Federal learning-based privacy protection data collaborative analysis platform

Abstract

The invention discloses a privacy protection data collaborative analysis platform based on federated learning, which relates to the technical field of privacy computing, and comprises a central server for initializing a global model, distributing the global model to a client, aggregating model update fed back by the client, obtaining an aggregation parameter, updating the global model, and sending the aggregation parameter to the client; distributing the updated global model to the client; each client receives the global model, carries out model training by utilizing local data, obtains model updating, and feeds back the model updating to the central server; and the intelligent routing gateway cluster randomly divides a plurality of routing groups, distributes one routing group for each client to construct an anonymous communication path, and transmits the model update and the global model through the anonymous communication path. According to the method, privacy protection during cross-mechanism data joint modeling can be realized, and the security of data collaborative analysis is improved.

Description

Technical Field

[0001] The present invention relates to the field of privacy computing technology, and more specifically to a privacy-preserving data collaborative analysis platform based on federated learning. Background Art

[0002] With the rapid development of information technology, data has become a vital resource in modern society. However, data silos exist between different industries, such as finance and healthcare. Privacy regulations (such as GDPR) prevent institutions from directly sharing raw data. Therefore, achieving data sharing and protecting personal privacy and data security during this process has become a global challenge.

[0003] Traditional data sharing methods often involve the risk of sensitive information leakage, limiting the maximum utilization of data value. Furthermore, using traditional encryption transmission or anonymization technologies during data sharing still carries the risk of data leakage and cannot support complex machine learning tasks. To address this issue, federated learning, an emerging machine learning technology, has attracted considerable attention in recent years for its ability to enable data sharing and model training while protecting user privacy.

[0004] Federated learning, as a distributed machine learning framework, enables multiple participants to collaboratively train global models without sharing training data, effectively ensuring client data privacy. However, federated learning still faces the risk of model parameter leakage and identity privacy threats during communication.

[0005] Therefore, how to improve the privacy and security of the data sharing process is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0006] In view of this, the present invention provides a privacy-preserving data collaborative analysis platform based on federated learning, which can achieve privacy protection during cross-institutional data joint modeling and improve the security of data collaborative analysis.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] A privacy-preserving data collaborative analysis platform based on federated learning, including:

[0009] The central server initializes the global model and distributes it to the clients, aggregates the model updates fed back by the clients, obtains the aggregated parameters and updates the global model, and distributes the updated global model to the clients;

[0010] Multiple clients, each client receives the global model and uses local data to train the model, obtains model updates, and feeds back to the central server;

[0011] The intelligent routing gateway cluster randomly divides several routing groups and assigns a set of routing groups to each client to build an anonymous communication path, through which model updates and global models are transmitted.

[0012] Preferably, the process of randomly dividing the intelligent routing gateway cluster into several routing groups is as follows:

[0013] Step 11: Get the gateway node list and grouping parameters, and initialize the cryptographic security entropy source;

[0014] The grouping parameters include the limit on the number of gateway nodes in the routing group, including the maximum and minimum values; the cryptographically secure entropy source uses a secure random number generator and establishes an entropy pool connection;

[0015] Step 12: Generate random numbers based on a cryptographically secure entropy source;

[0016] Step 13: Use the improved Fisher-Yates shuffling algorithm to randomly permute the gateway node list based on the random number to obtain a random sequence and implement shuffling;

[0017] Step 14: Randomly select the number of gateway nodes from the grouping parameters as the split threshold, divide the random sequence into several routing groups based on the split threshold, label each routing group with a routing ID, and label each gateway node in each routing group with a node ID; when the number of gateway nodes in the gateway node list cannot be divided evenly by the number of groups, adjust the number of gateway nodes in the last routing group.

[0018] Preferably, the specific process of step 13 is:

[0019] Step 131: Create an initial gateway node sequence according to the gateway node list;

[0020] Step 132: Using the random number as a key to set a random permutation function;

[0021] Step 133: permuting the gateway nodes in the initial gateway node sequence according to a random permutation function to obtain a random sequence.

[0022] Preferably, the process of assigning a routing group to each client to construct an anonymous communication path is as follows:

[0023] Step 21: Client C i Generate an access request and send it to the intelligent routing gateway cluster to apply for routing group access rights;

[0024] Step 22: The intelligent routing gateway cluster uses a weighted round-robin algorithm to dynamically distribute load balancing among routing groups and assigns a corresponding routing group to each client.

[0025] Step 23: Construct an anonymous communication path based on the assigned routing group, and use the anonymous communication path to transmit model updates and the global model.

[0026] Preferably, the specific process of step 22 is:

[0027] Step 221: The intelligent routing gateway cluster constructs a routing group weight matrix according to the real-time load status of each routing group;

[0028] Step 222: Construct a routing group status table based on the routing group ID, routing group weight matrix, and allocation times;

[0029] Step 223: For each client C i Perform cyclic allocation, select the routing group with the current maximum weight and assign it to the client, and update the number of allocations and the weight of the current routing group in the routing group status table until all clients are assigned the corresponding routing group.

[0030] Preferably, the specific process of step 23 is:

[0031] Step 231: For each routing group assigned to a client, each gateway node generates a corresponding public key PK and private key SK, and transmits the public keys PK of all gateway nodes to the corresponding client and central server;

[0032] Step 232: Arrange all gateway nodes in the routing group according to their communication distance to the client to construct an anonymous communication path;

[0033] Step 233: The client generates an encrypted transmission instruction using the public keys of all gateway nodes and the model update according to the communication direction of the anonymous communication path, and transmits the encrypted transmission instruction to the first gateway node. After being decrypted by each gateway node in the anonymous communication path, the model update is transmitted to the central server.

[0034] The central server generates update instructions using the public keys of all gateway nodes and the global model according to the reverse communication direction of the anonymous communication path, and transmits the update instructions to the last gateway node; after being decrypted by the gateway layers of each gateway node in the anonymous communication path, the global model is transmitted to the corresponding client.

[0035] Preferably, the encrypted transmission instruction includes several layers of shared keys and model update ciphertext; a shared key is generated based on the public key of a gateway node and the node ID of the next gateway node in the communication direction; the client encrypts the model update based on the local public key to obtain the model update ciphertext; the update instruction includes several layers of shared keys and the global model ciphertext; the central server encrypts the global model based on the server private key to obtain the global model ciphertext; the gateway node's private key is used to perform gateway decryption on the shared key of the corresponding layer to determine the ID of the next transmitted gateway node.

[0036] Preferably, initially, the client generates a registration request based on the local public key and sends it to the central server. The central server receives the registration request and saves the local public key, and at the same time feeds back the server public key to the client. The client receives and saves the server public key. The client transmits the model update ciphertext to the central server through an anonymous communication path. The central server uses the saved local public key to decrypt and obtain the model update. The central server transmits the global model ciphertext to the client through an anonymous communication path. The client uses the saved server public key to decrypt and obtain the global model.

[0037] Preferably, the central server also includes a heterogeneous adapter. When the client registers, the heterogeneous adapter collects and analyzes the hardware capabilities, network status and data characteristics of each client to generate a device performance profile; according to the device performance profile, a suitable model architecture is assigned to the client, a lightweight model is assigned to a low-computing power client, a pruned model is assigned to a medium-computing power client, and a complete global model is assigned to a high-computing power client; a structured pruning operation is performed on the complete global model, and the selected layer is frozen to obtain a pruned model; knowledge distillation is performed on the pruned model to obtain a lightweight model.

[0038] Preferably, the client also includes a differential privacy injection unit to add noise layer by layer during model training. The specific process is as follows:

[0039] Step 31: Preset the privacy noise distribution principle according to the global model structure;

[0040] Step 32: Calculate the sensitivity of each layer of the global model and perform gradient clipping to ensure that the sensitivity is bounded;

[0041] Step 33: Calculate the noise scale of each layer structure according to the privacy noise allocation principle and sensitivity, and construct the noise matrix corresponding to each layer of noise;

[0042] Step 34: In the forward propagation stage and the backward propagation stage, add corresponding noise matrices in different layer structures.

[0043] Preferably, the layer structure includes an input layer, an output layer, and an intermediate layer; the privacy noise distribution principle includes: the output layer is allocated the highest amount of noise, the intermediate layer is allocated a medium amount of noise, and the input layer is allocated the lowest amount of noise; at the beginning of the forward propagation phase, the noise matrix corresponding to the input layer is added to the input layer, when calculating the gradient in the backward propagation phase, the noise matrix corresponding to the intermediate layer is added to the intermediate layer, and when the gradient calculation in the backward propagation phase is completed, the noise matrix corresponding to the output layer is added to the output layer.

[0044] Preferably, a TEE (Trusted Execution Environment) module is integrated into the central server, and global model initialization, model update aggregation and global model update are performed in the TEE module.

[0045] Preferably, the client is also provided with a zero-knowledge proof verifier to verify the zero-knowledge proof generated by the client's model training process to verify whether the training complies with the protocol (such as no illegal data is used). If not, the learning is stopped.

[0046] Through the above technical solutions, it can be seen that compared with the existing technology, the present invention discloses a privacy-preserving data collaborative analysis platform based on federated learning, which dynamically constructs random routing groups through intelligent routing gateway clusters, adopts the improved Fisher-Yates shuffling algorithm to ensure the unpredictability of gateway allocation, and effectively hides the direct association between the client and the server; dynamically allocates routing groups based on the weighted polling algorithm, realizes efficient dynamic resource scheduling, responds to changes in gateway cluster load in real time, avoids single-point congestion, and improves communication efficiency; realizes layer-by-layer encrypted transmission based on anonymous communication paths, ensures that model parameters cannot be decrypted even if they are intercepted during transmission, and prevents eavesdropping by intermediate nodes; heterogeneous adapters dynamically allocate models according to device performance profiles (computing power / network status), significantly reduces the computing load of edge devices, and solves the efficiency bottleneck caused by device heterogeneity in federated learning; layered noise addition avoids global excessive disturbance and improves model accuracy under the same privacy budget; performs global model initialization, aggregation and update in the central server TEE module to ensure that sensitive operations run in a hardware-level security environment and prevent malicious code tampering. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0048] Figure 1 A schematic diagram of the structure of the privacy-preserving data collaborative analysis platform based on federated learning provided by the present invention;

[0049] Figure 2 A schematic diagram of the process of dividing routing groups provided by the present invention;

[0050] Figure 3 A schematic diagram of the process of constructing an anonymous communication path provided by the present invention. DETAILED DESCRIPTION

[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0052] The embodiment of the present invention discloses a privacy protection data collaborative analysis platform based on federated learning, such as Figure 1 Shown, including:

[0053] The central server initializes the global model and distributes it to the clients, aggregates the model updates fed back by the clients, obtains the aggregated parameters and updates the global model, and distributes the updated global model to the clients;

[0054] Multiple clients, each client receives the global model and uses local data to train the model, obtains model updates, and feeds back to the central server;

[0055] The intelligent routing gateway cluster randomly divides several routing groups and assigns a set of routing groups to each client to build an anonymous communication path, through which model updates and global models are transmitted.

[0056] Further, such as Figure 2 As shown in the figure, the process of randomly dividing the intelligent routing gateway cluster into several routing groups is as follows:

[0057] S11: Obtain the gateway node list and grouping parameters, and initialize the cryptographic security entropy source;

[0058] The grouping parameters include the limit on the number of gateway nodes in the routing group, including the maximum and minimum values; the cryptographically secure entropy source uses a secure random number generator and establishes an entropy pool connection;

[0059] S12: Generate random numbers based on a cryptographically secure entropy source;

[0060] Read cryptographically secure entropy source: read(entropy_fd,buffer,256);

[0061] Generate a cryptographically secure random number from a cryptographically secure entropy source: seed = SHA3-256(buffer);

[0062] S13: Randomly permute the gateway node list using the improved Fisher-Yates shuffling algorithm based on the random number to obtain a random sequence and implement shuffling;

[0063] S14: Randomly select the number of gateway nodes from the grouping parameters as a split threshold, divide the shuffled random sequence into several routing groups according to the split threshold, label each routing group with a routing ID, and label each gateway node in each routing group with a node ID; when the number of gateway nodes in the gateway node list cannot be divided evenly by the number of groups, adjust the number of gateway nodes in the last routing group.

[0064] Furthermore, the specific process of S13 is as follows:

[0065] S131: Create an initial gateway node sequence according to the gateway node list; seq = [1, 2, 3, ..., m] represents an ordered arrangement of gateway indexes;

[0066] S132: Using the random number as a key to set a random permutation function;

[0067] nonce = seed[16:24];

[0068] PRP=def PRP(key,nonce,i);

[0069] S133: Permuting the gateway nodes in the initial gateway node sequence according to a random permutation function to obtain a random sequence.

[0070] Further, such as Figure 3 As shown in the figure, the process of assigning a set of routing groups to each client to build an anonymous communication path is as follows:

[0071] S21: Client C i Generate an access request and send it to the intelligent routing gateway cluster to apply for routing group access rights;

[0072] S22: The intelligent routing gateway cluster uses a weighted round-robin algorithm to dynamically distribute load among routing groups, assigning a corresponding routing group to each client.

[0073] S23: Construct an anonymous communication path according to the assigned routing group, and use the anonymous communication path to transmit model updates and the global model.

[0074] Furthermore, the specific process of S22 is as follows:

[0075] S221: The intelligent routing gateway cluster constructs a routing group weight matrix W based on the real-time load status of each routing group:

[0076]

[0077] α+β=1;

[0078] represents the weight of the i-th routing group; represents the average delay within the i-th routing group, represents the available bandwidth of the i-th routing group; The maximum bandwidth is the sum of the physical and logical bandwidths of all gateway nodes in the routing group and can be dynamically configured based on SDN (Software Defined Network). α and β represent weights, with α and β taking values ​​of 0.3 and 0.7, respectively. The weights of all routing groups are normalized, and the routing group weight matrix is ​​constructed using the normalized weights.

[0079] S222: Constructing a routing group status table based on the routing group ID, the normalized weight in the routing group weight matrix, the weight of the current routing group, and the number of allocations;

[0080] S223: For each client C i Perform cyclic allocation, select the routing group with the current maximum weight and assign it to the client, and update the number of allocations and the weight of the current routing group in the routing group status table until all clients are assigned the corresponding routing group; subtract the sum of all normalized weights from the current weight of the assigned routing group, and increase the current weight of all routing groups by the corresponding normalized weight.

[0081] The weighted round-robin algorithm can better balance the load, allowing servers or processors with strong processing capabilities to handle more requests or tasks. At the same time, it can also ensure that even nodes with weaker processing capabilities will not be completely deprived of tasks, thereby improving resource utilization and service response speed.

[0082] Furthermore, the specific process of S23 is as follows:

[0083] S231: For each client, the routing group G is assigned k ={g1, g2, ..., g m}, m represents the number of gateway nodes in the routing group. Each gateway node generates a corresponding public key PK and private key SK, and transmits the public keys PK of all gateway nodes to the corresponding clients and central server;

[0084] S232: Arrange all gateway nodes in the routing group according to their communication distances to the client to construct an anonymous communication path;

[0085] S233: The client generates an encrypted transmission instruction using the public keys of all gateway nodes and the model update according to the communication direction of the anonymous communication path, and transmits the encrypted transmission instruction to the first gateway node; after decryption by the gateway layer by gateway of each gateway node in the anonymous communication path, the model update is transmitted to the central server; the central server generates an update instruction using the public keys of all gateway nodes and the global model according to the reverse communication direction of the anonymous communication path, and transmits the update instruction to the last gateway node; after decryption by the gateway layer by gateway of each gateway node in the anonymous communication path, the global model is transmitted to the corresponding client.

[0086] Furthermore, the encrypted transmission instruction includes several layers of shared keys and model update ciphertext; a shared key is generated based on the public key of a gateway node and the node ID of the next gateway node in the communication direction; the client encrypts the model update based on the local public key to obtain the model update ciphertext; the update instruction includes several layers of shared keys and global model ciphertext; the central server encrypts the global model based on the server private key to obtain the global model ciphertext; the gateway node's private key is used to perform gateway decryption on the shared key of the corresponding layer to determine the ID of the next transmitted gateway node.

[0087] Furthermore, initially, the client generates a registration request based on the local public key and sends it to the central server. The central server receives the registration request and saves the local public key, and at the same time feeds back the server public key to the client. The client receives and saves the server public key. The client transmits the model update ciphertext to the central server through an anonymous communication path. The central server uses the saved local public key to decrypt and obtain the model update. The central server transmits the global model ciphertext to the client through an anonymous communication path. The client uses the saved server public key to decrypt and obtain the global model.

[0088] On the other hand, in a specific embodiment, the central server also includes a heterogeneous adapter. When the client registers, the heterogeneous adapter collects and analyzes the hardware capabilities, network status and data characteristics of each client to generate a device performance profile; according to the device performance profile, a suitable model architecture is assigned to the client, a lightweight model is assigned to a low-computing power client, a pruned model is assigned to a medium-computing power client, and a complete global model is assigned to a high-computing power client; a structured pruning operation is performed on the complete global model, and the selected layer is frozen to obtain a pruned model; knowledge distillation is performed on the pruned model to obtain a lightweight model.

[0089] On the other hand, in a specific embodiment, the client further includes a differential privacy injection unit that adds noise (Gaussian noise distribution or Laplace noise) layer by layer during model training. The specific process is as follows:

[0090] S1: Preset privacy noise allocation principle according to the global model structure;

[0091] S2: Calculate the sensitivity of each layer of the global model and perform gradient clipping to ensure that the sensitivity is bounded;

[0092] S3: Calculate the noise scale of each layer of structure based on the privacy noise allocation principle and sensitivity, and construct the noise matrix corresponding to each layer of noise;

[0093] S4: In the forward propagation stage and the backward propagation stage, the corresponding noise matrix is ​​added in different layer structures.

[0094] Furthermore, the layer structure includes an input layer, an output layer, and an intermediate layer; the privacy noise distribution principle includes: the output layer is allocated the highest amount of noise, the intermediate layer is allocated a medium amount of noise, and the input layer is allocated the lowest amount of noise; at the beginning of the forward propagation phase, the noise matrix corresponding to the input layer is added to the input layer, when calculating the gradient in the backward propagation phase, the noise matrix corresponding to the intermediate layer is added to the intermediate layer, and when the gradient calculation in the backward propagation phase is completed, the noise matrix corresponding to the output layer is added to the output layer.

[0095] Noise is dynamically allocated according to the model layer structure (input layer, intermediate layer, and output layer). The output layer adds the strongest noise to protect the prediction results, while the input layer minimizes noise to preserve feature validity. Combined with gradient clipping to constrain sensitivity, this allows for fine-grained control of the privacy budget, ensuring model availability while meeting differential privacy requirements. This layered noise injection scheme deeply collaborates with the federated learning framework. The client completes noise injection during local training, and the intelligent routing gateway provides supplementary anonymity protection during transmission, forming a dual privacy protection system of "local noise addition + transmission anonymity."

[0096] In one embodiment, a central server integrates a TEE (Trusted Execution Environment) module, which performs global model initialization, model update aggregation, and global model updates. Global model initialization, aggregation, and updates are performed within the central server TEE module, ensuring that sensitive operations are executed in a hardware-level secure environment and preventing malicious code tampering.

[0097] On the other hand, in one specific embodiment, the client also has a zero-knowledge proof verifier to verify the zero-knowledge proof generated during the client's model training process to verify whether the training complies with the protocol. If not, the learning process is terminated. The client verifies the compliance of the training process (e.g., no illegal data is used) through zero-knowledge proof, ensuring the trustworthiness of the participants and eliminating malicious attacks at the source.

[0098] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0099] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A privacy-preserving data collaborative analysis platform based on federated learning, characterized by: include: The central server initializes the global model and distributes it to the clients, aggregates the model updates fed back by the clients, obtains the aggregated parameters and updates the global model, and distributes the updated global model to the clients; Multiple clients, each client receives the global model and uses local data to train the model, obtains model updates, and feeds back to the central server; The intelligent routing gateway cluster randomly divides several routing groups and assigns a set of routing groups to each client to build an anonymous communication path, through which model updates and global models are transmitted.

2. The privacy-preserving data collaborative analysis platform based on federated learning according to claim 1 is characterized in that: The process of randomly dividing the intelligent routing gateway cluster into several routing groups is as follows: Step 11: Obtain a list of gateway nodes and grouping parameters, and initialize a cryptographically secure entropy source; grouping parameters include a limit on the number of gateway nodes in a routing group; Step 12: Generate random numbers based on a cryptographically secure entropy source; Step 13: Use the improved Fisher-Yates shuffling algorithm to randomly permute the gateway node list according to the random number to obtain a random sequence; Step 14: Randomly select the number of gateway nodes from the grouping parameters as the split threshold, divide the random sequence into several routing groups based on the split threshold, label each routing group with a routing ID, and label each gateway node in each routing group with a node ID; when the number of gateway nodes in the gateway node list cannot be divided evenly by the number of groups, adjust the number of gateway nodes in the last routing group.

3. The privacy-preserving data collaborative analysis platform based on federated learning according to claim 2 is characterized in that: The specific process of step 13 is as follows: Step 131: Create an initial gateway node sequence according to the gateway node list; Step 132: Using the random number as a key to set a random permutation function; Step 133: permuting the gateway nodes in the initial gateway node sequence according to a random permutation function to obtain a random sequence.

4. The privacy-preserving data collaborative analysis platform based on federated learning according to claim 2 is characterized in that: The process of assigning a routing group to each client to build an anonymous communication path is as follows: Step 21: The client generates an access request and sends it to the intelligent routing gateway cluster to apply for routing group access rights. Step 22: The intelligent routing gateway cluster uses a weighted round-robin algorithm to dynamically distribute load balancing among routing groups and assigns a corresponding routing group to each client. Step 23: Construct an anonymous communication path based on the assigned routing group, and use the anonymous communication path to transmit model updates and the global model.

5. The privacy-preserving data collaborative analysis platform based on federated learning according to claim 4 is characterized in that: The specific process of step 22 is: Step 221: The intelligent routing gateway cluster constructs a routing group weight matrix according to the real-time load status of each routing group; Step 222: Construct a routing group status table based on the routing group ID, routing group weight matrix, and allocation times; Step 223: Perform cyclic allocation for each client, select the routing group with the current maximum weight and allocate it to the client, and update the number of allocations and the weight of the current routing group in the routing group status table until all clients are allocated the corresponding routing group.

6. The privacy-preserving data collaborative analysis platform based on federated learning according to claim 4 is characterized in that: The specific process of step 23 is: Step 231: For each routing group assigned to a client, each gateway node generates a corresponding public key and private key, and transmits the public keys of all gateway nodes to the corresponding client and central server; Step 232: Arrange all gateway nodes in the routing group according to their communication distances to the client from far to near, and construct an anonymous communication path; Step 233: The client generates an encrypted transmission instruction using the public keys of all gateway nodes and the model update according to the communication direction of the anonymous communication path, and transmits the encrypted transmission instruction to the first gateway node. After being decrypted by each gateway node in the anonymous communication path, the model update is transmitted to the central server. The central server generates update instructions using the public keys of all gateway nodes and the global model according to the reverse communication direction of the anonymous communication path, and transmits the update instructions to the last gateway node; after being decrypted by the gateway layers of each gateway node in the anonymous communication path, the global model is transmitted to the corresponding client.

7. The privacy-preserving data collaborative analysis platform based on federated learning according to claim 6 is characterized in that: The encrypted transmission instruction includes several layers of shared keys and model update ciphertext; a shared key is generated based on the public key of a gateway node and the node ID of the next gateway node in the communication direction; the client encrypts the model update based on the local public key to obtain the model update ciphertext; the update instruction includes several layers of shared keys and global model ciphertext; the central server encrypts the global model based on the server private key to obtain the global model ciphertext; the gateway node's private key is used to perform gateway decryption on the shared key of the corresponding layer to determine the ID of the next transmitted gateway node.

8. The privacy-preserving data collaborative analysis platform based on federated learning according to claim 7 is characterized in that: Initially, the client generates a registration request based on the local public key and sends it to the central server. The central server receives the registration request and saves the local public key. At the same time, it feeds back the server public key to the client. The client receives and saves the server public key. The client transmits the model update ciphertext to the central server through an anonymous communication path, and the central server uses the saved local public key to decrypt and obtain the model update; the central server transmits the global model ciphertext to the client through an anonymous communication path, and the client uses the saved server public key to decrypt and obtain the global model.

9. The privacy-preserving data collaborative analysis platform based on federated learning according to claim 8, characterized in that: The central server also includes heterogeneous adapters. When the client registers, the heterogeneous adapters collect and analyze the hardware capabilities, network status and data characteristics of each client to generate a device performance profile; based on the device performance profile, the appropriate model architecture is assigned to the client.

10. The privacy-preserving data collaborative analysis platform based on federated learning according to claim 1, characterized in that: The client also includes a differential privacy injection unit that adds noise layer by layer during model training. The specific process is as follows: Step 31: Preset the privacy noise distribution principle according to the global model structure; Step 32: Calculate the sensitivity of each layer of the global model and perform gradient clipping to ensure that the sensitivity is bounded; Step 33: Calculate the noise scale of each layer structure according to the privacy noise allocation principle and sensitivity, and construct the noise matrix corresponding to each layer of noise; Step 34: In the forward propagation stage and the backward propagation stage, add corresponding noise matrices in different layer structures.

Images