Data security and privacy computing method in edge computing environment

Abstract

The invention relates to the field of computers and privacy computing, and discloses a data security and privacy computing method in an edge computing environment. The method comprises the following steps: locally training edge nodes to generate a model parameter update quantity, generating an encrypted parameter packet by adopting lightweight homomorphic encryption, and embedding a zero-knowledge proof voucher based on a private key signature to verify parameter legality; after verifying the certificate, the aggregation server executes homomorphic aggregation, and broadcasts a global encryption model for node decryption; and synchronously writing an operation event into a distributed account book based on a Merkel directed acyclic graph, and realizing full-cycle tracking and tracing by combining dynamic identity binding and behavior fingerprint anomaly detection. The system comprises a local training unit, an encryption unit, a zero-knowledge proof generation unit, an uploading unit, a verification aggregation unit, a broadcasting unit, a decryption unit, an account book recording unit, an identity binding unit, an anomaly detection unit and the like. According to the method, on the premise of not depending on a centralized trust mechanism, collaborative enhancement of confidentiality, integrity, auditing performance and hostile attack resistance is realized.

Description

Technical Field

[0001] This invention belongs to the field of computer and privacy computing, specifically relating to data security and privacy computing methods in edge computing environments. Background Technology

[0002] With the rapid development of IoT, 5G communication, and AI technologies, edge computing, as a new architecture that pushes computing power down to the network edge, has been widely applied in smart cities, industrial internet, and connected vehicles. Edge computing reduces transmission latency and bandwidth pressure, and improves system response efficiency by performing real-time processing close to the data source.

[0003] Edge nodes are typically deployed in open or semi-controlled environments, making it difficult to guarantee their physical security and operational trustworthiness. This leads to severe security and privacy challenges for data processing tasks executed on such nodes. In particular, in distributed machine learning paradigms such as federated learning, edge devices collaborate to train a global model without sharing raw data. However, the model parameters they upload can still be maliciously tampered with or forged, thereby compromising model convergence or even triggering backdoor attacks.

[0004] Federated learning in edge computing environments relies on a central aggregation server to integrate gradients or model updates from multiple edge nodes. This process assumes that the behavior of participating nodes is generally trustworthy, but in real-world, untrusted edge environments, there is a lack of effective verification mechanisms for the authenticity, integrity, and compliance of parameter sources. Existing solutions often use digital signatures or simple hash verification to ensure transmission integrity, but these methods cannot provide fine-grained audit trail capabilities and are ill-equipped to handle collusion attacks or malicious actions by internal nodes. Furthermore, traditional centralized log systems are easily tampered with or deleted, failing to meet the core requirements of traceable and non-repudiable operational behavior in high-security scenarios.

[0005] Existing technologies in edge federated learning generally suffer from the following shortcomings: First, the parameter transmission process lacks an encrypted auditing mechanism decoupled from the underlying hardware or operating environment, making it difficult to achieve trust transfer across platforms and domains. Second, audit logs are mostly maintained by centralized servers, posing a single point of failure risk and lacking tamper-resistance. Third, while ensuring privacy, it is impossible to efficiently verify the compliance of encrypted gradients, making it difficult for regulatory agencies or aggregators to confirm whether participants have followed the established training protocol without decrypting the original data. These problems are particularly prominent in edge intelligence applications involving sensitive data (such as medical and financial data), severely restricting the deployment of federated learning in high-security scenarios. There is an urgent need for a new security architecture that integrates cryptographic primitives and distributed ledger technology to achieve trustworthiness, verifiability, and traceability throughout the entire edge federated learning process. Summary of the Invention

[0006] This invention provides a data security and privacy computing method for edge computing environments. By constructing a trusted parameter verification channel based on lightweight homomorphic encryption and zero-knowledge proof mechanisms, it ensures the integrity, confidentiality, and auditability of model parameters during transmission in federated learning processes, even when edge nodes are untrusted. This method introduces distributed ledger technology to immutably record the time sequence of each parameter upload and aggregation operation. Combined with dynamic identity binding and behavioral fingerprint extraction mechanisms, it achieves full lifecycle tracking of edge node behavior and precise tracing of abnormal operations, thereby establishing an end-to-end secure federated learning execution environment without relying on a centralized trust institution.

[0007] According to one aspect of the present invention, a data security and privacy computing method in an edge computing environment is provided, comprising: Train the federated learning model locally at the edge nodes to generate local model parameter updates; The local model parameter update amount is subjected to lightweight homomorphic encryption to generate an encrypted parameter package; A zero-knowledge proof credential signed by the edge node's private key is embedded in the encrypted parameter package. The zero-knowledge proof credential is used to prove that the local model parameter update satisfies the preset numerical range constraints and gradient sparsity rules, and does not reveal the original parameter values. Upload the encrypted parameter package and the zero-knowledge proof certificate together to the aggregation server; At the aggregation server, the validity of the zero-knowledge proof credential is verified. If the verification passes, a homomorphic aggregation operation is performed on the encrypted parameter packet to generate global encryption model parameters. Broadcast the global encryption model parameters to all edge nodes participating in this round of training; Each edge node uses its local private key to decrypt the global encryption model parameters to obtain the updated global model; The parameter upload event, the aggregated event, and the corresponding zero-knowledge proof credential hash value are written into the distributed ledger. The distributed ledger adopts a storage mechanism based on the Merkle directed acyclic graph structure to ensure the temporal irreversibility and tamper-proof nature of the event records. In the distributed ledger, each edge node is assigned a unique and dynamically updated identity identifier. The identity identifier is generated by the node's hardware characteristics, network access point information, and current timestamp, and is rebound before the start of each training round to prevent privacy leaks caused by long-term identity association. Before each parameter upload, the behavioral fingerprint of the edge node is extracted. The behavioral fingerprint includes the statistical moment features of local training time, memory usage fluctuation pattern, communication latency distribution and parameter update amount. The behavioral fingerprint is compared with the historical normal behavior baseline. If the deviation exceeds a preset threshold, an abnormal behavior mark is triggered and the mark event is recorded in the distributed ledger.

[0008] As one embodiment of the present invention, the lightweight homomorphic encryption process adopts an additive homomorphic encryption scheme based on integer rings. Its public key is uniformly distributed by the aggregation server during the training initialization phase, and the private key is generated locally by each edge node and kept strictly confidential. The encryption scheme supports encryption of floating-point model parameters after fixed-point quantization with a quantization precision of 6 decimal places. The encrypted parameters are encapsulated in the form of a hexadecimal byte stream.

[0009] As one embodiment of the present invention, the generation process of the zero-knowledge proof credential includes: First, mapping each dimension value of the local model parameter update to a preset valid interval, wherein the valid interval is predetermined by model convergence analysis and has an upper limit of ±0.5; Second, constructing a polynomial commitment, which encodes the parameter update as polynomial coefficients based on the Lagrange interpolation basis function; Third, using the Fiat-Shamir heuristic method to convert the interactive zero-knowledge proof into a non-interactive proof, generating a triple containing a challenge value, a response value, and a commitment hash; Finally, signing the triple using the elliptic curve digital signature algorithm private key of the edge node to form the final zero-knowledge proof credential.

[0010] As one embodiment of the present invention, in the Merkel directed acyclic graph structure of the distributed ledger, each event node includes an event type field, a timestamp field, an operation subject identifier field, an operation object hash field, a zero-knowledge proof credential hash field, and a parent node reference list. The parent node reference list includes at least the previous round of aggregated event nodes and the last uploaded event node of this node, so as to build a multi-path dependency relationship and prevent isolated events from being maliciously deleted or replayed.

[0011] As one embodiment of the present invention, the behavioral fingerprint extraction process specifically includes: during the local training phase, sampling the CPU utilization, random access memory usage, disk read / write speed, and network interface throughput at a frequency of 10 times per second; during the parameter upload phase, recording the time interval from the end of training to the completion of parameter transmission, and measuring the round-trip delay with the aggregation server three times, taking the median as the communication delay value; calculating the first moment, second central moment, and fourth kurtosis of the parameter update amount to form a statistical feature vector; and concatenating all the above features into a row vector of length 128 dimensions as the behavioral fingerprint for this round.

[0012] As one embodiment of the present invention, the historical normal behavior baseline is dynamically maintained through a sliding window mechanism. The window length is the most recent 30 training cycles. Principal component analysis is performed on all behavioral fingerprints within the window, and principal components with a cumulative variance contribution rate of not less than 95% are retained to form a low-dimensional feature subspace. In this low-dimensional feature subspace, the Mahalanobis distance between the current behavioral fingerprint and the window mean vector is calculated. If the Mahalanobis distance is greater than 3 times the window standard deviation, it is determined to be abnormal behavior.

[0013] As one embodiment of the present invention, when the aggregation server performs homomorphic aggregation operation, it only accumulates the encrypted parameter packets verified by zero-knowledge proof. The accumulation result remains encrypted. After the aggregation is completed, the aggregation server's digital signature is attached to the global encryption model parameters, and a unique event identifier for the aggregation event is generated. This event identifier is generated by hashing the aggregation round number, the number of participating nodes, and the aggregation timestamp.

[0014] As one embodiment of the present invention, the broadcast of the global encryption model parameters adopts a fragmented redundancy transmission mechanism, which divides the encryption parameters into several data blocks, each block is attached with a cyclic redundancy check code, and is sent through different network interfaces via a multipath transmission protocol. After receiving all the data blocks, the receiving node verifies the cyclic redundancy check code and reassembles the original encryption parameters. If any data block fails to be verified, it requests to retransmit the data block.

[0015] As one embodiment of the present invention, the dynamic update mechanism of the identity identifier is executed before the start of each training round. The specific steps are as follows: collecting the motherboard serial number, number of central processing unit cores, total memory capacity and current public network exit Internet Protocol address of the edge node; concatenating the above four hardware and network features into a string; performing a Chinese national cryptographic SM3 hash operation on the string to obtain a 256-bit digest; performing an XOR operation on the digest and the current Coordinated Universal Time timestamp to generate the final identity identifier, which is only valid within the current training cycle.

[0016] According to another aspect of the present invention, a data security and privacy computing system in an edge computing environment is provided, comprising: Local model training unit, used to train federated learning models locally on edge nodes to generate local model parameter updates; The parameter encryption unit is used to perform lightweight homomorphic encryption on the local model parameter update amount to generate an encrypted parameter package. A zero-knowledge proof generation unit is used to embed a zero-knowledge proof credential signed by the edge node's private key into the encrypted parameter package; The parameter uploading unit is used to upload the encrypted parameter package and the zero-knowledge proof certificate together to the aggregation server; The credential verification and homomorphic aggregation unit is used to verify the validity of the zero-knowledge proof credential at the aggregation server, and to perform homomorphic aggregation operation on the verified encrypted parameter packet to generate global encryption model parameters; The parameter broadcasting unit is used to broadcast the global encrypted model parameters to all edge nodes participating in this round of training. The model decryption unit is used by each edge node to decrypt the global encrypted model parameters using its local private key to obtain the updated global model. The distributed ledger recording unit is used to write the parameter upload event, the aggregation event, and the corresponding zero-knowledge proof credential hash value into the distributed ledger. The dynamic identity binding unit is used to assign a unique and dynamically updated identity identifier to each edge node; The behavior fingerprint extraction and anomaly detection unit is used to extract the behavior fingerprint of edge nodes before each parameter upload and compare it with the historical normal behavior baseline to trigger anomaly behavior marking.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention employs a collaborative mechanism of lightweight homomorphic encryption and zero-knowledge proof to ensure the confidentiality of model parameters while achieving strong verification of parameter validity, effectively preventing model poisoning attacks caused by malicious or abnormal parameters uploaded by edge nodes.

[0018] This invention provides immutable audit trail capabilities for all training operations by introducing a distributed ledger, addressing the technical deficiency of existing federated learning frameworks that lack post-event accountability mechanisms. The combination of dynamic identity binding and behavioral fingerprint analysis not only achieves anonymization of edge nodes to protect user privacy but also possesses the ability to accurately identify and track abnormal behavior over the long term.

[0019] This invention constructs a decentralized, verifiable, auditable, and malicious node-resistant edge federated learning security system through an overall solution without significantly increasing communication overhead and computational latency, thereby significantly improving the reliability and robustness of privacy computing in edge computing environments. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the trusted parameter verification channel based on lightweight homomorphic encryption and zero-knowledge proof in this invention. Figure 3 This is a flowchart illustrating the logical flow of the local model training, parameter encryption, and zero-knowledge proof generation stages in this invention. Figure 4This is a logical flow diagram of the credential verification, homomorphic aggregation, and global model broadcasting stages at the aggregation server in this invention; Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow between edge nodes, aggregation servers, and distributed ledgers in this invention; Figure 6 This is a logical flowchart of the dynamic identity binding, behavioral fingerprint extraction, and anomaly detection mechanism in this invention. Detailed Implementation

[0021] Please refer to the attached document. Figure 1 To be continued Figure 6 In edge computing environments, federated learning, as a distributed machine learning paradigm, allows multiple edge nodes to collaboratively train a global model without sharing raw data. However, when edge nodes are untrusted, the parameter updates they upload to their local models may be maliciously tampered with or injected with abnormal gradients, leading to model poisoning attacks. Furthermore, existing federated learning frameworks generally lack integrity verification mechanisms for parameter transmission and auditable tracing capabilities for operational behaviors, making it difficult to effectively identify malicious nodes or trace the source of abnormal events. To address these technical issues, this invention provides a data security and privacy computing method for edge computing environments. By constructing a trusted parameter verification channel based on lightweight homomorphic encryption and zero-knowledge proof mechanisms, and combining it with distributed ledgers, dynamic identity binding, and behavioral fingerprint analysis, an end-to-end secure, verifiable, and auditable federated learning execution process is achieved.

[0022] The method first performs the federated learning model training process locally at each edge node, generating local model parameter update values. These local model parameter update values ​​are typically represented as high-dimensional vectors, where each dimension corresponds to the gradient adjustment value of a certain weight or bias term in the model. Local training employs standard stochastic gradient descent or its variants, completing one or more rounds of forward and backward propagation calculations on the local dataset, and then outputting the parameter update values ​​obtained in this round of training. The local model parameter update values ​​should numerically satisfy preset convergence constraints, such as an absolute value not exceeding ±0.5, to ensure that the global model maintains stable convergence during the aggregation process.

[0023] Subsequently, the local model parameter updates are subjected to lightweight homomorphic encryption to generate an encrypted parameter package. This lightweight homomorphic encryption scheme employs an addition homomorphic encryption mechanism based on integer rings. Its mathematical foundation is built upon modular arithmetic, supporting the accumulation of multiple encrypted parameters in ciphertext without affecting the plaintext result after decryption. During the training initialization phase, the aggregation server uniformly distributes public keys to all participating nodes, while each edge node generates and strictly keeps its private key confidential locally. Since model parameters are typically floating-point numbers, they must first be converted to fixed-point form, with the quantization precision set to 6 decimal places (i.e., multiplied by 1 million and rounded down) before encryption. The encrypted parameters are encapsulated in hexadecimal byte stream format, forming a structured encrypted parameter package. This encrypted parameter package contains parameter dimension information, quantization coefficients, and the encrypted data body.

[0024] While generating the encrypted parameter package, the edge node constructs a zero-knowledge proof credential and embeds it into the metadata area of ​​the encrypted parameter package. The zero-knowledge proof credential is used to prove to the aggregation server that the value of each dimension of the local model parameter update falls within a preset legal value range, and the overall result satisfies the gradient sparsity rule, that is, the proportion of non-zero elements is lower than a certain threshold, but the entire proof process does not reveal the specific value of any original parameter.

[0025] The generation process of this proof includes the following steps: First, each dimension of the parameter update is mapped to a valid interval; second, a polynomial commitment is constructed, which encodes the parameter update as polynomial coefficients based on the Lagrange interpolation basis function; third, the zero-knowledge proof protocol, which originally required multiple rounds of interaction, is converted into a non-interactive form using the Fiat-Shamir heuristic, generating a triple containing a challenge value, a response value, and a commitment hash; finally, the triple is digitally signed using the elliptic curve digital signature algorithm private key of the edge node, forming the final zero-knowledge proof credential. Verification of this zero-knowledge proof credential only requires the aggregation server to hold the corresponding public key, without accessing the original parameters.

[0026] After encryption and proof generation are completed, the edge nodes upload the encrypted parameter packet and the zero-knowledge proof credential to the aggregation server. The upload process employs a secure transmission protocol to ensure that the data is not eavesdropped on or tampered with by a man-in-the-middle during transmission. Upon receiving the uploaded data from all participating nodes, the aggregation server first performs validity verification on each zero-knowledge proof credential. The verification process includes: parsing the signature in the credential and verifying its authenticity using the corresponding edge node's public key; verifying whether the challenge value and response value satisfy a preset algebraic relationship; and confirming that the commitment hash and the encrypted form of the parameter update are consistent. If any credential verification fails, the encrypted parameter packet of that node is discarded and it will not participate in subsequent aggregation.

[0027] For all verified encrypted parameter packets, the aggregation server performs a homomorphic aggregation operation. Due to the additive homomorphism of the encryption scheme, the aggregation operation only requires element-wise addition of the ciphertext in the corresponding dimension of each encrypted parameter packet. The result is still valid ciphertext, and its decryption equals the sum of the plaintext parameter updates. After aggregation, a global encryption model parameter is generated, representing the sum of contributions from all legitimate nodes in this round. The aggregation server then appends its own digital signature to the global encryption model parameter and generates a unique aggregation event identifier. This aggregation event identifier is generated by hashing the aggregation round number, the number of participating nodes, and the Coordinated Universal Time (UTC) timestamp of the aggregation completion time, ensuring its global uniqueness and temporal traceability.

[0028] The aggregation server broadcasts the global encryption model parameters to all edge nodes participating in this round of training. The broadcast employs a fragmented redundancy transmission mechanism: the encryption parameters are divided into several fixed-length data blocks, each with a 32-bit cyclic redundancy check (CRC) code; different data blocks are sent via different network interfaces or routing paths using a multipath transmission protocol; after receiving all data blocks, the receiving nodes sequentially verify the CRC code of each block. If all verifications pass, the original encryption parameters are reassembled in order; if any data block fails verification, the receiving node requests a retransmission of that specific data block from the aggregation server to avoid bandwidth waste caused by full retransmission.

[0029] After receiving the complete global encrypted model parameters, each edge node decrypts them using its local private key to obtain the plaintext global model parameter update. Then, it adds this global model parameter update to its current local model parameters, completing the update for this round of global model updates. This concludes one iteration of secure federated learning.

[0030] Simultaneously, the system writes the parameter upload event, the aggregated event, and the corresponding zero-knowledge proof credential hash value into the distributed ledger. This distributed ledger employs a storage mechanism based on a Merkel directed acyclic graph structure, rather than a traditional linear blockchain. In this Merkel directed acyclic graph structure, each event is recorded as an independent event node.

[0031] Each event node contains the following fields: an event type field to distinguish event categories such as parameter upload, aggregation operation, or anomaly flag; a timestamp field to record the exact time the event occurred; an operation subject identifier field to identify the edge node that performed the operation; an operation object hash field to store the hash digest of the data being operated on (such as encrypted parameter packets); a zero-knowledge proof credential hash field to store the hash value of the corresponding credential; and a parent node reference list containing at least two references: the previous aggregation event node and the previous parameter upload event node of this node. This multi-parent node reference mechanism constructs a complex dependency topology, ensuring that any attempt to delete or tamper with historical events will disrupt the consistency of the graph structure and be detected by the system.

[0032] To protect the long-term privacy of edge nodes and prevent user behavior tracking through fixed identity identifiers, this invention introduces a dynamic identity binding mechanism. Before each training round, a unique and temporary identity identifier is reassigned to each edge node. The generation process of this identity identifier is as follows: four hardware and network characteristics of the edge node are collected, including the motherboard serial number, the number of CPU cores, the total memory capacity, and the current public network exit Internet Protocol address; these four characteristics are concatenated into a single string in a fixed order; the SM3 hash algorithm is executed on this single string to generate a 256-bit digest; this digest is XORed with the current Coordinated Universal Time (UTC) timestamp to obtain the final identity identifier. This identity identifier is only valid within the current training cycle and will be regenerated at the start of the next training round, thereby severing cross-round identity associations and achieving anonymous communication.

[0033] Furthermore, to enhance the ability to identify malicious or abnormal edge nodes, this invention performs behavioral fingerprint extraction and anomaly detection before each parameter upload. The behavioral fingerprint extraction process covers two stages: local training and parameter upload. During the local training stage, the system samples the following system resource indicators at a frequency of 10 times per second: CPU utilization, random access memory usage, disk read / write speed, and network interface throughput, continuing throughout the entire training cycle. During the parameter upload stage, the time interval from the end of local training to the completion of parameter transmission is recorded, and the round-trip latency with the aggregation server is actively measured three times, with the median taken as the communication latency value. Simultaneously, the statistical moment features of the local model parameter updates are calculated, including the first moment (mean), the second central moment (variance), and the fourth kurtosis, forming a feature sub-vector describing the parameter distribution. After normalizing all the above sampled data and statistical features, they are concatenated to form a 128-dimensional row vector, which serves as the behavioral fingerprint for this round.

[0034] The system maintains a historical baseline of normal behavior for comparison with the current behavior fingerprint. This baseline is dynamically updated using a sliding window mechanism, with the window length set to the last 30 training cycles. Principal component analysis is performed on all historical behavior fingerprints stored within the window to calculate the eigenvalues ​​and eigenvectors of the covariance matrix. Principal components with a cumulative variance contribution rate of at least 95% are retained, forming a low-dimensional feature subspace. In this low-dimensional feature subspace, the Mahalanobis distance between the current behavior fingerprint and the mean vector of all historical fingerprints within the window is calculated. The formula for calculating the Mahalanobis distance is as follows: in, Let be the projection vector of the current behavior fingerprint in the principal component subspace. The mean vector of the window. This is the window covariance matrix. If the calculated Mahalanobis distance is greater than three times the window standard deviation, the current behavior is determined to significantly deviate from the normal pattern, triggering an abnormal behavior flag. This flag event, along with relevant contextual information (such as behavior fingerprint and deviation value), is written to the distributed ledger for subsequent audit analysis.

[0035] When performing homomorphic aggregation operations, the aggregation server strictly limits itself to accumulating only encrypted parameter packets verified through zero-knowledge proofs. After aggregation is completed, in addition to generating global encryption model parameters, it also needs to construct an aggregation event record, which includes a list of participating nodes, the identity identifier of each node, the aggregation timestamp, the hash of the aggregation result, and the signature of the aggregation server. This aggregation event record is also written to the distributed ledger, referencing all legitimate upload event nodes as its parent nodes, forming a complete causal chain.

[0036] In the Merkle directed acyclic graph (DAG) structure of a distributed ledger, the writing of event nodes follows a consensus mechanism to ensure that all participants agree on the order of events. The hash value of each new event node is calculated from all its field contents and serves as the parent reference target for its child nodes. Since the graph structure allows multiple branches to grow in parallel, the system maintains global consistency through periodic snapshots and state verification mechanisms to prevent state splitting caused by forks.

[0037] After completing model decryption and updates, each edge node enters the next training preparation phase. At this point, the dynamic identity binding unit regenerates its current identity identifier and registers it with the aggregation server. Registration information includes the new identity identifier, public key, and hardware feature digest, used for subsequent communication and verification. This entire process is repeated until the global model converges or the preset maximum number of training rounds is reached.

[0038] The method described in this invention constructs a multi-layered security protection system under the assumption that edge nodes are untrusted, through the aforementioned steps. Lightweight homomorphic encryption ensures the confidentiality of parameters, zero-knowledge proofs enable non-disclosure verification of parameter validity, distributed ledgers provide tamper-proof operation audit trails, dynamic identity binding protects user privacy, and behavioral fingerprint analysis enhances proactive defense against abnormal nodes. These mechanisms work together to ensure the security, reliability, and traceability of the federated learning process in an edge computing environment.

[0039] At the system level, this invention is implemented as a data security and privacy computing system in an edge computing environment. Deployed on various edge nodes and an aggregation server, the system comprises multiple functional units. The local model training unit is responsible for training the model on local data and outputting parameter updates. The parameter encryption unit calls a homomorphic encryption library to perform fixed-point quantization and encryption on the updates, generating an encrypted parameter package. The zero-knowledge proof generation unit implements multinomial commitment construction, Fiat-Shamir transformation, and digital signature to produce a zero-knowledge proof credential. The parameter upload unit manages network connections and securely transmits the encrypted parameter package and credential to the aggregation server.

[0040] On the aggregation server side, the credential verification and homomorphic aggregation unit is responsible for batch verification of zero-knowledge proofs and performing homomorphic accumulation on valid ciphertexts. The parameter broadcasting unit implements fragmentation, redundancy, and multipath transmission to ensure reliable distribution of the global model. The model decryption unit calls the local private key at the edge node to complete the global model update.

[0041] The distributed ledger recording unit maintains a Merkel directed acyclic graph structure, handling the creation, signing, referencing, and persistent storage of event nodes. The dynamic identity binding unit collects hardware and network characteristics at the beginning of each round, performs SM3 hashing and timestamp XOR, and generates temporary identity identifiers. The behavior fingerprint extraction and anomaly detection unit coordinates system resource monitoring, statistical feature calculation, principal component analysis, and Mahalanobis distance determination to achieve automatic identification and labeling of abnormal behavior.

[0042] All units exchange data through well-defined interfaces, employing message queues or remote procedure calls for asynchronous communication. During system runtime, units work in parallel, ensuring that the introduction of security mechanisms does not significantly increase end-to-end latency, even with limited edge computing resources. The entire system architecture is decentralized, independent of any third-party trusted institution, and all security attributes are guaranteed by cryptographic primitives and a distributed consensus mechanism.

Claims

1. A data security and privacy computing method in an edge computing environment, characterized in that, include: Train the federated learning model locally at the edge nodes to generate local model parameter updates; The local model parameter update amount is subjected to lightweight homomorphic encryption to generate an encrypted parameter package; A zero-knowledge proof credential signed by the edge node's private key is embedded in the encrypted parameter packet; Upload the encrypted parameter package and the zero-knowledge proof certificate together to the aggregation server; At the aggregation server, the validity of the zero-knowledge proof credential is verified. If the verification passes, a homomorphic aggregation operation is performed on the encrypted parameter packet to generate global encryption model parameters. Broadcast the global encryption model parameters to all edge nodes participating in this round of training; Each edge node uses its local private key to decrypt the global encryption model parameters to obtain the updated global model; The parameter upload event, the aggregated event, and the corresponding zero-knowledge proof credential hash value are written into the distributed ledger. The distributed ledger adopts a storage mechanism based on the Merkle directed acyclic graph structure to ensure the temporal irreversibility and tamper-proof nature of the event records. In the distributed ledger, each edge node is assigned a unique and dynamically updated identity identifier. The identity identifier is generated by the node's hardware characteristics, network access point information, and current timestamp, and is rebound before the start of each training round to prevent privacy leaks caused by long-term identity association. Before each parameter upload, the behavioral fingerprint of the edge node is extracted. The behavioral fingerprint includes the statistical moment features of local training time, memory usage fluctuation pattern, communication latency distribution and parameter update amount. The behavioral fingerprint is compared with the historical normal behavior baseline. If the deviation exceeds a preset threshold, an abnormal behavior mark is triggered and the mark event is recorded in the distributed ledger.

2. The data security and privacy computing method in an edge computing environment according to claim 1, characterized in that, The local model parameter update amount is subjected to lightweight homomorphic encryption to generate an encrypted parameter package, including: The floating-point parameters in the local model parameter update are fixed-point quantized with a quantization precision of 6 decimal places to obtain an integer parameter vector. The integer parameter vector is encrypted using an additive homomorphic encryption scheme based on an integer ring, wherein the public key is uniformly distributed by the aggregation server during the training initialization phase, and the private key is generated locally by each edge node and kept strictly confidential. The encrypted parameters are encapsulated in hexadecimal byte stream format to form an encrypted parameter package containing parameter dimension information, quantization coefficients, and encrypted data.

3. The data security and privacy computing method in an edge computing environment according to claim 2, characterized in that, Embedding a zero-knowledge proof credential signed by the edge node's private key in the encrypted parameter packet includes: Each dimension value of the local model parameter update is mapped to a preset valid range, the upper limit of which is ±0.5; Construct a polynomial commitment based on the Lagrange interpolation basis function and encode the parameter update as polynomial coefficients; We use the Fiat-Shamir heuristic to transform interactive zero-knowledge proofs into non-interactive proofs, generating triples containing challenge value, response value, and commitment hash; The triple is signed using the private key of the elliptic curve digital signature algorithm of the edge node to form a zero-knowledge proof credential.

4. The data security and privacy computing method in an edge computing environment according to claim 3, characterized in that, At the aggregation server, verifying the validity of the zero-knowledge proof credential includes: Parse the digital signature in the zero-knowledge proof credential and verify its authenticity using the public key of the corresponding edge node; Verify whether the challenge value and response value satisfy a preset algebraic relationship; Confirm that the commitment hash is consistent with the parameter dimension and quantization coefficient in the encrypted parameter package; If all the above verifications pass, the zero-knowledge proof certificate is deemed valid.

5. The data security and privacy computing method in an edge computing environment according to claim 4, characterized in that, Perform homomorphic aggregation operations on the encrypted parameter packet to generate global encryption model parameters, including: Add the ciphertext of the corresponding dimension in all encrypted parameter packets that have been verified by zero-knowledge proofs element by element; The sum is used as a parameter of the global encryption model, and its decryption result is equal to the sum of the update amounts of each plaintext parameter. The global encryption model parameters are appended with the digital signature of the aggregation server, and a unique event identifier for the aggregation event is generated. The unique event identifier is generated by hashing the aggregation round number, the number of participating nodes, and the aggregation timestamp.

6. The data security and privacy computing method in an edge computing environment according to claim 5, characterized in that, The global encryption model parameters are broadcast to all edge nodes participating in this round of training, including: The global encryption model parameters are divided into several fixed-length data blocks; Add a 32-bit cyclic redundancy check code to each data block; Different data blocks are sent via different network interfaces using a multipath transmission protocol; After receiving all data blocks, the receiving node verifies the cyclic redundancy check code of each block and reassembles the original encryption parameters. If any data block fails to be verified, it requests a retransmission of that data block.

7. The data security and privacy computing method in an edge computing environment according to claim 6, characterized in that, Write the parameter upload event, aggregation event, and corresponding zero-knowledge proof credential hash value into the distributed ledger, including: Create an event node for each event. The event node includes an event type field, a timestamp field, an operation subject identity identifier field, an operation object hash field, a zero-knowledge proof credential hash field, and a parent node reference list. The parent node reference list must contain at least the previous round of aggregated event nodes and the previous upload event node of this node; The event nodes are written into a distributed ledger based on a Merkel directed acyclic graph structure to ensure the temporal irreversibility and tamper-proof nature of the event records.

8. The data security and privacy computing method in an edge computing environment according to claim 7, characterized in that, Assign a unique and dynamically updated identity identifier to each edge node, including: Collect the motherboard serial number, number of CPU cores, total memory capacity, and current public network exit Internet Protocol address of the edge nodes; Concatenate the above four features into a string; Perform a Chinese national cryptographic SM3 hash operation on the string to obtain a 256-bit digest; The summary is XORed with the current Coordinated Universal Time timestamp to generate an identity identifier that is valid only within the current training cycle.

9. The data security and privacy computing method in an edge computing environment according to claim 8, characterized in that, Extracting behavioral fingerprints of edge nodes, including: During the local training phase, CPU utilization, random access memory usage, disk read / write speed and network interface throughput are sampled 10 times per second. During the parameter upload phase, the time interval from the end of training to the completion of parameter transmission is recorded, and the round-trip latency with the aggregation server is measured three times, with the median value taken as the communication latency value. The first moment, second central moment, and fourth kurtosis of the local model parameter update are calculated to form a statistical feature vector. All the above features are concatenated into a row vector of length 128, which serves as the behavior fingerprint for this round.

10. The data security and privacy computing method in an edge computing environment according to claim 9, characterized in that, The behavioral fingerprint is compared with a historical baseline of normal behavior, including: The historical behavioral fingerprints within the most recent 30 training cycles are maintained using a sliding window mechanism; Principal component analysis was performed on the historical behavioral fingerprints, and principal components with a cumulative variance contribution rate of not less than 95% were retained to form a low-dimensional feature subspace. Calculate the Mahalanobis distance between the current behavior fingerprint and the window mean vector in the low-dimensional feature subspace; If the Mahalanobis distance is greater than 3 times the window standard deviation, it is determined to be abnormal behavior and an abnormal behavior flag is triggered.

Images