Satellite internet model security data sharing method based on federated learning

Abstract

The invention discloses a federated learning-based satellite internet model security data sharing method, which comprises the following steps of: firstly, constructing a self-adaptive ground mobile terminal clustering strategy based on a dynamic graph attention mechanism, and establishing a space-time sensing ring topology by analyzing a node movement rule and communication characteristics to lay a network foundation for decentration calculation; secondly, a modular decomposition confusion encryption scheme is developed, homomorphic encryption of gradient parameters is realized by using a redundant modular component superposition and random number injection technology, the ciphertext size is compressed, and dependence of a third-party center is avoided through distributed private key management. And finally, creating a multi-sub-ring parallel computing architecture, splitting a modular decomposition task into independent computing units, and realizing local model combination based on a proximity communication principle, so that the bottleneck of a traditional server is eliminated, and the aggregation delay is also reduced. According to the whole scheme, through collaborative design of a topology layer, an encryption layer and a calculation layer, feasibility and efficiency balance of cross-domain collaborative training is achieved under the condition that data privacy is guaranteed.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of distributed satellite internet, and particularly relates to a satellite internet model security data sharing method based on federated learning. BACKGROUND

[0002] The wide-area coverage characteristics of low-orbit satellite internet and the data privacy and security challenges brought by the heterogeneous networking method. On the one hand, the spectrum feature information and communication data generated by heterogeneous ground mobile terminal devices (vehicle, aircraft, high-speed mobile terminal device such as high-speed rail) in the inter-satellite link or cross-layer transmission between satellite and ground, if lack of on-board end-to-end encryption measures, may lead to third-party attackers (high-orbit relay satellite or stationary orbit spy satellite) intercept sensitive information through man-in-the-middle attack; on the other hand, restricted by national spectrum sovereignty and orbit resource competition, different operators need to maintain the exclusivity of core data assets (such as terminal location trajectory and spectrum characteristics), and also need to realize knowledge fusion of multi-source data under the framework of satellite-ground collaborative computing, which puts higher requirements on the verifiability of distributed privacy computing protocol. Research shows that when the ground mobile terminal transmits encrypted positioning data to the low-orbit relay satellite through the inter-satellite link, even if differential privacy technology is used, there is still a possibility of reverse inference of terminal position and communication content through terminal moving trajectory and ephemeris data correlation analysis, highlighting the necessity of dynamic trust authentication mechanism on the sky in the sky-ground integrated network security architecture.

[0003] The implementation of intelligent driving includes three links of perception, decision and execution. Due to the limited computing resources of ground mobile terminals and the transmission delay problem caused by processing multi-modal remote sensing data, single-satellite autonomous intelligence faces serious challenges. These tasks usually cause serious queuing delay of inter-satellite link, affecting the safety of satellite-ground network, the overhead of laser communication and the efficiency of star swarm federated learning. With the continuous expansion of low-orbit constellation scale and satellite-ground integrated data, the future development of intelligent driving cannot be separated from the star-cloud-end collaborative computing architecture, and the networking technology of low-orbit mega constellation provides support for this goal. However, the wide-area heterogeneous topology and diversified access mode of low-orbit satellite Internet also bring serious privacy and security challenges. When the data related to the positioning trajectory of high-speed mobile terminals is transmitted to the server outside the local network, it may expose privacy leaks. If the data is not adequately protected during transmission, unauthorized static-orbit monitoring satellites may intercept, access or modify it, ultimately compromising the privacy of individuals related to the data. In addition, the satellite Internet contains heterogeneous satellite nodes operated by multiple countries and cross-military-commercial systems. These participants cannot directly share raw telemetry data due to orbit resource competition or restrictions of International Telecommunication Union (ITU) spectrum division rules, and lack of a unified inter-satellite trust anchor, making data privacy protection a core challenge for cross-trust domain collaboration. Satellite operators need to protect the strategic confidentiality of their own data, such as terminal frequency hopping communication features and satellite-borne SAR imaging parameters, while ensuring that the star swarm federated learning model maintains high accuracy in the complex global electromagnetic environment.

[0004] Federated Learning (FL) as a decentralized machine learning paradigm enables different computing nodes to securely collaborate and share models with the aggregation node rather than raw data. This provides a new approach to solving the privacy protection and data sharing problem in satellite Internet. However, relying solely on federated learning cannot completely solve the security problems in satellite Internet. Traditional protocols may need to be decrypted or rely on a unified key at the satellite-ground aggregation node, which may allow static-orbit spy satellites to reverse engineer or infer sensitive data from inter-satellite link parameters. And traditional federated learning relies on a central aggregation server, which has a single point of failure risk. Existing security enhancement mechanisms have limitations: differential privacy protects individual information by adding noise to data, but in some cases it may compromise model accuracy. In addition, You et al. constructed a sample reconstruction attack against local differential privacy (LDP), proving the insecurity of FL based on LDP; secure multi-party computation can protect the computing process, but the communication overhead is large and difficult to extend to large-scale federated learning nodes, and multiple trusted nodes are needed to cooperate. Using traditional encryption technology usually requires data to be decrypted before computing, which means that terminal sensitive data derived model parameters need to be exposed at the aggregation stage, and end-to-end privacy protection cannot be provided.

[0005] Homomorphic Encryption (HE) technology allows direct computation operations on encrypted data without prior decryption, so that the decrypted result is equivalent to the result of the same operation on the original data. This feature provides end-to-end privacy protection in satellite-ground, perfectly matches the principle of "data not leaving the edge node" of satellite constellation federated learning; at the same time, it avoids centralized risk, allows low-orbit satellite nodes to process data without knowing, reduces the trust dependence on the center node; and has natural compatibility with the core operations of satellite federated learning (such as linear operation). However, existing homomorphic encryption technology often has huge computational overhead and communication burden, making it difficult to meet the strict real-time requirements of sub-second inter-satellite collaborative reasoning in low-orbit constellation. Therefore, in the resource-constrained satellite Internet environment, how to balance the strength of privacy protection and system efficiency is still a problem to be solved. SUMMARY

[0006] The present application aims to at least partially solve one of the technical problems in the related art.

[0007] The present application proposes a satellite Internet model security data sharing method based on federated learning, integrating dynamic topology encryption and privacy enhancement technology. A multi-hop ring topology is constructed based on the dynamic grouping mechanism of the GAT architecture, and a double protection strategy of modular operation confusion and homomorphic encryption is used synchronously to realize privacy protection of gradient parameters in the ring transmission path. This scheme replaces the centralized aggregation of traditional federated learning with a multi-node relay verification mechanism, enabling each terminal node to complete local model collaborative optimization in the encrypted space while ensuring cross-domain data isolation, effectively solving the problem of trusted exchange of heterogeneous terminal data during joint training.

[0008] To achieve the above purpose, the present application proposes a satellite Internet model security data sharing method based on federated learning, comprising:

[0009] A low-orbit satellite Internet distributed federated learning system is constructed, including high-speed mobile terminals and satellite collaborative processing unit nodes. Each terminal is used to collect environmental data and perform machine learning. The system uses a ring topology communication mechanism, and the terminals transmit data and aggregate model parameters through the ring structure;

[0010] The terminal state feature vector defined by the system structure is standardized, the communication connection between terminals is defined based on communication distance and delay to construct an adjacency matrix, and the importance weight of different neighbor nodes is allocated through the GAT network using the attention mechanism to generate an embedding vector; the cosine similarity between terminals is calculated based on the embedding vector to form a similarity matrix, and a greedy algorithm is used to divide the terminals into multiple cluster information;

[0011] The plaintext data is amplified and random noise is added based on the cluster information, a set of mutually prime modulus bases is set, and real modulus components are calculated, an initial redundancy set is constructed to confuse the real modulus components, and the real modulus components are covered to the redundancy set through the key, and a confusion matrix is generated;

[0012] The confusion matrix generated by each terminal is aggregated through ring communication and element-by-element addition operation, the real component aggregation result is extracted from the aggregated confusion matrix, and the complete aggregation model is reconstructed by using the Chinese remainder theorem.

[0013] The satellite Internet model security data sharing method based on federated learning of the embodiment of the application constructs an adaptive ground mobile terminal clustering strategy based on a dynamic graph attention mechanism, establishes a space-time perception ring topology by analyzing node movement rules and communication characteristics, lays a network foundation for decentralized computing, develops a modulus decomposition confusion encryption scheme, realizes homomorphic encryption of gradient parameters by using superposition of redundant modulus components and random number injection technology, compresses the size of ciphertext, and avoids third-party center dependence through distributed private key management. A multi-subring parallel computing architecture is created, the modulus decomposition task is divided into independent computing units, local model combination is realized based on the principle of nearby communication, which not only eliminates the traditional server bottleneck, but also reduces the aggregation delay. The whole scheme realizes the feasibility and efficiency balance of cross-domain collaborative training under the condition of guaranteeing data privacy through the collaborative design of topology layer, encryption layer and computing layer.

[0014] Additional aspects and advantages of the application will be set forth in part in the description which follows, and in part will become apparent to those skilled in the art upon examination of the following and / or can be learned by practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0015] The above and / or additional aspects and advantages of the application will become apparent and be readily understood from the following description, taken in conjunction with the accompanying drawings, in which:

[0016] Figure 1 is a flowchart of the satellite Internet model security data sharing method based on federated learning according to the embodiment of the application;

[0017] Figure 2 is a whole architecture diagram of the satellite Internet model security data sharing method based on federated learning according to the embodiment of the application;

[0018] Figure 3 is an algorithm architecture diagram based on confusion modulus decomposition homomorphic encryption according to the embodiment of the application;

[0019] Figure 4 is a distributed federated learning algorithm architecture diagram based on ring communication according to the embodiment of the application;

[0020] Figure 5is a flowchart of an algorithm according to an embodiment of the present application. DETAILED DESCRIPTION

[0021] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0022] In order for those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings of the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the scope of protection of the present application.

[0023] A satellite Internet model secure data sharing method based on federated learning according to an embodiment of the present application will be described below with reference to the accompanying drawings.

[0024] Figure 1 is a flowchart of a satellite Internet model secure data sharing method based on federated learning according to an embodiment of the present application, as Figure 1 shown, comprising:

[0025] S1, a low-orbit satellite Internet distributed federated learning system is constructed, containing high-speed mobile terminals and on-board cooperative processing unit nodes, each terminal is used to collect environmental data and perform machine learning, the system adopts a ring topology communication mechanism, and the terminals perform data transmission and model parameter aggregation through the ring structure;

[0026] S2, the terminal state feature vector defined by the system structure is standardized, an adjacency matrix is constructed based on the communication distance and time delay to define the communication connection between terminals, an embedding vector is generated by using the attention mechanism to assign importance weights to different neighbor nodes through the GAT network; the cosine similarity between terminals is calculated based on the embedding vector to form a similarity matrix, and a greedy algorithm is used to divide the terminals into multiple cluster information;

[0027] S3, based on the cluster information, plaintext data is amplified and random noise is added, a set of mutually prime modules is set, and real module components are calculated, an initial redundancy set is constructed to confuse the real module components, and the real module components are covered to the redundancy set through a key to generate a confusion matrix;

[0028] S4, the confusion matrix generated by each terminal is aggregated through ring communication and element-by-element addition operation, the component aggregation of the confusion matrix is performed, the real component aggregation result is extracted from the aggregated confusion matrix, and the complete aggregation model is reconstructed by using the Chinese remainder theorem.

[0029] The application constructs a system model of a distributed federated learning algorithm based on homomorphic encryption of confusion module decomposition, as shown in Figure 2 The distributed federated learning model and the secure communication model involved in the algorithm are described respectively. The implementation details of the distributed homomorphic encryption federated learning are introduced in detail, including three important parts, i.e., the ground mobile terminal clustering and ring formation scheme design based on GAT, the homomorphic encryption scheme design based on confusion module decomposition, and the distributed federated learning algorithm design based on ring communication.

[0030] In an embodiment of the application, the system structure is as follows:

[0031] It is assumed that a satellite Internet system composed of 8 high-speed mobile terminals (VHSTs) and three satellite coordination unit (SCU) nodes is running in the coverage area of a low-orbit satellite, denoted as a set V = {v1, v2,..., v8}. Each ground mobile terminal vi i ∈ V has a perception module and local computing capability, can perceive and collect environmental data as a local data set D i i∈[1,8], and performs machine learning on the data set. At the same time, each terminal has a state vector hi i i∈[1,8], representing the current position, speed and heading angle of the terminal.

[0032] The satellite Internet system uses the distributed federated learning algorithm based on homomorphic encryption of confusion module decomposition proposed in the application. First, SCUa divides terminals 1, 2 and 3 into cluster A according to the state vectors of the terminals within the communication range, and SCUb divides terminals 5, 6 and 7 into cluster B. The terminals in the cluster have good communication performance and communicate in the form of a ring. Taking cluster A as an example, in t rounds of communication, the terminals have the terminal public key and private key fragments (pk, sk i ) issued by the key management center (KMC) i∈{1,2,3}, and each terminal trains to obtain a local model i∈{1,2,3}. Then, each terminal encrypts the local model using the public key pk to obtain the real module component i∈{1,2,3}, Then, based on the homomorphic encryption algorithm of confusion module decomposition, the terminal generates a redundant module component and uses the private key sk i to confuse to obtain a confusion matrix i∈{1,2,3} to mask the real modulus component. Since the real modulus component confusion position is related to the terminal private key, the real modulus component position of each terminal is random and cannot be known by other nodes. On this basis, terminals 1, 2, and 3 perform ring communication, sending data to the "next" neighbor in the clockwise direction while receiving data from the "previous" neighbor in the counterclockwise direction. During the ring communication process, the encrypted components are transmitted and aggregated among the terminals in the cluster along the ring topology, and the components are aggregated in modulus base units. Terminal 1 is responsible for component aggregation under modulus base n1, and after going through a complete communication round, terminal 1 obtains the aggregation result Correspondingly, terminal 2 obtains the aggregation result of component aggregation under aggregated modulus base n2 Terminal 3 obtains the aggregation result of component aggregation under aggregated modulus base n3 Then each terminal locally uses the private key to extract the aggregation result of the real modulus component under the modulus base it is responsible for And broadcasts this result to the remaining terminals in the cluster. Finally, the terminal locally decrypts to obtain the updated global model W using the Chinese Remainder Theorem t+1 And updates the local model for use in the next round of training. The entire process is completed in an encrypted state, and node 4 outside the cluster cannot obtain the parameter information of the terminals inside the cluster, ensuring the security of the model. At the same time, distributed aggregation is used, allowing terminals to participate in model training as training nodes and participate in distributed aggregation as aggregation nodes.

[0033] In an embodiment of the present application, the distributed federated learning model is as follows:

[0034] Based on the system model framework, intelligent networking terminals in the satellite coverage area are divided into multiple terminal clusters C={C1, C2,..., C m} and n terminals in the cluster form a ring communication structure. In the ring topology, each terminal v i ∈C m , i∈[1,n] has and only has two neighbor nodes, which send data to the "next" neighbor in the clockwise direction while receiving data from the "previous" neighbor in the counterclockwise direction. The specific process is as follows: in each iteration t, terminal v i uses local data D i to train and obtain a local model Subsequently, the terminal uses a homomorphic encryption algorithm based on confusion modulus decomposition to encrypt the local model and obtain a confusion matrix

[0035]

[0036] Where Encrypt represents the encryption operation, is the confusion modulus component under modulus base n. Then, the terminal transmits the components Send to "next" terminal node v j ∈C m ,j∈[1,n] and component aggregation, i.e. where denotes operation under modulo base n. This calculation result is passed to the next node along the direction of ring communication, and the component aggregation operation under modulo base n is repeated. Finally, terminal v i receives the component aggregation result from the "previous" neighbor node, which is expressed using the following formula.

[0037]

[0038] Then, terminal v i uses the private key fragment sk i locally to extract the real aggregation component result and broadcasts it to other intra-cluster terminal nodes. After each terminal receives the modulo component aggregation result from other terminals, it decrypts it locally using the Chinese remainder theorem, which is expressed using the formula

[0039]

[0040] where Decrypt denotes the decryption operation, and W t+1 is the global model obtained by aggregation. The terminal will use this global model for the next round of training. The entire process described above does not require the participation of a central server, and the parameter transmission and aggregation process of federated learning is always carried out in an encrypted environment. Table 1 summarizes the important symbols used in the present application.

[0041] In an embodiment of the present application, the secure communication model is as follows:

[0042] The terminal nodes in the network are distributed in different geographical locations and move dynamically. The state of each terminal node can be represented by a feature vector, i.e. h i = [x i , y i , υ i , θ i ]. Where (x i , y i ) is the location of the terminal, υ i is the current speed of the terminal, and θ i is the heading angle. The distance between terminal v i and terminal v j is expressed using the Euclidean distance as Each terminal is equipped with a wireless communication module and exchanges information with surrounding terminals based on (V2V) technology. The communication delay between terminals is denoted as

[0043]

[0044] where B denotes the communication bandwidth, P i is the transmit power of terminal v i , N0 is the noise power, g i,j is the channel gain, and |D ij | is the data size transmitted between terminals. The communication quality between terminals is negatively correlated with their distance d ij . The farther the distance, the lower the channel gain and the worse the communication quality.

[0045] In such a communication environment, the system faces the risk of hybrid attacks under the star-ground link. The present application adopts a hybrid adversary model that conforms to the real scenario to formalize the threat. The model considers two types of threat subjects: inter-satellite passive adversaries and external eavesdroppers. Passive adversaries conform to the "Honest-but-Curious" model, such as competing operator on-board computing nodes. Such adversaries strictly follow the inter-satellite communication protocol, but will attempt to derive private data of other participants from the information obtained legally. External eavesdroppers have the ability to intercept V2V communication traffic, but cannot directly participate in the protocol execution process. The present application makes the following assumptions about the ability boundary of the adversary: the adversary can listen to and record any proportion of network communication, but cannot modify the encrypted communication content between uncontrolled nodes. The adversary has probabilistic polynomial-time (PPT) computing power, which conforms to the standard cryptographic difficulty assumption. This means that the adversary cannot solve the discrete logarithm problem based on the homomorphic encryption of the confusion module within a feasible time. Specifically, in the homomorphic encryption scheme based on the confusion module decomposition proposed by the present application, the security relies on the following difficult problem:

[0046] Discrete logarithm problem (DLP):

[0047] For the large prime modulus q l selected by the present application, l∈1,2,...,n, given the generator g and g x modq i , the adversary cannot efficiently compute x within a polynomial time. Mathematically, it is expressed as

[0048]

[0049] where C is any PPT adversary algorithm, negl(λ) is a negligible function with respect to the security parameter λ, Zq l denotes the integer ring modulo q l , and x is uniformly randomly selected from the integer ring. In the implementation of the present application, the security parameter λ determines the bit length of the prime number q l , and |q l |≥2048 bits are usually required to ensure sufficient security.

[0050] Security of obfuscated modular decomposition based on Chinese remainder theorem (CRT):

[0051] The homomorphic encryption scheme of the present application is based on modular decomposition technology, in which the terminal model parameters are decomposed into the residue classes of different prime numbers q1, q2,..., q n . According to the Chinese remainder theorem, only when all the modulus and all the residue classes are known can the original data be restored. In the distributed design of the present application, each terminal v i only holds part of the modulus information, ensuring that even if part of the terminal is attacked, the enemy cannot obtain the complete encryption parameters, which is expressed by the formula as

[0052]

[0053] where k < n, indicating that the enemy cannot infer the residue value under the unknown modulus when knowing part of the modular residue and the corresponding modulus.

[0054] Security of homomorphic property:

[0055] The homomorphic encryption scheme designed in the present application supports additive homomorphism, that is, the operation performed in the ciphertext domain can be reflected in the plaintext domain.

[0056]

[0057] At the same time, the scheme meets the semantic security, that is, the enemy cannot distinguish the encryption results of any two different plaintexts, that is,

[0058]

[0059] where, and are any plaintext. The related symbols are shown in Table 1.

[0060] Table 1

[0061]

[0062]

[0063] In an embodiment of the present application, the terminal clustering based on GAT and the ring formation algorithm is as follows:

[0064] In order to optimize the communication efficiency and protect the privacy, based on the communication model established in the third section, the present application proposes a terminal clustering scheme based on GAT. This scheme can reasonably group the terminals according to the space-time attributes, movement characteristics and communication capabilities of the ground mobile terminals, and arrange them in a reasonable ring structure to realize subsequent secure aggregation.

[0065] First, the state feature vector h i = [x i , y i , u i , θ i ], i ∈ [1, N] of each terminal node is normalized, which is expressed by the formula

[0066]

[0067] where μ x , σ x , μ y , σ y , μ υ , σ υ , μ θ , σ θ are the mean and standard deviation of each feature dimension, respectively.

[0068] Next, the communication connection between terminals is defined based on the communication distance and delay between terminals, and the adjacency matrix A i ∈ R N×N is used to represent

[0069]

[0070] where A ij = 1 indicates that there is an effective communication link between terminal v i and terminal v j , and vice versa. d lim is the maximum distance limit between terminals, and τ lim is the maximum delay limit between terminals. When the distance or delay between terminals exceeds the maximum limit, the communication link is disconnected.

[0071] On the basis of the adjacency matrix, the GAT network is applied to convert the original features into embedded representations with more semantic information. The core of GAT is to adaptively assign importance weights to different neighbor nodes through attention mechanism. First, the normalized terminal feature matrix H 0 ∈ R N×4 is linearly transformed by the weight matrix to output Z 0 ∈ R N×d , d is the embedding space dimension. Next, the attention coefficients are calculated for the connected terminals, and for the Kth attention head, it is calculated as

[0072]

[0073] where a k ∈ R 2d is the learnable vector of the Kth attention head, and || represents the feature connection operation. For the unconnected terminals, i.e. Aij =0, set Then, the attention weights are normalized.

[0074]

[0075] Among them, N i For terminal v i The adjacent terminal set. Based on this attention weight, weighted aggregation is performed, that is,

[0076]

[0077] Finally, each terminal v i Get an embedding vector z i , which encodes the spatial position, mobility characteristics and interaction relationship of the terminal with the surrounding terminals. Based on the obtained embedding vector, the cosine similarity between the terminals is calculated to form the similarity matrix S ij .

[0078]

[0079] Based on this similarity matrix, a simple greedy algorithm is used to form terminal clusters. The specific process is as follows: first, all terminals are initially marked as unassigned and processed in order of node ID. For each unassigned terminal v i , create a new cluster C k And use it as the central node, and then traverse other unassigned terminals v j , the similarity is higher than the threshold S lim And A ij = 1 to join the cluster, ensuring that the addition operation does not destroy the connectivity of the cluster communication. All joined terminals are marked as allocated. If there are unallocated terminals, the above process is repeated. In order to ensure the stability and communication efficiency of the cluster, the cluster size constraint is set to C lim =[N min ,N max ], try to merge the clusters that are too small and split the clusters that are too large until all terminals are assigned, forming the final cluster list C = {C1, C2, ..., C L}, where L is the number of clusters.

[0080] After clustering, for each cluster C k Construct a ring communication structure and simply calculate the centroid coordinates of the cluster.

[0081]

[0082] Then, the coordinates of each terminal are converted into polar coordinates with the centroid as the origin, and the terminals in the cluster are sorted from small to large according to the size of the polar angle to obtain the sequence {v1, v2, ..., vn}, n is the number of terminals in the cluster. Connect the terminals in this order to form a closed loop. i The next neighbor is The previous neighbor is This ring structure simplifies the communication topology. Each terminal only needs to maintain connections with two neighbors. Even if a terminal leaves or joins, the ring structure only needs to adjust the connection relationship of the affected terminals.

[0083] Taking into account the dynamic joining and leaving of ground mobile terminals, the present invention stipulates that when a new terminal enters the network, its own characteristic information is first broadcast to the terminals within the communication range. The receiving terminal decides whether to accept the terminal based on the average similarity between the new terminal and the current cluster and the current size of the cluster. If the new terminal is accepted, the system will recalculate the center of mass and adjust the ring structure, insert the new terminal into the appropriate position of the polar angle sort, and update the connection relationship of the relevant terminals. When the terminal is about to leave, it will send a departure notification to the front and back neighbors in the ring in advance so that the neighboring terminals can reconnect to maintain the integrity of the ring. At the same time, the leaving terminal will complete the current federated learning round as much as possible to ensure that a safe exit does not affect the overall learning process. If leaving causes the cluster size to be less than the minimum threshold, the system will trigger a cluster merge operation. This dynamic processing mechanism ensures that the free entry and exit of terminals will not destroy the stability of the network topology and the federated learning process, while maintaining the high efficiency of communication and computing.

[0084] In one embodiment of the present invention, Figure 3 As shown, the algorithm for homomorphic encryption based on obfuscation module decomposition is as follows:

[0085] Ground mobile terminals conduct federated learning in clusters, and the terminals in the cluster {v1,v2,...,v n},v i ∈C k Both use the homomorphic encryption scheme based on obfuscation module decomposition to encrypt local models. During the encryption phase, all terminals in the cluster have the public key pk←(a,Q) and the private key fragment sk i ←(a,Q,Ι i ). Wherein, a is the magnification, Q is the modulus, I i For terminal v i Redundant set index. For terminal v i The specific encryption process includes the following steps:

[0086] Step 1: Amplify and perturb the plaintext. As the plaintext, the plaintext is amplified by a times and random noise η is added. For each plaintext data, the noise η is different, and a>>max(η)>>max(q).

[0087]

[0088] Step 2: Modular basis setting. Establish a set of modular basis Q←{q1,q2,...,q n}, where q1,q2,...,q n are positive integers and are mutually prime, satisfying the following conditions.

[0089] (q k ,q l )=1,k,l=1,2,...,n,k≠l (17)

[0090]

[0091] Step 3: Calculation of real modulus components. For each modulus basis q k , calculate the true modulus component. Get the true modulus component set

[0092]

[0093] Step 4: Generate an initial redundant set. Construct an initial redundant set to obfuscate the true modulus component.

[0094] Φ'={φ'1,φ'2,...,φ' m} (20)

[0095] Step 5: Calculate the redundant set

[0096]

[0097] Step 6: Set up the key i .

[0098] Ι i ←{I i,1 ,I i,2 ,...,I i,n},0≤I i,k ≤m,k=1,2,...,n (24)

[0099] Step 7: Generate ciphertext. Cover the true modulus component set with the redundant set according to the key to obtain a confusion matrix of size n×m

[0100]

[0101] Based on the above steps, the local training results of the terminal are encrypted and protected. Even if a malicious terminal node obtains the encryption result of a terminal, it cannot directly infer the true modulus component position of this terminal. Even if the true modulus component of a terminal is obtained, due to the existence of plaintext amplification disturbance, the malicious terminal node cannot directly obtain the plaintext information. Unlike traditional homomorphic encryption methods, the scheme of homomorphic encryption based on obfuscated modulus decomposition proposed in this invention is highly efficient, that is, it converts complex encryption operations into simple calculations in multiple modulus spaces through modulus decomposition, reducing computational overhead.

[0102] In one embodiment of the present invention, Figure 4 As shown in the figure, the distributed federated learning algorithm based on ring communication is as follows:

[0103] In the satellite Internet environment, the secure transmission and effective aggregation of information are the key to the successful implementation of federated learning. This section proposes a distributed federated learning algorithm based on a ring topology, which combines homomorphic encryption technology with a decentralized communication architecture to address the centralization risks and security issues in traditional federated learning.

[0104] Before the ring communication, each terminal v in the ring i ∈C k The confusion matrix is ​​obtained by encrypting the local model using a homomorphic encryption scheme based on confusion module decomposition. Indicates that in the module base q l During the ring communication process, each terminal is responsible for a module basis q k Specifically, the terminal v i The module base q l The confusion component under Send to the next neighbor terminal node in the direction of ring communication Then this node uses the module basis q l The true modulus component Confusion component with the received Add element by element, that is

[0105]

[0106] in, Represents modulo q l The linear element-by-element addition operation is then sent to the next terminal node in the ring communication direction. This terminal node then repeats the operation of adding the local true modulus component to the received component element by element. Finally, the terminal v i The "previous" neighbor node Complete the linear addition operation locally and send it to the terminal v i, get the aggregation result under the modulus q l

[0107]

[0108] The whole ring communication round is Γ, and multiple sub-rings are used in parallel communication mode, that is, the terminals in the cluster simultaneously perform component aggregation. i Use the private key fragment sk i Extract the aggregation result of the real component under the modulus q l And store this aggregation result locally and broadcast it to other terminal nodes in the ring, while receiving the aggregation result of the real modulus component under different moduli from other terminal nodes And store it locally.

[0109] Each terminal locally reconstructs the complete aggregation model using the Chinese Remainder Theorem (CRT). First, the terminal precomputes the CRT coefficient and stores it as a local lookup table.

[0110]

[0111] Q l = Q s / q l (32)

[0112]

[0113] Then calculate the aggregation model parameters.

[0114]

[0115] Determine the positive and negative of the aggregation result.

[0116]

[0117] Then update the global model as

[0118]

[0119] Each terminal uses this global model to continue the next round of federated learning training. In the ring communication process, the parameters are transmitted in the form of obfuscated components, so even if some information is obtained by malicious nodes, it is difficult to reconstruct the local training results of the terminals, and further cannot obtain the terminal privacy data through reverse inference attacks.

[0120] The distributed homomorphic encryption federated learning algorithm based on ring communication proposed in the application solves the privacy protection and decentralization challenges faced by federated learning in satellite internet environment. By innovatively combining ring topology and modulus decomposition homomorphic encryption scheme, the algorithm realizes efficient decentralized model aggregation while ensuring data privacy and security. Table 2 describes the implementation process.​​

[0121] The system first intelligently clusters the ground mobile terminals through the GAT network to form a ring-shaped communication structure. Then each terminal uses a homomorphic encryption scheme based on confusion module decomposition to encrypt and protect the local training model, and realizes distributed model aggregation through parallel communication in the ring topology. Finally, the decryption is completed through the Chinese remainder theorem, the global model is updated and the next round of training is continued, thereby realizing a decentralized federated learning system that protects data privacy.

[0122] Table 2

[0123]

[0124]

[0125]

[0126] The present application proposes a distributed satellite internet federated learning algorithm based on confusion module decomposition homomorphic encryption for the data privacy and security challenges in the low-orbit satellite internet environment. The algorithm combines the graph attention network (GAT) for terminal clustering and ring formation, uses an efficient confusion module decomposition homomorphic encryption mechanism to protect the model parameters in the federated learning process, and realizes a decentralized model aggregation strategy through a ring communication mechanism. The overall algorithm flow is as shown in Figure 5

[0127] Firstly, a terminal clustering and ring formation mechanism based on GAT is proposed, which can capture the terminal movement pattern and communication characteristics, model the spatio-temporal dependency relationship between ground mobile terminals through the graph attention network, and adjust the topology structure according to the terminal movement characteristics and network conditions. By organizing the terminals into a ring topology, a structural basis is provided for decentralized federated learning, enabling the model aggregation process to be carried out in the low-orbit satellite internet environment. Secondly, a lightweight confusion module decomposition homomorphic encryption framework is constructed, and a federated learning algorithm based on confusion module decomposition homomorphic encryption technology is proposed. The algorithm innovatively introduces redundant module components and random confusion technology to enhance security, enabling efficient encryption of federated learning model parameters. Compared with traditional homomorphic encryption, more compact ciphertext can be generated through module operation, effectively reducing communication overhead and storage requirements, and with the help of simple distributed key management, the single point failure risk is eliminated. Finally, an innovative ring communication distributed aggregation strategy is proposed, and a module component parallel processing mechanism based on ring topology is proposed. In this mechanism, each terminal only communicates with two adjacent nodes, breaking through the centralized limitation of traditional star topology and eliminating the single point failure risk of the central server. By allocating different module component calculation tasks to different terminals for synchronous execution, multiple functional sub-rings are formed to work in parallel, enabling the system to ensure data privacy while realizing efficient distributed computing. ​

[0128] Theoretical analysis shows that the calculation complexity of the algorithm is improved compared with the traditional homomorphic encryption method. Experimental analysis shows that compared with existing federated learning algorithms, the algorithm has good training stability and security while maintaining the training accuracy, can effectively resist model reverse inference attacks, the structural similarity index is significantly lower than the comparative algorithm, and can maintain good system throughput while controlling communication delay in the satellite internet environment. The satellite internet distributed homomorphic encryption federated learning method for cross-trust domain cooperation belongs to the field of research on Internet of Things communication system architecture design.

[0129] The beneficial effects of the present application are:

[0130] The present application can ensure data privacy and security in cross-trust domain cooperation in the satellite internet environment in a certain low-orbit satellite coverage area. By innovatively combining the graph attention network, lightweight homomorphic encryption and ring communication mechanism, a star group level efficient and secure decentralized model aggregation is realized. First, a terminal clustering and ring forming mechanism based on GAT is designed, which can construct a ring topology according to the terminal movement characteristics and communication conditions, providing a structural basis for star group decentralized federated learning. Secondly, a confusion module decomposition homomorphic encryption framework is introduced, which generates compact ciphertext through module operation, and introduces random confusion technology to improve security, which reduces the calculation complexity and communication overhead compared with the traditional homomorphic encryption method. Finally, a distributed aggregation strategy based on ring topology is designed, which eliminates the single point failure risk of the central server and improves the system robustness and attack resistance. Through theoretical analysis, the calculation complexity of the algorithm designed in the present research is improved compared with the traditional homomorphic encryption method, and can effectively resist the attack of passive enemy and external eavesdropper between stars. At the same time, the present research tests on two classic data sets MNIST and CIFAR10, the results show that compared with existing federated learning algorithms, the algorithm can effectively resist multi-model reverse inference attacks while maintaining model accuracy, and the structural similarity index is lower than the comparative algorithm. While ensuring data privacy, good system robustness is maintained. Overall, the present application provides a new technical path for low-orbit satellite internet cross-trust domain cooperation, and has a certain value for promoting the safe and efficient development of space-ground integrated intelligent transportation systems.

[0131] In the description of the specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in one or more embodiments or examples. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the specification and the features of the different embodiments or examples, without contradiction.

[0132] In addition, the terms "first", "second" are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "a plurality of" is at least two, for example, two, three, etc., unless otherwise specifically limited.

Claims

1. A method for secure data sharing in a satellite Internet model based on federated learning, characterized in that: Including: Construct a low-earth orbit satellite internet distributed federated learning system, including high-speed mobile terminals and on-board collaborative processing unit nodes. Each terminal is used to collect environmental data and perform machine learning. The system adopts a ring topology communication mechanism, and the terminals transmit data and aggregate model parameters through the ring structure. Standardize the terminal state feature vectors defined by the system structure, define the communication connection between terminals based on communication distance and delay to construct an adjacency matrix, and use the GAT network to assign importance weights to different neighbor nodes through the attention mechanism to generate embedding vectors. Calculate the cosine similarity between terminals based on the embedding vectors to form a similarity matrix, and use the greedy algorithm to divide the terminals into multiple cluster information. Amplify the plaintext data based on the cluster information and add random noise, set a group of pairwise relatively prime modulo bases, calculate the true modulo components, construct an initial redundant set to confuse the true modulo components, and cover the true modulo components into the redundant set through a key to generate a confusion matrix. Aggregate the components of the confusion matrix generated by each terminal through ring communication and element-wise addition operations, extract the true component aggregation result from the aggregated confusion matrix, and use the Chinese Remainder Theorem to reconstruct the complete aggregation model.

2. The method according to claim 1, characterized in that Based on the distributed federated learning system of low-orbit satellite Internet, the intelligent networked terminals in the satellite coverage area are divided into multiple terminal clusters C = {C1, C2, ..., C m }, n terminals in the cluster form a ring communication structure; in the ring topology, each terminal v i ∈C m ,i∈[1,n] has only two neighbor nodes, sends data to the next neighbor in the clockwise direction, and receives data from the previous neighbor in the counterclockwise direction; The specific process includes: In each iteration t, the terminal v i Use local data D i Perform training to obtain a local model The terminal uses a homomorphic encryption algorithm based on confusion module decomposition to encrypt the local model and obtain the confusion matrix Among them, Encrypt represents the encryption operation, is the aliased modulus component under the modulus basis n, and the terminal converts the component Send to the next terminal node v j ∈C m ,j∈[1,n] and perform component aggregation, that is, in Represents the operation under the module base n; the calculation result is passed to the next node along the direction of the ring communication, and the component aggregation operation under the module base n is repeated, and the terminal v i The component aggregation result received from the previous neighbor node is expressed using the following formula: Terminal v i Use private key fragment sk locally i Extracting the true aggregate component results And broadcast it to other terminal nodes in the cluster. After each terminal receives the modular component aggregation results from other terminals, it uses the Chinese remainder theorem to decrypt it locally. The formula is expressed as: Among them, Decrypt represents the decryption operation, W t+1 To aggregate the obtained global model, the terminal uses the global model for the next round of training.

3. The method according to claim 1, characterized in that The terminal nodes in the network are distributed in different geographical locations and move dynamically. The state of each terminal node is represented by a feature vector, namely h i =[x i ,y i ,υ i ,θ i ]; where (x i ,y i ) is the location of the terminal, υ i is the current speed of the terminal, θ i is the heading angle, terminal v i With terminal v j The distance between them is expressed as Euclidean distance Each terminal is equipped with a wireless communication module and exchanges information with surrounding terminals based on V2V technology. The communication delay between terminals is expressed as: Where B represents the communication bandwidth, P i For terminal v i The transmission power, N0 is the noise power, g i,j is the channel gain, |D ij | is the size of data transmitted between terminals; For the large prime modulus q selected in this paper l ,l∈1,2,...,n, given generators g and g x modq i , the adversary cannot efficiently calculate x in polynomial time, which can be expressed as: Where C is any PPT adversary algorithm, negl(λ) is a negligible function with respect to the security parameter λ, Represents modulo q l The integer ring below, x is selected uniformly randomly from the integer ring; The homomorphic encryption scheme is based on modular decomposition technology, where the terminal model parameters is decomposed into modules with different prime numbers q1,q2,...,q n The remaining class, in the distributed design, each terminal v i Only partial modulus information is held, expressed using the formula: Where k < n, it means that the adversary cannot infer the remaining value under the unknown modulus when knowing some modulo remainders and the corresponding moduli. The homomorphic encryption scheme supports additive homomorphy, that is, the operations performed in the ciphertext domain can be reflected in the plaintext domain: That is, the adversary cannot distinguish the encryption results of any two different plaintexts, that is in, and Any plaintext.

4. The method according to claim 1, wherein For each terminal node, the state feature vector h i =[x i ,y i ,υ i ,θ i ],i∈[1,N] are standardized and expressed as follows: Among them, μ x ,σ x ,μ y ,σ y ,μ υ ,σ υ ,μ θ ,σ θ are the mean and standard deviation of each feature dimension respectively; The communication connection between terminals is defined based on the communication distance and delay between terminals, and the adjacency matrix A is used. i ∈R N ×N express: Among them A ij =1 indicates terminal v i With terminal v j There is an effective communication link between them, otherwise there is no effective communication link, d lim is the maximum distance limit between terminals, τ lim The maximum delay limit between terminals; when the distance or delay between terminals exceeds the maximum limit, the communication link is disconnected; Based on the adjacency matrix, the GAT network is applied to convert the original features into more semantically informative embedding representations, and the normalized terminal feature matrix H 0 ∈R N×4 The linear change output is Z through the weight matrix 0 ∈R N×d , d is the embedding space dimension, then the attention coefficient is calculated for the connected terminals, and for the K-th attention head it is calculated as: Among them, a k ∈R 2d is the learnable vector of the K-th attention head, || represents the feature connection operation, for the unconnected terminal, i.e., A ij =0, set Then, the attention weights are normalized; Among them, N i For terminal v i The adjacent terminal set is weightedly aggregated based on this attention weight, that is, Finally, each terminal v i Get an embedding vector z i , which encodes the spatial position, mobility characteristics and interaction relationship of the terminal with the surrounding terminals; based on the obtained embedding vector, the cosine similarity between the terminals is calculated to form a similarity matrix S ij : Based on the similarity matrix, use the greedy algorithm to form terminal clusters.

5. The method according to claim 1, wherein Using the greedy algorithm to form terminal clusters includes: All terminals are initially marked as unassigned and processed in order of node ID. For each unassigned terminal v i , create a new cluster C k And as the central node, then traverse other unassigned terminals v j , the similarity is higher than the threshold S lim And A ij = 1 joins the cluster, and all joined terminals are marked as allocated. If there are unallocated terminals, repeat the above process; set the cluster size constraint to C lim =[N min ,N max ], try to merge the clusters that are too small and split the clusters that are too large until all terminals are assigned, forming the final cluster list C = {C1, C2, ..., C L }, where L is the number of clusters; After clustering, for each cluster C k Construct a ring communication structure and simply calculate the centroid coordinates of the cluster: Then, the coordinates of each terminal are converted into polar coordinates with the centroid as the origin, and the terminals in the cluster are sorted from small to large according to the size of the polar angle to obtain the sequence {v1, v2, ..., v n }, n is the number of terminals in the cluster; connect the terminals in this order to form a closed loop, then the terminal v i The next neighbor is The previous neighbor is 6. The method according to claim 1, characterized in that Ground mobile terminals conduct federated learning in clusters, and the terminals in the cluster {v1,v2,...,v n },v i ∈C k Both use the homomorphic encryption scheme based on obfuscation module decomposition to encrypt local models. Encryption; in the encryption phase, all terminals in the cluster have the public key pk←(a,Q) and the private key fragment sk i ←(a,Q,Ι i ); where a is the magnification, Q is the modulus, and i For terminal v i Redundant set index; for terminal v i The specific encryption process includes the following steps: Plaintext amplification and perturbation: the updated part As the plaintext, amplify the plaintext by a times and add random noise η; for each plaintext data, the noise η is different, and a>>max(η)>>max(q): Module basis setting: establish a set of module basis Q←{q1,q2,...,q n }, where q1,q2,...,q n are positive integers and are mutually prime, satisfying the following conditions: (q k ,q l )=1,k,l=1,2,...,n,k≠l(17) Calculation of real modulus components: For each modulus basis q k , calculate the true modulus component and get the true modulus component set Generation of redundant initial set: Construct an initial redundant set to confuse the true modulo components: Φ'={φ'1,φ'2,...,φ' m } (20) Computing redundant sets Setting the key i : Ι i ←{I i,1 ,I i,2 ,...,I i,n },0≤I i,k ≤m,k=1,2,...,n(24) Ciphertext generation: Cover the true modulus component set with the redundant set according to the key to obtain a confusion matrix of size n×m 7. The method according to claim 1, characterized in that Before the ring communication, each terminal v in the ring i ∈C k The confusion matrix is ​​obtained by encrypting the local model using a homomorphic encryption scheme based on confusion module decomposition. Indicates that in the module base q l During the ring communication process, each terminal is responsible for a module base q k Model aggregation under: Specifically, terminal v i The module base q l The confusion component under Send to the next neighbor terminal node in the direction of ring communication Then this node uses the module basis q l The true modulus component Confusion component with the received Add element by element, that is: in, Represents modulo q l The linear element-by-element addition operation is performed; the addition result is then sent to the next terminal node in the ring communication direction. The terminal node then repeats the operation of adding the local true modulus component to the received component element by element. Finally, the terminal v i The previous neighbor node Complete the linear addition operation locally and send it to the terminal v i , get the module basis q l The aggregation result under The overall ring communication round is Γ, and multiple sub-rings are used for parallel communication, that is, the terminals in the cluster perform component aggregation at the same time; the terminal v i Use private key fragment sk i Extract module base q l Aggregation results of the lower true component The aggregation result is stored locally and broadcast to other terminal nodes in the ring, and the aggregation results of the real modulus components under different modulus bases from other terminal nodes are received at the same time. and stored locally; Each terminal reconstructs the complete aggregation model locally using the Chinese Remainder Theorem CRT: First, the terminal pre-computes the CRT coefficients and stores them as a local lookup table: Q l =Q s / q l (32) Then calculate the aggregation model parameters: Judge the positive and negative of the aggregation result: Then the updated global model is:

Images