Unmanned aerial vehicle group security collaboration system based on block chain and federated learning

Abstract

The invention relates to the technical field of unmanned aerial vehicle networks, in particular to an unmanned aerial vehicle group security collaboration system based on a block chain and federated learning, which is provided with an identity management architecture module, an identity registration and authentication module, a communication optimization module and a dynamic selection module, the identity management architecture module designs a layered unmanned aerial vehicle identity management architecture, the identity registration and authentication module performs identity registration and identity authentication based on a distributed cross-domain unmanned aerial vehicle identity management mechanism, and the communication optimization module responds to an authentication success state. And determining that the teacher model performs knowledge distillation to update the local model, and dynamically selecting neighbors based on the actual communication environment for the transmission process of the embedded vector. According to the method, the safety of the identity management level and the communication optimization level is guaranteed, the whole training process is guaranteed, the calculation efficiency is improved, the pressure on the center node is reduced, and the overall communication efficiency and the system response speed of the unmanned aerial vehicle group are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle networks, and in particular to a secure collaborative system for unmanned aerial vehicle swarms based on blockchain and federated learning. Background Art

[0002] In the application of unmanned aerial vehicle networks, identity authentication and federated learning technologies are becoming key components for ensuring system security and improving collaboration efficiency. Current identity authentication technologies mostly rely on centralized mechanisms and traditional key management systems to verify the identities of unmanned aerial vehicles and their operators, ensuring that only authorized users can access the system and data. However, unmanned aerial vehicle networks relying on centralized mechanisms need to frequently transmit data between the central node and each node, which often leads to problems such as high data transmission costs and network bandwidth limitations. Federated learning allows each unmanned aerial vehicle to perform model training locally and then send the updated results to the central node for aggregation, thereby achieving distributed learning. These technologies provide basic support for the collaborative operation of unmanned aerial vehicle swarms.

[0003] Chinese Patent Publication No.: CN119316030A, discloses a method, system, device, medium and product for secure collaboration of unmanned aerial vehicle clusters, including: the unmanned aerial vehicle cluster includes several pre-divided clusters, each cluster includes a pre-selected first cluster head node and several in-cluster sub-nodes, and the trust degree of the first cluster head node meets a preset condition and is greater than that of the in-cluster sub-nodes. The method realizes secure communication and data sharing between the nodes of the unmanned aerial vehicle cluster through a consensus mechanism, improving the security and reliability of the unmanned aerial vehicle cluster; by designing a hierarchical network of the unmanned aerial vehicle cluster, the unmanned aerial vehicles are divided into several clusters, and multiple clusters work collaboratively. And by virtue of the hierarchical network characteristics of the unmanned aerial vehicle cluster, the consensus mechanism of the unmanned aerial vehicle cluster is divided into an in-cluster consensus mechanism and a lightweight inter-cluster consensus mechanism, reducing the communication pressure of the cluster head and alleviating the communication blockage between the cluster head nodes, reducing the communication complexity, and improving the communication efficiency of the unmanned aerial vehicle cluster.

[0004] Chinese Patent Publication No.: CN118301604A discloses a secure data sharing method based on blockchain and drone collaboration. The method includes: the drone and the ground control station perform two-way authentication and key negotiation to obtain a symmetric key. The drone constructs a secure communication link with the ground control station, and the drone symmetrically encrypts the collected data and sends it to the ground control station through the secure communication link; after the ground control station decrypts and processes the encrypted data sent by the drone, it encrypts part of the data and sends it to the cloud server, and sets up a data access control structure. The cloud server encrypts the data sent by the ground control station and stores the encrypted data. The inventive method performs local processing of sensitive data at the edge based on the cloud-edge-end architecture, reducing the risk of data exposure during network transmission. The invention outsources the calculation of encryption and decryption itself to the edge nodes, and can consume less resources based on the attribute-based encryption algorithm.

[0005] However, the following problems still exist in the prior art.

[0006] Current identity authentication technologies mostly rely on centralized mechanisms and traditional key management systems, which are prone to single points of failure. Sensitive data is easily accessed by unauthorized users or devices, increasing the risk of data leakage; at the same time, current federated learning technologies mostly rely on the central server to update and aggregate models, which easily leads to single points of failure and performance bottlenecks, resulting in reduced model training efficiency and differences in effects during the federated learning process. Summary of the Invention

[0007] Therefore, the present invention provides a secure collaboration system for a drone swarm based on blockchain and federated learning to solve the problems that current identity authentication technologies mostly rely on centralized mechanisms and traditional key management systems, are prone to single points of failure, sensitive data is easily accessed by unauthorized users or devices, increasing the risk of data leakage; at the same time, current federated learning technologies mostly rely on the central server to update and aggregate models, which easily leads to single points of failure and performance bottlenecks, resulting in reduced model training efficiency and differences in effects during the federated learning process.

[0008] To achieve the above object, the present invention provides a secure collaboration system for a drone swarm based on blockchain and federated learning, which includes:

[0009] An identity management architecture module, which designs a hierarchical drone identity management architecture based on a hierarchical blockchain to connect each sub-trust domain;

[0010] An identity registration and authentication module, which is connected to the identity management architecture module, designs a distributed cross-domain drone identity management mechanism to perform drone identity registration and identity authentication to ensure the identity interoperability and credibility among sub-domain drone swarms;

[0011] A communication optimization module, which is connected to the identity registration and authentication module, and in response to the authentication success status, determines that the teacher model performs knowledge distillation to update the local model, including,

[0012] Perform local training of the drone swarm, determine the embedding vector based on the results of the local training of the local training, perform filtering and aggregation operations on the embedding vector, and update the model parameters by combining the minimization of the distillation loss function and the stochastic gradient descent algorithm to perform global model aggregation and obtain a convergent model;

[0013] A dynamic selection module, which is connected to the communication optimization module, and for the transmission process of the embedding vector, dynamically selects neighbors based on the actual communication environment, including,

[0014] Calculate the user - end distance to construct a drone system distance matrix, determine the communication object based on the distance matrix, construct a temporary adjacency matrix, and transmit the soft label generated by training to adjacent drones.

[0015] Furthermore, the hierarchical drone identity management architecture includes an identity management system layer, a sub - blockchain layer, a cross - domain communication layer, and a main - blockchain layer.

[0016] Furthermore, the identity management architecture module connects each sub - trust domain, including,

[0017] To set identity attribute credentials based on the identity management system layer;

[0018] To store the identity attribute credentials in the sub - blockchains of each sub - domain in the sub - blockchain layer;

[0019] To perform communication protocol conversion between domains based on the cross - domain communication layer;

[0020] To complete identity authentication request communication between sub - domains based on the main - blockchain layer.

[0021] Furthermore, the identity registration and authentication module performs identity registration, including,

[0022] To generate a private key and the corresponding public key based on PUF;

[0023] To generate a DID document based on the identity management system that receives the public key;

[0024] To update the corresponding attribute accumulator in the blockchain based on the DID document and generate an attribute credential;

[0025] To store the attribute credential in the memory.

[0026] Furthermore, the identity registration and authentication module performs identity authentication, including,

[0027] for receiving an identity authentication request;

[0028] for parsing the request to determine the attribute credentials to be verified;

[0029] for determining that the attribute credentials are in the blockchain and the identity authentication request corresponding to the drone having this attribute can pass the authentication;

[0030] wherein the parameters of the identity authentication request include the service, the DID of the resource provider, the DID of the drone node, the access control policy, and the signature information.

[0031] Further, the communication optimization module determines a teacher model, wherein,

[0032] for determining that the drone model not participating in the current round of local training is the teacher model.

[0033] Further, the communication optimization module determines an embedding vector based on the results of the local training of the local training, including,

[0034] for initializing the local private model and training with the shared dataset to obtain a pre-trained model;

[0035] for obtaining the local training results after training is completed;

[0036] for generating an embedding vector on the shared dataset based on the local training results.

[0037] Further, the communication optimization module performs filtering and aggregation operations on the embedding vector, including,

[0038] for filtering out the invalid embedding vectors received by the drone;

[0039] for calculating the similarity scores between the valid embedding vectors;

[0040] for selecting a number of embedding vectors with the smallest anomaly scores to form a new set;

[0041] for aggregating the number of embedding vectors through an aggregation function to obtain an aggregated embedding vector;

[0042] wherein the invalid embedding vectors are the embedding vectors generated due to drone failures or malicious attacks.

[0043] Further, the communication optimization module updates the model parameters by combining the minimized distillation loss function and the stochastic gradient descent algorithm, including,

[0044] for based on the aggregated embedding vector;

[0045] for determining the distillation loss function;

[0046] is used to measure the difference in probability distributions based on the distillation loss function;

[0047] is used to determine model parameters by combining the minimization of the distillation loss function and the stochastic gradient descent algorithm;

[0048] is used to determine an updated model based on the parameters.

[0049] Furthermore, the dynamic selection module determines communication objects based on the distance matrix, where

[0050] is used to determine communication loss;

[0051] is used to determine that the communication object corresponding to the minimum communication loss is the communication object.

[0052] Compared with the prior art, the present invention provides an identity management architecture module, an identity registration and authentication module, a communication optimization module, and a dynamic selection module. The identity management architecture module designs a hierarchical UAV identity management architecture. The identity registration and authentication module performs identity registration and identity authentication based on a distributed cross-domain UAV identity management mechanism. The communication optimization module determines a teacher model for knowledge distillation to update the local model in response to the authentication success status. The dynamic selection module dynamically selects neighbors based on the actual communication environment for the transmission process of the embedding vector. The present invention provides guarantee for the entire training process by ensuring the security of the identity management level and the communication optimization level, improves the computing efficiency, reduces the pressure on the central node, and improves the overall communication efficiency and system response speed of the UAV swarm.

[0053] In particular, the present invention determines a method for UAV node identity management and authentication based on decentralized identity, designs a hierarchical UAV identity management architecture, uses a hierarchical blockchain to enable intercommunication between sub-trust domains, and utilizes the traceability and immutability of the blockchain to guarantee the credibility of UAV identities. In the existing UAV network, the training scheme lacks an effective identity authentication mechanism, resulting in unauthorized devices being able to access the network, thereby leaking privacy data or forging data to affect model training. Based on this, the present invention designs an identity management system layer, a sub-blockchain layer, a cross-domain communication layer, and a main blockchain layer to solve the cross-domain access efficiency problem between different identity management systems in a dynamic network, provides guarantee for the identity management level, improves the computing efficiency, and reduces the pressure on the central node.

[0054] In particular, the present invention combines knowledge distillation technology to perform local training on the UAV swarm, determine the embedding vectors, and update the model. In the prior art, most solutions highly rely on the central server for model aggregation and coordination. Once the central server fails, the entire system will face the risk of interruption or significant performance degradation. In addition, the computing and storage capabilities of the central server are limited and difficult to meet the requirements of large-scale UAV networks. When the number of nodes increases, the load on the central server will continuously increase, leading to performance bottlenecks and restricting the scalability of the system. Based on this, the present invention trains the UAVs to obtain a pre-trained model, then determines the embedding vectors, and updates the local model, providing guarantee for the communication optimization level and improving the overall communication efficiency and system response speed of the UAV swarm.

[0055] In particular, by proposing a new setting with robust properties, neighbors are dynamically selected during information transmission. In actual situations, in the scenario of decentralized federation, the number of nodes in the UAV network is huge, and each node needs to communicate with the central server frequently to upload local model updates and receive the global model. This frequent communication will significantly increase the communication overhead of the network. Especially in large-scale networks, it is easy to cause network congestion and delay, affecting the training efficiency. Based on this, the present invention determines the communication objects by calculating the distances between UAVs, providing guarantee for the communication optimization level and improving the overall communication efficiency and system response speed of the UAV swarm. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a schematic structural diagram of the UAV swarm security cooperation system based on blockchain and federated learning according to an embodiment of the invention;

[0057] Figure 2 It is a schematic structural diagram of the hierarchical UAV identity management architecture according to an embodiment of the invention;

[0058] Figure 3 It is a schematic structural diagram of the identity registration and authentication module according to an embodiment of the invention;

[0059] Figure 4 It is a schematic structural diagram of the communication optimization module according to an embodiment of the invention;

[0060] Figure 5 It is a schematic structural diagram of the dynamic selection module according to an embodiment of the invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0061] In order to make the objectives and advantages of the present invention clearer and more understandable, the present invention will be further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0062] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are only used to explain the technical principles of the present invention and do not limit the protection scope of the present invention.

[0063] It should be noted that in the description of the present invention, unless otherwise clearly specified and limited, the term "connection" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected, or indirectly connected through an intermediate medium, and it can be the communication inside two components. For those skilled in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0064] Please refer to Figure 1 , Figure 1 which is a schematic structural diagram of the UAV swarm security collaboration system based on blockchain and federated learning for an embodiment of the invention. A UAV swarm security collaboration system based on blockchain and federated learning of the present invention includes:

[0065] An identity management architecture module, which designs a hierarchical UAV identity management architecture based on a hierarchical blockchain to connect each sub-trust domain;

[0066] An identity registration and authentication module, which is connected to the identity management architecture module and designs a distributed cross-domain UAV identity management mechanism to perform UAV identity registration and identity authentication to ensure the identity intercommunication and credibility among sub-domain UAV swarms;

[0067] A communication optimization module, which is connected to the identity registration and authentication module and, in response to the authentication success status, determines a teacher model to perform knowledge distillation to update the local model, including,

[0068] Performing local training on the UAV swarm, determining an embedding vector based on the results of the local training of the local training, performing filtering and aggregation operations on the embedding vector, and combining the minimization of the distillation loss function and the stochastic gradient descent algorithm to update the model parameters to perform global model aggregation to obtain a convergent model;

[0069] A dynamic selection module, which is connected to the communication optimization module and, for the transmission process of the embedding vector, dynamically selects neighbors based on the actual communication environment, including,

[0070] Calculating the distance of the user end to construct a UAV system distance matrix, determining communication objects based on the distance matrix, constructing a temporary adjacency matrix, and transmitting the soft label generated by training to adjacent UAVs.

[0071] Specifically, the hierarchical blockchain is a design pattern that divides the blockchain network into different levels, aiming to improve the scalability, performance, and security of the blockchain system. By allocating different functions and tasks to different levels, the operation of the blockchain system can be better managed and optimized.

[0072] Specifically, the teacher model is a well-trained and well-performing model used to guide the learning process of another model to be trained or a simpler model.

[0073] Specifically, in this implementation, the specific steps of global model aggregation are as follows:

[0074] Train the drones for T rounds according to the knowledge distillation mechanism. It can be understood that no specific limit is imposed on the specific value of T, and those skilled in the art can determine it according to the actual situation, which will not be elaborated here.

[0075] Obtain the updated models of all drones.

[0076] Upload the model to the blockchain to complete the global model aggregation.

[0077] Specifically, no specific limit is imposed on the specific method of constructing the distance matrix of the drone system. For example, it can be the construction of the distance matrix based on multi-dimensional scaling analysis (MDS), or the construction of the distance matrix based on cooperative positioning, as long as the matrix construction is successful. Those skilled in the art can choose according to the actual situation, which will not be elaborated here.

[0078] Please refer to Figure 2 , Figure 2 which is a schematic structural diagram of the hierarchical drone identity management architecture of the invention embodiment. Specifically, the hierarchical drone identity management architecture includes an identity management system layer, a sub-blockchain layer, a cross-domain communication layer, and a main blockchain layer.

[0079] Specifically, the identity management system layer is composed of distributed identity management systems in each sub-domain. Based on the PUF-based drone public and private key generation scheme, calculate the accumulator corresponding to its identity attributes, and generate personal attribute credentials associated with the DID document to ensure the security and uniqueness of the drone key and the authenticity and immutability of the identity attributes.

[0080] Specifically, no specific limit is imposed on the method for the drone to calculate the accumulator. For example, the lightweight identity attribute authentication mechanism based on cryptographic accumulators can be used for calculation. This is prior art and will not be elaborated here.

[0081] Specifically, the DID document and the accumulator in the sub-blockchain layer are stored on the sub-chain to which the sub-domain belongs to ensure the integrity and immutability of the DID document and the accumulator. It can be understood that, in order to protect the privacy and security of the data on the sub-chain, the sub-blockchain can be designed as a consortium chain, and the drone node is only allowed to join the sub-blockchain when authorized by the consortium chain.

[0082] Specifically, the cross-domain communication layer is responsible for coordinating the communication between different communication domains and trust domains, performing conversions between domains to determine the interoperability between different trust domains, which will not be elaborated here.

[0083] Specifically, the main blockchain layer consists of a main chain. Through the main chain, the identity authentication request communication between sub-domains can be realized and recorded, ensuring the traceability of the authentication request.

[0084] Specifically, the identity management architecture module connects each sub-trust domain, including

[0085] for setting identity attribute credentials based on the identity management system layer;

[0086] for storing the identity attribute credentials in the sub-blockchains of each sub-domain in the sub-blockchain layer;

[0087] for performing communication protocol conversion between domains based on the cross-domain communication layer;

[0088] for enabling the completion of identity authentication request communication between sub-domains based on the main blockchain layer.

[0089] Specifically, the present invention determines a method for managing and authenticating the identity of drone nodes based on decentralized identity, designs a hierarchical drone identity management architecture, uses a hierarchical blockchain to enable communication between sub-trust domains, and utilizes the traceability and immutability of the blockchain to ensure the credibility of the drone identity. In the existing drone network, the training scheme lacks an effective identity authentication mechanism, resulting in unauthorized devices being able to access the network, thereby leaking privacy data or forging data and affecting model training. Based on this, the present invention designs an identity management system layer, a sub-blockchain layer, a cross-domain communication layer, and a main blockchain layer to solve the cross-domain access efficiency problem between different identity management systems in a dynamic network, provide a guarantee for the identity management level, improve the computing efficiency, and reduce the pressure on the central node.

[0090] Please refer to Figure 3 , Figure 3 which is a schematic structural diagram of the identity registration and authentication module of the invention embodiment. Specifically, the identity registration and authentication module performs identity registration, including

[0091] for generating a private key and the corresponding public key based on PUF;

[0092] An identity management system is used to generate a DID document based on the received public key;

[0093] Based on the DID document, it is used to update the corresponding attribute accumulator in the blockchain and generate an attribute credential;

[0094] The attribute credential is stored in a memory.

[0095] Specifically, the DID document includes attributes such as a DID number, the manufacturer to which the drone belongs, and a public key.

[0096] It can be understood that to ensure the integrity and immutability of identity information, the public key is set as the public key in the DID document, and the DID document is stored on the blockchain of the subdomain, which will not be elaborated here.

[0097] Specifically, the attribute credential needs to be encrypted with the public key of the drone and sent back to the drone. It can be understood that to ensure the traceability of operations, the hash value of the stored content after this attribute update is uploaded to the main blockchain.

[0098] Specifically, the identity registration and authentication module performs identity authentication, including,

[0099] Receiving an identity authentication request;

[0100] Parsing the request to determine the attribute credential to be verified;

[0101] Determining that the attribute credential is in the blockchain and the identity authentication request corresponding to the drone having this attribute can pass the authentication;

[0102] Among them, the parameters of the identity authentication request include a service, the DID of the resource provider, the DID of the drone node, an access control policy, and signature information.

[0103] Specifically, there is no limitation on the specific method for determining that the drone has this attribute. For example, the verification algorithm Verify(ACCattcode, ACattcode) can be used for verification. This is prior art and will not be elaborated here.

[0104] Specifically, if the identity authentication request passes the authentication, the corresponding service is provided for the drone.

[0105] Please refer to Figure 4 , Figure 4 It is a schematic structural diagram of the communication optimization module in an embodiment of the invention. Specifically, the communication optimization module determines a teacher model, where,

[0106] It is used to determine that the drone model that has not participated in the current round of local training is the teacher model;

[0107] Specifically, the communication optimization module determines the embedding vector based on the results of the local training, including:

[0108] Initializing the local private model and training it with the shared dataset to obtain a pre-trained model;

[0109] Obtaining the local training results after training is completed;

[0110] Generating an embedding vector on the shared dataset based on the local training results.

[0111] Specifically, the local training uses cross-entropy as the loss function, which is represented by formula (1):

[0112]

[0113] In formula (1), represents the loss function of the i-th human-machine, and Y i represents the private dataset D i 's true label, is the output of the local private model M i , and L CE represents the cross-entropy function.

[0114] Specifically, during the training process, by continuously adjusting the model parameters, the loss function is minimized to improve the model's fitting ability to the private data.

[0115] Specifically, the embedding vector is represented by formula (2):

[0116]

[0117] In formula (2), C represents the shared dataset, represents the embedding generation function based on the model Mi.

[0118] Specifically, the communication optimization module performs filtering and aggregation operations on the embedding vector, including:

[0119] Filtering the invalid embedding vectors received by the drone;

[0120] Calculating the similarity scores between the valid embedding vectors;

[0121] Selecting several embedding vectors with the smallest anomaly scores to form a new set;

[0122] Aggregating the several embedding vectors through an aggregation function to obtain an aggregated embedding vector;

[0123] Among them, the invalid embedding vectors are the embedding vectors generated due to drone failures or malicious attacks.

[0124] Specifically, the specific steps of the filtering and aggregation operations are as follows:

[0125] Step S1, calculate the Euclidean distance through formula (3):

[0126]

[0127] In formula (3), d (i,j) represents the Euclidean distance between drone i and drone j; represents the embedding vector of drone i; represents the embedding vector of drone j;

[0128] Step S2, assign an anomaly score to each embedding vector, and the anomaly score is represented by formula (4):

[0129]

[0130] In formula (4), γ i represents the anomaly score of;

[0131] Step S3, set a threshold, which is represented by formula (5):

[0132]

[0133] In formula (5), ρ k represents the threshold;

[0134] Step S4, select several embedding vectors with the smallest anomaly scores to form a new set Λ;

[0135] Step S5, aggregate the embedding vectors in through an aggregation function to obtain an aggregated embedding vector, which is represented by formula (6):

[0136]

[0137] In formula (6), represents the embedding vector in the set Λ.

[0138] Specifically, the present invention combines knowledge distillation technology to perform local training on the UAV swarm, determine the embedding vectors, and update the model. In the prior art, most solutions highly rely on the central server for model aggregation and coordination. Once the central server fails, the entire system will face the risk of interruption or serious performance degradation. In addition, the computing and storage capabilities of the central server are limited, making it difficult to meet the requirements of large-scale UAV networks. Moreover, when the number of nodes increases, the load on the central server will continuously increase, resulting in a performance bottleneck and limiting the scalability of the system. Based on this, the present invention trains the UAVs to obtain a pre-trained model, then determines the embedding vectors, and updates the local model, providing a guarantee for the communication optimization level and improving the overall communication efficiency and system response speed of the UAV swarm.

[0139] Specifically, the communication optimization module updates the model parameters by combining the minimized distillation loss function and the stochastic gradient descent algorithm, including,

[0140] for determining the aggregated embedding vector;

[0141] for determining the distillation loss function;

[0142] for measuring the probability distribution difference based on the distillation loss function;

[0143] for determining the model parameters by combining the minimized distillation loss function and the stochastic gradient descent algorithm;

[0144] for determining the updated model based on the parameters.

[0145] Specifically, the distillation loss function is represented by formula (7)

[0146]

[0147] In formula (7), D KL represents the Kullback-Leibler divergence, V k represents the aggregated embedding vector, represents the local embedding vector, and the local embedding vector is represented by formula (8),

[0148]

[0149] Specifically, the Kullback-Leibler divergence is used to measure the difference between two probability distributions.

[0150] Specifically, the specific steps of the stochastic gradient descent algorithm are as follows:

[0151] Initialize the parameters, including randomly initializing the model parameters, the learning rate, and setting the maximum number of iterations or the convergence condition;

[0152] In each iteration, a sample is randomly selected;

[0153] According to the selected sample, calculate the gradient of the loss function;

[0154] According to the calculated gradient, update the model parameters to reduce the value of the loss function;

[0155] Repeat the above steps until the stopping condition is met.

[0156] Specifically, the stopping condition can be reaching the maximum number of iterations, the change in the loss function being less than a certain threshold, or the performance on the validation set no longer improving, which will not be elaborated here.

[0157] Specifically, after T rounds of training according to the knowledge distillation mechanism, the drones upload their updated models to the blockchain respectively for global model aggregation to obtain a convergent model. This convergent model is the result of the collaborative training of several drones, integrating the local models of several drones, and has strong adaptability and robustness, improving the collaborative efficiency and learning results of the drones.

[0158] Please refer to Figure 5 , Figure 5 which is a schematic structural diagram of the dynamic selection module of the invention embodiment. Specifically, the dynamic selection module determines the communication object based on the distance matrix, where

[0159] is used to determine the communication loss;

[0160] The communication object corresponding to the minimum communication loss is determined as the communication object.

[0161] Specifically, by proposing a new setting with robust properties, neighbors are dynamically selected during information transmission. In actual situations, in the scenario of decentralized federation, the number of nodes in the drone network is huge, and each node needs to communicate with the central server frequently to upload local model updates and receive the global model. This frequent communication will significantly increase the communication overhead of the network. Especially in large-scale networks, it is easy to cause network congestion and delay, affecting the training efficiency. Based on this, the present invention determines the communication object by calculating the distance between drones, providing a guarantee for the communication optimization level and improving the overall communication efficiency and system response speed of the drone swarm.

[0162] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the protection scope of the present invention.

[0163] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention; for those skilled in the art, the present invention may have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A drone swarm safety collaboration system based on blockchain and federated learning, characterized in that: include: Identity management architecture module, based on the layered blockchain, designs a layered drone identity management architecture to connect each sub-trust domain; An identity registration and authentication module, which is connected to the identity management architecture module, designs a distributed cross-domain drone identity management mechanism to perform drone identity registration and identity authentication to ensure that the identities between sub-domain drone groups are interoperable and credible; a communication optimization module, which is connected to the identity registration and authentication module, and in response to the authentication success status, determines the teacher model to perform knowledge distillation to update the local model, include, Perform local training of the drone swarm, determine an embedding vector based on the results of the local training, filter and aggregate the embedding vector, and update the model parameters by combining the minimization of the distillation loss function and the stochastic gradient descent algorithm to perform global model aggregation to obtain a converged model; A dynamic selection module is connected to the communication optimization module, and dynamically selects neighbors based on the actual communication environment during the transmission process of the embedded vector. include, The user-side distance is calculated to construct a UAV system distance matrix, the communication object is determined based on the distance matrix, a temporary adjacency matrix is ​​constructed, and the soft labels generated by training are transferred to adjacent UAVs.

2. The drone swarm safety collaboration system based on blockchain and federated learning according to claim 1 is characterized in that: The layered drone identity management architecture includes an identity management system layer, a sub-blockchain layer, a cross-domain communication layer, and a main blockchain layer.

3. The drone swarm safety collaboration system based on blockchain and federated learning according to claim 2 is characterized in that: The identity management architecture module connects each sub-trust domain, including: Used to set identity attribute credentials based on the identity management system layer; Used to store the identity attribute certificate in the sub-blockchain of each sub-domain in the sub-blockchain layer; To perform communication protocol conversion between domains based on the cross-domain communication layer; It is used to complete identity authentication request communication between subdomains based on the main blockchain layer.

4. The drone swarm safety collaboration system based on blockchain and federated learning according to claim 1 is characterized in that: The identity registration and authentication module performs identity registration, include, Used to generate a private key and a corresponding public key based on PUF; Used to generate a DID document based on the identity management system that receives the public key; Used to update the corresponding attribute accumulator in the blockchain based on the DID document and generate attribute credentials; The attribute certificate is used to store the attribute certificate in a memory.

5. The drone swarm safety collaboration system based on blockchain and federated learning according to claim 4 is characterized in that: The identity registration and authentication module performs identity authentication, including: To receive identity authentication requests; Parsing the request to determine the attribute credentials that need to be verified; The identity authentication request to confirm that the attribute certificate is in the blockchain and the drone has the corresponding attribute can be authenticated; The parameters of the identity authentication request include the DID of the service and resource provider, the DID of the drone node, the access control policy and the signature information.

6. The drone swarm safety collaboration system based on blockchain and federated learning according to claim 1 is characterized in that: The communication optimization module determines a teacher model, where Used to determine the drone model that does not participate in the current round of local training as the teacher model.

7. The drone swarm safety collaboration system based on blockchain and federated learning according to claim 1 is characterized in that: The communication optimization module determines an embedding vector based on the result of the local training of the local training, including: Used to initialize the local private model and train it with the shared dataset to obtain a pre-trained model; Used to obtain local training results after training is completed; Used to generate an embedding vector on the shared dataset based on the local training result.

8. The drone swarm safety collaboration system based on blockchain and federated learning according to claim 1 is characterized in that: The communication optimization module performs filtering and aggregation operations on the embedding vectors, including: Used to filter invalid embedding vectors received by the drone; Used to calculate the similarity score between valid embedding vectors; Used to select several embedding vectors with the smallest anomaly score to form a new set; Aggregating the plurality of embedding vectors by using an aggregation function to obtain an aggregated embedding vector; The invalid embedding vector is an embedding vector generated by a drone failure or a malicious attack.

9. The drone swarm safety collaboration system based on blockchain and federated learning according to claim 8 is characterized in that: The communication optimization module combines the minimization of the distillation loss function and the stochastic gradient descent algorithm to update the model parameters, including: to embed a vector based on the aggregate; Used to determine the distillation loss function; Used to measure the probability distribution difference based on the distillation loss function; To determine the model parameters by combining the minimization distillation loss function and the stochastic gradient descent algorithm; to determine an updated model based on the parameters.

10. The drone swarm safety collaboration system based on blockchain and federated learning according to claim 1 is characterized in that: The dynamic selection module determines the communication object based on the distance matrix, wherein: To determine communication loss; The communication object corresponding to the minimum communication loss is determined as the communication object.

Images