A privacy protection distributed collaborative auditing system and method based on a blockchain and a federated graph neural network

Abstract

This invention discloses a privacy-preserving distributed collaborative auditing system and method based on blockchain and federated graph neural networks. The invention proposes a decentralized collaborative auditing architecture. First, local graph neural networks deployed on each audit node extract transaction subgraph features. Without sharing original data, a global audit anomaly detection model is collaboratively trained using a reputation-weighted federated learning algorithm. Second, a hybrid consensus mechanism combining dynamic proof-of-stake and reputation values ​​is employed to complete the on-chain recording of audit logs and the verification of model parameters. Finally, zero-knowledge proof technology is used to verify the authenticity of cross-institutional transactions while protecting trade secrets. This invention achieves the usability without visibility of audit data, significantly improving the ability to identify complex risks such as fraud and money laundering in cross-institutional related transactions and enhancing the credibility of audit results while ensuring data privacy and security.

Description

Technical Field

[0001] This invention relates to the fields of computer data processing and modern auditing technology, specifically to a privacy-preserving distributed collaborative auditing system and method based on blockchain and federated graph neural networks. Background Technology

[0002] With economic globalization and the development of corporate groups, modern audit targets often exhibit characteristics of cross-regional, cross-institutional, and cross-system distribution. Traditional audit models typically employ centralized data processing, requiring auditees to aggregate financial and operational data at a unified audit center for analysis. However, this centralized audit model faces significant challenges:

[0003] When cross-border audits or audits involving multiple legal entities are involved, due to restrictions imposed by laws and regulations such as the Data Security Law and GDPR, it is difficult to directly transmit raw sensitive data such as customer PII information and underlying transaction details to the central server, resulting in a serious problem of audit data silos.

[0004] Existing single-agency internal audit models, such as rule-based or local machine learning models, struggle to capture complex money laundering networks or related-party transaction fraud across agencies. For example, funds from Company A flowing to Company B and then back to Company A's related parties may appear legitimate from the perspective of a single node, but can only be identified from a global perspective.

[0005] Audit logs stored in traditional centralized databases are at risk of being tampered with by administrators or hackers, and lack decentralized trust endorsement.

[0006] Existing related technologies, such as the invention "An Auditable Privacy-Preserving Deep Learning Platform Based on Blockchain Incentive Mechanism" (CN109685501B), focus on general-purpose deep learning models. They process structured data in Euclidean space, such as images and text, and primarily rely on differential privacy techniques, using noise to protect privacy, which to some extent sacrifices model accuracy. This invention, however, introduces a Federated Graph Neural Network (FNN). The core of auditing often lies in relationship discovery, such as money laundering networks and guarantor circles. GNNs can extract the topological features of transactions, which is difficult for general-purpose deep learning models to achieve. Furthermore, it introduces zero-knowledge proofs (ZKP). ZKP allows verification of transaction authenticity without revealing specific data, representing a qualitative leap for financial auditing. It achieves logical verification rather than simple data masking without sacrificing model accuracy.

[0007] The invention "A Method for Multi-Party Power Sector Collaborative User Privacy Protection and Traceability" (CN114547677B) deals with power time-series data, focusing on the trajectory recording of data flow. This invention, however, deals with complex financial transaction subgraphs. Power data is typically uploaded from a single point, while financial audit data involves cross-institutional fund flow maps.

[0008] The invention "Data Sharing Method Based on Blockchain and Federated Learning" (CN114417398B) achieves data sharing and recording, thus solving the problem of whether data has been shared or not. This invention, however, through a smart contract audit scheduling module and ZKP, solves the problems of whether data has been shared correctly and whether it is compliant. It is not merely a sharing model, but rather the construction of an automated auditor capable of proactively identifying fraud and money laundering—a proactive defense capability not found in basic data sharing patents.

[0009] The purpose of this invention is to overcome the shortcomings of existing centralized auditing in terms of privacy protection and cross-institutional collaboration, and to provide a distributed auditing system and method that can achieve multi-party collaborative modeling without disclosing original data, and ensure that the auditing process is traceable and tamper-proof. Summary of the Invention

[0010] This invention addresses the fundamental shortcomings of traditional centralized auditing models in terms of data privacy protection, cross-institutional collaboration efficiency, and the credibility of audit evidence. It proposes a privacy-preserving distributed collaborative auditing system based on blockchain and federated graph neural networks. This system decouples "data ownership" from "data usage rights," constructing a globally high-precision anomaly detection model through multi-party collaborative computation without disclosing the auditee's original financial and business data. Furthermore, it utilizes blockchain technology to achieve tamper-proof evidence storage throughout the entire audit process.

[0011] The core technical architecture of this invention is divided into four logical layers from bottom to top: distributed data and privacy computing layer, federated graph collaborative learning layer, blockchain trust and consensus layer, and intelligent auditing application and scheduling layer.

[0012] I. Distributed Data and Privacy Computing Layer This layer, deployed on the local servers of each audited entity (such as group subsidiaries, supply chain partners, and bank branches), forms the physical foundation of the system. Its core task is to transform heterogeneous raw business data into graph-structured data suitable for graph neural network processing, while implementing strict privacy protection measures, specifically including the following methods: (1) Construction of dynamic graph structure Traditional audit data is typically stored in relational databases, which makes it difficult to directly depict complex networks of fund flows and related transactions. This system introduces a dynamic graph construction module based on time slices: Node Definition: The system identifies entity nodes and attribute nodes. Entity nodes represent the entities participating in the transaction, such as personal accounts, corporate accounts, and merchant terminals. Attribute nodes represent transaction environment characteristics, such as IP addresses, MAC addresses, and geographic coordinates. This helps detect characteristics of organized fraud, such as "multiple accounts sharing the same IP address."

[0013] Edge definition: Defined as a directed weighted edge. ,in Indicates the transaction amount. Represents a timestamp. Indicates the transaction type (transfer, payment, refund, etc.).

[0014] Time slice sequence: In order to capture timing patterns, the system divides the continuous time stream into discrete time windows. In each time window Within, construct a local subgraph. .

[0015] (2) Local feature engineering and preprocessing For unstructured text (such as transaction notes and contract summaries) in the original data, a BERT pre-trained model is used to extract semantic vectors, which are then concatenated with numerical features (amount, frequency) to form the initial feature vectors of the nodes. .

[0016] (3) Differential privacy noise injection To prevent model inversion attacks, this system employs Local Differential Privacy (LDP) before data leaves the local computing unit and enters model training. Assume the local gradient is... The system does not upload directly. Instead, it uploads the perturbed gradient. Using the Gaussian mechanism: ; in, S This refers to the gradient sensitivity, which is the maximum influence of a single sample on the gradient. For noise multipliers, determined by the privacy budget and Decision. Through strict control This invention guarantees that, in accordance with the requirements of Under the premise of differential privacy, the statistical properties of gradients are preserved to the greatest extent.

[0017] II. Federated Graph Collaborative Learning Layer This layer is considered the brain of the system, responsible for collaboratively training the globally optimal audit anomaly detection model in a distributed environment. To address the "missing neighbor" problem caused by the inability to physically stitch together cross-organizational graph data, this invention innovatively designs the FedGNN-Audit algorithm. In traditional graph convolutional networks (GCNs), nodes... The embedding vector update depends on all its neighbor nodes. However, in distributed auditing scenarios, cross-institutional transactions (such as transfers from Bank A to Bank B) can lead to the fragmentation of the graph structure, causing nodes to... Some of the nodes' neighbors are not visible locally. To address this, the system proposes a "boundary node alignment and encrypted interaction" mechanism: First, each node identifies counterparties involved in cross-institutional transactions, i.e., boundary nodes, by scanning its local transaction records; then, to obtain cross-institutional neighbor characteristic information without revealing their original identities, the system employs secret sharing technology from Multi-Party Secure Computation (MPC). Specifically, node A calculates its boundary nodes... u Intermediate layer embedding Split it into random pieces and ,Will Send to neighbor node B across institutions and retain When aggregating neighbor information, node B uses the received b3 to participate in the calculation. Ultimately, only through the cooperation of both parties can the complete aggregated features be restored, thus ensuring that neither party can peek into the other's feature distribution alone.

[0018] The training process of the federated model is divided into multiple iterative rounds, each with a compact and orderly process. First, the federated aggregation server (or the selected Leader node) processes the current global model parameters. Distribute to all active audit nodes. Upon receiving the parameters, each node utilizes its local graph data. G k implement E Local training for 1 epoch. To enhance sensitivity to sparse outliers, local training uses the sum of weighted cross-entropy loss and graph reconstruction loss as the loss function: ; : The total loss function value of the local graph neural network, used to guide the gradient descent optimization of the local model.

[0019] The local dataset contains a set of transaction nodes with clear labels (such as "normal" or "fraud"), representing samples confirmed in historical audits.

[0020] :node Ground Truth.

[0021] :node The original feature vector (including features such as transaction amount, frequency, and text summary).

[0022] : The trainable weight parameter matrix of the current local graph neural network.

[0023] Model in parameters Given the input features At that time, predict that it belongs to the label The probability. This term (the first term) is the weighted cross-entropy loss, used to optimize classification accuracy.

[0024] The hyperparameter is a weighting coefficient used to balance the supervised classification loss and the unsupervised reconstruction loss. The larger the value of $\lambda$, the more the model tends to focus on the self-reconstruction capability of the data (i.e., anomaly detection capability).

[0025] The set of all nodes in the local graph data (including labeled and unlabeled nodes).

[0026] Encoder / Decoder: The encoder and decoder functions of a graph autoencoder.

[0027] L2 norm (squared Euclidean distance). This term (the second term) represents the original feature. The difference between the reconstructed features and those after compression and decompression. The greater the difference, the more the transaction pattern deviates from the normal distribution, and the higher the probability of an anomaly.

[0028] The first term in the formula is the supervised classification loss for known historical fraud samples, and the second term is the unsupervised autoencoder reconstruction loss for unknown abnormal patterns. To balance the coefficients of the two.

[0029] The final stage of training is reputation-based weighted aggregation, which is also the core innovation of this scheme. Traditional FedAvg algorithms typically weight data only based on volume, making them susceptible to low-quality data or "free-riding" nodes. This invention introduces a reputation factor. To optimize the aggregation process, the specific global parameter update formula is as follows: ; : No. The parameters of the global audit model are updated after each iteration.

[0030] : No. The global audit model parameters for the previous round.

[0031] The global learning rate on the server side controls the step size at which the global model moves towards local updates, preventing model oscillations.

[0032] The total number of audit nodes participating in this round of federal training.

[0033] : No. The effective data sample size (or data volume) of each audit node participating in the training The larger the data volume of a node, the more information it theoretically contains.

[0034] : No. The model parameters of each audit node after local training is completed.

[0035] : indicates the first Model update amount (or pseudo gradient) for each node This refers to the direction of change of the local model relative to the previous round of the global model.

[0036] : No. The overall aggregation weight coefficient of each node is determined by both the reputation value and the model quality.

[0037] In this mechanism, aggregate weights Based on the node's reputation value and data quality score The decision is made jointly, and the calculation method is as follows: ,in Dynamic evaluation is performed using the F1-Score or AUC value of the local model on the validation set, thereby ensuring that the global model can preferentially absorb the contributions of high-quality nodes.

[0038] III. Blockchain Trust and Consensus Layer To address the potential single point of failure and malicious risks associated with federated learning center servers, this invention utilizes a consortium blockchain to replace the traditional central parameter server and employs smart contracts to automate the audit process.

[0039] The blockchain network consists of server nodes from various auditing institutions, jointly maintaining two parallel logical chains. The model chain stores metadata for the global model (such as structural descriptions and hyperparameters), hashes of model parameters from each iteration (IPFS addresses), and update gradient summaries from each node, ensuring the traceability and reproducibility of the AI ​​model's evolution. The audit evidence chain records trigger logs for audit tasks, hash fingerprints of abnormal transactions, and request and response logs for cross-institutional investigations. All data uploaded to the chain must be signed with the node's private key to ensure authenticity.

[0040] To address the drawbacks of traditional Proof-of-Work (PoW) consensus mechanism's high energy consumption and Proof-of-Stake (PoS) consensus mechanism's tendency to lead to "the rich getting richer," this system designs a hybrid consensus mechanism, Proof of Reputation (PoR), specifically adapted for auditing scenarios. The core of this mechanism lies in a reputation value model, where each node's reputation value... R k Its update logic changes dynamically over time. Executed automatically by a smart contract. In the formula... This is a time decay factor used to prevent nodes from depending on historical achievements; This represents the incremental contribution, which increases when the parameters submitted by a node optimize the global model loss. To penalize and deduct points, a significant deduction is triggered when a node is detected uploading malicious parameters (such as gradient explosion) or rejecting a legitimate request. Based on this reputation system, after each round of aggregation, the set of nodes with the highest reputation scores will take turns acting as the Leader to package blocks, while other validator nodes are responsible for verifying signature validity and the correctness of the aggregation logic. When the score exceeds a certain threshold... Once the node signature is confirmed, the new block is officially added to the chain and becomes effective.

[0041] Furthermore, the system deploys a smart contract called "AuditGovernance," which incorporates a token incentive mechanism to quantify audit contributions. Although operating within a consortium blockchain, by introducing a points token and combining it with the Shapley Value approximation algorithm, the system can calculate each node's marginal contribution to the global model's accuracy and automatically allocate rewards. Simultaneously, nodes must stake a certain amount of tokens upon joining the network. If a node is identified as malicious by the PoR consensus (e.g., by launching a poison attack), the smart contract will automatically forfeit its staked funds, thus achieving intrinsic security constraints based on the economic model.

[0042] IV. Intelligent Audit Application and Scheduling Layer This layer is geared towards end users such as auditors and regulatory agencies, and by providing a visual user interface and automated task scheduling capabilities, it has achieved a leap from "passive post-event auditing" to "proactive real-time monitoring".

[0043] When a local node detects a suspected cross-institutional abnormal fund flow (e.g., funds flowing back to executive accounts through multiple shell companies), the system automatically triggers a cross-institutional collaborative audit protocol. First, node A locks the abnormal transaction ID and generates a request for assistance containing a timestamp and risk type. To complete the verification without compromising privacy, this step employs a zero-knowledge proof (ZKP) mechanism. Node A does not need to directly query the specific ownership of cross-institutional accounts; instead, it constructs a proof circuit to verify whether "the account belongs to a high-risk blacklist" or "whether funds flowed back within a specific time period." Subsequently, node B runs the zk-SNARKs algorithm locally to generate proofs based on private data. And then put it on the blockchain, and the smart contract automatically executes the verification function. Once verification is successful, the risk status is confirmed, and the account's original identity information is not disclosed throughout the entire process.

[0044] In the report generation phase, the system integrates a finely tuned Audit-LLM (Audit-LLM) model tailored for the audit domain. This model takes as input the anomalous subgraph structure data discovered in this audit, the evidence chain hash on the blockchain, and risk assessment scores, and combines this with a built-in knowledge base of accounting standards and auditing law to perform in-depth analysis of anomalous patterns. Ultimately, the model can automatically generate a standardized draft audit report containing risk descriptions, regulatory references, and rectification recommendations, thereby significantly reducing manual drafting costs.

[0045] Furthermore, the system supports a dual operating mechanism of "observation mode" and "blocking mode". In particular, in blocking mode, the smart contract directly interfaces with the enterprise's ERP system API. Once the federated model determines that the risk value of an ongoing transfer exceeds a certain threshold and the determination is confirmed by blockchain consensus, the smart contract will immediately send an instruction to freeze the transaction, thereby moving the audit defense line forward and realizing "audit pre-emptive".

[0046] Compared with the prior art, the beneficial effects of the present invention include: (1) Construct a collaborative modeling mechanism for privacy protection and use federated learning technology to enable each audit node to jointly train a high-precision global anomaly detection model without leaving the local data domain.

[0047] (2) To realize cross-institutional risk mining, a graph neural network (GNN) is introduced to solve the problem of extracting and fusing transaction network features in a distributed environment.

[0048] (3) Establish a credible audit evidence storage and consensus system, use blockchain to record audit operation summaries and model update parameters, and ensure the fairness and immutability of audit results through a reputation consensus mechanism. Attached Figure Description

[0049] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0050] Figure 1 This is a diagram showing the overall system architecture of the method of the present invention.

[0051] Figure 2 This is a flowchart of the federated graph neural network training process of the method of the present invention.

[0052] Figure 3 This is a sequence diagram of the blockchain-based collaborative auditing method of the present invention.

[0053] Example 1 In the distributed data layer, distributed graph data construction and local feature extraction are performed, and each audit node... Graph construction is performed on the local data. The local adjacency matrix is ​​defined as follows: A k The feature matrix is X k Extracting node features using Graph Convolutional Networks (GCNs): ; in, To add a self-loop adjacency matrix, For degree matrix, For the first l The trainable weight matrix of the layer, The activation function is ReLU, and the initial state is... .

[0054] In the federated collaboration module, weighting is based on reputation. The training process for the federated graph neural network is as follows: Figure 2 As shown, each node is trained locally. E After one epoch, obtain the local model parameters. To prevent poisoning attacks from low-quality data or malicious nodes, a reputation scoring mechanism is introduced. Node definition... k Current reputation value Then the global model parameters The aggregation formula is: ; in, For nodes k The formula ensures that nodes with high reputation and large data volumes contribute more to the global model, based on the amount of data sampled.

[0055] : No. The global audit model parameters are updated after each round of iterations. This is the final result of collaborative training among all nodes, and it will be distributed to each node for the next round of training or anomaly detection.

[0056] The total number of audit nodes participating in this round of federal training.

[0057] : No. The audit node at the 1st The current reputation value at the time of the round. This value is maintained by the blockchain ledger; the higher the reputation of a node, the greater its contribution to the global model in terms of model updates.

[0058] : No. The audit node is responsible for the amount of data samples (Data Volume) used for local training. This parameter is introduced to ensure that nodes with more valid data can provide more stable gradient information.

[0059] : No. Each audit node completes locally. The local model parameters (or updated gradients) are uploaded after training for one epoch.

[0060] : Normalization factor (denominator), representing the sum of the "reputation-data volume" products of all participating nodes, used to ensure that the sum of the aggregation weights is 1.

[0061] Consensus and reputation are dynamically updated within the blockchain consensus network. The aggregated global model hash value and the contribution records of each node are packaged and uploaded to the chain. The smart contract dynamically updates the reputation value based on the performance of each node's local model on the validation set in this round. ; in, Historical reputation decay factor This is the model quality evaluation function. If malicious parameters are detected uploaded by a node, a penalty mechanism is triggered. ; Using a pre-trained global GNN model, anomaly scores are calculated for each transaction node. For suspected cross-institutional anomaly chains, such as nodes... With nodes An unusual fund transfer was detected, and the system generated a verification request using zero-knowledge proof: ; Audit node B received Then, verify the authenticity and compliance of the transaction without viewing the original data of node A: ; Zero-knowledge proof (Proof). It is an encrypted mathematical object that contains enough evidence to prove a certain assertion (such as "abnormal fund flow") to be true, but does not contain any original transaction details (such as specific amounts or account balances).

[0062] Private input / private key (Secret Witness). This refers to the original transaction data and private information held by node A. This information is kept confidential during the proof generation process and will not be transmitted to the verifier.

[0063] Suspected abnormal transaction chain. Nodes involved. (Initiator) and Nodes Records of specific financial transactions between (recipients).

[0064] Anomaly detection threshold. A preset risk scoring limit used as a logical benchmark to determine whether a transaction is in violation of regulations.

[0065] : Validation result (Boolean). The output is True or False, indicating whether the transaction actually carries a risk or violates regulations.

[0066] Verification Key: The public key corresponding to the proof generation circuit, used for verification. The effectiveness.

[0067] : Public parameters. Public parameters (such as elliptic curve parameters) generated during ZKP system initialization are shared by all nodes participating in the audit.

[0068] Example 2 Based on the method described in this invention, the following system is constructed, and the overall system architecture is as follows: Figure 1 As shown, an audit is conducted to verify the authenticity of the financing and trade background of the core enterprise and its multi-level suppliers.

[0069] A distributed data layer is used to store local raw data at each audit node and transform the raw data into graph structure data containing account nodes and transaction edges; A local computing unit, configured at each of the audit nodes, includes a graph neural network module for extracting features from the graph structure data; The federated coordination module coordinates multiple local computing units to perform encrypted transmission and secure aggregation of model parameters without transmitting the original data, in order to generate a global audit model. Each supplier node only uploads the encrypted transaction topology features, without disclosing the specific unit price and profit margin.

[0070] The blockchain consensus network, as a decentralized network connecting all the aforementioned audit nodes, is used to record model update logs, audit operation records, and node reputation ledgers in an immutable manner. The blockchain-based collaborative audit sequence is as follows: Figure 3 As shown; Constructing a complete supply chain map using graph neural networks to identify circular trade. This creates a closed loop of fraudulent transactions.

[0071] Blockchain records the review process and model predictions for each financing application, preventing subsequent denials.

[0072] The smart contract engine is deployed and runs on the blockchain consensus network to perform automated audit task scheduling, node reputation value calculation, and abnormal risk warning.

[0073] To address the common poisoning attacks in federated learning—where malicious nodes corrupt the model by uploading contaminated gradients—this system introduces a K-Norm pruning and geometric median aggregation mechanism during the aggregation phase. Specifically, if a node uploads a gradient norm... Exceeding the set threshold C The system will scale it to To limit its destructive power. At the same time, the aggregation method abandons the simple weighted average and instead uses the geometric median, which is highly robust to outliers, as the result, thus ensuring that even if a small number of malicious nodes upload extreme parameters, the aggregation result will not be significantly deflected.

[0074] To address the conflict between the massive number of parameters in graph neural networks and the limited communication bandwidth, this system employs gradient sparsity techniques to optimize communication efficiency. Before uploading data, each node retains only the top-k gradient elements with the largest absolute values ​​and sets the rest to zero. Simultaneously, quantization compression technology is used to compress 32-bit floating-point numbers into 8-bit integers for transmission. Experimental data shows that this strategy successfully reduces model accuracy loss by 90% while keeping it extremely low.

[0075] Considering the high cost of blockchain storage, this system adopts a hybrid storage model combining IPFS (InterPlanetary File System) and blockchain. Under this architecture, massive amounts of raw logs and encrypted model parameter files are stored in the IPFS distributed file system, while the blockchain only stores the hash indexes of the files. This design not only fully utilizes the large storage capacity of IPFS but also strictly guarantees data integrity through the immutability of blockchain.

Claims

1. A privacy-preserving distributed collaborative auditing system based on blockchain and federated graph neural networks, characterized in that, include: A distributed data layer is used to store local raw data at each audit node and transform the raw data into graph structure data containing account nodes and transaction edges; A local computing unit, configured at each of the audit nodes, includes a graph neural network module for extracting features from the graph structure data; The federated coordination module is used to coordinate multiple local computing units to perform encrypted transmission and secure aggregation of model parameters without transmitting the original data, so as to generate a global audit model. The blockchain consensus network, as a decentralized network, connects all the audit nodes and is used to record model update logs, audit operation records, and node reputation ledgers in an immutable manner. The smart contract engine is deployed and runs on the blockchain consensus network to perform automated audit task scheduling, node reputation value calculation, and abnormal risk warning.

2. The privacy-preserving distributed collaborative auditing system based on blockchain and federated graph neural networks according to claim 1, characterized in that, The federated collaboration module employs a reputation-weighted aggregation algorithm, which aggregates local model parameters uploaded by each node. To update the global audit model Its update rule is defined by the following formula: ； in, This is the amount of local model updates. The global learning rate is used to aggregate weights. With the k Reputation score of each node And it is directly proportional to the amount of effective data; Before performing the weighted average, there is also an outlier gradient removal step. This step calculates the geometric median of the local model update volume uploaded by all nodes, removes outlier update volumes that are more than a preset threshold away from the median, and aggregates only the remaining update volumes.

3. The privacy-preserving distributed collaborative auditing system based on blockchain and federated graph neural networks according to claim 1, characterized in that, The graph neural network module adopts a multi-layer graph convolutional structure and learns abnormal patterns in the transaction network by aggregating neighbor node information. Its single-layer feature transfer formula is as follows: ； in To add a self-loop adjacency matrix, Its degree matrix, and The first The node feature matrix and trainable weight matrix of the layer.

4. The privacy-preserving distributed collaborative auditing system based on blockchain and federated graph neural networks according to claim 1, characterized in that, The blockchain consensus network employs a reputation-based proof-of-stake consensus mechanism, in which the probability of a node being elected as a block producer and its voting weight during the verification process are correlated with its reputation value R recorded in the node's reputation ledger. k Directly related.

5. The privacy-preserving distributed collaborative auditing system based on blockchain and federated graph neural networks according to claim 1, characterized in that, The system adopts an on-chain and off-chain collaborative storage architecture. The blockchain consensus network itself only stores high-value, small-volume data, such as transaction hashes, model parameter hashes, reputation values, and audit conclusions. Large-volume data, such as complete original datasets and encrypted model parameter files, are stored in an off-chain distributed file system, and their corresponding file hash addresses are recorded on the blockchain to achieve a trusted index.

6. The privacy-preserving distributed collaborative auditing system based on blockchain and federated graph neural networks according to claim 4, characterized in that, The reputation value R k The update logic is dynamically adjusted by the smart contract engine according to the following rules: ； in, Historical reputation decay factor This measures the contribution of a node to the improvement of the global model accuracy. This is a penalty for malicious behavior or inaction by nodes.

7. A privacy-preserving distributed collaborative auditing method based on the system described in any one of claims 1-7, characterized in that, Includes the following steps: S1. Initialization and Distribution: The smart contract engine initializes the global audit model and distributes its parameters to each audit node; S2. Local Training and Encryption: Each audit node trains its graph neural network model using local graph data, generates local model update data, and performs privacy encryption processing on the update data; S3. Aggregation and Consensus: The federated collaboration module collects the encrypted update amount, performs reputation-based weighted aggregation to generate a global model update; the aggregation result and the contribution proofs of each node are packaged into a block and verified by the PoR consensus mechanism before being uploaded to the blockchain; S4. Iterative optimization: Repeat steps S1 to S3 until the performance of the global audit model converges; S5. Collaborative Audit Application: Utilizes the converged global audit model to detect anomalies in real-time transactions and initiates collaborative verification requests for suspected abnormal transactions across institutions.

8. The privacy-preserving distributed collaborative auditing method according to claim 7, characterized in that, The method also includes an audit tracing step: the global model hash of each model iteration, the abnormal event ID that triggers collaborative verification, and the final result receipt of collaborative verification are all written as transactions into the blockchain consensus network to form a digital audit evidence chain that is linked in chronological order and cannot be tampered with.

9. The privacy-preserving distributed collaborative auditing method according to claim 7, characterized in that, The collaborative verification request in step S5 is implemented using zero-knowledge proof technology. The requesting node A constructs a proof request circuit about its private data, and the receiving node B runs the circuit to generate a proof without exposing its own original data. The proof It is sufficient to confirm or refute an audit assertion without disclosing any additional information.

10. The privacy-preserving distributed collaborative auditing method according to claim 7, characterized in that, The privacy encryption process in step S2 employs a local differential privacy mechanism. Before uploading its model update, each local computing unit injects random noise following a Gaussian or Laplace distribution into the update. The amplitude of this noise is determined by a preset privacy budget. Controls are implemented to mathematically guarantee that no data from a single node will have a identifiable impact on the final aggregation result.

Images