homomorphic encryption by 1-2 orders of magnitude, shortening the computation of ciphertext models that might have taken hours to minutes. This breaks through the application bottleneck of privacy computing in scenarios such as real-time financial risk control and precision marketing. In terms of accuracy and functionality, the CKKS solution's excellent support for floating-point numbers ensures the computational accuracy of financial models (especially logistic regression involving probabilistic outputs and neural networks with complex nonlinear transformations). This keeps the error between ciphertext and plaintext computation results within an acceptable small range, ensuring the effectiveness of joint modeling. In terms of system scalability, the architecture based on hardware acceleration cards and parallel task scheduling allows the system to cope with the ever-increasing data scale and model complexity by linearly adding computing nodes, providing a scalable technical foundation for the continuous evolution of future commercial and financial big data analysis.

[0026] Specifically, before step S1, each data provider node adds Laplace noise or Gaussian noise that meets differential privacy requirements to the sensitive feature dimensions in its local feature vector that conform to preset rules, and then encrypts and uploads them.

[0027] In this embodiment, based on the strict mathematical definition of differential privacy, that is, under two adjacent scenarios where the data of any individual exists in the dataset, the probability distribution difference of the algorithm output is controlled within an exponential bound with ε (privacy budget) as the parameter. In specific implementation, before encrypting the feature vector locally, the data provider node will identify the dimensions that need to be protected (such as personal annual income, corporate debt ratio) according to preset rules (such as feature type, business sensitivity), and inject strictly calibrated random noise into these dimensions—Laplace noise is used for count-type features or features with large numerical ranges, and Gaussian noise is used for continuous features that are friendly to Gaussian distributions. The scale parameter of the noise (such as Δf / ε of the Laplace distribution) is accurately calculated based on the global sensitivity of the feature (i.e., the maximum change that a single record can cause) and the preset privacy budget ε, ensuring that sufficient privacy protection is provided while minimizing the impact on data availability.

[0028] Specifically, after step S7, the central coordinating node uploads the task identifier of this round of processing, the anonymized information of the participating nodes, and the digital fingerprint of the plaintext aggregation result to the blockchain network for immutable evidence storage via a smart contract.

[0029] In this embodiment, the blockchain evidence storage mechanism deeply integrates the immutability of distributed ledger technology with the automatic execution capability of smart contracts, constructing a trusted audit and traceability layer for the privacy computing process. After each round of joint computation, the central coordinating node submits key process metadata—including unique task identifiers (used to track specific modeling tasks), anonymized identifier hashes of participating nodes (recording contributors while protecting node identity privacy), and the cryptographic hash value of the final plaintext aggregation result (serving as a unique digital fingerprint of the data output)—as structured transaction data to the permissioned blockchain network. After being verified by consensus nodes, this data is encapsulated into new blocks and permanently linked to previous blocks via chained cryptographic hash pointers, forming a historical record that cannot be unilaterally modified or deleted by anyone (including system participants). The smart contract, as the carrier of the evidence storage logic, automatically performs operations such as data format verification, timestamp appending, and evidence storage status updates, ensuring the automation and non-repudiation of the entire process.

[0030] This invention also discloses a commercial financial big data security processing system, comprising: Multiple data provider nodes 100 are used to generate and encrypt local feature data and anonymous identifiers; Central coordination node 200, which is communicatively connected to data provider node 100, includes: The ciphertext alignment module 210 executes steps S2 and S3 and includes a random number generator 211, a homomorphic operator 212, and a secure adjudication communication interface 213. The task scheduling and matrix assembly module 220 executes step S4 and assembles the encrypted joint feature matrix according to the matching results; The secure computing scheduling module 230 executes step S5 and distributes the encrypted computing task to external computing resources; The security adjudication node 300 is deployed independently of the data provider node 100. It includes a joint private key share storage unit 310 and an adjudication logic execution unit 320. It performs the decryption and judgment operations in step S3 and communicates with the security adjudication communication interface 213 of the ciphertext alignment module 210. The secure computing cluster 400 is communicatively connected to the secure computing scheduling module 230 and consists of multiple computing nodes 410 equipped with homomorphic encryption hardware acceleration cards. It executes the encrypted computing tasks distributed by the secure computing scheduling module 230. The distributed key management service 500 communicates with the data provider node 100, the central coordinating node 200, the security adjudication node 300, and the secure computing cluster 400 to generate and maintain joint key pairs based on threshold homomorphic encryption and to manage the distribution and verification of private key shares.

[0031] In this embodiment, the data provider node 100 constitutes the system's trust starting point and data sovereignty boundary, ensuring that the original data is encrypted locally before leaving the domain. The central coordination node 200 serves as the process hub, with its internal ciphertext alignment module 210, task scheduling and matrix assembly module 220, and secure computation scheduling module 230 forming a pipeline to sequentially complete the secure alignment of ciphertext samples, the logical assembly of feature matrices, and the distribution and scheduling of computation tasks. However, this node itself does not possess any decryption capabilities, thus achieving the separation of control flow and data decryption permissions. The independently deployed security adjudication node 300 serves as a key "privacy arbitrator" in the system, specifically responsible for performing highly sensitive ciphertext equivalence determination operations. Its physical isolation ensures that the core adjudication logic's operating environment is protected from potential interference or attacks from other business modules; the secure computing cluster 400, as a high-performance computing resource pool, provides scalable and efficient encrypted arithmetic operations through hardware acceleration cards; while the distributed key management service 500, which runs throughout the entire system, serves as a cornerstone of trust, providing unified and secure key lifecycle management for all participants in a service-oriented manner; the entire system's working logic follows a strict order of "alignment first, computation second, and collaborative decryption," ensuring that data is in an encrypted state or under the protection of securely segmented keys at every stage of matching, fusion, and computation, and threshold decryption can only be triggered through multi-party collaboration at the final aggregation result stage; In terms of security, the architecture enforces the principle of least privilege and separation of duties through modular isolation. Even if the central coordinating node 200 is compromised, attackers cannot obtain the decryption key or tamper with the security adjudication logic. The independent security adjudication node 300 and distributed key management 500 further prevent single points of failure or internal collusion risks, constructing a defense-in-depth system. Regarding performance and scalability, the combination of the dedicated hardware-accelerated computing cluster 400 and the task scheduling capabilities of the central coordinating node enables the system to scale linearly to meet the computational demands of massive amounts of financial data, transforming the computational overhead of homomorphic encryption from a theoretical bottleneck into a manageable cost. In terms of reliability and maintainability, clear module boundaries allow individual components (such as hardware acceleration card drivers and key management algorithms) to be independently upgraded, replaced, or expanded without refactoring the entire system, greatly improving the system's operational stability and the flexibility of technological iteration. In terms of compliance and deployment flexibility, the architecture supports the independent operation of the 300 secure adjudication nodes and the 500 key management services by trusted third parties. It can flexibly adapt to the trust models and data compliance requirements of different business cooperation scenarios (such as meeting the data localization regulations of different jurisdictions), providing a solid and universal infrastructure platform for building cross-institutional, auditable, and high-performance commercial and financial data collaboration alliances.

[0032] Specifically, the secure computing scheduling module 230 allocates computing subtasks using a dynamic load balancing algorithm based on the size of the encrypted joint feature matrix and the real-time load of each computing node 410 in the secure computing cluster 400; the communication channel between the adjudication logic execution unit 320 of the secure adjudication node 300 and the ciphertext alignment module 210 of the central coordination node 200 is encrypted using a transport layer security protocol.

[0033] In this embodiment, the dynamic load balancing algorithm performs real-time analysis of the computational complexity of the encrypted joint feature matrix (such as the number of rows and columns, and the number of model layers), and combines it with multi-dimensional indicators such as the current working status of each computing node 410 in the secure computing cluster 400 (such as CPU / accelerator card utilization, memory usage, and task queue length). It then uses heuristic or predictive scheduling strategies (such as minimum load priority and latency-aware scheduling) to intelligently decompose and allocate computational subtasks to the most suitable nodes, ensuring full utilization of the cluster's overall computing resources and avoiding performance bottlenecks where some nodes are overloaded while others are idle. Simultaneously, the Transport Layer Security (TLS / SSL) protocol establishes an end-to-end encrypted tunnel between the secure adjudication node 300 and the central coordinating node 200. This protocol uses asymmetric encryption for authentication and key exchange, followed by high-strength encryption of all data in transmission (including encrypted matching factors, adjudication results, etc.) using a symmetric encryption algorithm. Combined with Message Authentication Code (MAC), it prevents data from being tampered with or replayed during transmission, thus providing communication layer security for the most critical privacy judgment interaction in a complex network environment. In terms of performance and efficiency, dynamic load balancing enables near-optimal parallelization of large-scale encrypted computation tasks in a distributed cluster. This not only significantly reduces the total time required for a single joint modeling session but also improves the return on investment for hardware resources by avoiding local hotspots, allowing the system to economically and efficiently handle the ever-growing massive datasets in commercial and financial settings. Regarding security and reliability, the TLS encrypted channel eliminates the risk of sensitive information leakage due to network eavesdropping or man-in-the-middle attacks (for example, even if an attacker intercepts channel data packets, they cannot obtain the decryptable matching decision factor), ensuring the confidentiality, integrity, and authenticity of communication in the core privacy alignment process. Simultaneously, the dynamic scheduling strategy itself enhances the system's fault tolerance. When a computing node 410 fails, the scheduling module automatically reassigns unfinished tasks to other healthy nodes, ensuring the stable operation of long-term, multi-round joint learning tasks.

[0034] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0035] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for securely processing commercial and financial big data, characterized in that, Includes the following steps: S1, receive uploaded data from multiple data provider nodes. The uploaded data includes a local feature vector encrypted using a joint public key based on a threshold homomorphic encryption mechanism, and an anonymous identifier of a commercial financial entity corresponding to the feature vector sample, encrypted using the same joint public key. S2, without decrypting any encrypted anonymous identifiers, perform a predetermined homomorphic operation on any two encrypted anonymous identifiers Enc(ID_i) and Enc(ID_j) from different nodes to generate an encrypted matching decision factor Enc(F_ij); S3, send the encrypted matching decision factor Enc(F_ij) to a designated security adjudication node; the security adjudication node uses its share of the joint private key to perform a partial decryption operation on Enc(F_ij), obtains a result, and generates a binary judgment result of "match" or "not match" based on whether the result is a predetermined value; during this process, the security adjudication node cannot know the specific value of ID_i or ID_j; S4. Based on all matching results, identify the set of encrypted samples representing the same commercial and financial entity. For each set, concatenate the encrypted feature vectors from different data provider nodes in the ciphertext domain to form an encrypted joint feature vector, and summarize the encrypted joint feature vectors of all entities to generate an encrypted joint feature matrix. S5 distributes the encrypted joint feature matrix and encrypted model parameters to one or more secure computing nodes. Each secure computing node executes the specified computational subtask in parallel in the encrypted state, generates encrypted intermediate results and returns them. S6. Homomorphically aggregate the returned intermediate encrypted results to form the final encrypted result; coordinate multiple data provider nodes holding a share of the joint private key to partially decrypt the final encrypted result under the condition of meeting a preset threshold number, and finally synthesize the plaintext aggregated result. S7 securely distributes the plaintext aggregation results or the updated parameters of the model generated based on them to each data provider node.

2. A method for securely processing commercial and financial big data according to claim 1, characterized in that, The pre-defined homomorphic operation in step S2 and the judgment logic in step S3 are as follows: In step S2, the central coordination node locally generates a random blinding factor r and a random offset s, where s≠0, and calculates: Enc(F_ij)=Enc(r*(ID_i-ID_j)+s); In step S3, after the security adjudication node decrypts, it obtains F_ij=r*(ID_i-ID_j)+s, and determines: if F_ij=s, then ID_i and ID_j are the same, and "match" is returned; otherwise, "mismatch" is returned.

3. A method for secure processing of commercial financial big data according to claim 1, characterized in that: The threshold homomorphic encryption mechanism is as follows: the joint public key is generated jointly by all data provider nodes; the corresponding joint private key is divided into n shares through secret sharing technology, of which n-1 shares are distributed to each data provider node, and the remaining 1 share is distributed to the security adjudication node; in step S6, at least t shares of data provider nodes and shares of security adjudication node are required to participate in decryption, where t < n.

4. A method for securely processing commercial financial big data according to claim 1, characterized in that: In step S5, the specified computational subtask is logistic regression or forward propagation computation of a neural network; the secure computing node is equipped with a dedicated hardware acceleration card that supports the CKKS homomorphic encryption scheme and SIMD operations, which is used to perform parallel batch processing operations on the packaged and encoded ciphertext data.

5. A method for securely processing commercial financial big data according to claim 1, characterized in that: Before step S1, each data provider node adds Laplace noise or Gaussian noise that meets differential privacy requirements to the sensitive feature dimensions in its local feature vector that conform to preset rules, and then encrypts and uploads them.

6. A method for securely processing commercial financial big data according to claim 1, characterized in that: After step S7, the central coordinating node uploads the task identifier of this round of processing, the anonymized information of participating nodes, and the digital fingerprint of the plaintext aggregation result to the blockchain network for immutable evidence storage via a smart contract.

7. A commercial financial big data security processing system, characterized in that: A commercial financial big data security processing method according to any one of claims 1-6 includes: Multiple data provider nodes (100) are used to generate and encrypt local feature data and anonymous identifiers; A central coordination node (200), which is communicatively connected to a data provider node (100), includes: The ciphertext alignment module (210) executes steps S2 and S3 and includes a random number generator (211), a homomorphic operator (212), and a secure adjudication communication interface (213). The task scheduling and matrix assembly module (220) executes step S4 and assembles the encrypted joint feature matrix according to the matching results; The secure computing scheduling module (230) executes step S5 and distributes the ciphertext computing task to external computing resources; The security adjudication node (300), deployed independently of the data provider node (100), includes a joint private key share storage unit (310) and an adjudication logic execution unit (320), performs the decryption and judgment operations in step S3, and communicates with the security adjudication communication interface (213) of the ciphertext alignment module (210); The secure computing cluster (400) is connected to the secure computing scheduling module (230) and consists of multiple computing nodes (410) equipped with homomorphic encryption hardware acceleration cards. It executes the encrypted computing tasks distributed by the secure computing scheduling module (230). The distributed key management service (500) communicates with the data provider node (100), the central coordinating node (200), the security adjudication node (300), and the secure computing cluster (400) to generate and maintain joint key pairs based on threshold homomorphic encryption and to manage the distribution and verification of private key shares.

8. A commercial financial big data security processing system according to claim 7, characterized in that: The secure computing scheduling module (230) allocates computing subtasks using a dynamic load balancing algorithm based on the size of the encrypted joint feature matrix and the real-time load of each computing node (410) in the secure computing cluster (400). The communication channel between the adjudication logic execution unit (320) of the secure adjudication node (300) and the ciphertext alignment module (210) of the central coordination node (200) is encrypted using a transport layer security protocol.

Images