Cross-border trade data processing method and system based on smart port

Abstract

The invention relates to the field of intelligent data processing, and particularly discloses a cross-border trade data processing method and system based on an intelligent port, and the method comprises the steps: processing original data in a local trusted execution environment of each port node, and generating a privacy protection feature vector, generating a verifiable voucher for verifying the calculation compliance by using zero-knowledge proof; verifying the voucher through a distributed consensus network, and fusing each node feature vector by adopting secure multi-party calculation to generate a cross-domain joint feature; the features are input into a decentralized agent network for asynchronous negotiation, and a collaborative risk research and judgment report is output; each node finally optimizes a local model under differential privacy protection according to report feedback; the system comprises four corresponding modules. According to the method, credible collaboration and deep risk insight on the premise that the original data is not out of the domain are realized, and continuous autonomous evolution of the system is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of intelligent data processing technology, and in particular to a method and system for cross-border trade data processing based on smart ports. Background Technology

[0002] In recent years, the construction of smart ports has become crucial for improving the efficiency and security of cross-border trade. Existing technologies mainly focus on data automation and process optimization within a single port, such as automatically entering customs declarations through optical character recognition (OCR) technology or using IoT sensors to track cargo status. However, existing solutions face fundamental bottlenecks when dealing with complex regulatory scenarios requiring coordination among multiple ports and regulatory departments.

[0003] The core contradiction lies in the conflict between the rigid requirements of data sovereignty and the need for cross-border collaborative supervision. On the one hand, regulations on local data storage in various countries are becoming increasingly stringent, and original trade data should, in principle, not flow across borders; on the other hand, high-risk activities such as smuggling and document laundering often exhibit cross-customs area linkages, urgently requiring collaborative data analysis to identify hidden risks. Existing mainstream technological approaches are insufficient to adequately resolve this contradiction. Centralized data pool solution: This requires each port to upload data to a central server for analysis. Although it can achieve collaboration, it directly violates the principle of data sovereignty and poses significant compliance risks, making it difficult to implement in practice.

[0004] Traditional federated learning schemes, while avoiding the concentration of raw datasets, may still implicitly leak sensitive information through gradient exchanges during model training, and lack an effective verification mechanism for the compliance of data preprocessing by participating parties, resulting in a weak foundation of trust.

[0005] Simple data anonymization and exchange schemes: Data is shared after processing through hashing, generalization and other methods. This results in significant loss of data value, makes it difficult to support complex correlation risk analysis, and cannot prove the reliability and consistency of the anonymization process.

[0006] Therefore, existing technologies either sacrifice compliance or analytical efficiency, resulting in smart port systems remaining in a state of data silos for a long time, making it impossible to form a cross-domain regulatory synergy.

[0007] There is an urgent need for a new method for collaborative processing of cross-border trade data that can achieve data availability without visibility, process credibility and auditability, and system learning and evolution within a legal framework. Summary of the Invention

[0008] The purpose of this invention is to solve the problem of cross-domain regulatory collaboration under data sovereignty constraints in the prior art, and to propose a cross-border trade data processing method and system based on smart ports.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A method for processing cross-border trade data based on smart ports, comprising: S1. Process the raw data in the local trusted environment of each node to generate privacy-preserving feature vectors and corresponding verifiable credentials; S2. Verify the verifiable credential in the distributed network and fuse the privacy-preserving feature vectors of each node to generate cross-domain joint features; S3. Input the cross-domain joint features into the decentralized intelligent agent network for asynchronous negotiation, and output a collaborative risk assessment report; S4. Based on the collaborative risk assessment report, each node generates and applies an optimized gradient locally to update the feature extraction model.

[0010] As a further technical solution of the present invention, S1 specifically includes: S11. In a trusted execution environment, the original cross-border trade data is parsed to extract structured fields including at least commodity description, declared value, logistics entity, and consignor / consignee; and the values ​​of each field are mapped to a predefined standard terminology dictionary to generate standardized structured data records. S12. Input the structured data record into a preset feature extraction model. The model maps the input into a fixed-length, desensitized numerical vector through a multi-layer perceptual network and attention mechanism, which serves as a privacy-preserving feature vector. S13. Based on the zero-knowledge proof method, generate verifiable credentials for the privacy-preserving feature vector to verify its computational integrity and compliance.

[0011] As a further technical solution of the present invention, S13 specifically includes: S131. Based on the compliance check record of the feature extraction model, the definition of the predetermined format and range, and the calculation logic of the model, construct an arithmetic circuit. The public input of the arithmetic circuit is the cryptographic commitment of the privacy-preserving feature vector, and the secret input is the structured data record and the intermediate variables of the model. The arithmetic circuit is satisfied if and only if the secret input data conforms to the predetermined format and range and the model calculation process is correct. S132. Within the trusted execution environment, using the structured data records and secret variables in the model calculation process as witnesses, the proof generation algorithm of the zero-knowledge proof method is run to output a zero-knowledge proof that proves the arithmetic circuit is correctly satisfied. S133. The zero-knowledge proof, the cryptographic commitment of the privacy-preserving feature vector, and the public parameters used for verification are jointly encapsulated to generate the verifiable credential, and then bound to the corresponding privacy-preserving feature vector for output.

[0012] As a further technical solution of the present invention, S2 specifically includes: S21. Each node broadcasts its own privacy-preserving feature vector and its bound verifiable credentials as a transaction to the distributed consensus network through the trusted data delivery protocol. The trusted data delivery protocol uses digital signatures to ensure the credibility of the publishing node's identity and the immutability of the data. S22. After receiving multiple related transactions, the collaborative nodes in the distributed consensus network perform parallel verification of all verifiable credentials based on a preset zero-knowledge proof verification contract. Only when all credentials have been verified successfully will the corresponding privacy-preserving feature vector be marked as valid data to be fused. S23. The collaborative node will mark all privacy-preserving feature vectors as valid, input a preset secure multi-party computation fusion function, maintain the encrypted or secret-sharing state of each vector during the computation process, and finally output a cross-domain joint feature that fuses the features of all nodes.

[0013] As a further technical solution of the present invention, S23 specifically includes: S231. Each node uses a secret sharing algorithm to split the effective privacy-preserving feature vector into encrypted shares and distribute them to all collaborating nodes; S232. All collaborating nodes locally based on their held share sets. The secure multi-party computation fusion function is executed, specifically using a weighted average fusion method: Set nodes The weight is Joint characteristics of weighted average for: ,in: For privacy-preserving feature vectors, The number of collaborating nodes; node Calculate local results on a share : ,in: For plaintext shares stored locally; Ultimately, each node A result share of generating cross-domain joint features ,satisfy ; S233. The nodes aggregate the result shares through the security protocol, jointly recover and output the complete plaintext cross-domain joint characteristics.

[0014] As a further technical solution of the present invention, S3 specifically includes: S31. Input the cross-domain joint features into a pre-trained decentralized risk assessment agent network. Multiple dedicated agents in the network correspond to different risk assessment dimensions. Each agent performs reasoning asynchronously and independently based on the cross-domain joint features to generate preliminary risk assessment sub-results for its respective dimension. S32. Each agent exchanges its preliminary risk assessment sub-results and conducts multiple rounds of negotiation through the network's built-in communication channel based on predetermined rules. The negotiation aims to reach a consensus on the existence, main types, and correlations of risks, forming a negotiated risk assessment set that integrates the reasoning of all parties. S33. Based on the negotiated risk assessment set, automatically generate a collaborative risk assessment report, including cross-regional associated risk labels, risk transmission path inferences, and collaborative handling priority suggestions; wherein, the cross-regional associated risk labels are generated based on the consensus risk types in the assessment set, the risk transmission path inferences are derived based on the entity association relationships in the sub-results of each intelligent agent, and the collaborative handling priority suggestions are sorted and output according to the comprehensive threat level of various risks in the assessment set.

[0015] As a further technical solution of the present invention, S32 specifically includes: S321. Each agent broadcasts its preliminary risk assessment sub-results and corresponding confidence levels through a communication channel; S322. Based on the preset consensus rules, each agent adjusts its own judgment in each round of negotiation according to the sub-results and confidence levels received from other agents through a Bayesian update or consensus algorithm. After multiple rounds of interaction, when all agents' judgments on the existence, type and correlation of risks reach the preset consensus threshold, the final judgment is locked and aggregated to form a negotiated risk judgment set.

[0016] As a further technical solution of the present invention, S4 specifically includes: S41. Each smart port node subscribes to and obtains a collaborative risk assessment report containing its own node identifier and subsequent handling verification results from the distributed consensus network; S42. In a local trusted execution environment, based on differential privacy technology, the processing verification results in the report are subjected to security correlation analysis with the privacy-preserving feature vectors generated in the past, and an optimized gradient for the local feature extraction model is generated. S43. Within the local trusted execution environment, the optimized gradient is used to update the parameters of the locally stored feature extraction model through a secure aggregation algorithm, so that the privacy-preserving feature vector generated after the model update can more effectively serve cross-domain risk collaborative identification.

[0017] As a further technical solution of the present invention, S42 specifically includes: S421. Within the local trusted execution environment, the historically stored privacy-preserving feature vectors are paired with their corresponding processing verification results obtained from the collaborative risk assessment report to form a set of encrypted training sample pairs for model optimization; S422. Let the parameters of the local feature extraction model be... The predicted value is calculated using forward propagation. ,in: The mapping function for the feature extraction model. For output layer weights, It is the Sigmoid activation function. For risk prediction probability, For privacy-preserving feature vectors; The loss function uses binary cross-entropy: ,in: For loss function, To handle verification labels; Calculate the original gradient ; Calculate the L2 norm of the gradient ; Apply clipping threshold Obtain the gradient after clipping. ; Injecting Gaussian noise yields ; in: For privacy budget, The gradient after noise injection; It is multivariate Gaussian noise with zero mean; S423. The initial noise gradient is truncated according to its norm, and its magnitude is limited to a preset threshold. Then, the truncated gradient is standardized to finally output an optimized gradient that meets privacy protection requirements and is numerically stable.

[0018] A cross-border trade data processing system based on smart ports, used to implement a cross-border trade data processing method based on smart ports, including: The local trusted feature and credential generation module is configured to: process the raw cross-border trade data stored locally in the Trusted Execution Environment (TEE) within the data sovereignty domain of any smart port node participating in collaborative analysis; the processing includes: generating a privacy-preserving feature vector corresponding to the raw data using a preset feature extraction model; and generating a verifiable credential for the privacy-preserving feature vector to verify its computational integrity and compliance based on a zero-knowledge proof method. The cross-domain trusted feature fusion module is configured to: publish the privacy-preserving feature vectors generated by each node and their accompanying verifiable credentials to a distributed consensus network through a trusted data delivery protocol; and in the distributed consensus network, after verifying the validity of all received verifiable credentials, fuse the privacy-preserving feature vectors from multiple sources to generate a global cross-domain joint feature. A decentralized risk intelligent assessment module is configured to: input the cross-domain joint features into a pre-trained decentralized risk assessment agent network; multiple dedicated agents in the agent network perform asynchronous reasoning and negotiation based on the cross-domain joint features, and output a collaborative risk assessment report, which includes at least cross-regional associated risk labels, risk transmission path inference, and collaborative handling priority suggestions; The privacy and security feedback optimization module is configured to: enable each smart port node to obtain a collaborative risk assessment report related to itself from the distributed consensus network; in the local trusted execution environment (TEE), based on differential privacy technology, perform a security correlation analysis between the processing verification results in the report and its own historically generated privacy protection feature vectors to generate an optimization gradient for the local feature extraction model; and use the optimization gradient to update the local model.

[0019] The beneficial effects of this invention are as follows: 1. By combining a trusted execution environment with zero-knowledge proofs, privacy-preserving feature vectors and verifiable credentials that combine data value and verifiable trust are generated locally, fundamentally resolving the contradiction between data sovereignty compliance and cross-domain data value utilization.

[0020] 2. By generating cross-domain joint features through secure multi-party computation and privacy fusion, and by using a decentralized intelligent agent network for asynchronous negotiation and judgment, it can accurately identify complex cross-regional risk patterns that cannot be detected by a single node, greatly improving risk perception capabilities.

[0021] 3. Introduce a feedback optimization closed loop based on differential privacy, enabling each node to securely optimize its local model by utilizing business feedback generated through global collaboration, thereby driving the intelligent judgment capability of the entire system to continuously and autonomously improve under the premise of privacy protection. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the structure of a cross-border trade data processing method based on a smart port proposed in Embodiment 1 of the present invention; Figure 2 This is an architecture diagram of the feature extraction model in Embodiment 1 of the present invention; Figure 3 This is a block diagram of a cross-border trade data processing system based on a smart port, as proposed in Embodiment 2 of the present invention. Figure 4 This is a comparison diagram of the effects of the present invention and the prior art in Embodiment 3 of the present invention. Detailed Implementation

[0023] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0024] Example 1 Please see the appendix Figure 1 - Appendix Figure 2 A method for processing cross-border trade data based on smart ports, comprising: S1. Process the raw data in the local trusted environment of each node to generate privacy-preserving feature vectors and corresponding verifiable credentials; specifically including: S11. Raw Data Parsing and Standardization Mapping: S111. Within the Trusted Execution Environment (TEE) in the data sovereignty domain, the system first performs in-depth analysis of the locally stored raw cross-border trade data. This raw data typically exists in unstructured or semi-structured form, such as scanned copies of customs declarations, electronic data interchange (EDI) messages, and logistics ledgers.

[0025] The parsing process uses natural language processing and pattern recognition technologies to extract structured fields from at least four core dimensions: product description, including HS code, product name, and specifications; declared value, including currency, unit price, and total price; logistics entity, including means of transport, manifest number, and container number; and consignor / consignee, including company name, place of registration, and type of trading entity.

[0026] S112. After extraction, the system maps these field values ​​to a predefined standard terminology dictionary. This dictionary is jointly maintained by port regulatory authorities and industry experts and covers standardized commodity classification codes, enterprise qualification identifiers, logistics status terms, etc.

[0027] The mapping process achieves semantic alignment and format unification, eliminating heterogeneity issues caused by differences in data standards across different ports, and ultimately generating a complete, standardized, structured data record. This record serves as input for subsequent feature extraction, and its completeness and standardization have been verified within the TEE, ensuring compliant preprocessing within the data sovereignty domain.

[0028] S12. Privacy-preserving feature vector generation: S121. Input the standardized structured data generated in S11 into the feature extraction model deployed in the TEE. The model adopts an architecture that combines a multilayer perceptual network (MLP) with an attention mechanism: the MLP is responsible for performing deep nonlinear transformations on the semantic information of different fields to capture high-order interaction relationships; the attention mechanism dynamically evaluates the contribution weight of each field to the risk assessment task and highlights key signals (such as sensitive commodity categories, abnormal value fluctuations, and characteristics of high-risk consignors and consignees).

[0029] S122. The model uses an encoder-decoder structure to map variable-length structured data records into a fixed-length, anonymized numerical vector, i.e., a privacy-preserving feature vector. The "privacy-preserving" attribute of this vector is reflected in: First, the vector itself does not contain any original text or directly identifiable field values, but only retains abstract, high-dimensional semantic representations; Secondly, the vector generation process is completely isolated within the TEE. The original data is destroyed immediately after feature extraction, and only the vector results are retained in memory. Third, the numerical distribution of the vectors is normalized and subjected to noise perturbation to prevent the original input from being inferred through inversion attacks. Vectors not only meet the needs of cross-domain collaborative analysis, but also ensure that the original local data does not leave the country or leak.

[0030] S13. Based on zero-knowledge proof methods, a cryptographic digital birth certificate is constructed for privacy-preserving feature vectors, enabling cross-domain collaborative nodes to mathematically determine the integrity and compliance of the vector generation process without accessing any original data. This mechanism transforms the computational logic of feature extraction and compliance rules into verifiable arithmetic circuits, thereby generating a cryptographic proof with zero-knowledge properties, which is ultimately encapsulated as a standardized credential. Specifically: S131. Formal Construction of Arithmetic Circuits: The construction of arithmetic circuits is the cornerstone of verifiable computation. Their design requires precise mapping of the computational semantics and compliance constraints of the feature extraction model. The specific construction process follows these principles: a. Cryptographic definition of circuit input: The circuit is set with two types of input interfaces. This design decomposes the originally black-box deep learning inference process into basic arithmetic gate circuits that can be verified one by one. The public input interface only accepts cryptographic commitments to privacy-preserving feature vectors. These commitments are generated using collision-resistant hash functions (such as SHA-3) or the Pedersen commitment scheme based on the discrete logarithm problem, binding a fixed-length feature vector into a concise numerical digest. This commitment is used during the verification phase to ensure that the proven feature vector and the proof itself have not been tampered with.

[0031] The secret input interface then carries all the private data required for verification in a structured manner, including the standardized structured data records generated by S11 (in plain text as the circuit's initial variables), the weight matrices of each layer of the MLP in the feature extraction model (as constant parameters of the circuit), and all intermediate activation values ​​generated during the attention mechanism calculation process (as intermediate variables of the circuit).

[0032] b. Circuit coding of compliance constraints: The circuit logic contains two major constraint systems. The two types of constraints are coupled through Boolean logic gates to form an "if and only if" relationship: the circuit output will only generate a satisfaction signal when all inputs pass the compliance check and all calculation steps are accurately reproduced. One is input validity constraints, which enforce predefined rules by using range check gates and membership verification gates to ensure that the values ​​of each field meet the predefined rules. For example, the declared value field must be a positive number and not exceed a preset upper limit, and the consignor / consignee identifier must be in the whitelist hash set.

[0033] Secondly, there is the constraint of computational correctness. The matrix multiplication of MLP is encoded as an ordered combination of multiplication gates and addition gates, and the attention weight is normalized and encoded as division and exponentiation circuits to ensure that the value of each dimension of the feature vector can be traced back to a valid arithmetic operation sequence.

[0034] c. Commitment binding of circuit output: The final output of the circuit is designed with a dual verification structure. This design unifies the proof of the correctness of the vector content and the proof of the integrity of the computation process into a single circuit satisfiability proposition. The first output is a cryptographic commitment to the plaintext value of the feature vector, which requires that the commitment be completely consistent with the commitment value of the public input, thereby binding the scope of the proof.

[0035] The second output is the compliance flag, which outputs 1 if and only if all constraints are satisfied, otherwise it outputs 0.

[0036] S132. Generation and execution of zero-knowledge proofs: Proof generation is a cryptographic process in which witness data within the TEE interacts with public algorithms. Its execution must be completed in a strictly isolated environment to maintain the confidentiality of secret inputs. a. Preparation and encoding of witness data: In the trusted execution environment, the system aggregates the structured data records (serialized in binary form) in step S11 and the activation values ​​of each layer generated during the forward propagation of the feature extraction model (including the output of the MLP hidden layer, attention distribution vector, and final feature vector) into a complete witness.

[0037] The witness data is encoded as finite field elements, satisfying the input format requirements of the target zero-knowledge proof system (such as Groth16 or PLONK). TEE adds a randomization blinding factor to the witness using a hardware random number generator to prevent correlation attacks caused by multiple proofs using the same witness.

[0038] b. Execution of the proof algorithm within the TEE: Utilizing the Trusted Computation Base (TCB) provided by the TEE, the proof generation algorithm for zero-knowledge proofs is run. This algorithm uses the arithmetic circuit description file constructed by S131 as a public parameter and the encoded witness as private input, generating the proof object through multiple rounds of cryptographic transformations. Specifically, the algorithm performs operations such as polynomial commitment, elliptic curve scalar multiplication, and homomorphic concealment, ultimately outputting a compact proof string (typically only a few hundred bytes).

[0039] Throughout the entire execution process, the witness data does not leave the TEE memory boundary. After the proof is generated, the intermediate state is immediately cleared. The TEE confirms to the outside world through remote authentication that the proof was indeed generated by a legitimate model in a trusted environment.

[0040] c. Timeliness and Uniqueness of Proofs: The proof generation process embeds a timestamp provided by the TEE's trusted clock and uses this timestamp as part of the circuit's public input, ensuring that each proof has strict timeliness and preventing replay attacks. Simultaneously, the TEE generates a unique session identifier for each proof, associating it with the proof string to achieve the one-time use characteristic of the proof.

[0041] S133. Encapsulation and Binding of Verifiable Credentials: Credential encapsulation aims to transform zero-knowledge proofs into a standardized data structure that is transmissible, verifiable, and auditable, and to achieve cryptographic binding with feature vectors; a. Three-tier data structure of vouchers: The generated verifiable vouchers adopt a layered encapsulation format; The bottom layer is the proof ontology layer, which contains the complete zero-knowledge proof string, the cryptographic commitment value of the feature vector, and the timestamp.

[0042] The middle layer is the metadata layer, which records the TEE identity public key of the credential issuer, the version number of the feature extraction model, and the version number of the compliance policy, which are used by the verifier to identify the source of the credential and the timeliness of the policy.

[0043] The top layer is the binding signature layer, which uses the TEE's private key to digitally sign the entire credential structure, ensuring the non-repudiation and integrity of the evidence.

[0044] b. Cryptographic Binding to Feature Vector: The credential is associated with the privacy-preserving feature vector through an implicit binding mechanism. Specifically, the feature vector is transmitted along with its cryptographic commitment in plaintext form, which also appears in the credential's proof ontology layer.

[0045] Upon receiving the vector and the credential, the verifier first verifies whether the commitment value of the vector matches the commitment value within the credential, and then verifies the validity of the credential signature and zero-knowledge proof, thereby establishing a strong binding relationship between the two. Any tampering with the feature vector will cause a change in the commitment value, leading to credential verification failure.

[0046] c. Verification and auditing process of credentials: After receiving the feature vector and credentials, the cross-domain collaborative node performs triple verification: First, it verifies the TEE signature to confirm the credibility of the credential source; second, it verifies the zero-knowledge proof to confirm the compliance of the calculation process through public parameters and commitment values; finally, it extracts the model version and policy version from the credential metadata and compares them with the local policy library to ensure that the feature generation meets the current regulatory requirements.

[0047] Verified credentials will be recorded in the immutable log of the distributed consensus network, forming a complete audit traceability chain to support post-event supervision and dispute arbitration.

[0048] S2. Verify verifiable credentials in a distributed network and fuse the privacy-preserving feature vectors of each node to generate cross-domain joint features; specifically including: S21. Trusted Broadcasting and Transaction Encapsulation of Feature Vectors: S211. After completing the local S13 step, each smart port node needs to publish the results to the cross-domain collaboration layer. Specifically, each node encapsulates its own privacy-preserving feature vector and its cryptographically bound verifiable credentials into a structured transaction through a standardized data trusted delivery protocol.

[0049] The transaction payload employs a layered encoding structure: the core data layer contains the ciphertext form of the feature vector (encrypted using the recipient's collaborating node's temporary public key) and the complete binary sequence of the credentials; the metadata layer contains the node identity, timestamp, data sovereignty domain label, and target collaborative task identifier.

[0050] S212. Security Mechanism of the Trusted Data Delivery Protocol: The protocol mandates that publishing nodes digitally sign the entire transaction using the private key of their Data Sovereign Domain (TEE). This private key is generated by the TEE during remote authentication, and its public key is registered in the identity contract of the distributed consensus network. The signature covers all fields of the transaction payload, ensuring that any bit-level tampering can be detected.

[0051] Meanwhile, the protocol uses a hash chain mechanism to associate the current transaction with the node's historical transaction sequence, forming an immutable publication trajectory. Before relaying a transaction, the propagating nodes in the network verify the signature validity and discard any malicious transactions that fail verification, thus achieving dual protection of identity trustworthiness and data immutability at the transmission layer.

[0052] S22. Parallel Verification and Validity Marking of Zero-Knowledge Proofs: S221. In a distributed consensus network, cooperating nodes (usually validator nodes with strong computing power) continuously monitor the transaction pool and collect multiple characteristic broadcast transactions related to a specific risk assessment task.

[0053] After accumulating a sufficient number of transactions (reaching the preset cross-domain analysis threshold), the collaborating nodes initiate a parallel verification process. This process calls a zero-knowledge proof verification contract deployed in the consensus network state machine. This contract pre-compiles and solidifies the verification algorithm of the proof system used in step S13 (such as the verification key circuit of Groth16).

[0054] S222. Execution logic of parallel verification: The cooperating nodes decompose the verifiable credentials in each transaction, extract the zero-knowledge proof string, feature vector cryptographic commitment and public parameters, and submit them to the verification contract in batches.

[0055] The contract executes proof verification in parallel in a multi-core execution environment, independently checking each credential: whether the commitment value is consistent with the declaration in the credential; whether the TEE signature is valid; and whether the zero-knowledge proof passes the algebraic verification equation.

[0056] If any check fails, the corresponding feature vector is marked as invalid and permanently rejected for fusion; if all checks pass, the vector is written into the valid dataset to be fused, and the timestamp and block height of the verification are recorded to provide a time-series anchor point for subsequent fusion calculations to prevent replay and reordering.

[0057] S223. Verification State Consensus: After all collaborating nodes independently complete verification, they vote on the verification results using a Byzantine Fault Tolerance consensus algorithm (such as PBFT or HotStuff). Only when more than 2 / 3 of the nodes reach a Merkle root consensus on the validity of the same batch of feature vectors is the batch finally confirmed, triggering the fusion computation in stage S23. This design decouples the verification process from the fusion computation, ensuring that the credibility of the fusion input is endorsed by the network-wide consensus, rather than decided by a single node.

[0058] S23. The collaborative node will mark all valid privacy-preserving feature vectors, input a pre-defined secure multi-party computation fusion function, and maintain the encrypted or secret-sharing state of each vector during the computation process, and finally output a cross-domain joint feature that fuses the features of all nodes; specifically: Assume there is a total The first collaborative node participates in the fusion, the second The effective privacy protection feature vector held by each node is: ,in The feature dimension is used for calculating joint features across domains. ,in This can be achieved using a weighted average, PCA, or neural network. This can be accomplished through a three-phase protocol: S231. Secret Sharing Generation and Distribution: Each collaborative node Locally, its feature vector Split into One additive share. Specifically, ( ) Threshold secret sharing scheme (usually) (This requires rebuilding all shares), generating a share set: ,in It is a random vector; To enhance security, nodes Use collaborative nodes public key Shares Encrypt and generate ciphertext. And transmit to the node via a private channel. After the transmission is complete, the node Immediately destroy locally stored With plain text share Only retain the share that should be held by itself. .

[0059] Ultimately, each node A set of shares holding all eigenvectors: .

[0060] This design ensures that no single node can be restored independently. This is because each node only possesses a single encrypted share of each vector.

[0061] S232. Multi-party secure computation fusion: All cooperating nodes locally based on their held share sets. The secure multi-party computation fusion function is executed; the computation is completed in a secret sharing form, that is, without decrypting any share, arithmetic operations are directly performed on the shares; specifically, a weighted average fusion is adopted, and each node finally obtains a result share of the cross-domain joint features; Set nodes The weight is (Weights can be dynamically determined based on the historical data quality or trade volume of nodes, and are pre-disclosed through the consensus network.) The joint characteristics of the weighted average are: node Calculate partial results on a share basis: Each node A result share of generating cross-domain joint features ,satisfy This calculation requires only linear combination, no interaction is needed, and it is extremely efficient.

[0062] S233. Joint Feature Decryption and Output: All cooperating nodes aggregate the results of their respective calculated cross-domain joint features to a designated aggregation node, or through a secure decryption protocol, jointly recover and output the complete, plaintext cross-domain joint features for use by the subsequent risk assessment agent network.

[0063] After completing S232, each cooperating node A share of the result that holds cross-domain joint features To obtain plaintext union features The node executes the secure decryption protocol: a. Specifying the aggregation node mode: The network elects an aggregation node (usually the cooperating node with the highest reputation), and all nodes share the quota. Encrypt the data using the aggregator node's public key and send it to the aggregator node. The aggregator node then decrypts and aggregates the data. ; To prevent malicious behavior by aggregation nodes, a verifiable secret sharing mechanism is adopted, with each node carrying a proof of the correctness of its share (such as the Schwartz-Zippel test). Aggregation nodes only accept shares that have passed verification.

[0064] b. Decentralized Collaborative Recovery Mode: If there is no trusted aggregation node, nodes execute a collaborative decryption protocol. This protocol is based on the linear property of additive secret sharing, and nodes directly reconstruct the plaintext through a round of secure communication. Node 1 generates a random mask And send it to node 2, node 2 will Send to node 3, and so on, until the final node. Summarize all masks and shares, and calculate The results are then broadcast. This process ensures that no single node possesses complete, unencrypted data at any given time. Until it is finally made public.

[0065] The final output cross-domain joint features ( The merged dimension is marked as an "auditable merge result," and its hash value and the list of participating nodes are recorded in the distributed consensus network for public access. The design enables intelligent agent network invocation at each stage. It achieves the goal of cross-domain collaboration, ensuring that the original data remains within its domain, the computation process is verifiable, and the fusion results are auditable.

[0066] S3. Input cross-domain joint features into a decentralized agent network for asynchronous negotiation, and output a collaborative risk assessment report; specifically including: S31. Asynchronous Independent Reasoning by Multiple Agents: The cross-domain joint features output from stages S311 and S2 are input into a pre-trained decentralized risk assessment agent network. This network consists of multiple heterogeneous specialized agents, each corresponding to an independent risk assessment dimension, such as trade authenticity assessment, logistics anomaly detection, entity association analysis, commodity compliance assessment, and capital flow risk identification.

[0067] Each agent performs independent reasoning in parallel and asynchronously within a local trusted execution environment based on cross-domain joint features. Internally, it adopts a pre-trained deep neural network architecture and extracts risk representations of specific dimensions through multi-layer perception and attention mechanisms.

[0068] S312. Each agent generates a corresponding preliminary risk assessment sub-result, which contains two core components: one is a risk probability distribution vector, which represents the agent's confidence assessment of the possibility of each risk type; the other is a risk-related entity set, which records the trade entities (such as enterprises, logistics nodes, and commodity categories) related to the agent's assessment dimension and their association strength and confidence weight.

[0069] After all agents complete their reasoning, a preliminary set of judgment results is formed, which serves as the input basis for subsequent negotiations. This asynchronous design avoids single-point performance bottlenecks, ensuring that each agent can make independent decisions based on its own computing resources and data characteristics.

[0070] S32. Multi-round negotiation and consensus formation: Each agent exchanges preliminary judgment sub-results through the network's built-in encrypted communication channel. While protecting the confidentiality of inference details, a group consensus is reached through iterative negotiation. The negotiation process employs a Bayesian update mechanism to gradually correct individual biases, ultimately forming a consistent judgment. Specifically: S321. Confidence Broadcast and Information Aggregation: In each round of negotiation, each agent broadcasts its current risk assessment information to the entire network, including the risk probability distribution, the mean vector of entity association strength, and the uncertainty covariance matrix of association strength.

[0071] To prevent the leakage of specific reasoning data, the broadcast content is processed using homomorphic encryption technology to ensure that other intelligent agents cannot decrypt the original information, but can perform consensus calculations based on the ciphertext.

[0072] Meanwhile, each agent appends its dynamically adjusted confidence weights based on its historical accuracy to the broadcast message in plaintext for subsequent aggregation. This design strikes a balance between information confidentiality and consensus efficiency.

[0073] S322. Bayesian Consensus Iteration and Decision Locking: Upon receiving encrypted judgment information from other agents, each agent performs a Bayesian update to adjust its own risk probability distribution. This process treats the joint information from other agents as evidence, combining its confidence weights with the correlation coefficient of historical judgment consistency among agents to iteratively calculate the posterior distribution, thereby dynamically correcting the existence, type, and correlation of risks. Entity correlations are updated through a distributed Kalman fusion mechanism, utilizing the correlation strength information from other agents to optimize its own estimation and dynamically adjusting the fusion gain based on the uncertainty covariance matrix.

[0074] Negotiation iterations continue until the consensus metric reaches a preset threshold or the maximum number of rounds. The consensus metric is calculated by taking the average statistical distance between the risk probability distributions of each agent. When this distance falls below a preset minimum threshold, it indicates that the agents' decisions are converging, and the negotiation terminates. At this point, the final decision set is locked, containing the final risk probability distribution and entity association set of each agent, serving as the direct basis for report generation. This locking mechanism ensures the determinism and stability of the negotiation results.

[0075] S33. Generation of Collaborative Risk Assessment Report: Based on the Post-Negotiation Risk Assessment Set The system automatically generates a structured collaborative risk assessment report, which includes three core components: S331. Cross-regional risk tag generation: The system generates risk tags from each intelligent agent. The final probability distribution Extract the risk type with the highest confidence level to form a type set. : ,in: This is the threshold for risk assessment.

[0076] For each risk type in the set The system calculates its cross-agent support. : ,in For each intelligent agent The confidence weight.

[0077] A list of tags is generated and sorted according to support; for tags with support exceeding the cross-regional association threshold... The risk is marked as "cross-customs area related risk", indicating that the risk is collaboratively related among multiple port nodes.

[0078] S332. Risk Transmission Path Prediction: System Integration of Entity Association Sets of All Agents Construct a cross-agent entity association graph ,vertex Representing trade entities (enterprises, logistics nodes, commodity categories), edges The weights are determined by the mean of the association strength. Fusion confirmed: in: Entity Relationship Diagram Mid-vertices and vertex The formed edge The weight, The cosine similarity function is used. and respectively intelligent agents Records about entities in the entity association set and entity The correlation strength vector.

[0079] Based on this graph, the maximum spanning tree or shortest path algorithm is used to identify the transmission chains from high-risk source nodes to target nodes, generating a list of predicted risk transmission paths. Each path is accompanied by a transmission probability. : .

[0080] S333. Collaborative Response Priority Recommendation: The system generates a priority ranking based on risk threat level and urgency. For each identified risk instance, a comprehensive threat score is calculated. in: For the preset weighting coefficients, satisfy ; This represents the maximum propagation probability. Risk type The preset impact coefficient is set based on historical loss data; according to Risk items are sorted in descending order to generate a priority queue for collaborative handling. The report also includes suggested handling measures for each priority (such as enhanced inspection, cross-customs area early warning, and enterprise interviews) and is linked to the original release node, supporting feedback optimization in the S4 stage.

[0081] The final report is output in a structured JSON format, containing three main modules: a risk label array, a transmission path diagram, and a priority queue. It is also signed with the private key of the agent network to ensure the trustworthiness and immutability of the report generation process.

[0082] S4. Each node generates and applies optimized gradients locally based on the collaborative risk assessment report to update the feature extraction model; specifically including: S41. Secure subscription and access to collaborative analysis reports: S411. Each smart port node, as the holder of data sovereignty, proactively initiates a subscription request to the distributed consensus network. The subscription mode adopts a topic filtering mechanism, and the subscription condition is a collaborative risk assessment report containing the node's identity identifier. This identifier is the node's unique and verifiable public key hash in the consensus network, which was publicly registered when the feature vector was broadcast in step S21. The event notification engine in the distributed consensus network continuously monitors newly confirmed risk assessment reports. When the handling verification result in the report is associated with a certain node (the list of handling suggestions embedded in the report contains the node's port code or enterprise identifier), a targeted push is automatically triggered.

[0083] S412. Nodes must complete two-way authentication before obtaining reports: Nodes sign subscription requests using their local TEE private key. Access is only authorized after the consensus network verifies the signature and confirms the node's identity is on the list of valid participants. Report transmission uses an end-to-end encrypted channel, encrypting the report content using a session key temporarily generated by the node in its TEE to ensure the report is not leaked or tampered with during transmission.

[0084] Upon receiving the report, the node first verifies the agent network signature of the report within the TEE to confirm that the report originates from a legitimate decentralized risk assessment agent network and has not been maliciously modified. Subsequently, the node parses the report structure and extracts the handling verification results directly related to the node, including: the specific handling measures implemented for a certain batch of goods submitted by the node (such as inspection, release, and control), the post-event verification label of the handling measures (whether a real risk was found), and the confidence score of the risk association.

[0085] These validation labels serve as key supervisory signals for the S42 stage model optimization, and together with the historically generated privacy-preserving feature vectors, they form a closed-loop feedback data pair.

[0086] S42. In a local trusted execution environment, based on differential privacy technology, the processing verification results in the report are security-correlated with the privacy-preserving feature vectors generated historically to generate an optimized gradient for the local feature extraction model; specifically: S421. Construction of secure sample pairs: Within the isolated memory space of the TEE, the system retrieves the historical privacy-preserving feature vector sequence from persistent storage, denoted as starting from time [time]. arrive The generated vector set, where This refers to the feedback time window length (e.g., 30 days). Simultaneously, the corresponding handling verification results are extracted from the collaborative risk assessment report obtained from S41. These results are discrete classification labels, for example: This indicates that the handling and verification process confirms the existence of risks. This indicates that the processing and verification have confirmed there is no risk. The system performs strict data alignment checks: based on the timestamp, batch number, and trade document hash used when the feature vector was generated, ensuring that each feature vector... A unique match for a processing verification tag Aligned data pairs Encrypted training sample pairs are constructed within the TEE, and their encryption properties are reflected in the following: the feature vectors themselves are already desensitized numerical representations, and the label information is limited to use within the TEE and will never be leaked. The sample pairs are loaded in batches into the GPU secure computing area of ​​the TEE, ready for gradient calculation under noise injection.

[0087] S422. Noise Injection and Gradient Calculation: The system is based on a differential privacy mechanism, which injects controllable noise during the gradient calculation process to ensure that the presence or absence of a single sample pair does not significantly affect the gradient distribution, thereby protecting the privacy of the original trade data corresponding to that sample.

[0088] Let the parameters of the local feature extraction model be... The model copy deployed within the TEE (with the same model structure as used in step S12) uses forward propagation to compute predictions: in: The mapping function for the feature extraction model. For output layer weights, It is the Sigmoid activation function. This is the probability of risk prediction.

[0089] The loss function uses binary cross-entropy: ; Calculate the original gradient : ; To achieve - Differential privacy: The system performs gradient pruning to control sensitivity. First, the L2 norm of the gradient is calculated. ; Apply clipping threshold (Preset value, such as 1.0), to obtain the gradient after clipping. : Pruning ensures that the upper bound of the L2 sensitivity of each sample to the gradient is... That is, the maximum impact of adding or deleting a sample on the gradient vector is no more than [a certain value]. .

[0090] Then, Gaussian noise is injected: Noise Standard Deviation Calculated by the privacy budget: in: To manage privacy budgets, control the strength of privacy protection (the smaller the value, the stronger the protection; typically, a value of [value missing] is used). ); Let be the failure probability, representing the probability that the differential privacy inequality does not hold (usually taken as ). ); The gradient after noise injection; for identity matrix; It is a multivariate Gaussian noise with zero mean.

[0091] S423. Gradient Clipping and Normalization: This involves processing the noisy gradient... Secondary pruning and standardization are performed to ensure numerical stability and convergence.

[0092] a. First, apply norm clipping: in: This is the final pruning threshold (e.g., 1.5).

[0093] b. Then, standardization is performed, and the mean of each dimension of the gradient is calculated. with standard deviation : Obtain the standardized gradient : ,in: Stabilization constant (e.g.) ), to prevent division by zero.

[0094] Final output To optimize the gradient, satisfy differential privacy requirements, and ensure stable numerical distribution.

[0095] S43. Within the TEE, use optimized gradients to update the feature extraction model parameters, ensuring that model evolution does not leak any intermediate computational states. Specifically: S431. Secure Aggregation Algorithm Execution: A secure parameter update protocol is executed within the TEE, employing homomorphic encryption or secret sharing to protect the update process. Assume the current value of the model parameters is... The optimized gradient is The learning rate is (Default value, such as 0.01), parameter update value for: To prevent side-channel attacks during parameter updates, mask arithmetic is used within the TEE: SecureUpdate is implemented through mask multiplication and addition to generate a random mask. ,calculate: in: This represents the model parameters after masking; This represents the gradient after masking; Indicates the updated parameters after the mask; This indicates addition / subtraction operations under a mask, ensuring that intermediate results are always randomized and preventing memory sniffing attacks.

[0096] S432. Atomic replacement after model update: new parameters After computation is completed within the TEE, the old model parameters are immediately replaced via an atomic write operation. The old parameters are securely erased, and the TEE generates a model update log, recording the update version number, timestamp, and privacy budget used. The batch sample size and logs are signed with the TEE private key and stored locally on the node for regulatory auditing.

[0097] S433. Performance Verification and Monitoring: After the update, the TEE uses a small batch of validation samples (not used in training) in an isolated environment to evaluate the feature extraction quality of the new model. It calculates the separability index and information entropy of the feature vectors in the cross-domain joint feature space to ensure that the model update does not lead to performance degradation. The index results are fed back to the node operation and maintenance system through a secure channel, achieving closed-loop monitoring of model optimization.

[0098] This process ensures that the feature extraction models of each smart port node, while continuously receiving cross-domain business feedback, always prioritize privacy protection in their targeted optimization. The generated privacy-protected feature vectors continuously evolve, more accurately capturing cross-border trade risk patterns, thereby improving the collaborative identification efficiency of the overall S1-S3 process.

[0099] Example 2 Please see the appendix Figure 3 A cross-border trade data processing system based on smart ports, used to implement a cross-border trade data processing method based on smart ports, including: The local trusted features and credential generation module specifically includes: Data Preprocessing and Standardization Unit: Configured to parse and clean raw cross-border trade data in TEE, map entity information to standard terms, and generate standardized structured data records.

[0100] Privacy-preserving feature generation unit: configured to input the structured data records into the feature extraction model and generate desensitized privacy-preserving feature vectors through neural network calculation.

[0101] Verifiable credential generation unit: configured to generate, within the TEE, a proof that the feature vector is correctly calculated by the compliance model from the compliance data based on zero-knowledge proof, and output it as a verifiable credential.

[0102] The cross-domain trusted feature fusion module specifically includes: Trusted Publishing Unit: Configured to broadcast privacy-preserving feature vectors and verifiable credentials as trusted transactions to the distributed consensus network via a digital signature protocol.

[0103] Credential Consensus Verification Unit: Configured to perform batch validity verification of all received verifiable credentials in the consensus network via smart contracts, and to mark valid data.

[0104] Security Feature Fusion Unit: Configured to perform fusion calculations on valid privacy-preserving feature vectors in an encrypted or secret-sharing state via a secure multi-party computation (MPC) protocol, ultimately generating cross-domain joint features.

[0105] The decentralized risk intelligent assessment module specifically includes: Feature allocation and asynchronous inference unit: Configured to distribute cross-domain joint features to different risk dimensions in the network, and each agent asynchronously generates preliminary risk assessment sub-results.

[0106] Inter-agent negotiation and consensus unit: configured to provide communication channels so that each agent can exchange sub-results and conduct multiple rounds of negotiation based on predetermined rules (such as Bayesian updates) to form a consensus-based risk assessment set after negotiation.

[0107] Report generation and structured output unit: configured to automatically generate a structured collaborative risk assessment report containing associated risk tags, transmission paths and handling suggestions based on the negotiated risk assessment set.

[0108] The privacy and security feedback optimization module specifically includes: Targeted Report Acquisition Unit: Configured to enable each node to subscribe to and obtain collaborative risk assessment reports and handling verification results containing its own node identifier from the distributed consensus network.

[0109] Privacy-preserving correlation analysis and gradient generation unit: This unit is configured to perform correlation analysis between the disposal verification results as a supervision signal and the historical privacy-preserving feature vectors under differential privacy protection in the TEE, thereby generating a secure optimization gradient.

[0110] Local Model Secure Update Unit: Configured to update local feature extraction model parameters within the TEE using the optimized gradient via a secure aggregation algorithm.

[0111] Example 3 1. Scene Setup To verify the effectiveness of the method of this invention, a simulation environment was constructed containing three main smart port nodes (A, B, and C). Each node locally stores one year's worth of cross-border trade data, covering general trade, express delivery, cross-border e-commerce, and other models. Fifteen known risk patterns across five categories are preset, with some risks exhibiting cross-customs area correlation characteristics. The simulation comparison object is the current mainstream existing technical solutions based on centralized data analysis.

[0112] 2. Implementation process of the embodiments of the present invention Step 1: Each node, within its local TEE, standardizes entities such as goods and enterprises using unified standard terminology and generates privacy-preserving feature vectors. Simultaneously, zero-knowledge proofs are used to generate verifiable credentials for each feature vector. In the simulation, a single node processed 100,000 data points in an average of approximately 2.1 minutes.

[0113] Step Two: Nodes publish their feature vectors and credentials to a simulated permissioned blockchain network (as a distributed consensus network). The network automatically verifies the validity of all credentials and fuses valid features through secure multi-party computation (MPC). A single consensus fusion process involving three nodes takes an average of approximately 8 seconds.

[0114] Step 3: The fused cross-domain joint features are input into a network consisting of five specialized agents (each responsible for the dimensions of goods flow, capital flow, corporate reputation, product attributes, and historical violations). After an average of three rounds of negotiation, the agents reach a consensus on the risks and generate a report.

[0115] Step 4: Based on the processing verification results reported in the report, each node uses differential privacy technology locally to perform security optimization on the feature vectors and models generated in the past week. The model update process is completed within the TEE, and the data does not leave the domain.

[0116] 3. Implementation process of the existing technology comparison scheme A centralized data pool approach was used for comparison: each port node was required to upload its anonymized raw data to a central server. On the server side, the same machine learning model was used for feature extraction, fusion, and risk analysis, and the results were then distributed. This approach also underwent multiple rounds of iterative optimization.

[0117] 4. Simulation Results and Comparative Analysis After multiple rounds of simulation testing, the key performance indicators of existing technologies and this invention are compared as shown in Table 1 below. Figure 4As shown: Table 1: Comparison of Results 5. Explanation of the technical principles behind key data Data utilization efficiency (92% vs 35%): Due to strict restrictions on data leaving the country, existing technologies can only collect very little data in practice. This invention allows 100% of local data to be transformed into usable features without leaving the domain, with only a small amount of information lost, resulting in a significant increase in utilization.

[0118] Cross-regional association risk identification rate (89% vs 22%): This invention generates a global view through security feature fusion, making the association patterns hidden in multi-node data explicit, while the data silo effect of existing technologies leads to the omission of most association risks.

[0119] Improved system evolution speed (391%): The privacy and security feedback optimization module (S4) of this invention enables each node to receive "personalized guidance" based on its own data quality, achieving precise evolution. Existing technologies' global model updates cannot take into account node differences, resulting in slow evolution.

[0120] Conclusion: Simulation results show that, under the premise of strictly adhering to the regulations on local data storage, this invention, through an innovative technical approach, surpasses existing centralized solutions by orders of magnitude in two core dimensions: the depth of data value release (cross-domain association identification) and the system's continuous evolution capability. At the same time, it achieves a higher level of privacy and security protection, verifying the advanced nature and practicality of its solution.

[0121] As can be seen from the above description, the embodiments of the present invention achieve the following technical effects: By strictly adhering to data sovereignty, this approach achieves efficient and reliable release of data value, resolving the fundamental contradiction between data silos and compliant utilization. By processing raw data within a Trusted Execution Environment (TEE) on each node and generating verifiable credentials using zero-knowledge proofs, it ensures that raw data never leaves the local sovereign domain, fundamentally meeting the regulatory requirements for localized data storage. The credentials cryptographically prove the compliance and correctness of the generation process of the privacy-preserving feature vector, thus establishing a solid foundation of trust for cross-domain collaboration without transmitting any sensitive raw information. This overcomes the trust bottleneck of traditional centralized or federated learning, which struggles to prove compliance, enabling data previously locked in silos due to compliance concerns to circulate securely and generate collaborative value in a highly reliable and standardized manner.

[0122] By deeply integrating privacy computing with decentralized intelligence, we have achieved accurate and in-depth insights into hidden risks across customs regions, solving the problem of insufficient capabilities of traditional centralized analysis models.

[0123] By leveraging secure multi-party computation (MPC) to fuse privacy-preserving feature vectors from multiple parties, a global, cross-domain joint feature can be generated without any party being able to access the original data of others. Through the asynchronous reasoning and negotiation mechanism of a decentralized risk assessment agent network, a multi-department expert consultation process was simulated, enabling the identification of complex interconnected risk patterns from the fused global features—patterns that a single node could not detect. Simulation results show that this mechanism increases the identification rate of cross-regional interconnected risks from approximately 22% in existing technologies to over 89%, achieving a qualitative leap from local risk monitoring to global situational awareness, and significantly enhancing the ability of smart ports to detect complex illegal activities.

[0124] A closed-loop local model security evolution mechanism based on global business feedback was constructed, enabling continuous, autonomous, and privacy-secure improvement of the system's overall judgment capabilities. A privacy-secure feedback optimization module was designed, allowing each node to use the processing verification results generated in collaborative judgment as a gold standard, securely optimized for its local feature extraction model under the strict protection of differential privacy technology. This mechanism creates a positive cycle: the more accurate the collaborative judgment, the more effective the feedback; the more effective the feedback, the higher the quality of the privacy-preserving feature vectors generated by the local model, thus making the next collaborative judgment more accurate. Simulation verification shows that this closed loop enables the system to achieve an overall risk identification accuracy evolution speed nearly four times that of traditional centralized solutions after multiple iterations, while ensuring that the business privacy and sensitive business details of each node are not leaked throughout the process, achieving a sustainable intelligent upgrade that unifies security and efficiency.

[0125] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in detail for the sake of brevity.

[0126] This invention is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this specification. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for processing cross-border trade data based on smart ports, characterized in that, include: The raw data is processed in the local trusted environment of each node to generate privacy-preserving feature vectors and corresponding verifiable credentials. The verifiable credentials are verified in a distributed network, and the privacy-preserving feature vectors of each node are fused to generate cross-domain joint features; The cross-domain joint features are input into a decentralized intelligent agent network for asynchronous negotiation, and a collaborative risk assessment report is output. Based on the collaborative risk assessment report, each node generates and applies an optimized gradient locally to update the feature extraction model.

2. The cross-border trade data processing method based on smart ports according to claim 1, characterized in that, The process of processing raw data in the local trusted environment of each node to generate privacy-preserving feature vectors and corresponding verifiable credentials specifically includes: In a trusted execution environment, raw cross-border trade data is parsed to extract structured fields; and the values ​​of each field are mapped to a predefined standard terminology dictionary to generate standardized structured data records. The structured data record is input into a preset feature extraction model. The model maps the input into a numerical vector through a multilayer perceptual network and an attention mechanism, which serves as a privacy-preserving feature vector. Based on the zero-knowledge proof method, verifiable credentials are generated for the privacy-preserving feature vector to verify its computational integrity and compliance.

3. The cross-border trade data processing method based on smart ports according to claim 2, characterized in that, Based on zero-knowledge proof methods, verifiable credentials are generated for the privacy-preserving feature vector to verify its computational integrity and compliance, specifically including: An arithmetic circuit is constructed based on the feature extraction model. The public input of the arithmetic circuit is the cryptographic commitment of the privacy-preserving feature vector, and the secret input is the structured data record and intermediate variables of the model. The arithmetic circuit is satisfied if and only if the secret input data conforms to a predetermined format and range and the model calculation process is correct. Within the trusted execution environment, the structured data records and secret variables in the model calculation process are used as witnesses to run the proof generation algorithm of the zero-knowledge proof method and output a zero-knowledge proof to prove that the arithmetic circuit is correctly satisfied. The zero-knowledge proof, the cryptographic commitment of the privacy-preserving feature vector, and the public parameters used for verification are jointly encapsulated to generate the verifiable credential, which is then bound to the corresponding privacy-preserving feature vector and output.

4. The cross-border trade data processing method based on smart ports according to claim 3, characterized in that, The verifiable credentials are verified in a distributed network, and the privacy-preserving feature vectors of each node are fused to generate cross-domain joint features, specifically including: Each node broadcasts its own privacy-preserving feature vector and its bound verifiable credentials as a transaction to the distributed consensus network through a trusted data delivery protocol. After receiving multiple related transactions, the collaborative nodes in the distributed consensus network perform parallel verification of all verifiable credentials based on a preset zero-knowledge proof verification contract. Only when all credentials have been verified are the corresponding privacy-preserving feature vectors marked as valid data to be fused. The collaborative node will be marked as all valid privacy-preserving feature vectors, input to a pre-defined secure multi-party computation fusion function, and output a cross-domain joint feature that fuses the features of all nodes.

5. A cross-border trade data processing method based on a smart port according to claim 4, characterized in that, The collaborative nodes will be labeled with all valid privacy-preserving feature vectors, input to a preset secure multi-party computation fusion function, and output a cross-domain joint feature that fuses the features of all nodes, specifically including: Each node uses a secret sharing algorithm to split the effective privacy-preserving feature vector into encrypted shares and distribute them to all collaborating nodes; All collaborative nodes locally are based on their held share sets. The secure multi-party computation fusion function is executed, specifically using a weighted average fusion method: Set nodes The weight is Joint characteristics of weighted average for: ,in: For privacy-preserving feature vectors, The number of collaborating nodes; node Calculate local results on a share : ,in: The plaintext share stored locally; Ultimately, each node A result share of generating cross-domain joint features ,satisfy ; The nodes aggregate the results shares through a security protocol, jointly recover and output the complete plaintext cross-domain joint characteristics.

6. The cross-border trade data processing method based on smart ports according to claim 5, characterized in that, The cross-domain joint features are input into a decentralized intelligent agent network for asynchronous negotiation, and a collaborative risk assessment report is output, specifically including: The cross-domain joint features are input into a pre-trained decentralized risk assessment agent network. Multiple specialized agents in the network correspond to different risk assessment dimensions. Each agent performs reasoning based on the cross-domain joint features to generate preliminary risk assessment sub-results for its respective dimension. Each intelligent agent exchanges its preliminary risk assessment sub-results and conducts multiple rounds of negotiation through the network's built-in communication channels, based on predetermined rules, to form a negotiated risk assessment set. Based on the agreed-upon risk assessment set, a collaborative risk assessment report is automatically generated, including cross-customs area related risk labels, risk transmission path predictions, and collaborative handling priority suggestions.

7. A method for processing cross-border trade data based on a smart port according to claim 6, characterized in that, Each intelligent agent exchanges its preliminary risk assessment sub-results through the network's built-in communication channel based on predetermined rules and conducts multiple rounds of negotiation to form a negotiated risk assessment set, specifically including: Each agent broadcasts its preliminary risk assessment results and corresponding confidence levels through the communication channel; Based on the preset consensus rules, each agent adjusts its own judgment in each round of negotiation according to the sub-results and confidence levels received from other agents, through a Bayesian update or consensus algorithm. After multiple rounds of interaction, when all agents reach a preset consensus threshold in their judgments of the existence, type, and correlation of risks, the final judgment is locked and aggregated to form a negotiated risk judgment set.

8. A method for processing cross-border trade data based on a smart port according to claim 7, characterized in that, Based on the collaborative risk assessment report, each node generates and applies optimized gradients locally to update the feature extraction model, specifically including: Each smart port node subscribes to and obtains a collaborative risk assessment report containing its own node identifier and subsequent handling verification results from the distributed consensus network; In a local trusted execution environment, based on differential privacy technology, the processing verification results in the report are subjected to a secure correlation analysis with the privacy-preserving feature vectors generated in the past, and an optimized gradient for the local feature extraction model is generated. Within the local trusted execution environment, the optimized gradient is used to update the parameters of the locally stored feature extraction model via a secure aggregation algorithm.

9. A method for processing cross-border trade data based on a smart port according to claim 8, characterized in that, The generation of optimized gradients for the local feature extraction model specifically includes: Within the local trusted execution environment, the historically stored privacy-preserving feature vectors are paired with the processing verification results to form a set of encrypted training sample pairs for model optimization; Let the parameters of the local feature extraction model be... The predicted value is calculated using forward propagation. ,in: The mapping function for the feature extraction model. For output layer weights, It is the Sigmoid activation function. For risk prediction probability, For privacy-preserving feature vectors; The loss function uses binary cross-entropy: ,in: For loss function, To handle verification labels; Calculate the original gradient ; Calculate the L2 norm of the gradient ; Apply clipping threshold Obtain the gradient after clipping. ; Injecting Gaussian noise yields ; in: For privacy budget, The gradient after noise injection; It is multivariate Gaussian noise with zero mean; The initial noise gradient is truncated according to its norm, and then the truncated gradient is standardized to finally output an optimized gradient that meets privacy protection requirements and is numerically stable.

10. A cross-border trade data processing system based on smart ports, characterized in that, A method for processing cross-border trade data based on a smart port, as described in any one of the claims, includes: The local trusted features and credential generation module is configured to: process the raw cross-border trade data stored locally in a trusted execution environment within the data sovereignty domain of any smart port node participating in collaborative analysis. The cross-domain trusted feature fusion module is configured to: publish the privacy-preserving feature vectors generated by each node and their accompanying verifiable credentials to a distributed consensus network through a trusted data delivery protocol; and in the distributed consensus network, after verifying the validity of all received verifiable credentials, fuse the privacy-preserving feature vectors from multiple sources to generate a global cross-domain joint feature. A decentralized risk intelligent assessment module is configured to: input the cross-domain joint features into a pre-trained decentralized risk assessment agent network; and have multiple dedicated agents in the agent network perform asynchronous reasoning and negotiation based on the cross-domain joint features to output a collaborative risk assessment report. The privacy and security feedback optimization module is configured to: enable each smart port node to obtain a collaborative risk assessment report related to itself from the distributed consensus network; in the local trusted execution environment, based on differential privacy technology, perform a security correlation analysis between the processing verification results in the report and its own historically generated privacy protection feature vectors to generate an optimization gradient for the local feature extraction model; and use the optimization gradient to update the local model.

Images