A space information network bidirectional personalized federated learning method, system and device

Abstract

The present application belongs to the technical field of federated learning, and particularly relates to a space information network bidirectional individualized federated learning method, system and device, which comprises the following steps: firstly, aggregating pre-training models of each ground station to obtain a global initial model and broadcasting the global initial model to clients; in subsequent rounds, a satellite distributes ground station models to clients in a targeted manner; the clients aggregate multiple ground station models by weighting with historical similarity as the weight, generate an individualized initial model for training, calculate the similarity metric and membership of the client model and the ground station model, fuse the historical similarity to update the similarity metric and membership, assign a set of collaborative ground stations to the clients and forward the model, and the ground stations aggregate the client models by weighting with historical similarity as the weight to obtain updated specialized models. Through the cooperative bidirectional individualized mechanism, the present application solves the model adaptation problem caused by data heterogeneity in the space information network, ensures the stability of training, and improves the performance of the individualized models of the clients and the ground stations.

Description

Technical Field

[0001] This invention belongs to the field of federated learning technology, specifically relating to a bidirectional personalized federated learning method, system, and device for spatial information networks. Background Technology

[0002] Spatial information networks can efficiently collect and transmit large amounts of heterogeneous data over a wide area, serving as a key infrastructure for applications such as meteorological monitoring and intelligent transportation. Federated learning, as a distributed machine learning framework, enables global model sharing while ensuring local data privacy. Its performance in independent, co-distributed scenarios is similar to that of centralized training, providing a feasible path for the privacy-compliant use of multi-source data from spatial information networks.

[0003] However, in the actual deployment of spatial information networks, the diversity of information and the complex relationships between data providers hinder the establishment of a trust framework. Although existing personalized and clustered federated learning has alleviated the statistical heterogeneity problem on the client side, it still has significant shortcomings.

[0004] The traditional architecture of a single ground station serving multiple clients cannot fully unleash the inherent value of space information network infrastructure, nor does it address the personalized needs of ground stations for specific scenarios and regional data characteristics. At the same time, the open and dynamic network topology of space information networks brings multiple security challenges, such as model transmission eavesdropping and tampering, insufficient system robustness under high downtime, and lack of traceability of training records. Traditional federated learning frameworks are mostly designed for static network environments and lack the ability to adapt to dynamic topologies and a sound security protection mechanism. They are unable to achieve efficient convergence while ensuring security and robustness, which seriously limits their practical application value in space information networks. Summary of the Invention

[0005] The purpose of this invention is to provide a bidirectional personalized federated learning method, system, and apparatus for spatial information networks.

[0006] A bidirectional personalized federated learning method for spatial information networks includes the following steps: S1. Each ground station uses its local private dataset to pre-train and obtain an initial model, which is then broadcast to the satellite. The satellite averages and aggregates all the initial models to obtain a global initial model, which is then broadcast to all clients. S2. Each client uses its local private dataset to train a personalized initial model based on the global initial model. S3. The satellite calculates the distance between each personalized initial model and the initial model, and performs extreme value standardization to obtain a similarity metric. Based on the similarity metric, it calculates the membership degree of each client to each ground station in the current round, calculates the historical similarity based on the membership degree, and matches several ground stations for each client based on the historical similarity, which serve as the ground station set for that client in this round. S4. Each ground station calculates the aggregation weight of the matched client based on historical similarity, performs weighted aggregation on the personalized initial model of the matched client, obtains the aggregated model corresponding to the ground station, and uses the aggregated model as the new initial model for each ground station to execute S1 until the preset iteration round is reached.

[0007] When the number of iterations is greater than or equal to 2, perform the following steps: S5. Broadcast the new initial model of each ground station via satellite to the clients that were matched with that ground station in the previous round; S6. Each client calculates the weight coefficient based on the historical similarity of the previous round, performs weighted aggregation on the new initial model of all matched ground stations, obtains the updated model, and uses its local private dataset for local training to obtain the personalized model. Configure appropriate local training rounds on the client side. Learning rate Regularization coefficient The model is trained iteratively using the Adam optimizer and other hyperparameters on a local dataset. After training, the model hash value is calculated and uploaded. The model is also encrypted and sent to the current service satellite.

[0008] S7. The satellite calculates the distance between each personalized model and the new initial models of all ground stations, and performs extreme value standardization to obtain a new similarity metric. Based on the new similarity metric, it calculates the membership degree of each client to each ground station in the current round, and performs weighted fusion of the historical similarity of the previous round and the membership degree of the current round to obtain the new historical similarity of the current round. Based on the new historical similarity, several ground stations are matched for each client, which serve as the set of ground stations for that client in this round; S8. In this round of ground station collection, each ground station obtains new historical similarity, which is used as weight to perform weighted aggregation on multiple client-side personalized models to obtain a new aggregated model; The new aggregated model is used as the new initial model for each ground station to execute S5 until the preset iteration round is reached.

[0009] S3 performs extreme value standardization to obtain a similarity measure, specifically: To obtain the global maximum and global minimum distances, the following calculations are performed: , in, This is a similarity metric between the i-th client and the j-th ground station. The distance between the personalized initial model for each client and the initial model for each ground station. and These are the global minimum distance and the global maximum distance.

[0010] In S3, membership degree is calculated based on similarity metric, specifically as follows: The membership degree of each client to each ground station is calculated as follows: , in, This is a similarity metric between the i-th client and the j-th ground station. This is a similarity metric between the i-th client and the k-th ground station. For the first The client and the first Membership degree between ground stations >1 is the fuzzy weight coefficient, M is the total number of ground stations, and k is the kth ground station.

[0011] The sum of the membership degrees of each client to all ground stations equals 1.

[0012] In S3, historical similarity is calculated based on membership degree, specifically as follows: Round t The client and the first Historical similarities between ground stations for: , in, For the first The client and the first The membership degree between ground stations, where t is the training round. The attenuation coefficient is... , It is located in An increasing function between them It is the total number of training rounds. For round t-1 The client and the first Historical similarities between ground stations.

[0013] In S3, several ground stations are matched for each client based on historical similarity, forming the ground station set for that client in this round, specifically: Each client can match up to [number] times. The client is determined by the following calculations for each ground station. No. The set of ground stations corresponding to the wheel : , in, The historical similarity matrix represents the composition of historical similarities. The Row vectors Indicates the first Wheel Client The assigned ground station index forwards the client's personalized initial model to the ground station set.

[0014] In S4, each ground station calculates the aggregate weight of matched clients based on historical similarity, specifically as follows: ground station Client The ratio of the historical similarity to the sum of the historical similarities of all clients matched by that ground station is used as the client's... Corresponding aggregate weight .

[0015] A bidirectional personalized federated learning system for spatial information networks, used in a bidirectional personalized federated learning method for spatial information networks, comprising: The global model generation module allows each ground station to pre-train its model using its local private dataset, obtain an initial model, and broadcast it to the satellite. The satellite then averages and aggregates all the initial models to obtain a global initial model, which is then broadcast to all clients. The personalization module allows each client to train its own personalized initial model based on the global initial model and its local private dataset. The matching module calculates the distance between each personalized initial model and the initial model, performs extreme value standardization, and obtains a similarity metric. Based on the similarity metric, it calculates the membership degree of each client to each ground station in the current round, calculates the historical similarity based on the membership degree, and matches several ground stations for each client based on the historical similarity, which serve as the ground station set for that client in this round. The aggregation module calculates the aggregation weight of the matched clients based on historical similarity for each ground station, performs weighted aggregation on the personalized initial models of the matched clients to obtain the aggregated model corresponding to the ground station, and uses the aggregated model as the new initial model for each ground station to be processed by the global model generation module until the preset iteration round is reached.

[0016] A bidirectional personalized federated learning device for a spatial information network includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement a bidirectional personalized federated learning method for a spatial information network.

[0017] Compared with the prior art, the beneficial effects of this application are as follows: Traditional federated learning is mostly a one-way aggregation mode. This application has a strong ability to adapt to heterogeneous mixed data in spatial information networks. It can solve the problems of difficult model convergence and insufficient personalized adaptation at the same time. The performance of the finally trained model is significantly improved compared with traditional methods.

[0018] The bidirectional framework of this invention allows ground stations to continuously iterate and optimize specialized models for specific scenarios, while enabling clients to adaptively integrate models from multiple ground stations to generate personalized models that fit their local mixed data distribution. This satisfies both the specialized needs of ground stations for specific scenarios and the personalized adaptation of clients, breaking the limitations of traditional one-way aggregation and achieving a bidirectional balance between global knowledge transfer and local personalized adaptation.

[0019] By employing a bidirectional dynamic grouping strategy, a global standardization metric method is proposed. In traditional federated learning, a client can typically only be assigned to one cluster, which is unsuitable for the mixed distribution data resulting from cross-regional data collection in spatial information networks, easily leading to model adaptation bias. This invention allows a single client to simultaneously integrate information from multiple ground stations, thereby more accurately matching the characteristics of local mixed data.

[0020] It can unify the distance metric standard while sharing only a small amount of distance data, avoid deviations in clustering results for different satellite regions, make the results of dynamic grouping more closely match the actual data distribution of the client, and make the model converge faster and have a better fitting effect.

[0021] This invention combines the historical similarity between the client and the ground station with the instantaneous membership calculated in the current round, dynamically adjusting the weights of both. In the early stages of training, instantaneous membership is emphasized to allow the model to quickly adapt to the data distribution; in the later stages of training, historical similarity is given more weight to stabilize the model state, avoid oscillations caused by frequent grouping, and make the model convergence process smoother and the fitting accuracy higher. Detailed Implementation

[0022] To further understand the content of this invention, the invention will be described in detail with reference to the embodiments.

[0023] To address the collaborative optimization problem of global models in a distributed environment, federated learning reconstructs the traditional machine learning model of data being concentrated in the model into a distributed paradigm where the model flows into the data. Under the premise of strictly protecting data privacy, it achieves value synergy of multi-source data.

[0024] In the FedAvg federated learning algorithm, it is assumed that there exists A collection of clients And a central server responsible for model aggregation and distribution. Each client Holding local private datasets And based on this dataset, a local model is trained, denoted as . The global sharing model is The global optimization objective of federated learning is to minimize the average of the local empirical loss functions of all clients, achieving the best fit of the global model on all global data. The optimization objective of federated learning can be expressed as: , in, The optimal global model obtained by solving for the global optimization objective. Indicates the first Each client is based on a local dataset. The empirical loss function is calculated.

[0025] To solve the aforementioned distributed optimization problem, federated learning employs a multi-round iterative training approach. Each round of communication iteration mainly executes four standard steps, forming a complete training loop: (1) Global model distribution: The central server will distribute the first model. Global model of round iteration It broadcasts to all clients participating in this round of training, providing a unified initial benchmark for local training on each client; (2) Local update on the client: client Receive global model Then, using it as the initial model for training, local private datasets were utilized. Perform multiple rounds of stochastic gradient descent (SGD) iterations to complete the personalized update of the local model.

[0026] The update process of the local model is shown in the formula: , in, The learning rate for local training. For the client Based on local dataset The calculated gradient vector of the loss function; (3) Local parameter upload: After completing local training, the client... Updated local model Uploaded to the central server, without transmitting any original data throughout the process, thus avoiding the leakage of original data; (4) Global Model Aggregation: After receiving all local models uploaded by clients, the central server uses an average aggregation strategy to generate the first model. Global model of round iteration The polymerization process is shown in the formula.

[0027] , The FedAvg algorithm's average aggregation mechanism achieves stable convergence and excellent model performance when client data satisfies the independent and identically distributed (IID) assumption. However, in spatial information network scenarios, this assumption is often difficult to uphold. Due to factors such as wide geographical coverage, differences in observed objects, and heterogeneity of terminal types, data collected by terminals distributed across different regions naturally exhibits significant non-independent and identically distributed (Non-IID) characteristics. This data heterogeneity causes a significant deviation between the local optimization direction of each client and the global optimal direction, leading to model drift. This not only severely slows down the convergence speed of the global model and reduces the overall prediction accuracy, but also makes it impossible for a single global model to adapt to the local data characteristics of different clients, resulting in inconsistent model performance among nodes.

[0028] To address the model drift and performance degradation issues caused by data heterogeneity, Personalized Federated Learning (PFL) has become a core technological direction. Clustered Federated Learning (CFL) is one of the representative solutions. Its core idea stems from the logic of clustering similarities. By dynamically detecting the similarity of data distribution among clients, the set of clients participating in training is divided into several subsets (clusters) whose internal data approximately follow an independent and identically distributed (IID) pattern. Instead of training a single global model, a dedicated intra-cluster optimal model is trained in parallel for each cluster.

[0029] This divide-and-conquer strategy alleviates the optimization conflicts caused by non-independent identically distributed (Non-IID) data through personalized model customization, breaking through the limitations of traditional federated learning in adapting to heterogeneous data.

[0030] In the clustering federated learning system, it is assumed that the global existence... A subset of potential clients Satisfying the set covering relationship (That is, all clients belong to a certain cluster), and the distribution of client data within each cluster satisfies similarity constraints. ,in Indicates the client The local data distribution. Each cluster will train its own dedicated model independently. Its global optimization objective is to minimize the sum of intra-cluster losses for all clusters, as shown in the following mathematical expression: , In the formula, For the first The number of clients per cluster, For the first The total intra-cluster loss function for each cluster. This refers to the dimension of the model parameters.

[0031] Traditional clustering federated learning often employs hard clustering strategies (such as K-means), requiring each client to strictly belong to a unique cluster and participate only in model training and parameter aggregation for that cluster.

[0032] However, in the spatial information network scenario focused on in this application, clients often move over a wide area, and the data they collect naturally exhibits a multi-source mixed distribution characteristic. The either-or partitioning rules of hard clustering cannot adapt to this fuzzy distribution characteristic, easily causing client models to get trapped in information cocoons, limiting their generalization ability and personalized adaptation effects. In contrast, soft clustering methods allow clients to belong to multiple clusters simultaneously in the form of membership degrees, which can more accurately capture the inherent characteristics of mixed data distribution and become a key technical path to solve the problem of adapting to mixed data distribution.

[0033] Based on the above considerations, this invention relates to a bidirectional personalized federated learning method for spatial information networks, which includes the following steps: S1. Each ground station uses its local private dataset for pre-training to obtain an initial model, which is then broadcast to the satellite. The satellite averages and aggregates all the initial models to obtain a global initial model, which is then broadcast to all clients.

[0034] This proposal suggests a dual-personalized federated learning method and system suitable for space information networks, establishing bidirectional collaboration between clients and ground stations. Ground stations aggregate models from clients with similar data distributions for specific feature extraction; clients dynamically connect to multiple ground stations via satellite relay, adaptively fusing multi-source models based on similarity fuzzy clustering to generate personalized models.

[0035] To ensure the security and auditability of the model, implement model hash storage and historical similarity tracing, threshold signature is used to verify the consistency of the initial model, and broadcast encryption is used to ensure the confidentiality of the model.

[0036] Hash functions are sensitive to input and random to output, making them ideal for data integrity verification.

[0037] A hash function, also known as a hashing function or hashing algorithm, is one of the cornerstones of modern cryptography. It is a type of mathematical algorithm that maps input data of arbitrary length to a fixed-length output string, typically called a hash value or digest. Due to its high computational efficiency and unique one-way and collision-resistant properties, hash functions have become crucial for saving storage space and ensuring data integrity, thus becoming key to constructing core data structures such as Merkle trees.

[0038] From a mathematical perspective, a secure cryptographic hash function This can be represented as a mapping relationship. That is, for binary input strings of arbitrary length A fixed-length hash can be obtained by using a hash function. Bit-based binary output Hash function It possesses the following key cryptographic properties: (1) High efficiency and computability: Given msg, calculate It is efficient and feasible; (2) Unidirectionality: given Calculating msg is computationally infeasible; (3) Collision resistance: given Search Make and It is not feasible to calculate.

[0039] Hash functions are sensitive to input and random in output; a small change in the input `msg` can lead to a drastic and unpredictable change in the output. This characteristic makes hash functions an ideal tool for data integrity verification, by comparing the original hash values. The hash value recalculated after receiving The consistency of the message (msg) can determine whether it has been tampered with. This mechanism is widely used in software distribution (such as verifying the integrity of installation packages) and digital evidence storage.

[0040] In traditional public-key cryptography, if the sender needs to send... When multiple recipients send the same plaintext message, it is usually necessary to decrypt the message using each recipient's public key. Encryption is performed twice, and the data is sent. This method involves creating a separate encrypted message. However, this leads to significant computational overhead and wasted communication bandwidth in scenarios with a large number of receivers. To address the efficiency issues in one-to-many secure communication, broadcast encryption (BE) technology was developed.

[0041] Broadcast encryption allows the sender to perform an encryption operation only once to generate a ciphertext, which can then be decrypted individually by multiple pre-specified target receivers. Unauthorized users outside the target receivers cannot obtain the original message even if they intercept the ciphertext.

[0042] This technology comprises the following four core steps: (1) Initialization With a security parameter The input is the common parameter pp, and the output is the same.

[0043] (2) Key generation : Taking the common parameter pp as input, output the first... The public and private key pair of the message recipient ( Among them, public key Used to build broadcast groups, private key The recipient shall keep it securely for subsequent decryption operations.

[0044] (3) Broadcast encryption : The set of public keys of the target group to be broadcast to Take the message msg as input and output a ciphertext. .

[0045] (4) Broadcast decryption : with the first The private key of the message recipient target group public key set and ciphertext As input, output a plaintext message msg.

[0046] When the number of target recipients becomes 1, broadcast encryption remains consistent with general public-key encryption. In terms of security, broadcast encryption schemes must satisfy collusion resistance. This means that even if all recipients are not in the target set... Even if unauthorized users collude and share their private keys, they still cannot computationally crack the ciphertext. Receive plaintext message msg.

[0047] Furthermore, digital signatures (DS) are one of the core technologies in modern public-key cryptography used to maintain information security. They establish a mathematical binding relationship between digital information and a specific subject through cryptographic algorithms, aiming to achieve identity authentication, data integrity verification, and non-repudiation of actions in open network environments.

[0048] A typical digital signature scheme involves two main entities: the signer and the verifier. The signer uses their unique private key to perform a mathematical transformation on the message to generate a signature, while the verifier uses the signer's published public key to verify the signature. A standard digital signature scheme consists of the following four polynomial-time processes: (1) Initialization process With a security parameter The input is the common parameter pp, and the output is the same.

[0049] (2) Key generation process Input the public parameter pp to generate the signer's public-private key pair. The public key pk is publicly distributed, while the private key sk is securely kept by the signer.

[0050] (3) Signature generation process Generate a signature using the signer's private key and the message as input. .

[0051] (4) Signature verification processing After receiving the message and signature, the verifier uses the signer's public key to verify the message and signature. Perform a verification. Output 1 if the verification passes, otherwise output 0.

[0052] To ensure security, digital signatures must be unforgeable, meaning they cannot be forged without the private key. In this situation, attackers cannot forge a valid signature for any new message. , making Established.

[0053] Threshold signatures (TS) are a special type of digital signature based on the theory of secure multi-party computation and secret sharing. Unlike traditional digital signatures where the private key is held by a single entity, threshold signatures distribute signing authority among multiple participants, aiming to solve the single point of failure problem in key management and provide a higher level of security and fault tolerance.

[0054] In one Threshold signature scheme There are a total of Each signing participant's private key is divided into... Each share is secretly held by a different participant. Only when at least... Each participant collaborates, partially signing the message and then aggregating the signed signatures to generate a valid signature that can be verified across the entire network. Any number less than [a certain number] Even with collusion among all participants, a valid signature cannot be forged. This technique enhances the robustness of the system by distributing trust. A standard threshold signature scheme typically includes the following six core steps: (1) Initialization steps: with a security parameter The input is the common parameter pp, and the output is the same.

[0055] (2) Distributed key generation : Using the common parameter pp, the number of threshold signers Q, and the threshold value As input, the Q signers jointly execute a distributed key generation protocol to generate a global public key pk. Each signer generates a unique share of the public and private keys for signing.

[0056] (3) Generation of signature shares : with the first Each participant's private key share Take the message to be signed as input and output a signature share. .

[0057] (4) Signature share verification : with the first Each participant's public key share Message and signature shares The input is 1 if the validation passes, and 0 otherwise.

[0058] (5) Signature synthesis : Input any One legitimate signature share Output synthesized signature .

[0059] (6) Global signature verification Using the global public key, the message to be signed, and the synthesized signature The input is 1. Output 1 if the signature verification is successful, otherwise output 0.

[0060] Threshold signature technology possesses two key security properties: unforgeability and robustness. Unforgeability ensures that even if an attacker has less than [a certain amount of] control, the signature remains valid. In the case of one participant, a valid signature cannot be generated; robustness guarantees that even with [a certain number of participants], [the following is true]. If one participant crashes or is compromised, the remaining ones... Honest participants can still maintain the normal operation of the signature system. This governance model greatly enhances the survivability of the distributed system when faced with the failure of some nodes or malicious hijacking.

[0061] This application consists of four core entities, whose specific functions and responsibilities are described below: (1) Network Control Center (NCC): The NCC is the management center and task initiator of the entire space information network. Before the federated learning task begins, the NCC is responsible for the system initialization, including generating and distributing global training parameters (such as learning rate, iteration rounds, etc.), and registering and granting access permissions to clients, ground stations, and satellites applying to participate in the training. During the training process, the NCC is responsible for monitoring the topology connection status of the entire network, ensuring the stability of the satellite-ground link, and providing basic network environment support for multi-party collaborative computing.

[0062] (2) Client: The system contains There are 1 client, and the client set is 1 These clients are distributed across a wide geographical area and are typically terminal devices (such as satellite terminals, drones, and ground sensors). Client Having its own private dataset Its data distribution exhibits a significant non-independent and identically distributed (Non-IID) characteristic. Each client aims to train a personalized model adapted to its own data distribution through participation in federated learning. .

[0063] (3) Ground station: The system contains A set of ground stations with sufficient computing resources is represented as follows: Unlike the client, the j-th ground station Have datasets for specific tasks or regions And maintains the corresponding initial model. This model is from a ground station. Based on dataset It is obtained through pre-training. In the federated learning process, the ground station, as the main model aggregator, is responsible for collecting client models with similar data distributions to generate specialized models capable of handling specific domain features. .

[0064] Low Earth Orbit satellites: Existing in the system A set of satellites with sufficient coverage area serves as air relay nodes, represented as follows: Limited by onboard energy and heat dissipation conditions, satellites do not undertake heavy deep learning training tasks, but instead serve as communication hubs and computation offloading nodes connecting ground stations and clients. In the algorithm proposed in this invention, the satellite utilizes its wide-area coverage to perform lightweight fuzzy clustering calculations and route distribution tasks to solve the node association and model forwarding problems under dynamic topology.

[0065] As a trusted authority, the Network Control Center (NCC) operates the algorithm. , ,as well as This generates the global parameters needed to create the broadcast encryption scheme, digital signature scheme, and threshold signature scheme. Subsequently, the Network Control Center (NCC) executes the key generation algorithm for the digital signature. This is used to generate a digital signature key pair. Here, the private key... The public key is used to issue entity registration certificates and is publicly distributed. All parameters that need to be made public are uploaded by the NCC.

[0066] Entity registration: Part One entity (client) Ground station or satellite Run the broadcast encryption key generation algorithm Generate its own encryption key pair .

[0067] Specifically, client The key pair is ( ), ground station The key pair is ( ),satellite The key pair is ( ).

[0068] No. Each entity will have access to the private key. Securely store the information, and then transmit your identity information through a secure channel. (Client identity information) Ground station identity information Or satellite identity information ) and public key Send to the Network Control Center (NCC).

[0069] Network Control Center Verification legality and and If the binding relationship fails to be verified, the entity will be refused entry into the system; if the verification is successful, the credentials will be calculated. Entity-related information Stored in an internal database, and Upload and generate publicly verifiable registration records.

[0070] Satellite threshold signature public key registration: After all Q satellites have completed entity registration, joint execution is performed. Generate the global public key for the threshold signature scheme and the unique threshold signature key share for each satellite. The satellite will Confidential storage, transmitted via secure channel Send to the network control center. The network control center calculates the extended credentials. ,Will Stored in an internal database, and Upload and generate publicly verifiable registration records.

[0071] Any entity can verify the legitimacy of the target party based on registration records (by verifying certificates). and While ensuring the validity of the public key, it cannot link the public key information to the entity's real identity. When an entity exhibits abnormal behavior, the Network Control Center (NCC) can quickly trace its real identity through pre-stored mapping records in its secure database, thereby achieving accountability. This architecture protects the identity privacy of participating parties while ensuring system auditability.

[0072] Ground station pre-training: Before the first round of federal training, the... ground stations Utilizing local private datasets Pre-training is performed, where the ground station initializes the parameters to prevent the initial parameter gradients of the neural network model from vanishing or exploding, thus achieving rapid convergence in subsequent training and ultimately obtaining the initial model. The pre-training objective function is: , in Indicates ground station The model, For ground station Based on dataset The empirical loss function.

[0073] Through this pre-training process, each ground station generates an initial model adapted to the distribution of its specialized scene features. .although The amount of data may be much smaller than the total amount of global data, but these initial models with specific domain knowledge will serve as the cornerstone for subsequent "multi-source knowledge fusion" in the client, providing a solid guarantee for solving the client cold start problem and improving personalized adaptation capabilities.

[0074] After the above work is completed, the first Each ground station will initialize the model. Hash And submit it, then broadcast it to all satellites. This is to facilitate subsequent federal training, among which .

[0075] In space information networks, specialized models pre-trained by ground stations contain prior knowledge of specific domains. The core task of the model distribution phase is to efficiently and accurately transfer this knowledge to widely distributed clients via satellite relay networks. Considering the unknown and heterogeneous nature of client data distribution, this solution designs two model distribution strategies for satellites based on different training cycles. First round of federal training : The qth satellite After receiving the ciphertext, use the private key to perform a decryption operation to obtain the model. The satellite calculates its hash value and verifies whether a corresponding record exists. If no corresponding record exists, the model is immediately discarded; if a matching record exists, the model can be confirmed. It maintains integrity throughout the entire transmission process and is authorized to participate in subsequent aggregation calculations.

[0076] After that, the satellite The global initial model is obtained by averaging all the initial ground station models that pass hash verification. , in This represents the currently available set of ground station indexes. Represents a set The number of elements; satellite implement Generate a threshold signature share and combine the threshold signature share with... Upload. For the first round of training. , for The public key set of all clients currently covered by the service, used by the satellite. Encapsulate the global initial model into ciphertext. and to The client broadcasts the encrypted message.

[0077] Round t of Federal Training : After receiving the encrypted broadcast message, the satellite decrypts it using its private key. Subsequently, the satellite calculates the hash value. It then queries for matching records; if no match is found, the model is discarded.

[0078] After a successful match, set for The first in the current service area Wheel and the j-th ground station The public key set of all clients who have collaborated, i.e. Satellite execution Generate model ciphertext And send the encrypted message to the corresponding client group.

[0079] S2. Each client uses its local private dataset to train a personalized initial model based on the global initial model.

[0080] Before performing local training, the client needs to verify the validity of the received global initial model information, and then perform personalized aggregation on the valid models to obtain the personalized initial model for local training.

[0081] Client Two types of encryption models are received, corresponding to the first round of training and subsequent training rounds, respectively.

[0082] For the first round of training, if The client received the encrypted message. Using private key Decryption Compare its hash value with the corresponding record; After successful verification, collect arbitrary threshold signature shares. indivual After setting the threshold signature share, run the signature combination algorithm to generate a combined signature. ,confirm Once effective, the global initial model can be used for the first round of local training to obtain a personalized initial model.

[0083] for Round training, setting up the client No. The set of ground stations for wheel collaboration is as follows The client will receive (in and ), using private key Decryption The hash value is calculated sequentially and the corresponding record is queried. If no corresponding record is found, the model is deemed invalid and discarded. If a matching hash value is found, it indicates that the model has not been tampered with during the transmission from the ground station to the satellite relay and then to the client, and can be included in the subsequent weighted aggregation operation.

[0084] In the first Before local training, an initial model adapted to its data distribution is obtained on the client side. Based on the The weighting coefficients are calculated based on the historical similarity between the wheel and various surface stations, thereby enabling personalized aggregation of effective surface station models.

[0085] This step is skipped in the first round of training because the client directly uses the global initial model in this round.

[0086] For the Round training ( ≥2), let For the client In carrying out the first Before the local training round, aggregate the j-th ground station. model The weighting coefficients used .

[0087] , in, For the first Wheel Client With ground station Historical similarity; Client These weights are used to perform weighted aggregation on the effective surface station model, thereby obtaining the first... The initial model trained locally in one round: , Client In the The initial model definition for the first round of local training is as follows: , Local model update and upload: Client Obtain the After the initial global model is generated, based on its dataset... Perform multiple local iterative training iterations, with the optimization objective being: , in, For the client The local empirical loss function, Indicates the client The model parameters to be optimized in this round of training. The regularization coefficient is used to control the difference between the updated model and the initial model.

[0088] The pseudocode for Algorithm 1 is shown below. After local training is complete, the client calculates the personalized initial model. hash value At the same time, the model is encrypted using the public key of the satellite currently responsible for its area, and the ciphertext is then uploaded to the satellite.

[0089] Algorithm 1 Pseudocode Input: Initial model, number of local training iterations Client dataset Model learning rate Regularization coefficient .

[0090] Output: Personalized initial model .

[0091] ; For each local training iteration round do For each batch in each dataset ( ) do ; End for

[0092] End for ; return: .

[0093] After local training is complete, the client Personalized initial model Packaged and ready to be sent to the currently responsible client via the uplink. The satellite nodes in the region then proceed to the next step of dynamic grouping calculation.

[0094] S3. The satellite calculates the distance between each personalized initial model and the initial model, and performs extreme value standardization to obtain a similarity metric. Based on the similarity metric, it calculates the membership degree of each client to each ground station in the current round, calculates the historical similarity based on the membership degree, and matches several ground stations for each client based on the historical similarity, which serve as the ground station set for that client in this round.

[0095] In round t, the q-th satellite For the received ciphertext Perform decryption and verification operations to obtain the set of clients within its current service area. Corresponding client model and the All ground station models .

[0096] To facilitate personalized model aggregation at ground stations, a fuzzy C-means clustering method is used to select matching initial personalized models for clients based on the Euclidean distance between client models and ground station models. This is a soft partitioning clustering method based on fuzzy set theory. Its idea breaks through the rigid constraint of traditional hard clustering that data objects belong to only a single cluster. By quantifying the degree of belonging of data objects to multiple clusters through membership functions, it allows data to belong to multiple clusters simultaneously in a probabilistic form. It is more suitable for complex data classification scenarios with overlapping categories and fuzzy boundaries, and is highly compatible with the processing needs of mixed distributed client data in spatial information networks.

[0097] The core mathematical representation of fuzzy clustering is the membership matrix. ,in The total number of samples, The total number of clusters, matrix elements Indicates the first Each sample corresponds to a membership degree in the j-th cluster, and satisfies the normalization constraint. (That is, the sum of the membership degrees of a single sample to all clusters is always 1). This constraint ensures both the physical meaning of the membership degree and the rationality of the clustering results.

[0098] Taking the Fuzzy C-Mean (FCM) clustering algorithm as an example, this algorithm is the most classic and widely used algorithm in the field of fuzzy clustering. It achieves soft partitioning of data by alternately optimizing the membership matrix and cluster centers and minimizing the weighted squared error objective function. The objective function of FCM is defined as shown in the following formula.

[0099] , In the formula, For the first One input data sample, Let j be the cluster center of the j-th cluster. For the first The membership degree of each data sample to the j-th cluster center, and It is the set of all cluster centers.

[0100] The FCM algorithm finds the minimum value of the objective function through iterative optimization. Its core consists of two steps: membership update and cluster center update. These two steps are executed alternately until the algorithm converges. The specific solution process is as follows: Membership update: Fixed cluster center By solving the constrained optimization problem using the Lagrange multiplier method, the membership update formula that minimizes the objective function is obtained, as shown in the following equation: , In the formula, For the sample To the j-th cluster center The Euclidean distance is used to quantify the similarity between a sample and the cluster center. The smaller the distance, the greater the membership degree, and the stronger the sample's association with that cluster. These are fuzzy weighting coefficients used to control the membership matrix. Smoothness, when At that time, the membership degree tends to become binary ( FCM degenerates into the K-means algorithm for hard clustering; when At that time, the membership degrees of all samples to each cluster tend to be uniformly distributed. ), the cluster boundaries are completely blurred.

[0101] Cluster center update: fixed membership matrix Regarding the objective function By taking the partial derivatives and setting them to 0, we can obtain the update formula for the cluster centers, as shown in the following equation.

[0102] , This formula indicates that cluster centers It is the weighted average of all input samples, with the weight being the degree of membership of the sample to the cluster. Samples with high membership have a dominant role in the cluster center. The specific process is shown in Algorithm 2.

[0103] For the client Update model With ground station Aggregation model The Euclidean distance between the two is defined as: , (in It is a client-side model The parameter vector, It is a ground station model The parameter vector, the dimension of the parameter vector is 1. 。 We employ parameter-based Euclidean distance, rather than resource-intensive pre-trained models (such as BERT), to adapt to the stringent computational constraints of satellite relay. Since local model parameters implicitly encode data distribution, this metric serves as an efficient alternative measure of distribution similarity, effectively capturing offsets without incurring the overhead of complex feature extraction.

[0104] The satellite selects the maximum distance within its service area. and minimum distance It then exchanges these extreme values ​​with other satellites to determine the global maximum distance. and global minimum distance .

[0105] Subsequently, for the client With ground station The original distances between them are normalized to obtain a similarity metric, specifically: After obtaining the global maximum and global minimum distances, the following extreme value standardization calculations are performed: , in, This is a similarity metric between the i-th client and the j-th ground station. The distance between the personalized initial model for each client and the initial model for each ground station. and These are the global minimum distance and the global maximum distance.

[0106] Innovatively, low-Earth orbit satellites are used as dynamic clustering hubs, and a dynamic grouping strategy based on fuzzy C-means clustering (FCM) is used to propose a global standardized metric method for multi-satellite collaboration.

[0107] In traditional federated learning, a client can typically only be assigned to one cluster, which is ill-suited for the mixed distribution of data collected across regions in spatial information networks, easily leading to model adaptation bias. However, this invention allows a single client to simultaneously integrate information from multiple ground stations, thereby more accurately matching the characteristics of local mixed data.

[0108] We propose a global standardization method based on multi-satellite collaboration. This method can unify the distance measurement standards of all satellite coverage areas while sharing only a small amount of distance data. This avoids deviations in clustering results from different satellite regions, making the dynamic grouping results more closely match the actual data distribution of the client, resulting in faster model convergence and better fitting performance.

[0109] Through this standardization process, satellite The similarity matrix can be obtained to obtain the similarity measure. The similarity metric quantifies the standardized similarity relationship between clients and ground stations within its service coverage area. It is worth noting that through the above operations, the satellite can complete the global standardization of distance values ​​without revealing any of its own computational information.

[0110] Based on similarity metrics, membership degree is calculated as follows: The membership degree of each client to each ground station is calculated as follows: , in, This is a similarity metric between the i-th client and the j-th ground station. This is a similarity metric between the i-th client and the k-th ground station. For the first The membership degree between a client and the j-th ground station >1 is the fuzzy weight coefficient, M is the total number of ground stations, and k is the kth ground station.

[0111] The sum of the membership degrees of each client to all ground stations equals 1.

[0112] Construct a fuzzy clustering membership matrix. For the first Wheel Client ground station The membership degree only includes membership information for the current round.

[0113] To mitigate the impact of outliers and improve the accuracy of the clustering strategy, we incorporate membership into the historical similarity between the client and the ground station. Specifically, in round t... Historical similarity between client j and ground station j for: , , For the first The membership degree between a client and the j-th ground station, where t is the training round. The attenuation coefficient is... , It is located in An increasing function between them It is the total number of training rounds. For round t-1 Historical similarity between client j and ground station j.

[0114] It is an output located at An increasing function between them This is the total number of rounds of federal training. (The number of training rounds...) The increase, It keeps increasing, making The values ​​gradually decrease from larger initial values ​​to smaller values ​​later. This way, in the early stages of training, When the historical similarity is large, it is mainly dominated by information from the current round; in the later stages of training, When the data size is small, historical similarity relies more on long-term relationships that have been established beforehand. This avoids drastic changes to the grouping strategy due to a single disturbance, allowing the grouping strategy to better reflect the data similarity between the client and the ground station, thus ensuring the stability and accuracy of the grouping.

[0115] Since the clients covered by each satellite in different rounds are different, all satellites upload their calculated fuzzy clustering membership matrix and historical similarity matrix in plaintext, thereby ensuring that the entire network can continuously track the current status and historical evolution trend of the matching relationship between the client and the ground station.

[0116] This creates a global membership record on the chain. Similarity matrix with global history This includes all client-to-ground station affiliation information and historical similarity information. Missing items are marked as... .

[0117] A dynamic similarity update mechanism that integrates instantaneous membership degree and historical similarity.

[0118] If only the instantaneous membership degree of a single round is used for grouping, it is easily affected by the fluctuation of data in a single round, which will cause frequent changes in grouping, leading to oscillation of model training data and difficulty in convergence.

[0119] This mechanism combines the historical similarity between the client and the ground station with the instantaneous membership calculated in the current round, dynamically adjusting the weights of both. In the early stages of training, instantaneous membership is emphasized to allow the model to quickly adapt to the data distribution; in the later stages of training, historical similarity is given more weight to stabilize the model state, avoid oscillations caused by frequent grouping, and make the model convergence process smoother and the fitting accuracy higher.

[0120] To fully utilize multi-source data while avoiding model oscillations caused by redundant associations, this invention proposes a method based on historical similarity matrices. The associated threshold constraint mechanism.

[0121] This mechanism stipulates that each client can match a maximum of [number] clients. The client is determined by the following calculations for each ground station. No. The set of ground stations corresponding to the wheel : , in, The historical similarity matrix represents the composition of historical similarities. The Row vectors Indicates the first Wheel Client The assigned ground station index forwards the client's personalized initial model to the ground station set.

[0122] Meanwhile, if the historical similarity is lower than a preset threshold, it will not be included in the set.

[0123] satellite Perform encryption operation Obtain the ciphertext and sent to All ground stations. (Among them) ).

[0124] Algorithm 2 pseudocode: Input: Set of client models within the service area, set of ground station models, historical similarity matrix .

[0125] Output: Set of ground stations to be assigned.

[0126] 1. Calculation ; 2. ; 3. ; 4. Send to other satellites and ; 5. Receive the maximum and minimum values ​​from other satellites; 6. ; 7. ; 8. Calculation ; 9. Calculate and upload ; 10. Update the historical similarity matrix All and The corresponding value; 11. For each do 12. From the historical similarity matrix Extract ; 13. Obtain the set of ground stations to be assigned. ; 14. End for 15. Return: .

[0127] S4. Each ground station calculates the aggregation weight of the matched client based on historical similarity, performs weighted aggregation on the personalized initial model of the matched client, obtains the aggregated model corresponding to the ground station, and uses the aggregated model as the new initial model for each ground station to execute S1 until the preset iteration round is reached.

[0128] The specific steps are as follows: For all containing ciphertext Ground station Use private key Run the decryption algorithm Get the client Model parameters ; The ground station will calculate the hash value of the model. The system queries the corresponding records. If no match is found, the model is discarded. If a match is found, it indicates that the model has remained consistent during the multi-hop transmission from the client to the satellite and then to the ground station, and can be safely incorporated into subsequent operations. In the collection from clients After all receiving models have been verified, the ground station Based on historical similarity matrix The updated matching relationships obtained are then aggregated using the model.

[0129] To assign greater weight to clients with higher similarity to themselves, each ground station calculates the aggregate weight of matched clients based on historical similarity, specifically: ground station Client The ratio of the historical similarity to the sum of the historical similarities of all clients matched by that ground station is used as the client's... Corresponding aggregate weight .

[0130] ground station The client's personalized initial model is calculated using the following formula. Corresponding aggregate weight : , in, Indicates the first Wheel Client With ground station Historical similarities between them Indicates the first Wheel Client With ground station The historical similarity, where k is the kth client. Indicates for ground station Matching clients.

[0131] After determining the aggregation weight coefficients, the ground station For the Client-side model retained in rounds Perform weighted aggregation to obtain the aggregated model corresponding to the ground station. Use the aggregated model as the new initial model for each ground station and execute S1 until the preset iteration round is reached. , By using this weighted aggregation based on historical similarity, ground station models can better adapt to the data distribution of their specific scenarios, thereby improving personalized performance.

[0132] Thus, the first The training session is complete.

[0133] All ground stations should proceed to the model download phase and begin the first... Rotational training.

[0134] Algorithm 3: Ground Station Aggregation Input: Client collection Client model in Historical similarity matrix .

[0135] Output: Ground station aggregation model .

[0136] 1. For each do 2. From the historical similarity matrix Extracting clients At the corresponding ground station Historical similarity ; 3. Computing Client Allocation of aggregate weight coefficients ; 4. End for 5. Calculate the ground station aggregation model ; 6. Return: .

[0137] Thus, the first The federal training round is complete. All ground stations need to re-enter the model distribution phase and initiate the next... The process of moving from one round of training to the next is as follows: When the number of iterations is greater than or equal to 2, perform the following steps: S5. Broadcast the new initial model of each ground station via satellite to the clients that were matched with that ground station in the previous round; S6. Each client calculates the weight coefficient based on the historical similarity of the previous round, performs weighted aggregation on the new initial model of all matched ground stations, obtains the updated model, and uses its local private dataset for local training to obtain the personalized model. S7. The satellite calculates the distance between each personalized model and the new initial models of all ground stations, and performs extreme value standardization to obtain a new similarity metric. Based on the new similarity metric, it calculates the membership degree of each client to each ground station in the current round, and performs weighted fusion of the historical similarity of the previous round and the membership degree of the current round to obtain the new historical similarity of the current round. Based on the new historical similarity, several ground stations are matched for each client, which serve as the set of ground stations for that client in this round; S8. In this round of ground station collection, each ground station obtains new historical similarity, which is used as weight to perform weighted aggregation on multiple client-side personalized models to obtain a new aggregated model; The new aggregated model is used as the new initial model for each ground station to execute S5 until the preset iteration round is reached.

[0138] Algorithm 4, as shown below, details the complete process of bidirectional personalized federated training involving the collaborative participation of the client, ground station, and satellite.

[0139] Algorithm 4: Bidirectional Personalized Federated Learning Framework Input: Total number of training rounds Local training iteration rounds Client collection Client dataset Ground station assembly Pre-trained model set Regularization coefficient Satellite collection Initial historical similarity matrix Model learning rate .

[0140] Output: Final client model Final ground station model .

[0141] 7. For each do 8. / / Model distribution phase 9. For each do 10. Assemble the satellites All satellite broadcast ground station models ; 11. End for 12. For each do 13. If

[0142] 14. Aggregate and generate a global initial model; 15. Broadcast to all clients within the service area; 16. Else 17. To the set Client broadcast model in

[0143] 18. End if 19. End for 20. / / Client-side personalized model update 21. For each do 22. Generate the initial model for local training; twenty three. Algorithm 1: Local Model Update ; 24. Send to the nearest satellite; 25. End for 26. / / Satellite dynamic grouping phase 27. For each do 28. Extract the historical similarity matrix from distributed storage nodes. ; 29. Algorithm 2: Dynamic Grouping ; 30. For each do 31. Client-side model Send to collection All ground stations ; 32. End for 33. End for 34. / / Ground station aggregation phase 35. For each do 36. Algorithm 3: Ground Station Aggregation

[0144] 37. End for 38. End for Traditional federated learning is mostly a one-way aggregation model, either focusing only on global model optimization or only providing one-way personalized customization for clients, completely ignoring the professional needs of ground stations for specific scenarios. Moreover, the architecture of a single ground station serving multiple clients also traps clients in information cocoons, making it impossible to adapt to the mixed distributed data brought about by cross-regional collection.

[0145] The bidirectional framework of this invention allows ground stations to continuously iterate and optimize specialized models for specific scenarios, while enabling clients to adaptively integrate models from multiple ground stations to generate personalized models that fit their local mixed data distribution. This satisfies both the specialized needs of ground stations for specific scenarios and the personalized adaptation of clients, breaking the limitations of traditional one-way aggregation and achieving a bidirectional balance between global knowledge transfer and local personalized adaptation.

[0146] Furthermore, in the dynamic and open environment of Space Information Networks (SINs), client disconnections caused by satellite switching or signal attenuation pose a significant challenge to federated learning systems. To maintain training continuity and ensure robust model convergence, this application proposes a disconnection handling mechanism based on historical similarity prediction.

[0147] In response to the Reconnect to satellite client Because it was not updated during the disconnection period, the existing historical similarity cannot accurately reflect the current data similarity between the client and various ground stations. To solve this problem, satellite... With the client Perform the following recovery procedure: Historical similarity extraction: satellite Extract client With each ground station recent Wheel (default) If the client has participated in fewer than 5 rounds, then the actual number of rounds is used (the historical similarity sequence is used). ,in Indicates that a client has been detected. The number of rounds to disconnect. This represents the historical similarity value from the last round before disconnection; Similarity trend calculation: For each ground station's historical similarity sequence, the satellite uses a piecewise difference strategy to calculate the approximate rate of change of historical similarity. , Subsequently, the satellite processed the resulting sequence. Taking the average, we get This value reflects the client With ground station recent The average trend of historical similarity.

[0148] Similarity update: based on reconnecting client Satellite Superimposed with the corresponding historical average change value After the client reconnects, the first Wheels and ground stations Historical similarity .

[0149] When the client reconnects, it communicates with the ground station. Historical similarity between The specific expression is as follows: , After completing the historical similarity update, the satellite Will Send to the client .

[0150] Obtaining the initial model for local training: Client Calculations are performed to obtain its value in the th order. The initial model trained locally in one round: , , At this point, the client... Then, following the detailed process described in the client-side local training phase and subsequent steps, the first step can be executed. Through local training and subsequent operations, they successfully rejoined the federated learning training process.

[0151] By following the steps above, reconnect the client. Able to perform the first based on the corrected historical similarity The local training and subsequent operations (such as Algorithm 4) of the rounds were successfully reintegrated into the federated training process.

[0152] Therefore, a layered multi-party trust model adapted to spatial information network scenarios was constructed based on the consortium blockchain. Through on-chain evidence storage of model hash and historical similarity, the entire federated training process was made traceable and auditable.

[0153] It integrates broadcast encryption and threshold signature technologies to ensure the confidentiality and authentication of model transmission and enhances anti-eavesdropping capabilities.

[0154] Furthermore, this application employs a disconnection recovery mechanism based on historical similarity difference prediction, which enables the method of this invention to maintain the convergence stability of federated training even in scenarios with extremely high disconnection rates, thereby enhancing the robustness of the framework to dynamic network disturbances.

[0155] A bidirectional personalized federated learning system for spatial information networks, used in a bidirectional personalized federated learning method for spatial information networks, comprising: The global model generation module allows each ground station to pre-train its model using its local private dataset, obtain an initial model, and broadcast it to the satellite. The satellite then averages and aggregates all the initial models to obtain a global initial model, which is then broadcast to all clients. The personalization module allows each client to train its own personalized initial model based on the global initial model and its local private dataset. The matching module calculates the distance between each personalized initial model and the initial model, performs extreme value standardization, and obtains a similarity metric. Based on the similarity metric, it calculates the membership degree of each client to each ground station in the current round, calculates the historical similarity based on the membership degree, and matches several ground stations for each client based on the historical similarity, which serve as the ground station set for that client in this round. The aggregation module calculates the aggregation weight of the matched clients based on historical similarity for each ground station, performs weighted aggregation on the personalized initial models of the matched clients to obtain the aggregated model corresponding to the ground station, and uses the aggregated model as the new initial model for each ground station to be processed by the global model generation module until the preset iteration round is reached.

[0156] A bidirectional personalized federated learning device for a spatial information network includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement a bidirectional personalized federated learning method for a spatial information network.

Claims

1. A bidirectional personalized federated learning method for spatial information networks, characterized in that, Includes the following steps: S1. Each ground station uses its local private dataset to pre-train and obtain an initial model, which is then broadcast to the satellite. The satellite averages and aggregates all the initial models to obtain a global initial model, which is then broadcast to all clients. S2. Each client uses its local private dataset to train a personalized initial model based on the global initial model. S3. The satellite calculates the distance between each personalized initial model and the initial model, and performs extreme value standardization to obtain a similarity metric. Based on the similarity metric, it calculates the membership degree of each client to each ground station in the current round, calculates the historical similarity based on the membership degree, and matches several ground stations for each client based on the historical similarity, which serve as the ground station set for that client in this round. In S3, historical similarity is calculated based on membership degree, specifically as follows: Round t Historical similarity between client j and ground station j for: ， in, For the first The membership degree between a client and the j-th ground station, where t is the training round. The attenuation coefficient is... , It is located in An increasing function between them It is the total number of training rounds. For round t-1 Historical similarity between client and j-th ground station; In S3, several ground stations are matched for each client based on historical similarity, forming the ground station set for that client in this round, specifically: Each client can match up to [number] times. The client is determined by the following calculations for each ground station. No. The set of ground stations corresponding to the wheel : ， in, The historical similarity matrix represents the composition of historical similarities. The Row vectors Indicates the first Wheel Client The assigned ground station index forwards the client's personalized initial model to the ground station set; S4. Each ground station calculates the aggregation weight of the matched clients based on historical similarity, performs weighted aggregation on the personalized initial model of the matched clients, obtains the aggregated model corresponding to the ground station, and uses the aggregated model as the new initial model for each ground station. S5. Broadcast the new initial model of each ground station via satellite to the clients that were matched with that ground station in the previous round; S6. Each client calculates the weight coefficient based on the historical similarity of the previous round, performs weighted aggregation on the new initial model of all matched ground stations, obtains the updated model, and uses its local private dataset for local training to obtain the personalized model. S7. The satellite calculates the distance between each personalized model and the new initial models of all ground stations, and performs extreme value standardization to obtain a new similarity metric. Based on the new similarity metric, it calculates the membership degree of each client to each ground station in the current round, and performs weighted fusion of the historical similarity of the previous round and the membership degree of the current round to obtain the new historical similarity of the current round. Based on the new historical similarity, several ground stations are matched for each client, which serve as the set of ground stations for that client in this round; S8. In this round of ground station collection, each ground station obtains new historical similarity, which is used as weight to perform weighted aggregation on multiple client-side personalized models to obtain a new aggregated model; The new aggregated model is used as the new initial model for each ground station to execute S5 until the preset iteration round is reached.

2. The method according to claim 1, characterized in that, S3 performs extreme value standardization to obtain a similarity measure, specifically: To obtain the global maximum and global minimum distances, the following calculations are performed: ， in, This is a similarity metric between the i-th client and the j-th ground station. The distance between the personalized initial model for each client and the initial model for each ground station. and These are the global minimum distance and the global maximum distance.

3. The method according to claim 1, characterized in that, In S3, membership degree is calculated based on similarity metric, specifically as follows: The membership degree of each client to each ground station is calculated as follows: ， in, This is a similarity metric between the i-th client and the j-th ground station. This is a similarity metric between the i-th client and the k-th ground station. For the first The membership degree between a client and the j-th ground station >1 is the fuzzy weight coefficient, M is the total number of ground stations, and k is the kth ground station.

4. The method according to claim 3, characterized in that, The sum of the membership degrees of each client to all ground stations equals 1.

5. The method according to claim 1, characterized in that, In S4, each ground station calculates the aggregate weight of matched clients based on historical similarity, specifically as follows: ground station Client The ratio of the historical similarity to the sum of the historical similarities of all clients matched by that ground station is used as the client's... Corresponding aggregate weight .

6. A bidirectional personalized federated learning system for spatial information networks, used to implement the bidirectional personalized federated learning method for spatial information networks as described in any one of claims 1-5, characterized in that, include: The global model generation module allows each ground station to pre-train its model using its local private dataset, obtain an initial model, and broadcast it to the satellite. The satellite then averages and aggregates all the initial models to obtain a global initial model, which is then broadcast to all clients. The personalization module allows each client to train its own personalized initial model based on the global initial model and its local private dataset. The matching module calculates the distance between each personalized initial model and the initial model, performs extreme value standardization, and obtains a similarity metric. Based on the similarity metric, it calculates the membership degree of each client to each ground station in the current round, calculates the historical similarity based on the membership degree, and matches several ground stations for each client based on the historical similarity, which serve as the ground station set for that client in this round. The aggregation module calculates the aggregation weight of the matched clients based on historical similarity for each ground station, performs weighted aggregation on the personalized initial models of the matched clients to obtain the aggregated model corresponding to the ground station, and uses the aggregated model as the new initial model for each ground station to be processed by the global model generation module until the preset iteration round is reached.

7. A two-way personalized federated learning device for spatial information networks, characterized in that, It includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement a bidirectional personalized federated learning method for spatial information networks as described in any one of claims 1-5.

Images