Federal learning excitation system based on block chain

Abstract

The invention discloses a federated learning excitation system based on a block chain, and relates to the technical field of privacy computing, and the system comprises a task module which receives a federated learning task request, and generates a neuron-level contribution mark matrix through a fractal probability function; the verification module is used for performing logic consistency verification on the gradient homomorphic commitment and the zero-knowledge proof, analyzing a gradient norm in the gradient homomorphic commitment after the verification is passed, calculating the number of effective contribution neurons in the neuron-level contribution marking matrix, and calculating a neuron-level contribution value by combining a preset weight coefficient; generating a federal learning reward value according to a secondary incentive rule; the aggregation module is used for calculating the gradient similarity of the blinding gradients, discarding the blinding gradients lower than a similarity threshold value, and carrying out weighted average on the residual blinding gradients to generate a global weight parameter; according to the method, the privacy protection capability and the traceability credibility of the gradient data are remarkably enhanced, so that the overall improvement of the federated learning system in the aspects of safety, fairness and credible cooperation is realized.

Description

Technical Field

[0001] This invention relates to the field of privacy computing technology, and in particular to a blockchain-based federated learning incentive system. Background Technology

[0002] With increasing awareness of data privacy, traditional centralized machine learning faces numerous challenges. Federated learning (FL), as a distributed collaborative training method, has been widely researched and applied. In recent years, to enhance the scalability and security of the system, more and more research has begun to explore the introduction of blockchain technology into federated learning systems to achieve advantages such as decentralized management of the training process, data traceability, and tamper-proof results. Among these advantages, using blockchain to record model updates and to transparently distribute incentives among participants has become a current research hotspot.

[0003] While existing research has attempted to design incentive mechanisms for federated learning using blockchain, certain limitations remain. Most current methods use changes in parameters across the entire model or at each level as the basis for contribution evaluation, lacking the ability to identify fine-grained structures (such as neuron-level) and failing to accurately reflect the true impact of different participants on model performance. Furthermore, existing methods often process incentives directly based on the original gradients, lacking sufficient protection for gradient privacy and failing to adequately address the potential "free-riding" problem among participants. Therefore, how to achieve a more refined and trustworthy incentive mechanism for federated learning while ensuring privacy and security has become a crucial issue that urgently needs to be addressed in this field. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a blockchain-based federated learning incentive system to address the problems of difficulty in accurately identifying the contributions of participants and insufficient protection of gradient privacy.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a blockchain-based federated learning incentive system, which includes a task module that receives federated learning task requests and generates a neuron-level contribution label matrix through a fractal probability function.

[0008] The gradient blinding module uses a neuron-level contribution label matrix to screen the initial backpropagation gradient to obtain the screened gradient, injects Gaussian noise to form a noisy gradient, and performs gradient blinding processing through a random orthogonal matrix to output the blinded gradient. It also uses elliptic curve bilinear mapping to generate gradient homomorphic commitments and constructs zero-knowledge proofs.

[0009] The verification module performs logical consistency verification on gradient homomorphic commitment and zero-knowledge proof. After the verification is successful, it parses the gradient norm in the gradient homomorphic commitment, calculates the number of effective contributing neurons in the neuron-level contribution label matrix, calculates the neuron-level contribution value in combination with the preset weight coefficients, and generates the federated learning reward value according to the secondary incentive rule.

[0010] The aggregation module calculates gradient similarity for blinded gradients, discards blinded gradients below the similarity threshold, and performs a weighted average of the remaining blinded gradients to generate global weight parameters.

[0011] The token distribution module mints incentive tokens through the blockchain network based on global weight parameters, federated learning reward values, and neuron-level contribution values, and distributes the incentive tokens to the federated learning client digital wallets through a cross-chain bridge, outputting token distribution records.

[0012] As a preferred embodiment of the blockchain-based federated learning incentive system of this invention, the specific steps for receiving federated learning task requests and generating a neuron-level contribution label matrix using a fractal probability function are as follows:

[0013] The task request is received through the blockchain event listening interface. After verifying the legality of the task through digital signature authentication, the sensitivity parameters and gradient magnitude threshold in the federated learning task are extracted.

[0014] The federated learning task is backpropagated using the backpropagation algorithm to obtain the neuron weight tensors, and the gradient magnitude of each neuron weight is calculated.

[0015] The sensitivity parameter is used as the initial perturbation factor to generate a nonlinear adjustment factor, and the gradient magnitude of the neuron weights is dynamically normalized by combining the gradient magnitude threshold. The contribution probability value of each neuron weight is generated by the fractal probability function and binarized using a quantum random number generator.

[0016] Based on the neural network connection relationships, a neural network connection topology graph is constructed, and the binarized labeling results are compressed and encoded into a sparse matrix format to output a neuron-level contribution labeling matrix.

[0017] As a preferred embodiment of the blockchain-based federated learning incentive system described in this invention, the specific steps for using a neuron-level contribution label matrix to screen neurons in the initial backpropagation gradient to obtain the screened gradient are as follows:

[0018] Based on the neuron-level contribution label matrix and the neural network connection topology, a fractal-topology joint encoder is constructed to generate a differentially differentiable selection matrix, which is then encoded into a quantum bit superposition state through quantum rotation gate operation.

[0019] Quantum entanglement is performed on the neurons of the superposition state of qubits based on the connection relationship of the neural network connection topology graph, and quantum measurement is performed on the entangled qubits. A quantum selection mask matrix is ​​generated by annealing threshold processing.

[0020] The initial backpropagation gradient is masked using a quantum selection masking matrix, and the unselected neuron weights in the masking operation are redirected using a graph convolutional network. The masking operation result and the gradient redirection operation are then fused to output the filtered gradient.

[0021] As a preferred embodiment of the blockchain-based federated learning incentive system described in this invention, the injection of Gaussian noise to form a noise gradient, and the gradient blinding processing using a random orthogonal matrix to output a blinded gradient, are detailed below.

[0022] The inner product matrix is ​​generated by transposing and multiplying the screening gradient, and the maximum and minimum eigenvalues ​​are obtained through eigenvalue decomposition. The geometric features of the gradient manifold are calculated, and conformal adaptation noise is generated.

[0023] The conformal adaptation noise and the screening gradient are added element-wise to generate a noise gradient. A true random seed is obtained from the blockchain oracle to initialize a random matrix. The random matrix is ​​then subjected to verifiable Jacobi iterative orthogonalization to generate a random orthogonal matrix. The verifiable orthogonalization parameters of the Jacobi iterative process are extracted.

[0024] Partial decryption is performed on the homomorphic encrypted weights uploaded by the client from the blockchain to generate a decryption scalar factor, which is then multiplied by a random diagonal matrix to generate a gradient confusion matrix.

[0025] Calculate the eigenvectors of the random orthogonal matrix and combine them with verifiable orthogonalization parameters to generate a bilinear binding signature. Combine the random orthogonal matrix with the gradient confusion matrix to perform a double-blinding transformation on the noisy gradient, obtain the double-blinding gradient, attach the bilinear binding signature, and output the blinded gradient.

[0026] As a preferred embodiment of the blockchain-based federated learning incentive system described in this invention, the specific steps for generating gradient homomorphic commitments using elliptic curve bilinear mapping and simultaneously constructing zero-knowledge proofs are as follows.

[0027] Singular value decomposition is performed on the blinded gradient to extract the left singular vector matrix, the right singular vector matrix, and the singular value diagonal matrix. Non-zero singular values ​​are extracted from the singular value diagonal matrix through a preset singular value truncation threshold to form a simplified singular value vector.

[0028] The simplified singular value vector is encoded into a polynomial, superimposed with the hash perturbation term of the bilinear binding signature, and the linear combination of the polynomial and the hash perturbation term is encrypted using the NTRU public key obtained through the blockchain smart contract to generate a gradient homomorphic commitment.

[0029] The orthogonality constraint between the left and right singular vector matrices is verified using the first-level proof. The validity of the first-level proof is recursively verified using the second-level proof. The first two-level proofs are then aggregated based on the third-level proof to generate a zero-knowledge proof.

[0030] As a preferred embodiment of the blockchain-based federated learning incentive system described in this invention, the specific steps for verifying the logical consistency of gradient homomorphic commitments and zero-knowledge proofs are as follows:

[0031] Extract the hash perturbation terms embedded in the gradient homomorphic commitment and the perturbation path information in the zero-knowledge proof, parse the corresponding index mapping relationship, and perform layer-by-layer backtracking verification of the constraint structure of each layer in the zero-knowledge proof to verify the reversible consistency of the gradient commitment generation chain.

[0032] After the verification is passed, the semantic consistency between the gradient homomorphic commitment and the zero-knowledge proof is jointly verified based on the bilinear binding signature contained in the gradient homomorphic commitment.

[0033] As a preferred embodiment of the blockchain-based federated learning incentive system described in this invention, the curvature parameters are calculated by eigenvalue decomposition based on the geometric features of the gradient manifold, and the gradient homomorphic commitment is decrypted in an anti-interference manner to output a denoised gradient singular value vector. The curvature weighted norm is constructed by combining the curvature parameters, and the gradient norm is calculated.

[0034] An entangled Hamiltonian is constructed based on the neuron-level contribution marker matrix and curvature parameters. Quantum annealing evolution is performed to output the neuron ground state probability distribution. The number of effective contributing neurons is obtained by performing non-zero probability amplitude statistics.

[0035] As a preferred embodiment of the blockchain-based federated learning incentive system of the present invention, the number of effective contributing neurons is matched with preset weight coefficients, the neuron-level contribution value of each effective contributing neuron is calculated, and the total contribution value is generated by accumulating them. According to the secondary incentive rule, the gradient norm is used as an adjustment factor and combined with the total contribution value to calculate the final federated learning reward value.

[0036] As a preferred embodiment of the blockchain-based federated learning incentive system described in this invention, the steps of calculating gradient similarity for blinded gradients, discarding blinded gradients below a similarity threshold, and performing a weighted average of the remaining blinded gradients to generate global weight parameters are as follows.

[0037] For each blinded gradient vector, perform random projection to generate a corresponding low-dimensional vector, and calculate the average similarity between each low-dimensional vector and other low-dimensional vectors. Compare the average similarity with a preset similarity threshold, discard blinded gradients that are below the similarity threshold, and select a set of reliable blinded gradient vectors.

[0038] Assign relative weights to the set of trustworthy blinding gradient vectors, and then sum up all the trustworthy blinding gradients with assigned relative weights to obtain the global weight parameters.

[0039] As a preferred embodiment of the blockchain-based federated learning incentive system described in this invention, the following steps are taken: based on global weight parameters, federated learning reward values, and neuron-level contribution values, a minting request is submitted to the blockchain through an off-chain aggregation service to trigger the minting operation of incentive tokens, and a cross-chain transfer certificate is generated and locked. The incentive tokens are then distributed to the digital wallets of the federated learning client through a cross-chain bridge, and a token distribution record is output.

[0040] The beneficial effects of this invention are as follows: By introducing a neuron-level contribution labeling mechanism based on fractal probability and quantum perturbation, accurate identification and sparse representation of key local neurons in the model are achieved, providing a high-granular basis for differentiated incentives in federated learning and improving the fairness and efficiency of incentive allocation; at the same time, by constructing a gradient double-blind mechanism combining conformal noise perturbation and a biorthogonal structure, coupled with bilinear binding signatures and blockchain true random seeds, the privacy protection capability and traceability credibility of gradient data are significantly enhanced, thereby achieving an overall improvement in the security, fairness, and trustworthy collaboration of the federated learning system. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a schematic diagram of a blockchain-based federated learning incentive system.

[0043] Figure 2 A flowchart for generating the neuron-level contribution label matrix.

[0044] Figure 3 A flowchart for integrating gradient selection and blinding processes.

[0045] Figure 4 A flowchart for token distribution and reward calculation. Detailed Implementation

[0046] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0047] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0048] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0049] Reference Figures 1-4 This is one embodiment of the present invention, which provides a blockchain-based federated learning incentive system, comprising the following steps:

[0050] The task module receives federated learning task requests and generates a neuron-level contribution label matrix using a fractal probability function.

[0051] The task request is received through the blockchain event listening interface. After verifying the legality of the task through digital signature authentication, the sensitivity parameters and gradient magnitude thresholds in the federated learning task are extracted.

[0052] Furthermore, after receiving the task request through the blockchain event listening interface and completing the digital signature authentication to verify the legality of the federated learning task, the sensitivity parameters and gradient magnitude thresholds for calculating the neuron weight tensors are extracted based on the structural information of the received federated learning task and the dimensions of the client's local training data and the neural network structure contained in the federated learning task.

[0053] The sensitivity parameter is used to measure the perturbation sensitivity of each neuron to the backpropagation gradient in the current training round. Specifically, by statistically analyzing the range of changes in the backpropagation gradient in each round of historical training, the mean and variance of the gradient change corresponding to each neuron are calculated. Then, combined with the weight distribution density of the layer where the neuron is located, the perturbation sensitivity weight of each neuron is constructed.

[0054] The gradient magnitude threshold is obtained by statistically analyzing the upper bound of the backpropagation gradient tensor of each neuron in the previous round or several rounds of training of all clients in the federated learning task, and dynamically truncating it in combination with the allowable error range specified by the federated learning task, so as to obtain the effective gradient screening threshold for each neuron.

[0055] It should also be noted that neurons originate from the neural network structure used in the federated learning task. Specifically, by analyzing the dimensional information of the neural network structure contained in the federated learning task, the number of input and output nodes in each layer is determined, and a complete set of neurons is configured according to these structural dimensions. The position of each neuron in the network is determined by its layer and its connection relationship.

[0056] The federated learning task is backpropagated using the backpropagation algorithm to obtain the neuron weight tensors, and the gradient magnitude of each neuron weight is calculated.

[0057] It should be noted that after extracting the sensitivity parameters and gradient magnitude thresholds in the federated learning task, the backpropagation algorithm is used to perform chain-like differentiation on the loss function of the neural network in the federated learning task, and the gradient values ​​of the weight parameters corresponding to each layer of neurons with respect to the loss function are calculated in turn. The set of gradient values ​​of all neurons is then represented as the neuron weight tensor.

[0058] This neuron weight tensor maintains structural consistency with the neural network weight tensor used in federated learning tasks, and its dimensions match the input and output structure of each layer of neurons. Based on the absolute value of each element in this neuron weight tensor, the magnitude of the gradient weight of each neuron in the current training round is calculated. Specifically, the magnitude is obtained by performing L2 norm operation on the gradient weight vector corresponding to the neuron in the tensor. The set of magnitudes of all neuron weights constitutes the complete gradient magnitude.

[0059] The sensitivity parameter is used as the initial perturbation factor to generate a nonlinear adjustment factor, and the gradient magnitude of the neuron weights is dynamically normalized by combining the gradient magnitude threshold. The contribution probability value of each neuron weight is generated by the fractal probability function and binarized using a quantum random number generator.

[0060] It should be noted that after obtaining the neuron weight tensor and calculating the gradient magnitude of each neuron weight, the sensitivity parameter in the federated learning task is used as the initial perturbation factor. Combined with a preset gradient magnitude threshold, the gradient magnitude of each neuron weight is dynamically normalized. The expression is as follows:

[0061]

[0062] In the formula, It is the normalized result of the gradient magnitude from the weight i of the previous layer to the weight j of the current layer neuron. It is the L2 norm of the gradient weight vector from the weight i of the previous layer to the weight j of the current layer neuron. is the gradient magnitude threshold, i is the index of the weight of the previous layer neuron, and j is the index of the weight of the current layer neuron;

[0063] Based on the dynamically normalized neuron gradient magnitude, the contribution probability value corresponding to each neuron is calculated using a fractal probability function with nonlinear adjustment capability, expressed as:

[0064]

[0065] In the formula, P ij λ is the contribution probability value from the weight i of the previous layer neuron to the weight j of the current layer neuron, and λ is the adjustment factor of the sensitivity parameter under nonlinear perturbation.

[0066] Based on the contribution probability value of each neuron, a quantum random number generator is used to generate a quantum random number between [0,1]. The quantum random number is then compared with the corresponding contribution probability value. If the quantum random number is less than or equal to the contribution probability value, the corresponding neuron weight is marked as "participating in uploading"; otherwise, it is marked as "not uploading".

[0067] It should also be noted that the gradient magnitude threshold is obtained by recording the L2 norm of the weights of each neuron in multiple communication rounds, calculating the mean and standard deviation of their distribution, and dynamically adjusting it in the form of "for example, the example value is the mean weighting coefficient multiplied by the standard deviation".

[0068] Based on the neural network connection relationships, a neural network connection topology graph is constructed, and the binarized labeling results are compressed and encoded into a sparse matrix format to output a neuron-level contribution labeling matrix.

[0069] It should be noted that, based on the layer number and intra-layer index position of each neuron in the neural network structure, the connection paths between neurons are identified, and directed edges are established in the graph structure through these connection paths to construct a complete neural network connection topology graph. The binarized labeling results of the neuron weights obtained by the quantum random number generator in the previous step are mapped to the corresponding nodes in the topology graph according to the neuron index position. Nodes with a binary label of "1" indicate that they have a significant contribution to the federated learning task, and nodes with a binary label of "0" indicate that they have no significant contribution.

[0070] To optimize storage and computation efficiency, the binarized labeling results on the nodes in the neural network connection topology are compressed and encoded using a sparse matrix format. Specifically, the positions and values ​​corresponding to all non-zero binary labels are arranged in order according to the neuron index position, and the position and value information are recorded using existing sparse matrix encoding methods such as COO format or CSR format, thereby forming a neuron-level contribution label matrix represented in sparse matrix format.

[0071] The gradient blinding module uses a neuron-level contribution label matrix to screen the initial backpropagation gradient to obtain the screened gradient, injects Gaussian noise to form a noisy gradient, and performs gradient blinding processing through a random orthogonal matrix to output the blinded gradient. It also uses elliptic curve bilinear mapping to generate gradient homomorphic commitments and constructs zero-knowledge proofs.

[0072] Based on the neuron-level contribution label matrix and the neural network connection topology, a fractal-topology joint encoder is constructed to generate a differentially differentiable selection matrix, which is then encoded into a superposition state of qubits through a quantum rotation gate operation.

[0073] It should be noted that, based on the neuron-level contribution label matrix and the neural network connection topology, by traversing the neuron index and its connection path, each pair of binary neurons with effective connections in the neural network is extracted, and combined with the corresponding binarization results in the neuron-level contribution label matrix, a joint weight map is constructed.

[0074] Based on the joint weight graph, a fractal-topology joint encoder is constructed by extracting the spatial self-similar structure of neuron distribution through an encoding method based on fractal geometric features and combining the structural sparsity of the connection relationship in the neural network connection topology graph.

[0075] The fractal-topology co-encoder uses non-zero elements in the neuron-level contribution label matrix as salient feature nodes and connection paths in the neural network connection topology graph as combined edge structures. It encodes and transforms salient feature nodes and associated edges through a combined mapping method to form a differentially differentiable selection matrix that jointly expresses structure and contribution. Each element in this differentially differentiable selection matrix represents the retention probability value of the neuron connection pair under the dual features of structure and contribution. This retention probability value can be used for quantum state encoding.

[0076] Each element in the differentially differentiable choice matrix is ​​mapped to a quantum rotation angle parameter. Based on common quantum rotation gate operations such as the RX rotation gate or the RY rotation gate, these rotation angles are applied to the initial qubit state, and the angle rotation is completed in the quantum circuit, so that the information represented by the differentially differentiable choice matrix is ​​encoded into the qubit superposition state.

[0077] It should also be noted that the connection relationships in the neural network connection topology graph are obtained by parsing the structural information of the neural network in the federated learning task, combining the number of input and output nodes and weight matrices of neurons in each layer, identifying which neurons have direct connections, and thus constructing a set of directed edges between neurons to form the basic connection relationships of the topology graph.

[0078] The connection paths in the neural network connection topology graph are obtained by continuously searching for the sequence of paths between neuron nodes along these connections using graph traversal algorithms (such as depth-first search or breadth-first search) based on the established connection relationships. This is used to represent the possible paths for the propagation of neural signals or gradients and to reflect the indirect connection pathways between nodes.

[0079] Quantum entanglement is performed on the neurons of the superposition state of qubits based on the connection relationship of the neural network connection topology graph, and quantum measurement is performed on the entangled qubits. A quantum selection mask matrix is ​​generated by annealing threshold processing.

[0080] It should be noted that, based on the connection relationship of the neural network connection topology, the corresponding qubit superposition state neurons are entangled through quantum entanglement operation according to the adjacency relationship of neurons in the topology graph. Specifically, quantum gates such as controlled NOT gates (CNOT gates) are used to realize the entanglement between qubits.

[0081] Quantum measurements are performed on entangled qubits, and the measurement results reflect the joint state information between neurons. The frequency or expected value of the output result after the measurement of each entangled qubit is extracted, and an annealing threshold is set, such as a dynamic critical value determined by simulated annealing or quantum annealing. Each measurement result is compared with the annealing threshold. If the frequency or expected value of a certain measurement result is greater than or equal to the annealing threshold, the corresponding position in the quantum selection mask matrix is ​​assigned a value of 1, otherwise it is assigned a value of 0, thereby generating a quantum selection mask matrix that only retains the results of high-fidelity qubit measurements.

[0082] The initial backpropagation gradient is masked using a quantum selection masking matrix, and the unselected neuron weights in the masking operation are redirected using a graph convolutional network. The masking operation result and the gradient redirection operation are then fused to output the filtered gradient.

[0083] It should be noted that the initial backpropagation gradient is masked element-wise using a quantum selection mask matrix, retaining the original backpropagation gradient value at the position corresponding to the value 1 in the mask matrix, and setting the gradient at the position corresponding to the value 0 in the mask matrix to zero.

[0084] The weight regions of neurons masked in the masking operation are processed by a graph convolutional network. The graph convolutional network takes the neural network connection topology graph as the input structure, and preserves the gradient information through the adjacency relationship between nodes. It calculates the semantic relevance and gradient context weight of the unselected neurons in the graph structure, and generates the gradient retargeting value corresponding to each unselected neuron based on the weight combination between adjacent neurons and the convolution result.

[0085] The initial backpropagation gradient preserved by the masking operation is weighted and fused with the redirected gradient output by the graph convolutional network according to the neuron index position, and the updated gradient corresponding to the neuron structure is output as the filtering gradient.

[0086] It should also be noted that the training of the graph convolutional network is based on the adjacency matrix constructed from the topology graph of the neural network connection and the gradient retained after the masking operation as input features. It adopts a supervised approach with the original complete gradient as the target output. Backpropagation training is performed by minimizing the mean squared error loss between the output of the graph convolutional network and the complete gradient, and the convolution weight parameters in the graph convolutional network are updated, so that the graph convolutional network can learn an effective gradient redirection strategy under the condition of missing mask information.

[0087] It should also be noted that the initial backpropagation gradient is the gradient information calculated by performing the standard backpropagation algorithm during the training of neural network parameters.

[0088] The inner product matrix is ​​generated by transposing and multiplying the screening gradient, and the maximum and minimum eigenvalues ​​are obtained through eigenvalue decomposition. The geometric features of the gradient manifold are calculated, and conformal fitting noise is generated.

[0089] It should be noted that performing a matrix transpose operation on the screening gradient and multiplying it by itself generates an inner product matrix with consistent dimensions. This inner product matrix represents the cooperative change relationship of the screening gradient across dimensions.

[0090] After the inner product matrix is ​​constructed, eigenvalue decomposition is performed to obtain all eigenvalues ​​and extract the largest and smallest eigenvalues. The largest eigenvalue represents the magnitude of change of the strongest principal axis in the gradient direction, and the smallest eigenvalue represents the magnitude of change of the weakest principal axis in the gradient direction.

[0091] Based on the ratio between the maximum and minimum eigenvalues, a manifold geometric feature quantity is constructed in the gradient space to represent the scalability of the selected gradient in the feature subspace. Based on the manifold geometric feature quantity, conformal fitting noise with the same principal axis direction but different scale perturbation characteristics is constructed through conformal mapping relationship.

[0092] The conformal adaptation noise and the screening gradient are added element-wise to generate a noise gradient. A true random seed is obtained from the blockchain oracle to initialize a random matrix. The random matrix is ​​then subjected to verifiable Jacobi iterative orthogonalization to generate a random orthogonal matrix. The verifiable orthogonalization parameters of the Jacobi iterative process are extracted.

[0093] It should be noted that the conformal adaptation noise and the screening gradient are added element-wise at corresponding positions to generate a noise gradient with spatial perturbation characteristics.

[0094] A true random seed is obtained from a blockchain oracle and used as the initial seed to generate a random matrix, ensuring that the random matrix has global unpredictability;

[0095] After the random matrix is ​​generated, the Jacobi iterative algorithm is used to gradually orthogonalize the random matrix. Through multiple rounds of matrix rotation operations, the random matrix is ​​gradually approximated to the identity orthogonal matrix, thereby improving the orthogonality of the random matrix in high-dimensional space.

[0096] During the Jacobi iteration process, verifiable orthogonalization parameters such as the angle of the rotation matrix, the position of the rotation axis, and the update error range are recorded in real time for each step. These parameters are used for subsequent orthogonality and randomness verification, ultimately obtaining a random orthogonal matrix with strict orthogonality properties.

[0097] Partial decryption is performed on the homomorphic encrypted weights uploaded by the client from the blockchain to generate a decryption scalar factor, which is then multiplied by a random diagonal matrix to generate a gradient confusion matrix.

[0098] It should be noted that the homomorphic encryption weights obtained from the blockchain and uploaded by the client are partially decrypted using verifiable homomorphic ciphertext processing technology. The corresponding decryption scalar factor is generated using the key generation function and the authorization information provided by the key distribution protocol. This decryption scalar factor represents the linear decryption estimate of the hidden gradient structure in the ciphertext.

[0099] Using a true random seed obtained from a blockchain oracle as initialization input, a random diagonal matrix with the same dimension as the decryption scalar factor is constructed using a pseudo-random number generation function. The diagonal elements are generated by sampling from a normal distribution controlled by the random seed, and all off-diagonal elements are set to zero to ensure that the matrix has orthogonal interference characteristics.

[0100] The random diagonal matrix is ​​multiplied element-wise by the generated decryption scalar factor according to the element index position. Multiplication is performed only on the corresponding element, while the other elements are always zero, forming a gradient confusion matrix that retains the relevant information of the original decryption structure while injecting controllable perturbations.

[0101] Calculate the eigenvectors of the random orthogonal matrix and combine them with verifiable orthogonalization parameters to generate a bilinear binding signature. Combine the random orthogonal matrix with the gradient confusion matrix to perform a double-blinding transformation on the noisy gradient, obtain the double-blinding gradient, attach the bilinear binding signature, and output the blinded gradient.

[0102] It should be noted that eigenvalue decomposition is performed on the random orthogonal matrix to extract eigenvectors as the intrinsic representation of the random orthogonal matrix in high-dimensional space; then, combined with the verifiable orthogonalization parameters generated by the Jacobi iteration process, a bilinear mapping structure is used to construct a bilinear binding signature, where the eigenvectors and orthogonalization parameters are used as the input vector pair and group mapping parameters of the signature generation function, respectively, to establish the corresponding bilinear relationship verification basis;

[0103] The random orthogonal matrix is ​​multiplied with the gradient confusion matrix obtained by the completed Hadamard product, and applied to the noise gradient output by the previous step. A linear projection transformation based on the random orthogonal matrix structure is performed on the noise gradient to form the first stage of blinding.

[0104] Then, through the orthogonal transformation direction generated by the feature vector, a second perturbation projection is applied to the first blinding result to form a second blinding, thus constituting a double blinding transformation process;

[0105] The result of the double-blinding transformation is output as the double-blinding gradient, and a bilinear binding signature generated by the eigenvectors of the random orthogonal matrix and the verifiable orthogonalization parameters is appended to output the blinding gradient.

[0106] Singular value decomposition is performed on the blinded gradient to extract the left singular vector matrix, the right singular vector matrix, and the singular value diagonal matrix. Non-zero singular values ​​are extracted from the singular value diagonal matrix through a preset singular value truncation threshold to form a simplified singular value vector.

[0107] It should be noted that the blinded gradient obtained after the double blinding transformation is subjected to singular value decomposition, and the blinded gradient is expressed as a matrix product of the left singular vector matrix, the singular value diagonal matrix, and the right singular vector matrix; where the left singular vector matrix is ​​used to represent the orthogonal basis of the blinded gradient in the column space, the right singular vector matrix is ​​used to represent the orthogonal basis of the blinded gradient in the row space, and the diagonal elements in the singular value diagonal matrix represent the scaling factor of each corresponding singular vector pair in the mapping direction;

[0108] After completing the singular value decomposition, each singular value in the singular value diagonal matrix is ​​filtered according to the preset singular value truncation threshold. Singular values ​​less than the singular value truncation threshold are set to zero, and only non-zero singular values ​​greater than or equal to the singular value truncation threshold are retained. All non-zero singular values ​​are extracted and arranged in their original order to form a one-dimensional simplified singular value vector.

[0109] It should also be noted that the singular value truncation threshold is selected by analyzing the magnitude of singular values ​​and their proportion in the overall energy contribution. This selects a singular value truncation threshold that can effectively remove noise and redundant information while retaining the main features.

[0110] The simplified singular value vector is encoded into a polynomial, superimposed with the hash perturbation term of the bilinear binding signature, and the linear combination of the polynomial and the hash perturbation term is encrypted using the NTRU public key obtained through the blockchain smart contract to generate a gradient homomorphic commitment.

[0111] It should be noted that the simplified singular value vector is regarded as the coefficient sequence of a polynomial. Each singular value in the vector is taken as the coefficient of the corresponding power term in the polynomial and arranged in order from low to high order. All singular value coefficients are combined with the corresponding power terms of the polynomial to construct a complete polynomial expression.

[0112] This polynomial expression is superimposed with the hash perturbation term in the bilinear binding signature according to the element correspondence to form a linear combination polynomial; the NTRU public key obtained by the blockchain smart contract is used to perform encryption operation on the linear combination polynomial, and homomorphic encryption is achieved through the NTRU encryption algorithm, finally generating a gradient homomorphic commitment with homomorphic operation capability.

[0113] It should also be noted that the hash perturbation term in the bilinear binding signature is obtained by hashing the eigenvectors of the random orthogonal matrix and the verifiable orthogonalization parameter.

[0114] The orthogonality constraint between the left and right singular vector matrices is verified using the first-level proof. The validity of the first-level proof is recursively verified using the second-level proof. The first two-level proofs are then aggregated based on the third-level proof to generate a zero-knowledge proof.

[0115] It should be noted that the first-level proof is achieved by calculating the product of the left singular vector matrix and its transpose, and the product of the right singular vector matrix and its transpose, comparing them with the corresponding identity matrix, obtaining the norm of the difference between the two, and generating the corresponding orthogonality constraint proof evidence, thus proving that the left singular vector matrix and the right singular vector matrix satisfy the orthogonality condition.

[0116] The second-level proof receives the evidence generated by the first-level proof and uses a recursive verification method to check the validity and consistency of the first-level proof through a kernel function or hash chain mechanism, ensuring the integrity and accuracy of the orthogonality constraint verification process.

[0117] The third-level proof merges the proof evidence from the first and second levels according to a preset aggregation algorithm to generate a unified zero-knowledge proof output.

[0118] The verification module performs logical consistency verification on gradient homomorphic commitment and zero-knowledge proof. After the verification is successful, it parses the gradient norm in the gradient homomorphic commitment, calculates the number of effective contributing neurons in the neuron-level contribution label matrix, calculates the neuron-level contribution value in combination with the preset weight coefficients, and generates the federated learning reward value according to the secondary incentive rule.

[0119] Extract the hash perturbation terms embedded in the gradient homomorphic commitment and the perturbation path information in the zero-knowledge proof, parse the corresponding index mapping relationship, and perform layer-by-layer backtracking verification of the constraint structure of each layer in the zero-knowledge proof to verify the reversible consistency of the gradient commitment generation chain.

[0120] It should be noted that the embedded hash perturbation term is extracted from the gradient homomorphic commitment, and the perturbation path information is extracted from the zero-knowledge proof. By parsing the index mapping relationship between the two, the index corresponding to the hash perturbation term is matched with the node index of each layer of constraint structure in the zero-knowledge proof. The constraint structure of each layer in the zero-knowledge proof is verified by backtracking layer by layer. By comparing the consistency between the perturbation path information and the constraint node index, the reversibility and consistency of the gradient homomorphic commitment generation chain under each layer of constraint conditions are verified.

[0121] After the verification is passed, the semantic consistency between the gradient homomorphic commitment and the zero-knowledge proof is jointly verified based on the bilinear binding signature contained in the gradient homomorphic commitment.

[0122] It should be noted that, based on the bilinear binding signature contained in the gradient homomorphic commitment, the bilinear binding signature is decoded and the signature content is extracted. The extracted signature is then matched and verified with the gradient homomorphic commitment data using a signature verification algorithm to confirm the consistency between the signature and the commitment data.

[0123] The verified bilinear binding signature is jointly compared with the relevant semantic information in the zero-knowledge proof. The matching relationship between the two in terms of expression content, structural features and constraints is verified item by item to ensure that the gradient homomorphic commitment and the zero-knowledge proof are completely consistent at the semantic level, thereby realizing the credibility and completeness of the joint verification result.

[0124] Based on the geometric features of the gradient manifold, the curvature parameters are calculated through eigenvalue decomposition, and the gradient homomorphic commitment is decrypted in an anti-interference manner to output the denoised gradient singular value vector. The curvature weighted norm is constructed by combining the curvature parameters, and the gradient norm is calculated.

[0125] It should be noted that the original gradient tensor before denoising in the gradient homomorphic commitment is regarded as a gradient dataset. The covariance matrix of the gradient dataset is constructed by multiplying the gradient dataset with its own transpose. The covariance matrix of the gradient in the manifold space is then calculated.

[0126] Perform eigenvalue decomposition on the covariance matrix to extract the largest and smallest eigenvalues. Use the trace of the covariance matrix as the numerator and the difference between the largest and smallest eigenvalues ​​as the denominator to calculate the curvature parameter, expressed as:

[0127]

[0128] In the formula, κ is the curvature parameter, T is the matrix trace operation, and G... σ It is the original gradient tensor, ω max It is the largest eigenvalue, ω min It is the smallest eigenvalue;

[0129] After calculating the curvature parameters, the gradient homomorphic commitment is decrypted. This decryption process has anti-interference characteristics and can effectively remove abnormal information introduced by environmental noise, external attacks or disturbances. After completing the decryption and obtaining the denoised true gradient tensor, the first few dominant singular values ​​are extracted using the singular value decomposition method and a singular value vector is formed.

[0130] The denoised singular value vector is combined with the obtained curvature parameters to construct the curvature weighted norm. Based on the sum of squared singular values, a weighted term between the curvature parameters and the absolute values ​​of the singular values ​​is further introduced, and finally the gradient norm is obtained through square root operation.

[0131] An entangled Hamiltonian is constructed based on the neuron-level contribution marker matrix and curvature parameters. Quantum annealing evolution is performed to output the neuron ground state probability distribution. The number of effective contributing neurons is obtained by performing non-zero probability amplitude statistics.

[0132] It should be noted that the neuron-level contribution label matrix is ​​analyzed to clarify the quantified contribution value of each neuron in the gradient propagation process. Combining the curvature parameters obtained in the previous steps, the curvature parameters are applied as weighting factors to each weight position in the neuron-level contribution label matrix, so that the overall contribution information not only reflects the local influence of the neuron, but also considers its degree of nonlinear change in the gradient manifold space, and constructs the corresponding entangled Hamiltonian.

[0133] Using the constructed entangled Hamiltonian as the initial state for quantum annealing, a quantum annealing evolution process is executed. This evolution, through controlling quantum fluctuations, iteratively searches for the lowest energy state corresponding to the entangled Hamiltonian. After the evolution is complete, the corresponding neuron ground state probability distribution is output.

[0134] The ground state probability distribution of neurons is statistically processed to screen out the locations of all neurons with a probability amplitude greater than zero. The number of neurons that meet the non-zero probability amplitude condition is counted to obtain the number of effective contributing neurons.

[0135] It should also be noted that the entangled Hamiltonian is represented as an energy function in a quantum state space. Its structure includes quantum entanglement terms between neurons and energy offset terms corresponding to each neuron. The entanglement terms are used to represent the interaction relationship between neurons, and the energy offset terms are used to reflect the local energy state of individual neurons.

[0136] The number of effective contributing neurons is matched with the preset weight coefficients, the neuron-level contribution value of each effective contributing neuron is calculated, and the total contribution value is generated by accumulating them. According to the secondary incentive rule, the gradient norm is used as an adjustment factor and combined with the total contribution value to calculate the final federated learning reward value.

[0137] It should be noted that the number of effective contributing neurons obtained through non-zero probability amplitude statistics will be correlated with the corresponding preset weight coefficients. Based on the weight coefficients corresponding to each neuron, the neuron-level contribution value of each effective contributing neuron will be calculated.

[0138] The total contribution value is obtained by summing up the contribution values ​​of all neurons.

[0139] According to the secondary incentive rule, the calculated gradient norm is used as an adjustment factor and input into the total contribution value for functional combination calculation. Finally, the combined value is used as the federated learning reward value for allocating incentive resources.

[0140] It should also be noted that the secondary incentive rule is a strategy based on the nonlinear enhancement relationship between the number of effective contributing neurons and their corresponding neuron-level contribution values. The secondary incentive rule uses the gradient norm as a moderating factor on the basis of the initial reward mechanism, and nonlinearly combines this moderating factor with the total contribution value to achieve a weighted increase in rewards for high-contributing participants.

[0141] The secondary incentive rule is based on the differences in the gradients provided by different participants in the geometric space during the federated learning process. Combined with the curvature weighted norm determined by the geometric characteristics of the gradient manifold, it is used to dynamically regulate the matching relationship between contributions and incentives.

[0142] The aggregation module calculates gradient similarity for blinded gradients, discards blinded gradients below a similarity threshold, and performs a weighted average on the remaining blinded gradients to generate global weight parameters.

[0143] For each blinding gradient, a random projection is performed to generate a corresponding low-dimensional vector. The average similarity between each low-dimensional vector and other low-dimensional vectors is calculated. The average similarity is compared with a preset similarity threshold. Blinding gradients below the similarity threshold are discarded, and a set of reliable blinding gradient vectors is selected.

[0144] It should be noted that a dimensionality compression process is performed on each blinding gradient, and the original high-dimensional blinding gradient is converted into a corresponding low-dimensional vector using a random projection method.

[0145] After completing the projection operation of all blinded gradients, for each low-dimensional vector, calculate the similarity index between each low-dimensional vector and all other low-dimensional vectors, and average this similarity index to obtain the average similarity corresponding to each low-dimensional vector.

[0146] Each average similarity value obtained is compared with a preset similarity threshold, and only blinded gradients with average similarity values ​​higher than the similarity threshold are retained, forming a set of credible blinded gradient vectors.

[0147] It should also be noted that the similarity threshold is determined by statistically analyzing the average similarity distribution among the blinded gradients uploaded by each participant in the historical multi-round federated learning, combined with the similarity stability characteristics that a trustworthy gradient vector should possess.

[0148] Assign relative weights to the set of trustworthy blinding gradient vectors, and then sum up all the trustworthy blinding gradients with assigned relative weights to obtain the global weight parameters.

[0149] It should be noted that a relative weight is assigned to each reliable blinded gradient vector. This relative weight is determined based on the reliability or contribution of each blinded gradient. All reliable blinded gradient vectors assigned relative weights are weighted and accumulated according to their corresponding relative weights. After weighted accumulation calculation, the global weight parameters used for subsequent federated learning are obtained.

[0150] The token distribution module mints incentive tokens through the blockchain network based on global weight parameters, federated learning reward values, and neuron-level contribution values, and distributes the incentive tokens to the federated learning client digital wallets through a cross-chain bridge, outputting token distribution records.

[0151] Based on global weight parameters, federated learning reward values, and neuron-level contribution values, the minting request is submitted to the blockchain through an off-chain aggregation service, triggering the incentive token minting operation. A cross-chain transfer certificate is then generated and distributed to the federated learning client digital wallet via a cross-chain bridge, outputting the token distribution record.

[0152] The global weight parameters, federated learning reward value, and neuron-level contribution value are used as inputs to generate minting request information. The minting request information is then uniformly aggregated and formatted by the off-chain aggregation service and submitted to the smart contract of the blockchain network.

[0153] The smart contract executes the incentive token minting operation according to preset rules, generates a corresponding number of incentive tokens, creates cross-chain transfer certificates and locks their state to prevent reuse, listens for certificate generation events on the blockchain network through the cross-chain bridge, verifies the validity of the certificates, and sends the incentive tokens to the corresponding digital wallet address of the federated learning client through the cross-chain communication protocol, thus completing the secure transmission and distribution of the tokens.

[0154] In summary, this invention achieves accurate identification and sparse representation of key local neurons in the model by introducing a neuron-level contribution labeling mechanism based on fractal probability and quantum perturbation. This provides a high-granular basis for differentiated incentives in federated learning, improving the fairness and efficiency of incentive allocation. Simultaneously, by constructing a gradient double-blind mechanism combining conformal noise perturbation and a biorthogonal structure, coupled with bilinear binding signatures and blockchain true random seeds, the privacy protection and traceability of gradient data are significantly enhanced. This results in an overall improvement in the security, fairness, and trustworthy collaboration of the federated learning system.

[0155] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A blockchain-based federated learning incentive system, characterized in that: include, The task module receives federated learning task requests and generates a neuron-level contribution label matrix using a fractal probability function. The gradient blinding module uses a neuron-level contribution label matrix to screen the initial backpropagation gradient to obtain the screened gradient, injects Gaussian noise to form a noisy gradient, and performs gradient blinding processing through a random orthogonal matrix to output the blinded gradient. It also uses elliptic curve bilinear mapping to generate gradient homomorphic commitments and constructs zero-knowledge proofs. The verification module performs logical consistency verification on gradient homomorphic commitment and zero-knowledge proof. After the verification is successful, it parses the gradient norm in the gradient homomorphic commitment, calculates the number of effective contributing neurons in the neuron-level contribution label matrix, calculates the neuron-level contribution value in combination with the preset weight coefficients, and generates the federated learning reward value according to the secondary incentive rule. The aggregation module calculates gradient similarity for blinded gradients, discards blinded gradients below the similarity threshold, and performs a weighted average of the remaining blinded gradients to generate global weight parameters. The token distribution module mints incentive tokens through the blockchain network based on global weight parameters, federated learning reward values, and neuron-level contribution values, and distributes the incentive tokens to the federated learning client digital wallets through a cross-chain bridge, outputting token distribution records.

2. The blockchain-based federated learning incentive system as described in claim 1, characterized in that: The specific steps for generating the neuron-level contribution label matrix are as follows: The task request is received through the blockchain event listening interface. After verifying the legality of the task through digital signature authentication, the sensitivity parameters and gradient magnitude threshold in the federated learning task are extracted. The federated learning task is backpropagated using the backpropagation algorithm to obtain the neuron weight tensors, and the gradient magnitude of each neuron weight is calculated. The sensitivity parameter is used as the initial perturbation factor to generate a nonlinear adjustment factor, and the gradient magnitude of the neuron weights is dynamically normalized by combining the gradient magnitude threshold. The contribution probability value of each neuron weight is generated by the fractal probability function and binarized using a quantum random number generator. Based on the neural network connection relationships, a neural network connection topology graph is constructed, and the binarized labeling results are compressed and encoded into a sparse matrix format to output a neuron-level contribution labeling matrix.

3. The blockchain-based federated learning incentive system as described in claim 1, characterized in that: The specific steps for obtaining the screening gradient are as follows. Based on the neuron-level contribution label matrix and the neural network connection topology, a fractal-topology joint encoder is constructed to generate a differentially differentiable selection matrix, which is then encoded into a quantum bit superposition state through quantum rotation gate operation. Quantum entanglement is performed on the neurons of the superposition state of qubits based on the connection relationship of the neural network connection topology graph, and quantum measurement is performed on the entangled qubits. A quantum selection mask matrix is ​​generated by annealing threshold processing. The initial backpropagation gradient is masked using a quantum selection masking matrix, and the unselected neuron weights in the masking operation are redirected using a graph convolutional network. The masking operation result and the gradient redirection operation are then fused to output the filtered gradient.

4. The blockchain-based federated learning incentive system as described in claim 1, characterized in that: The specific steps for blinding the output gradient are as follows. The inner product matrix is ​​generated by transposing and multiplying the screening gradient, and the maximum and minimum eigenvalues ​​are obtained through eigenvalue decomposition. The geometric features of the gradient manifold are calculated, and conformal adaptation noise is generated. The conformal adaptation noise and the screening gradient are added element-wise to generate a noise gradient. A true random seed is obtained from the blockchain oracle to initialize a random matrix. The random matrix is ​​then subjected to verifiable Jacobi iterative orthogonalization to generate a random orthogonal matrix. The verifiable orthogonalization parameters of the Jacobi iterative process are extracted. Partial decryption is performed on the homomorphic encrypted weights uploaded by the client from the blockchain to generate a decryption scalar factor, which is then multiplied by a random diagonal matrix to generate a gradient confusion matrix. Calculate the eigenvectors of the random orthogonal matrix and combine them with verifiable orthogonalization parameters to generate a bilinear binding signature. Combine the random orthogonal matrix with the gradient confusion matrix to perform a double-blinding transformation on the noisy gradient, obtain the double-blinding gradient, attach the bilinear binding signature, and output the blinded gradient.

5. The blockchain-based federated learning incentive system as described in claim 1, characterized in that: The method of generating gradient homomorphic commitments using elliptic curve bilinear mapping and constructing zero-knowledge proofs involves the following specific steps. Singular value decomposition is performed on the blinded gradient to extract the left singular vector matrix, the right singular vector matrix, and the singular value diagonal matrix. Non-zero singular values ​​are extracted from the singular value diagonal matrix through a preset singular value truncation threshold to form a simplified singular value vector. The simplified singular value vector is encoded into a polynomial, superimposed with the hash perturbation term of the bilinear binding signature, and the linear combination of the polynomial and the hash perturbation term is encrypted using the NTRU public key obtained through the blockchain smart contract to generate a gradient homomorphic commitment. The orthogonality constraint between the left and right singular vector matrices is verified using the first-level proof. The validity of the first-level proof is recursively verified using the second-level proof. The first two-level proofs are then aggregated based on the third-level proof to generate a zero-knowledge proof.

6. The blockchain-based federated learning incentive system as described in claim 5, characterized in that: The specific steps for verifying the logical consistency of gradient homomorphic commitments and zero-knowledge proofs are as follows. Extract the hash perturbation terms embedded in the gradient homomorphic commitment and the perturbation path information in the zero-knowledge proof, parse the corresponding index mapping relationship, and perform layer-by-layer backtracking verification of the constraint structure of each layer in the zero-knowledge proof to verify the reversible consistency of the gradient commitment generation chain. After the verification is passed, the semantic consistency between the gradient homomorphic commitment and the zero-knowledge proof is jointly verified based on the bilinear binding signature contained in the gradient homomorphic commitment.

7. The blockchain-based federated learning incentive system as described in claim 1, characterized in that: Based on the geometric features of the gradient manifold, the curvature parameters are calculated through eigenvalue decomposition, and the gradient homomorphic commitment is decrypted in an anti-interference manner to output a denoised gradient singular value vector. The curvature weighted norm is constructed by combining the curvature parameters, and the gradient norm is calculated. An entangled Hamiltonian is constructed based on the neuron-level contribution marker matrix and curvature parameters. Quantum annealing evolution is performed to output the neuron ground state probability distribution. The number of effective contributing neurons is obtained by performing non-zero probability amplitude statistics.

8. The blockchain-based federated learning incentive system as described in claim 1, characterized in that: The number of effective contributing neurons is matched with the preset weight coefficients, the neuron-level contribution value of each effective contributing neuron is calculated, and the total contribution value is generated by accumulating them. According to the secondary incentive rule, the gradient norm is used as an adjustment factor and combined with the total contribution value to calculate the final federated learning reward value.

9. The blockchain-based federated learning incentive system as described in claim 1, characterized in that: The specific steps for generating global weight parameters are as follows: For each blinded gradient vector, perform random projection to generate a corresponding low-dimensional vector, and calculate the average similarity between each low-dimensional vector and other low-dimensional vectors. Compare the average similarity with a preset similarity threshold, discard blinded gradients that are below the similarity threshold, and select a set of reliable blinded gradient vectors. Assign relative weights to the set of trustworthy blinding gradient vectors, and then sum up all the trustworthy blinding gradients with assigned relative weights to obtain the global weight parameters.

10. The blockchain-based federated learning incentive system as described in claim 1, characterized in that: Based on global weight parameters, federated learning reward values, and neuron-level contribution values, the minting request is submitted to the blockchain through an off-chain aggregation service, triggering the incentive token minting operation. A cross-chain transfer certificate is then generated and distributed to the federated learning client digital wallet via a cross-chain bridge, outputting the token distribution record.

Images