A blockchain-based trusted distributed hybrid expert model edge computing system
By introducing blockchain technology into the distributed MoE system to record and verify data interaction behavior, the problem of data tampering caused by lack of trust between distributed edge nodes is solved, thereby improving the system's security and learning performance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV OF SCI & TECH
- Filing Date
- 2025-09-16
- Publication Date
- 2026-05-26
Smart Images
Figure CN121168529B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of machine learning and edge computing technology, specifically to a blockchain-based trusted distributed hybrid expert model edge computing system. Background Technology
[0002] In fields such as artificial intelligence, computer vision, and natural language processing, improving model performance often relies on expanding the number of parameters and optimizing computational efficiency. Currently, Mixture of Experts (MoE) models are widely used in neural network architecture design. This model consists of a set of independent neural networks, which are divided into a gating network and multiple expert networks based on their functions. Specifically, the core working mechanism of this MoE model is as follows: The gating network first calculates the weight values of each expert network for each input task, and then performs sparse activation based on these weight values, selecting the top K expert networks with the largest weight values for activation. Subsequently, the activated expert networks process the input task, and the gating network or related processing modules weight and aggregate the calculation results of each expert network according to their corresponding weight values, ultimately generating the output of the MoE model. In recent years, MoE models have become widely adopted in mainstream large language model architectures, such as the Mistral model and the DeepSeek-MoE model. This is mainly due to the technical advantages of hybrid expert models: they can expand the number of model parameters to improve the model's expressive power while avoiding the introduction of too much additional computational overhead, thus enabling the model to achieve learning performance comparable to densely activated neural network models under the premise of controllable computational cost.
[0003] Currently, mainstream MoE systems are typically deployed using cloud servers equipped with sufficient computing resources. However, such cloud-based MoE frameworks suffer from the following technical drawbacks in practical applications: First, data transmission and processing are prone to prolonged response latency; second, interaction between the cloud and the terminal leads to high bandwidth consumption; third, centralized data processing in the cloud poses a potential risk of privacy leaks; and fourth, the hardware configuration and maintenance costs of cloud servers are high, resulting in high overall deployment costs. To address these issues, reference 1 (Wang J, Du H, Niyato D, et al. Toward scalable generative ai via mixture of experts in mobile edge networks[J]. IEEE Wireless Communications, 2024.) has proposed a technical solution for deploying MoE systems in distributed edge intelligent networks. The core advantages of this technical solution are: the nodes of the edge intelligent network are deployed closer to the data source, enabling near-end data processing, thereby significantly reducing response latency and bandwidth consumption; simultaneously, the near-end processing mode reduces the number of cross-node transmission steps for raw data, improving data privacy protection capabilities; furthermore, the procurement and maintenance costs of computing power for edge nodes are significantly lower than those of cloud servers, effectively reducing the economic cost of system deployment. Figure 2 The diagram illustrates a distributed MoE architecture for edge intelligent networks. It should be noted that edge nodes typically have limited computing power and storage capacity. Therefore, in this architecture, the gating network and expert network are deployed in a distributed manner, i.e., dispersed across a group of edge nodes. Data interaction between the gating network and expert network in the MoE system is achieved through a cross-node communication mechanism.
[0004] However, the aforementioned distributed MoE architecture for edge intelligent networks still faces specific security challenges in practical applications: in scenarios lacking supervision from trusted authorities, data interactions between distributed edge nodes are vulnerable to data tampering attacks (such as malicious perturbation attacks and model poisoning attacks) due to the lack of trust mechanisms. Specifically, malicious edge nodes can launch attacks to impair the learning performance of the MoE system in the following ways: first, by maliciously manipulating the computation results of gating and experts; second, by illegally tampering with the model parameters of gating and experts. Currently, research on security issues arising from untrusted data interactions in distributed MoE networks remains relatively scarce. Summary of the Invention
[0005] To address the aforementioned problems, this invention provides a blockchain-based trusted distributed hybrid expert model edge computing system, comprising a task publisher, a blockchain network, an edge layer, and a storage layer, with a gating network deployed within the blockchain network; wherein:
[0006] The task publisher is used to publish learning tasks, which are either training tasks or inference tasks.
[0007] The blockchain network is used to receive learning tasks and perform expert weight evaluation on the learning tasks to obtain sparse activation expert index results; it is also used to receive the learning task processing results of the edge layer and perform consensus.
[0008] The edge layer is used to download the corresponding expert model from the storage layer to process the learning task based on the sparse activation expert index result, and upload the learning task processing result to the blockchain network.
[0009] The storage layer is used to store all expert networks, and each expert network has a CID.
[0010] The beneficial effects of this invention are:
[0011] This invention introduces blockchain technology, enabling the data interaction behavior of each node in the distributed MoE network to be recorded on the blockchain in the form of blocks. Because blockchain is immutable and traceable, it can effectively record and track the data interaction process, providing support for the security of the distributed MoE network from the perspective of data interaction recording.
[0012] The blockchain network in this invention enables coordinated control of the edge layer and storage layer. It can verify the computation results of off-chain edge nodes and reach a global consensus on the verified and trusted computation results. Through this approach, the distributed MoE network's defense capabilities against malicious edge node data tampering attacks can be significantly improved, thereby ensuring the learning performance of the MoE system. Attached Figure Description
[0013] Figure 1 This is a schematic diagram of the method of the present invention;
[0014] Figure 2 This is a schematic diagram of a distributed MoE architecture for edge intelligent networks;
[0015] Figure 3 This is a schematic diagram of the system architecture of the present invention;
[0016] Figure 4 This is a bar chart showing the activation ratio of each expert in a traditional distributed MoE, where 'Y' represents data tampering attacks and 'N' represents no data tampering attacks.
[0017] Figure 5 This is a comparison chart of the training performance of the present invention and the traditional distributed MoE;
[0018] Figure 6 This is a comparison chart of the latency of the present invention and traditional distributed MoE;
[0019] Figure 7 This is a comparison chart of the inference performance of this invention and traditional distributed MoE. Detailed Implementation
[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0021] Figure 3 This is a structural diagram of a blockchain-based trusted distributed hybrid expert model edge computing system, as shown in some embodiments of the present invention.
[0022] Some embodiments of this invention propose a blockchain-based trusted distributed hybrid expert model edge computing system, such as... Figure 3 As shown, it includes task publishers, a blockchain network, an edge layer, and a storage layer, with a gating network deployed within the blockchain network.
[0023] In a trusted distributed hybrid expert model edge computing system:
[0024] The task publisher is used to publish learning tasks, which are either training tasks or inference tasks.
[0025] In some embodiments, the training task is the training data used in the training process. The training process refers to the process of adjusting the model parameters of the MoE through a large amount of training data to minimize the error between the predicted output of the MoE and the true label. The model parameters of the MoE include the trainable parameters of the gating network and the expert network, such as weight parameters and bias parameters. The training process typically uses the backpropagation algorithm and gradient descent optimizer to iteratively update the model parameters of the MoE, allowing the MoE to gradually learn the patterns in the training data. Typically, a training task includes multiple training samples, which can be original sample data, preprocessed sample features, or the output results after processing by a preceding neural network.
[0026] Inference tasks are the inference data used in the inference process. The inference process refers to the process of using a trained Model of Effect (MoE) with fixed model parameters to perform forward computation on the input data to obtain a prediction result. During inference, the MoE is used only as a function mapping. Inference tasks may include new sample data, user requests to be processed, or prompts, etc.
[0027] The blockchain network is used to receive learning tasks and perform expert weight evaluation on the learning tasks to obtain sparse activation expert index results; it is also used to receive the learning task processing results of the edge layer and perform consensus.
[0028] In some embodiments, after receiving a learning task, the blockchain network performs expert weight evaluation on the learning task to obtain a sparse activation expert index result, including:
[0029] The deployed gating network generates an expert weight for each expert network based on the learning task, which is used to characterize the expert network's suitability for handling the learning task. The larger the expert weight value, the more suitable the expert network is.
[0030] By employing a sparse selection strategy, several expert networks with high expert weights are selected from all expert networks for activation and participation in subsequent calculations. The CIDs of the selected expert networks are recorded to obtain the sparse activation expert index result.
[0031] The edge layer has sufficient computing resources to download the corresponding expert model from the storage layer based on the sparse activation expert index result to process the learning task, and upload the learning task processing result to the blockchain network.
[0032] The storage layer stores all expert networks, each with a unique Content Identifier (CID). Specifically, the blockchain network stores CIDs, which can be used to query or download expert networks within the storage layer.
[0033] In some embodiments, the storage layer consists of a set of distributed storage nodes to provide decentralized storage services. It may employ Interplanetary File System (IPFS) technology for distributed storage, or other storage technologies may be adopted or developed according to the developer's needs.
[0034] In some embodiments, data interactions between different layers are automatically triggered by smart contracts, such as task downloading and computation result uploading between the edge layer and the blockchain network, expert downloading and uploading between the edge layer and the storage layer, and CID generation of the expert network within the storage layer.
[0035] In some embodiments, such as Figure 1As shown, the workflow of the trusted distributed hybrid expert model edge computing system includes:
[0036] S1. The task publisher posts a learning task;
[0037] S2. Gated Evaluation: The blockchain network receives the learning task and uses an on-chain gating network to evaluate expert weights to obtain the sparse activation expert index result. The blockchain records the learning task and the sparse activation expert index result; the sparse activation expert index result includes the CID of all activated expert networks.
[0038] S3. Expert Computation: The edge layer downloads the learning task and sparse activation expert index results from the blockchain network, downloads the expert network from the storage layer to process the learning task based on the sparse activation expert index results, and uploads the learning task processing results to the blockchain network.
[0039] S4.MoE Output: The blockchain network receives the learning task processing results uploaded by the edge layer and performs distributed consensus to obtain the aggregated results of the hybrid expert model;
[0040] S5.MoE Update: The blockchain network generates loss function values and updates the model parameters of the on-chain gating network using the gradient descent algorithm; the edge layer downloads the loss function values and updates the model parameters of all activated expert networks using the gradient descent algorithm, generating expert hash values and uploading them to the blockchain network.
[0041] S6. Expert Storage: The blockchain network receives expert hash values and performs distributed consensus to obtain trusted expert hash values and uploads them to the storage layer; the storage layer stores the trusted expert hash values and generates CIDs for them.
[0042] S7. Block Generation: The blockchain network generates new blocks and uploads them to the chain.
[0043] In some embodiments, the edge layer includes a set of edge nodes, and step S3 specifically includes:
[0044] Each edge node downloads all activated expert network processing learning tasks from the storage layer based on the sparse activation expert index results, generates a learning task processing result, and uploads the learning task processing result to the blockchain network.
[0045] The learning task processing result generated by an edge node includes the processing results of all activated expert networks in the current edge node processing the learning task.
[0046] In some embodiments, the blockchain network includes a set of blockchain nodes. The blockchain network receives the learning task processing results of all edge nodes and broadcasts them to all blockchain nodes. Based on the learning task processing results, a distributed consensus is performed to sequentially obtain the trusted processing results of all activated expert networks. Each blockchain node aggregates the trusted processing results of all activated expert networks to obtain a hybrid expert model aggregation result.
[0047] Edge nodes are vulnerable to data tampering attacks. The calculation results of malicious edge nodes that are attacked are usually different from those of honest edge nodes that are not attacked, while the calculation results of different honest edge nodes that are not attacked are consistent. Therefore, distributed consensus can be used to enable the blockchain network to distinguish between malicious edge nodes and honest edge nodes, filter out untrusted calculation results given by malicious edge nodes, and reach a global consensus on trustworthy calculation results.
[0048] For example, the distributed consensus adopted by blockchain networks is compatible with mainstream consensus protocols, such as Proof of Work (PoW) and Practical Byzantine Fault Tolerance (PBFT); developers can also design their own consensus protocols according to their needs.
[0049] For example, the process of obtaining a trusted processing result from an activated expert network through distributed consensus can be represented as:
[0050] A specific blockchain node compares the processing results of the expert network across different edge nodes, and uses the processing result that is the same for most edge nodes as the candidate trusted processing result of the expert network, thus initiating a proposal. It is important to note that the role of blockchain nodes differs in different consensus algorithms. For example, in PBFT, nodes are divided into primary and replica nodes, with consensus typically initiated by the primary node. In PBFT, a specific blockchain node refers to the primary node.
[0051] The remaining blockchain nodes validate the proposal by sending approval messages (such as Raft's "AppendEntries" response and PBFT's "prepare" message) to express their support for the proposal.
[0052] When a proposal receives support from a sufficient number of blockchain nodes (e.g., more than half or 2 / 3), the system determines that the proposal has reached a consensus, and all blockchain nodes write the proposed candidate trusted processing result as the trusted processing result of the expert network into their local state.
[0053] In some embodiments, in step S5, each edge node of the edge layer downloads the loss function value and updates the model parameters of all activated expert networks using the gradient descent algorithm, while generating an expert hash value for the model parameters of each activated expert network.
[0054] For example, suppose K' expert networks are selected according to the current learning task, and each edge node downloads these K' expert networks; when performing step S5, an edge node downloads the loss function value and uses the gradient descent algorithm to update the model parameters of the K' expert networks it has downloaded, thereby generating K' expert hash values.
[0055] In some embodiments, the blockchain network receives all expert hash values from all edge nodes and broadcasts them to all blockchain nodes. Based on distributed consensus, it sequentially obtains the trusted expert hash values of all activated expert networks and uploads them to the storage layer for storage. The storage layer also generates a CID for each trusted expert hash value of an activated expert network.
[0056] For example, the process of obtaining the trusted expert hash value of an activated expert network through distributed consensus can be represented as:
[0057] A specific blockchain node initiates a proposal by comparing the expert hash values of the expert network on different edge nodes and using the expert hash value that is the same on most edge nodes as the candidate trusted expert hash value of the expert network.
[0058] The remaining blockchain nodes verify the proposal, expressing their support for it by sending an approval message.
[0059] When a proposal gains support from a sufficient number of blockchain nodes, the system determines that the proposal has reached consensus, and all blockchain nodes write the hash value of the candidate trusted experts of the proposal as the trusted expert hash value of the expert network into their local state.
[0060] In some embodiments, the workflow of a trusted distributed hybrid expert model edge computing system encompasses both training and inference processes. The training process involves forward computation and backward training, while the inference process involves only forward computation. Specifically:
[0061] During the training process, the learning task is the training task, and steps S1 to S7 are executed. In step S7, each blockchain node packages a block, which includes the trusted processing results of all activated expert networks, the aggregation results of the hybrid expert model, the model parameters of the gated network (updated in step S5), the trusted hash values of all activated expert networks (obtained in step S6), and the CIDs of all activated expert networks (updated in step S6). A new block is generated competitively under distributed consensus. During the inference process, the learning task is the inference task, and steps S1 to S4 and S7 are executed. In step S7, each blockchain node packages a block, which includes the trusted processing results of all activated expert networks, the aggregation results of the hybrid expert model, the model parameters of the gated network, and the trusted hash values of all activated expert networks. A new block is generated competitively under distributed consensus.
[0062] In some embodiments, the training and inference processes can be considered as independent processes. The inference process may refer to deploying a pre-trained MoE (e.g., trained in the cloud) in the edge layer and then processing the received inference tasks; while the training process involves directly deploying the initialized MoE in the edge layer and training it from scratch. In some embodiments, simulation results are used to illustrate the performance of the proposed method.
[0063] The simulation environment is as follows:
[0064] Windows 11 Pro operating system, Intel-13900KF (3.0Hz, 64GB RAM) CPU, NVIDIA RTX4090 (24GB RAM) graphics card and PyTorch platform.
[0065] The datasets selected are as follows:
[0066] The performance of the invention was tested in simulations on the Fashion-MNIST and CIFAR-10 datasets. The Fashion-MNIST dataset contains 60,000 training samples and 10,000 test samples, each of which is a 28×28 grayscale image of clothing. The CIFAR-10 dataset contains 50,000 training samples and 10,000 test samples, each of which is a 32×32 RGB color image, and the labels include airplane, car, bird, cat, deer, dog, frog, horse, boat and truck.
[0067] The simulation parameters are set as follows:
[0068] This embodiment sets N=10 experts and M=10 edge nodes. For traditional distributed MoE, the simulation deploys a different expert for each edge node; that is, expert i is deployed on edge node i. During training, the learning rates for Fashion-MNIST and CIFAR-10 are set to 0.01 and 0.1, respectively. This invention and traditional MoE set T=1500 training rounds, and the task publisher releases 1000 samples (i.e., learning tasks) in each round.
[0069] The neural network model selected is as follows:
[0070] This invention uses a linear model as the gating network and two types of neural network models as experts for different datasets. For the Fashion-MNIST dataset, this invention selects a Multi-Layer Perceptron (MLP) as the expert neural network model, where each MLP consists of two fully connected layers with 256 hidden neurons and ReLU activation function. For the CIFAR-10 dataset, this invention selects a Convolutional Neural Network (CNN) as the expert neural network model, where each CNN consists of three convolutional layers and two fully connected layers.
[0071] The simulation results and analysis are as follows:
[0072] Simulations demonstrate the training and inference performance of this invention and traditional distributed MoE under data tampering attacks. The simulation sets the number of activation experts per sample to K=3. For this invention, the simulation considers malicious edge nodes colluding to publish the same data tampering results.
[0073] Figure 4 A bar chart showing the activation ratios of various experts in a traditional distributed MoE is presented, where 'Y' represents data tampering attacks and 'N' represents no data tampering attacks. In this simulation, edge nodes 7-9 are malicious. According to... Figure 4During the training process, it can be observed that the activation frequency of experts 7-9 significantly decreased under data tampering attacks compared to before the attacks. This indicates that the gating system can detect tampered expert calculation results during training, thereby reducing the activation frequency of experts deployed on malicious edge nodes. This is because the gating system can self-update during the backpropagation process of training. Therefore, in order to minimize the MoE loss function, gating systems mostly select experts that can reduce the MoE loss function value. However, this raises two key issues: 1) Inefficient utilization of expert resources (i.e., unbalanced load), which, according to reference 2 (Shazeer N, Mirhoseini A, Maziarz K, et al. Outrageously large neural networks: Thesparsely-gated mixture-of-experts layer[C], 2017.), exacerbates expert overfitting and impairs the generalization ability of MoE; 2) Deterioration of learning performance, as Figure 5 As shown, during the inference process, it can also be observed that under data manipulation attacks, the activation rates of experts 7-9 are significantly higher during inference than during training. This is because the parameters of MoE are no longer updated during inference, and the gating cannot detect the tampered calculation results of the experts. This further leads to a severe decline in the inference performance of MoE, as... Figure 7 As shown.
[0074] Figure 5 The training accuracy of this invention and traditional distributed MoE under data tampering attacks was compared, where r represents the malicious ratio (the ratio of malicious edge nodes to the total number of edge nodes), and edge nodes 7-9 are set as malicious. Figure 5 It was observed that the present invention consistently outperforms traditional distributed MoE. Furthermore, the present invention achieves nearly the same accuracy as traditional distributed MoE without data tampering attacks. These observations demonstrate that the present invention can effectively improve the robustness of distributed MoE to data tampering attacks during training.
[0075] Figure 6 The training accuracy and latency of this invention and traditional distributed MoE were compared under data tampering attacks. Figure 6 It was observed that the present invention achieved better accuracy (e.g., 45% improvement on the Fashion-MNIST dataset and 67% improvement on the CIFAR-10 dataset) compared to traditional distributed MoE, but at the cost of higher latency.
[0076] Figure 7The inference accuracy of this invention and a traditional distributed MoE, both well-trained, was compared under different malware ratios. Figure 7 It was observed that when the malicious ratio is below 50%, the present invention achieves higher accuracy compared to traditional distributed MoE (e.g., a 66% improvement on the Fashion-MNIST dataset and a 44% improvement on the CIFAR-10 dataset). This observation indicates that the present invention is more resistant to data tampering attacks during inference compared to traditional distributed MoE. However, simulation results also show that when the malicious ratio exceeds 50%, the present invention experiences a significant performance degradation. This is because when the number of malicious edge nodes exceeds the number of honest edge nodes, the blockchain reaches global consensus on the tampered computation results with a very high probability, leading to a decrease in the inference performance of MoE.
[0077] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "setting," "connection," "fixing," "rotation," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Unless otherwise explicitly limited, those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.
[0078] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A blockchain-based trusted distributed hybrid expert model edge computing system, characterized in that, Including task publishers, blockchain network, edge layer, and storage layer, the gating network is deployed within the blockchain network, where: The task publisher is used to publish learning tasks, which are either training tasks or inference tasks. The blockchain network is used to receive learning tasks and perform expert weight evaluation on the learning tasks to obtain sparse activation expert index results; it is also used to receive the learning task processing results of the edge layer and perform consensus. The edge layer is used to download the corresponding expert model from the storage layer to process the learning task based on the sparse activation expert index result, and upload the learning task processing result to the blockchain network. The storage layer is used to store all expert networks, and each expert network has a CID; The system workflow includes: S1. The task publisher posts a learning task; S2. The blockchain network receives the learning task and uses an on-chain gating network to evaluate expert weights to obtain a sparse activation expert index result. The blockchain network records the learning task and the sparse activation expert index result; the sparse activation expert index result includes the CID of all activated expert networks. S3. The edge layer downloads the learning task and sparse activation expert index results from the blockchain network, downloads the expert network to process the learning task from the storage layer based on the sparse activation expert index results, and uploads the learning task processing results to the blockchain network. S4. The blockchain network receives the learning task processing results uploaded by the edge layer and performs distributed consensus to obtain the aggregated results of the hybrid expert model; S5. The blockchain network generates loss function values and updates the model parameters of the on-chain gating network using the gradient descent algorithm; the edge layer downloads the loss function values and updates the model parameters of all activated expert networks using the gradient descent algorithm, generating expert hash values and uploading them to the blockchain network. S6. The blockchain network receives expert hash values and performs distributed consensus to obtain trusted expert hash values and uploads them to the storage layer; the storage layer stores the trusted expert hash values and generates CIDs for them. S7. The blockchain network generates new blocks and uploads them to the chain.
2. The blockchain-based trusted distributed hybrid expert model edge computing system according to claim 1, characterized in that, The edge layer includes a set of edge nodes, and step S3 specifically includes: Each edge node downloads all activated expert networks from the storage layer based on the sparse activation expert index results to generate a learning task processing result, and uploads the learning task processing result to the blockchain network. The learning task processing results include the processing results of all activated expert networks in the current edge node processing the learning task.
3. The blockchain-based trusted distributed hybrid expert model edge computing system according to claim 2, characterized in that, The blockchain network includes a set of blockchain nodes. The blockchain network receives the learning task processing results from all edge nodes and broadcasts them to all blockchain nodes. Based on the learning task processing results, a distributed consensus is performed to obtain the trusted processing results of each activated expert network. Each blockchain node aggregates the trusted processing results of all activated expert networks to obtain the hybrid expert model aggregation result.
4. The blockchain-based trusted distributed hybrid expert model edge computing system according to claim 3, characterized in that, The process of obtaining the trusted processing result of an activated expert network through distributed consensus is as follows: A blockchain node compares the processing results of the expert network on different edge nodes, and uses the processing result that is the same on most edge nodes as the candidate trusted processing result of the expert network, and uses it as the proposal to initiate; The other blockchain nodes verify the proposal, and if the verification is successful, they send an approval message to express their support for the proposal. When a proposal gains the support of a sufficient number of blockchain nodes, it is determined that the proposal has reached a consensus. All blockchain nodes then write the proposed candidate trusted processing result into their local state as the trusted processing result of the expert network.
5. The blockchain-based trusted distributed hybrid expert model edge computing system according to claim 1, characterized in that, The edge layer includes a set of edge nodes. In step S5, each edge node of the edge layer downloads the loss function value and updates the model parameters of all activated expert networks using the gradient descent algorithm. At the same time, an expert hash value is generated for the model parameters of each activated expert network.
6. The blockchain-based trusted distributed hybrid expert model edge computing system according to claim 1, characterized in that, When the learning task is a training task, steps S1 to S7 are executed. In step S7, each blockchain node packages a block, which includes the trusted processing results of all activated expert networks, the aggregation results of hybrid expert models, the model parameters of the gated network, the trusted hash values of all activated expert networks, and the CIDs of all activated expert networks. A new block is generated competitively under distributed consensus. When the learning task is a reasoning task, steps S1 to S4 and step S7 are executed. In step S7, each blockchain node packages a block, which includes the trusted processing results of all activated expert networks, the aggregation results of hybrid expert models, the model parameters of the gated network, and the trusted hash values of all activated expert networks. A new block is generated competitively under distributed consensus.
7. The blockchain-based trusted distributed hybrid expert model edge computing system according to claim 1, characterized in that, The storage layer consists of a set of distributed storage nodes, and the storage layer uses IPFS technology for distributed storage.