Block chain enabled zero-trust low-altitude Internet of Things routing method

Abstract

The invention discloses a blockchain-enabled zero-trust low-altitude networking of intelligence routing method, which divides an unmanned aerial vehicle into a plurality of clusters to cooperatively transmit demands from sensing equipment to a base station, and proposes a zero-trust network architecture based on a software-defined boundary controller for identity authentication of the unmanned aerial vehicle and dynamic management of joining and exiting of the unmanned aerial vehicle. Designing an adaptive weight trust model to evaluate the reputation value of the unmanned aerial vehicle, proposing a lightweight alliance block chain storage transaction record and an unmanned aerial vehicle node credible state, and establishing an integer nonlinear programming routing problem which simultaneously optimizes end-to-end time delay and a transmission success rate; the routing problem is remodeled into a distributed partially observable Markov decision process, and a deep reinforcement learning algorithm based on a soft hierarchical experience playback pool and a priority experience playback mechanism is designed to solve a routing path. The method can adapt to a dynamically changing network topology environment, and minimizes the harm of potential malicious nodes to the network.

Description

Technical Field

[0001] This invention relates to the field of trusted routing technology for low-altitude intelligent networks, specifically to a blockchain-enabled zero-trust low-altitude intelligent network routing method. Background Technology

[0002] As a crucial component of 6G communication networks, low-altitude intelligent networks (LAIs) are widely used in various tasks such as disaster relief and real-time monitoring. In these applications, drones can collect data from sensor devices and then transmit it to ground base stations. Furthermore, within LAIs, drone networks can collaboratively complete complex tasks, providing low-cost, flexible, and multi-purpose services. Routing is a critical issue when transmitting the required data. However, due to the complex application environment, high dynamics, and distributed topology of drones, they are vulnerable and unreliable when subjected to security threats such as attacks and node failures.

[0003] To mitigate the impact of potential attacks, identity authentication mechanisms are needed. In particular, traditional authentication strategies based on network boundaries, such as firewalls and virtual private networks (VPNs), have been extensively studied. However, when a node passes through the security authentication of a traditional boundary, there is a lack of dynamic monitoring of the node's state; the node remains trusted, and malicious behavior is ignored. Such static strategies are unsuitable for the highly dynamic low-altitude intelligent network. In contrast, a zero-trust network security architecture is based on dynamic verification of nodes and behaviors, adhering to the principle of "never trust, always verify." In a zero-trust architecture, software-defined boundaries play a crucial role. They achieve secure access control through authentication, dynamic authorization, and detailed management, reducing unauthorized access to network resources and ensuring the security of data transmission.

[0004] Furthermore, the reliability and availability of communication links in low-altitude intelligent networks (LAHs) are variable due to the malicious intent and malfunctions of drones, which can degrade routing performance. Therefore, efficiently assessing drone trust and managing drone mobility is crucial in LHAs. Specifically, some works have introduced ground control stations as central managers to manage drone trust; however, these lack fault tolerance and are susceptible to tampering, long distances, and interference. Thus, effectively managing node trust and ensuring routing security in distributed and unreliable networks remains a challenging problem. In particular, blockchain, as a distributed ledger, provides verifiable and traceable records of interactions, storing transaction details and records, and is difficult to modify. Therefore, blockchain can be applied to record drone opinions and build decentralized trust management mechanisms. Meanwhile, some works related to node trust assessment use fixed weights for calculation, lacking adaptability and flexibility. Therefore, an evaluation method based on adaptive weights is needed to improve timeliness.

[0005] Furthermore, end-to-end latency and transmission success rate are key foundations for optimizing routing in low-altitude intelligent networks, as low latency and high transmission success rate can significantly enhance the reliability and timeliness of various emergency applications. Therefore, it is necessary to propose dynamic routing algorithms to achieve efficient, reliable, and timely data transmission in changing network topologies. However, most traditional routing algorithms (e.g., A*, Floyd-Warshall, and Dijkstra) are designed for static environments or fixed rules and cannot directly adapt to moving drones and changing topologies. Therefore, multi-agent deep reinforcement learning is an effective algorithm for dynamically interacting with the environment.

[0006] The invention disclosed in CN119341967A proposes a blockchain-enabled intelligent trusted routing method for drone networks. This method does not rely on difficult-to-obtain global information, can adapt to dynamic environments, has low computational costs, and can obtain trusted routing paths in a short time, effectively solving the trusted routing problem caused by malicious nodes in drone networks due to deliberate attacks. However, this invention does not address the issue of continuous verification, has high consensus costs, and still lacks in security and network efficiency. Summary of the Invention

[0007] The purpose of this invention is to propose a blockchain-enabled zero-trust low-altitude intelligent network routing method that can adapt to dynamically changing network topology environments, does not rely on global information that is difficult to obtain in low-altitude intelligent networks, and proposes an adaptive weighted trust evaluation method to update the trust status of nodes at a faster speed. Through continuous verification of the zero-trust network, the harm of potential malicious nodes to the network can be minimized.

[0008] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0009] A blockchain-enabled zero-trust low-altitude intelligent network routing method, the method comprising the following steps:

[0010] S1. Design a multi-source-multi-purpose low-altitude intelligent network model. The multi-source-multi-purpose low-altitude intelligent network model includes multiple nodes, including sensing devices, drones, and base stations. Construct a set of communication link states between all nodes. Divide the time period into T time slots. In each time slot, the demand data collected from the sensing devices will be relayed to the base station for reception and processing by the drone swarm.

[0011] S2 employs a zero-trust low-altitude intelligent network architecture to authenticate and authorize drone joining or leaving requests. This architecture comprises four basic units: a ground control station, drones, a software-defined boundary controller (SDB), and a blockchain. The ground control station is responsible for registering and initializing drones at the start of a mission. The SDB dynamically manages and controls drone access to network resources based on identity information. The blockchain stores immutable and trusted information and enables the SDB to handle authentication and requests in a zero-trust environment.

[0012] S3 assigns a reputation value representing trustworthiness to the drone, and updates the drone's reputation value by combining the drone's demand forwarding rate, trust interaction degree, probe packet reception rate, and opinions of neighboring drones; malicious drones are filtered out based on security thresholds.

[0013] S4 employs a lightweight consortium blockchain to record drone transactions and credit value, assisting a software-defined boundary controller in managing drone movement. Drones meeting specific credit, storage, and computing power requirements are allocated to participate in the overall consensus process. The drone with the highest credit value and strongest storage and computing power is designated as the master drone, acting as the blockchain client to collect all transactions, verify the signature and message authentication code of each transaction, and compile the collected transactions into a block. The total consensus latency of the lightweight blockchain consensus process is calculated.

[0014] S5 sets up multiple demand data streams from different source sensors to different destination base stations; introduces a queue buffer for each node, models the channel, and calculates the channel transmission rate; establishes a routing transmission model, and calculates the transmission delay of the demand between nodes, as well as the total end-to-end delay of transmitting the demand from the source sensor node to the target base station, based on the channel transmission speed. The number of unsuccessful transmission requests is obtained by combining the transmission delay between nodes, the total consensus delay, and the total end-to-end delay. The transmission success rate is then calculated by combining the number of unsuccessful transmission requests with the total number of transmissions.

[0015] S6 is designed based on an integer nonlinear optimization problem and corresponding constraints. It aims to maximize the transmission success rate during the transmission process, taking into account the mobility of drones and malicious drones. Total end-to-end delay The ratio;

[0016] S7 is a problem reconstruction based on a distributed partially observable Markov decision process. It treats each drone as an independent agent to make routing decisions and then sends the requirements to the next available drone to obtain the routing paths for all requirements.

[0017] S8, in the case of multiple sources and multiple destinations, utilizes a multi-agent deep reinforcement learning algorithm based on a soft-layered experience replay pool and a priority experience replay mechanism to solve the strategy that optimizes the end-to-end latency and transmission success rate, so as to optimize the total end-to-end latency and transmission success rate.

[0018] Compared with existing technologies, this invention proposes a blockchain-enabled zero-trust low-altitude intelligent network architecture. This architecture employs software-defined boundaries and blockchain technology to manage the identity and mobility of drones. Based on this architecture, an improved Multi-Agent Deep Q-Network (MADDQN) algorithm (enabled by the proposed soft-layered experience replay caching and priority experience replay mechanism) is designed to optimize the routing problem. The beneficial effects of this invention are as follows:

[0019] First, the blockchain-enabled zero-trust low-altitude intelligent network routing method of the present invention improves the transmission efficiency of demand through multi-cluster drones and designs a new network model based on zero-trust architecture. It applies software-defined boundary controllers to the identity management and authentication of drones, and minimizes the damage caused by potential security threats through continuous verification, thus solving the problems of security threats and the dynamic joining and leaving of drones.

[0020] Secondly, the blockchain-enabled zero-trust low-altitude intelligent network routing method of this invention designs an adaptive weighted trust model that comprehensively evaluates the credit value of drones by combining direct and indirect factors, thereby mitigating the impact of low-trust drones on network performance. Simultaneously, by introducing blockchain technology to record drone transactions and reputation values, it supports software-defined boundary controllers in managing drone mobility. Furthermore, it proposes the use of a lightweight consortium blockchain and improved Practical Byzantine fault tolerance. The lightweight consortium blockchain reduces consensus latency costs and improves the efficiency of the low-altitude intelligent network. The improved Practical Byzantine fault tolerance mechanism enhances network security—this mechanism comprehensively considers the drone's credit value and storage and computing resources when selecting the master drone, thereby effectively reducing consensus latency costs, enhancing network security, and improving the efficiency of the low-altitude intelligent network.

[0021] Third, the blockchain-enabled zero-trust low-altitude intelligent network routing method of this invention constructs an integer nonlinear routing problem to optimize the total end-to-end latency and transmission success rate of transmission requirements. This routing problem is reformulated as a partially observable Markov decision process to address the challenge of obtaining global information in decentralized low-altitude intelligent networks. It learns the dynamically changing network topology of UAVs through a multi-agent deep double-Q network algorithm based on a soft-layered experience replay pool and priority experience replay, and makes optimization decisions. Extensive simulations have evaluated the multi-agent deep double-Q network algorithm based on the soft-layered experience replay pool and priority experience replay, demonstrating that the adaptive weighted trust method identifies malicious UAVs faster than average and random evaluation methods. Simulation results show that the adaptive weighted trust model can identify malicious UAVs more quickly than methods based on multi-agent deep double-Q networks and deep double-Q networks without soft-layered experience replay pools and priority experience replay mechanisms. Furthermore, experiments demonstrate that the designed SP-MADDQN algorithm outperforms other comparative algorithms in terms of average end-to-end latency, transmission success rate, convergence, and training performance loss. Attached Figure Description

[0022] Figure 1 This is a routing example of a blockchain-enabled zero-trust low-altitude intelligent network involved in this invention; Part 1) introduces the routing scenarios and security threats in the low-altitude intelligent network based on the zero-trust architecture; Part 2) introduces the mobile management of drones through blockchain and software-defined boundary technology; and Part 3) explains the end-to-end latency model.

[0023] Figure 2 Simulation diagrams showing the minimum time step required to identify malicious drones under different simulation parameters involved in this invention;

[0024] Figure 3 The simulation diagram shows the convergence performance of the algorithm proposed in this invention under different demand quantities.

[0025] Figure 4 The simulation diagram shows the relationship between the end-to-end routing delay and the number of drones for the algorithm proposed in this invention and three other algorithms.

[0026] Figure 5 The figure shows the experimental simulation of the transmission success rate of the algorithm proposed in this invention under different numbers of malicious drones. Detailed Implementation

[0027] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0028] This invention discloses a blockchain-enabled zero-trust low-altitude intelligent network routing method, the method comprising the following steps:

[0029] S1. Design a multi-source-multi-purpose low-altitude intelligent network model. The multi-source-multi-purpose low-altitude intelligent network model includes multiple nodes, including sensing devices, drones, and base stations. Construct a set of communication link states between all nodes. Divide the time period into T time slots. In each time slot, the demand data collected from the sensing devices will be relayed to the base station for reception and processing by the drone swarm.

[0030] S2 employs a zero-trust low-altitude intelligent network architecture to authenticate and authorize drone joining or leaving requests. This architecture comprises four basic units: a ground control station, drones, a software-defined boundary controller (SDB), and a blockchain. The ground control station is responsible for registering and initializing drones at the start of a mission. The SDB dynamically manages and controls drone access to network resources based on identity information. The blockchain stores immutable and trusted information and enables the SDB to handle authentication and requests in a zero-trust environment.

[0031] S3 assigns a reputation value representing trustworthiness to the drone, and updates the drone's reputation value by combining the drone's demand forwarding rate, trust interaction degree, probe packet reception rate, and opinions of neighboring drones; malicious drones are filtered out based on security thresholds.

[0032] S4 employs a lightweight consortium blockchain to record drone transactions and credit value, assisting a software-defined boundary controller in managing drone movement. Drones meeting specific credit, storage, and computing power requirements are allocated to participate in the overall consensus process. The drone with the highest credit value and strongest storage and computing power is designated as the master drone, acting as the blockchain client to collect all transactions, verify the signature and message authentication code of each transaction, and compile the collected transactions into a block. The total consensus latency of the lightweight blockchain consensus process is calculated.

[0033] S5 sets up multiple demand data streams from different source sensors to different destination base stations; introduces a queue buffer for each node, models the channel, and calculates the channel transmission rate; establishes a routing transmission model, and calculates the transmission delay of the demand between nodes, as well as the total end-to-end delay of transmitting the demand from the source sensor node to the target base station, based on the channel transmission speed. The number of unsuccessful transmission requests is obtained by combining the transmission delay between nodes, the total consensus delay, and the total end-to-end delay. The transmission success rate is then calculated by combining the number of unsuccessful transmission requests with the total number of transmissions.

[0034] S6 is designed based on an integer nonlinear optimization problem and corresponding constraints. It aims to maximize the transmission success rate during the transmission process, taking into account the mobility of drones and malicious drones. Total end-to-end delay The ratio;

[0035] S7 is a problem reconstruction based on a distributed partially observable Markov decision process. It treats each drone as an independent agent to make routing decisions and then sends the requirements to the next available drone to obtain the routing paths for all requirements.

[0036] S8, in the case of multiple sources and multiple destinations, utilizes a multi-agent deep reinforcement learning algorithm based on a soft-layered experience replay pool and a priority experience replay mechanism to solve the strategy that optimizes the end-to-end latency and transmission success rate, so as to optimize the total end-to-end latency and transmission success rate.

[0037] Figure 1 This invention demonstrates a routing example in a blockchain-enabled zero-trust low-altitude intelligent network (LAH) based on a specific embodiment. The invention discloses a blockchain-enabled zero-trust LHA routing method, comprising the following steps:

[0038] Step S1: Design a multi-source, multi-purpose low-altitude intelligent interconnection network model.

[0039] In the network, the total number of nodes is I+U+B, and each node has a unique identity; the network includes I sensor nodes, U drones, and B base stations. and These represent collections of base stations, sensor nodes, and drones, respectively; furthermore, drone collections are categorized according to their different capabilities. It can be divided into and Three types are used for data collection, relay forwarding, and downlink, respectively; specifically, each type of drone ensemble is divided into multiple clusters, and the drone with the strongest energy and computing power in each cluster is selected as the cluster head; ε = ε iu ∪ε uu ∪ε ub Represents the set of communication link states between all nodes, where e iu ∈ε iu e ub ∈ε ub and e uu ∈ε uu These represent the links between sensor nodes and drones, base stations and drones, and drones and drones, respectively; e nm ∈ε represents a node and Does a communicable link exist between them? Specifically, e nm =1 indicates the link is active, e nm =0 indicates that there is no direct link;

[0040] The length of each time slot is τ, and T = {1, ..., t, ..., T} represents the set of time slots; at the beginning of time slot t, the demand r ∈ R is generated from the source sensing device. Transmitted to the destination base station Information is represented as Where the source node s r =i, destination node dr=b, required size r is Bit, This represents the maximum latency tolerance for demand transmission; furthermore, the demand originates from the sensing device. Upload to drone By drone Relay, from drone Download to base station This indicates that the demand r will originate from the source sensing device. Transmitted to the corresponding destination base station The complete route path.

[0041] The coordinates of sensor i and base station b are kept constant, and are denoted as Θ. i =(x i y i ,0) and Θ b =(x b y b ,0); In time slot t, the position of UAV u in the three-dimensional rectangular coordinate system is Θ u (t)=(x u (t), y u (t), z u (t)); Furthermore, in time slot t, the Euclidean distance d between node n∈I∪U and node m∈U∪B nm (t) is represented as:

[0042]

[0043] Specifically, the distance between drones must meet the following requirements:

[0044]

[0045] Where, d min This represents the safe distance to avoid collisions between drones; at time t, the set of drones u that can communicate with each other is denoted as Γ. u (t), UAV u and UAV κ∈Γ u The distance d between (t) uκ (t) satisfies:

[0046]

[0047] Where, d u，maxThe maximum communication distance for the UAV u is given.

[0048] Step S2: Construct a zero-trust low-altitude intelligent network architecture solution.

[0049] In distributed low-altitude intelligent networks, malicious drones are more likely to survive and disrupt network performance. To ensure the security of the low-altitude intelligent network, a zero-trust architecture scheme is proposed, combining blockchain and software-defined boundary technology, to authenticate and authorize join or leave requests. The zero-trust architecture consists of four basic units: a ground control station, a drone, a software-defined boundary controller, and a blockchain, such as... Figure 1 As shown;

[0050] In this system, drones act as clients joining or leaving the Low-Altitude Intelligent Network (LAI); the ground station is responsible for registering and initializing the drones at the start of the mission, and the drones need to obtain valid registration information from the ground station; the software-defined boundary controller dynamically manages and controls the drones' access to network resources based on their identity information. Specifically, only authenticated and authorized drones have the authority to join the LAI; the blockchain stores immutable and trustworthy information and enables the software-defined boundary controller to handle identity authentication and requests in a zero-trust environment.

[0051] Drone joining and leaving: Before applying to join the cluster, a drone must register with the ground control station to obtain unique identity information. Then, the drone packages the identity information, the ground control station's signature, and the identification document for joining the cluster into a transaction and broadcasts the transaction to the Low-Altitude Intelligent Network (LAI), which is then verified by the selected cluster leader drone. After consensus is reached, the drone is allowed to join the cluster network. When a drone leaves, the cluster leader drone that leaves the drone needs to send the corresponding transaction to the LAI to dynamically update the status and changes of link availability.

[0052] S3 assigns a reputation value representing trustworthiness to the drone, and updates the drone's reputation value by combining the drone's request forwarding rate, trust interaction level, probe packet reception rate, and opinions from neighboring drones.

[0053] In low-altitude intelligent networks, to improve routing efficiency, reputation values ​​representing trustworthiness are assigned to drones. An adaptive weighted trust model is designed to update the reputation values ​​of drones, including direct trust that considers drone demand forwarding rate, trust interaction degree and probe packet reception rate, as well as indirect trust by aggregating opinions from neighboring drones. It is a set of reputation values, for drone u, The value is in the range [0, 1]; a given security threshold determined according to security requirements. Drones fall into two categories: trusted and malicious; u (t) indicates whether the drone u is trustworthy, i.e.:

[0054]

[0055] Therefore, the set and number of malicious drones are respectively represented as: The comprehensive formula for calculating credit score is:

[0056]

[0057] In the formula, This represents the initial reputation value of drone u. As a direct trust value, This is an indirect trust value; and They represent and The weights; In addition, an adaptive model for dynamically adjusting weights is proposed, as follows:

[0058]

[0059] along with The increase of weight Reduce, thus and Increasing the value amplifies the impact of malicious behavior on reputation assessment; furthermore, the auxiliary parameter β∈(0,1] is used to limit... The range of values ​​for; when hour, The adaptive weighted trust model includes direct trust based on demand forwarding rate, trust interaction degree, and probe packet reception rate, as well as indirect trust based on aggregating opinions from neighboring drones, thus avoiding the inadequacy of relying on a single evaluation factor.

[0060] S4 uses a lightweight consortium blockchain to record the transaction and credit value of drones, assisting the software-defined boundary controller in managing drone movement; it calculates the total consensus latency of the lightweight blockchain consensus process.

[0061] Centralized trust managers rely on third-party trusted centers, making them vulnerable to diverse attacks that could lead to network collapse. To overcome these security threats, a consortium blockchain is designed to record drone transactions and credit value, providing a distributed and reliable manager for reputation values. This assists a software-defined boundary controller in managing drone movement, creating a secure and reliable communication network and ensuring secure routing within the low-altitude intelligent network. Simultaneously, the Practical Byzantine Fault Tolerance (PBT) method, as the primary consensus mechanism of the consortium blockchain, is applied to the consensus process, tolerating up to one-third of malicious drones. Furthermore, to enhance the security of the low-altitude intelligent network and the reliability of the PBT method, an improved PBT method is proposed. This method designates the drone with the highest reputation value and strongest capabilities as the master drone, and this designation is updated periodically. Only drones with high trust values, storage, and computing power are assigned as master drones to participate in the consensus, while other drones are designated as light drones, which helps reduce resource consumption.

[0062] Regarding role selection, in a lightweight blockchain, drone ensembles Divided into a complete collection of drones and light drone collection in:

[0063] a) Fully drone-based: Fully drone-based systems possess a full blockchain ledger and are primarily responsible for blockchain consensus, including the broadcasting and verification of transactions;

[0064] b) Light drones: In the low-altitude intelligent network, apart from fully automated drones, other light drones can only store the header of blocks and are mainly responsible for generating local transactions and forwarding transactions;

[0065] c) Master Drone: The master drone acts as a blockchain client, collecting all transactions, verifying the signature and message authentication code of each transaction, and then compiling the collected transactions into a block.

[0066] d) Secondary drones: In addition to the primary drone, other fully unmanned aerial vehicles are considered as non-primary (secondary) drones.

[0067] For a consensus process based on improved practical Byzantine fault tolerance:

[0068] At time slot t, the number of total consensus drones and malicious consensus drones are expressed as K(t) and K(t), respectively. This indicates that, in order to reach consensus, each node needs to aggregate at least [number] replicas from different replicas. A consistent preparation message, where F(t) = (K(t) - 1) / 3, if Consensus based on improved practical Byzantine fault tolerance may fail, leading to untimely information updates; Figure 1The document details the consensus process, where the consensus time cost includes generating a signature, verifying the signature, and generating / verifying a message verification code, with costs ∈ [missing information]. s ,∈ v and ∈ m The consensus process consists of four phases: collection, pre-preparation, preparation, and response. The corresponding delays are discussed below:

[0069] a) During the collection phase, the master drone collects routing information from other drones and then verifies it; specifically, after transmitting the request, the light drone sends its information to the nearest full drone; then, each secondary drone generates a transaction based on the information received from the surrounding environment, the transaction is signed by the secondary drone's private key and forwarded to the master drone through a message verification code;

[0070] In this step, the delay is primarily due to the main drone verifying all packaged transactions; specifically, to verify the transaction signature and message verification code from drone K(t)-1, it is necessary for (K(t)-1)(∈ v +∈ m )CP∪period; in addition, (∈) is required v +∈ m The CPU cycles are used to verify the request message; the latency of this step is:

[0071]

[0072] in This refers to the computing resources allocated by the main drone for block consensus, namely CPU speed;

[0073] b) During the pre-preparation phase, the master drone first packages the verified transactions into a block, then generates signatures for each block, and generates (K(t)-1) message verification codes for (K(t)-1) slave drones. The master drone's latency is:

[0074]

[0075] Subsequently, each secondary drone consumes ∈ v +∈ m CPU cycles are used to verify the prepared message and consume K(t)(∈ v +∈ m The CPU cycles are used to verify the signature and message verification code of transaction K(t), with a latency of:

[0076]

[0077] This refers to the CPU resources allocated by the replica drone for blockchain consensus; therefore, the total latency of the pre-preparation step is...

[0078] c) During the preparation phase, since the main UAV only needs to be verified The message verification code and the received preparation message signature, therefore the delay of the main drone is:

[0079]

[0080] Compared to the main drone, the secondary drone consumes additional ∈ s +(K(t)-1)∈ m The CPU cycles are used to generate the signature for the preparation message, while the other K(t)-1 full drones require additional K(t)-1 message verification codes. Therefore, the latency for each secondary drone is:

[0081]

[0082] Based on the above analysis, the delay in this stage is It is a collection of all secondary drones;

[0083] d) During the response phase, each drone receives After completing the consistency preparation message, a commit message is broadcast to all other consensus drones, and at least one consensus drone must be accumulated. A consistent commit message is generated; in this step, each drone generates a signature for the commit message and generates a K(t)-1 message verification code for the K(t)-1 full drones; in addition, each full drone verifies... Signature and message verification code. Therefore, the total delay for this step is:

[0084]

[0085] Therefore, the overall consensus delay for:

[0086]

[0087] After routing is completed, the drone periodically generates transactions awaiting consensus, including the drone's status (i.e., queue buffer), location, and information about neighboring drones. Then, all generated transactions are broadcast to the entire network for verification via the Gossip protocol. Once verified, the transaction is added to the blockchain's transaction pool and then synchronized to the entire blockchain. Once consensus is reached, the data transactions are recorded in the blockchain, making them tamper-proof.

[0088] S5, calculate the total end-to-end delay of transmitting the demand from the source sensor node to the target base station. and transmission success rate

[0089] In designing a low-altitude intelligent network, a total transmission routing delay model is used to represent multiple requirements from different source sensing devices to different destination base stations, using binary variables. Indicates whether demand r passes through the link. Right now:

[0090]

[0091] binary variables The drone that represents the demand is:

[0092]

[0093] Since transmission delay is constrained by the channel transmission rate G, this paper analyzes the channel characteristics of ground-to-air and air-to-air wireless communication links; specifically, the node and The path loss between them is represented by L. nm (t), which remains constant within time slot t, i.e.:

[0094]

[0095] Where λ represents the carrier frequency and c is the speed of light; and These represent the additional path loss under line-of-sight and non-line-of-sight propagation, respectively; when parameter ω = 0, L nm (t) represents the path loss between drones; otherwise, if ω = 1, then L nm (t) indicates from arrive Or from arrive Path loss; in addition, Pr nm (t) represents the probability of a line-of-sight link between the sensor node / base station and the drone, i.e.

[0096]

[0097] Where h nm (t) and χ nm (t) represents the difference between the height and horizontal distance between the two nodes, respectively; and It is a constant parameter;

[0098] Due to the high mobility and unstable data traffic fluctuations of the low-altitude intelligent network, a queue buffer is introduced for each node n to alleviate network congestion; specifically, at time slot t-1, the demand received by node n is... Furthermore, at time slot t, C n (t), and Representing nodes respectively The queue length, maximum queue capacity, queuing demand set, and transmission demand quantity are as follows:

[0099]

[0100] Here, the drone transmits requests in parallel, indicating that all requests received from time slot t-1 are forwarded in time slot t. To reduce the transmission latency difference between different demands within a time slot, a demand-based... and total bandwidth The adaptive channel bandwidth allocation scheme allocates a total bandwidth. for:

[0101]

[0102] Based on Shannon's theory, at time slot t, the demand r from node arrive transmission rate for:

[0103]

[0104] in, and These are the transmission power and noise power between nodes n and m, respectively; the delay in transmitting the demand from node n to node m at time slot t is:

[0105]

[0106] Wherein, demand r is the delay of transmission from node n to node m. The routing path for transmitting demand r from source sensor node i to target base station b. At that time, total end-to-end delay for:

[0107]

[0108] S6 is designed based on an integer nonlinear optimization problem and corresponding constraints. It aims to maximize the transmission success rate during the transmission process, taking into account the mobility of drones and malicious drones. Total end-to-end delay The ratio of .

[0109] The optimization objective is to minimize the total end-to-end latency in the presence of malicious drones and mobility. And improve the transmission success rate of low-altitude intelligent networks. Therefore, the optimization problem is formulated as follows:

[0110]

[0111] in, Indicates whether requirement r is present. and Link e between nm ∈ε transmission; Indicates whether demand r is located at Above; it is worth noting that from the node The requirement r can only be met by another node take over.

[0112] Specifically The number N of requests that failed to be transmitted fail (t) and total number of transmissions N total (t) = |R| is calculated as follows: In particular, transmission failure may be caused by the following: a) Compare large; b) or Greater than τ; c) Greater than τ; d) The request will not be transmitted to the corresponding destination base station, represented as arrive = 0; According to the above definition, the total number of failures by time slot t is:

[0113] Where, N fail (t)=N fail (t-1)+ψ(t);

[0114]

[0115] In the formula, N fail (t) is initialized to 0. When t = T, N fail =N fail (t).

[0116] S7 is a problem reconstruction based on a distributed partially observable Markov decision process. It treats each drone as an independent agent to make routing decisions, and then sends the requirements to the next available drone to obtain the routing paths for all requirements.

[0117] It takes the form of integer nonlinear programming, which is difficult to handle; therefore, in this invention, a feasible solution is proposed based on a reformulation of a distributed partially observable Markov decision process; to adapt to the dynamically changing network environment and the local observations of each UAV, Rewritten as a distributed partially observable Markov decision process; in the distributed partially observable Markov decision process, each drone u is an independent agent, making routing decisions and sending the demand to the next available drone; therefore, the set of agents is the set of drones u; in each time slot t, agent u∈u observes the local state o from the environment.u (t), and according to o u (t) Execute action a u (t), and receive a reward immediately. The environment shifts to the next observation state. u (t+1), the transition marker for agent u is f. u (t) indicates; specifically This refers to the conversion of experience, specifically as follows:

[0118] 1) State Space In time slot t, state o u (t) is the drone n∈{u}∪Γ u The position of (t) Θ n (t), Queuing demand set and reputation value Set, that is:

[0119]

[0120] In time slot t, the joint state s(t) aggregates the observations of all agents, denoted as Therefore, the state space is represented as

[0121] 2) Action Space: The agent independently addresses the needs of each carrier. Make a decision, Let be the queuing demand set of drone u in time slot t. Sub-action This indicates that the next-hop neighbor node selected by drone u is used to relay the request r, i.e.:

[0122]

[0123] Wherein, in time slot t, Z u (t) is related to drones A set of directly connected base station nodes; if a drone Connect to demand If the target base station b is to be reached, the data will be transmitted directly from drone u to base station b; conversely, if the demand needs to be relayed by a nearby drone, in time slot t, Represents the set of actions for all agents. Represents the action space.

[0124] 3) Rewards: in This is the reward that agent u receives after transmitting the demand r in time slot t. The routing path is determined hop-by-hop; therefore, the reward function is designed with one hop as the minimum period. Considering... For the data transmission requirement r from drone u to drone κ, design the reward value and link e.uκ Delay It is negatively correlated with Positive correlation; in particular, Indicates link e uκ The probability of successful transmission; this is reasonable because it is small. and This can lead to incorrect transmissions and random loss of requests, directly reducing the transmission success rate; therefore, the transmission success rate of a single hop... It can be used Therefore, the reward is:

[0125]

[0126] The setting ι is to ensure and They are of the same order of magnitude; It is a hyperparameter that balances the impact of latency and reputation score; furthermore, in time slot t, the total reward for all demands is:

[0127]

[0128] 4) Transition flag: f(t) = {f r (t)|r∈R} is a flag set for all requirements, f r (t) indicates whether demand r has reached the target base station. Defined as:

[0129]

[0130] 5) Discount factor: γ u This is the cumulative reward for drone u; the reward set for all drones is...

[0131] At time slot t, the strategy Drive the intelligent agent u to observe O u The demand r under (t) takes action The joint policy for demanding agent u is represented as: The strategy of all agents is: Once the distributed partially observable Markov decision process and specific policies are determined, the routing paths for all requirements can be obtained; therefore, the objective becomes finding the optimal policy. This represents the optimal joint policy for agent u.

[0132] S8, in the case of multiple sources and multiple destinations, utilizes a multi-agent deep reinforcement learning algorithm based on a soft-layered experience replay pool and a priority experience replay mechanism to solve the strategy that optimizes the end-to-end latency and transmission success rate, so as to optimize the total end-to-end latency and transmission success rate.

[0133] The priority-based experience replay mechanism allows agents to sample experiences of higher importance more frequently, enabling them to focus on the most informative experiences and achieve faster and more efficient learning when dealing with large-scale state and action spaces. The soft-hierarchical experience replay pool constructs an experience buffer for each agent by embedding distance into the reward, i.e.:

[0134]

[0135] Where d uκ (t) and d κb (t) represents the distances between UAV κ and UAV u and the target base station b, respectively. Indicates whether demand r is on link e uκ Transmission; due to the large d uκ (t) requires fewer hop counts, saving time costs; therefore d uκ (t) is designed to be positively correlated with rewards; meanwhile, d uκ (t)+d κb (t) is the minimum transmission distance from the selected next-hop drone κ to drone u and base station b; when d uκ (t)+d κb (t) is closer to the shortest straight-line distance d ub When (t), agent u may receive a larger reward value, and vice versa;

[0136] In DDQN, for each agent u, there is a parameter θ. u The online network and parameters are The target network; historical experience is represented as Stored in the experience replay buffer of agent u; based on importance priority, from Extract historical experience transformation tuples from D, and then transform L(θ) into L(θ). u ) is calculated as the Q-value function Q(o) u (t), a u (t); θ u ) and Q objective function y u The mean square error between (t) is:

[0137]

[0138] in, Indicates from Samples v are obtained by sampling from the middle, and their corresponding loss values ​​are weighted and averaged. Tuple v is The transformed data, y u (t) is:

[0139]

[0140] also, L(θ) u Find the gradient of θ and use it to update θ via gradient descent. u , such that L(θ) u Minimize, that is:

[0141]

[0142] α u It's the learning rate, a parameter. Updated periodically every W steps to match the online network parameter θ u μ is the soft update coefficient, i.e.:

[0143]

[0144] exist Figure 2 In order to evaluate the performance of the proposed adaptive weighting mechanism, two benchmark methods are introduced, described as follows:

[0145] 1) Average weighting method: and for The average value is expressed as:

[0146]

[0147] 2) Random weight method: Modeled as uniformly distributed in Random variables within a certain range, i.e.:

[0148]

[0149] To demonstrate the effectiveness of the simulation trust evaluation method, the experimental setup is configured as follows:

[0150] 1) The number of drones is 12, and the number of malicious drones is 2;

[0151] 2) The reputation value is initialized to 1, indicating that the drone is fully trusted when it leaves the ground control area;

[0152] 3) Considering the stringent safety requirements, The probability of malicious behavior is set to 0.8, and the probability range is set to [0.1, 0.5]. Specifically, this paper selects 0.4 and 0.2 as the probabilities of all malicious behaviors for simulation. Therefore, the probability of malicious drone demand forwarding, the probability of trusted interaction with high-trust drones, and the probability of receiving probe packets are p1∈{0.6,0.8}, p2∈{0.6,0.8}, and p3∈{0.6,0.8}, respectively. Furthermore, the simulation parameter space A of (p1, p2, p3) is:

[0153]

[0154] The evaluation metric is defined as the minimum time step required to identify all malicious drones, i.e.:

[0155]

[0156] Figure 2 To evaluate the efficiency of the proposed adaptive weighting method in calculating reputation values, the results were compared with those of the average weighting and random weighting methods. Specifically, this invention evaluated the performance across various simulation parameters. When identifying malicious drones, the time step required is minimal when (p1, p2, p3) = (0.6, 0.6, 0.6). Conversely, the time step required is maximum when (p1, p2, p3) = (0.8, 0.8, 0.8). Decreasing values ​​of p1, p2, and p3 indicate a higher frequency of erroneous behavior by malicious drones, leading to an increase in the assessed erroneous behavior ratio. Furthermore, compared to other methods, the proposed method consistently achieves the minimum time step across different simulation parameters. This is because the trust weights in the adaptive weighted trust model are proportional to the erroneous behavior rate, and the drone's reputation value is dynamically adjusted at a quadratic rate. Therefore, the penalty for erroneous behavior is more severe, and the detection of malicious drones is accelerated. Under the same conditions, the proposed method outperforms the average weighted trust method and the random weighted trust method.

[0157] Figure 3 The convergence performance of the proposed algorithm in transmitting requests under different request quantities is shown. It is noteworthy that different request quantities have different impacts on the convergence and value of the reward. As the number of requests increases, the reward converges rapidly in the early datasets. This is because higher request values ​​can meet the batch size requirements for training more quickly. Furthermore, more requests yield higher rewards because the total reward of the Low Altitude Intelligent Network is defined as the sum of all transmitted requests.

[0158] exist Figure 4In the study, as the number of requests increased, the average end-to-end latency of all algorithms showed an upward trend, consistent with the designed latency model. Among them, the Multi-Agent Deep Q-Network algorithm (SP-MADDQN) based on a soft-layered experience replay pool and a priority experience replay mechanism achieved the lowest latency cost compared to the other three algorithms (MADQN (Multi-Agent Deep Q-Network), MADDQN (Multi-Agent Deep Deterministic Policy Gradient), and SP-MADQN (StatePrediction-based Multi-Agent Deep Q-Network)). This indicates that the variant of the method based on the soft-layered experience replay pool and priority experience replay mechanism is more effective in reducing end-to-end latency.

[0159] Figure 5 The figure shows the simulation results of the proposed algorithm's transmission success rate under different numbers of malicious drones. It can be observed that the average transmission success rate is highest when the number of malicious drones is 0. When malicious drones are present, the transmission success rate drops significantly, indicating that malicious drones reduce transmission performance during routing. Furthermore, the average transmission success rate drops sharply as the number of malicious drones increases. This is because malicious drones may reject transmission requests, thus losing the request. As the number of drones increases, the total network resources become richer, such as channel bandwidth, storage, and computing power, leading to a higher transmission success rate. Overall, however, this invention can identify malicious drones more quickly and achieves better performance in terms of average end-to-end latency, transmission success rate, convergence, and training performance loss.

[0160] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0161] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A blockchain-enabled zero-trust low-altitude intelligent network routing method, characterized in that, The method includes the following steps: S1. Design a multi-source-multi-purpose low-altitude intelligent network model. The multi-source-multi-purpose low-altitude intelligent network model includes multiple nodes, including sensing devices, drones, and base stations. Construct a set of communication link states between all nodes. Divide the time period into T time slots. In each time slot, the demand data collected from the sensing devices will be relayed to the base station for reception and processing by the drone swarm. S2 employs a zero-trust low-altitude intelligent network architecture to authenticate and authorize drone joining or leaving requests. This architecture comprises four basic units: a ground control station, drones, a software-defined boundary controller (SDB), and a blockchain. The ground control station is responsible for registering and initializing drones at the start of a mission. The SDB dynamically manages and controls drone access to network resources based on identity information. The blockchain stores immutable and trusted information and enables the SDB to handle authentication and requests in a zero-trust environment. S3 assigns a reputation value representing trustworthiness to the drone, and updates the drone's reputation value by combining the drone's demand forwarding rate, trust interaction degree, probe packet reception rate, and opinions of neighboring drones; malicious drones are filtered out based on security thresholds. S4 employs a lightweight consortium blockchain to record drone transactions and credit value, assisting a software-defined boundary controller in managing drone movement. Drones meeting specific credit, storage, and computing power requirements are allocated to participate in the overall consensus process. The drone with the highest credit value and strongest storage and computing power is designated as the master drone, acting as the blockchain client to collect all transactions, verify the signature and message authentication code of each transaction, and compile the collected transactions into a block. The total consensus latency of the lightweight blockchain consensus process is calculated. S5 sets up multiple demand data streams from different source sensors to different destination base stations; introduces a queue buffer for each node, models the channel, and calculates the channel transmission rate; establishes a routing transmission model, and calculates the transmission delay of the demand between nodes, as well as the total end-to-end delay of transmitting the demand from the source sensor node to the target base station, based on the channel transmission speed. The number of unsuccessful transmission requests is obtained by combining the transmission delay between nodes, the total consensus delay, and the total end-to-end delay. The transmission success rate is then calculated by combining the number of unsuccessful transmission requests with the total number of transmissions. S6 is designed based on an integer nonlinear optimization problem and corresponding constraints. It aims to maximize the transmission success rate during the transmission process, taking into account the mobility of drones and malicious drones. Total end-to-end delay The ratio; S7 is a problem reconstruction based on a distributed partially observable Markov decision process. It treats each drone as an independent agent to make routing decisions and then sends the requirements to the next available drone to obtain the routing paths for all requirements. S8, in the case of multiple sources and multiple destinations, utilizes a multi-agent deep reinforcement learning algorithm based on a soft-layered experience replay pool and a priority experience replay mechanism to solve the strategy that optimizes the end-to-end latency and transmission success rate, so as to optimize the total end-to-end latency and transmission success rate.

2. The blockchain-enabled zero-trust low-altitude intelligent network routing method according to claim 1, characterized in that, Step S1 further includes: Construct a multi-source, multi-purpose low-altitude intelligent interconnection network model, which includes a set of sensor nodes consisting of I sensing devices. A drone ensemble consisting of U drones and a set of base stations consisting of B base stations The total number of nodes is I+U+B, and each node has a unique identity; Based on their different capabilities, drones are grouped together Divided into and Three types of drones are used for data collection, relay forwarding, and downlink, respectively; Each type of drone collection is divided into multiple clusters, and the drone with the strongest energy and computing power in each cluster is selected as the cluster head; Construct a set of communication link states ε = ε between all nodes iu ∪ε uu ∪ε ub , where e iu ∈ε iu e ub ∈ε ub and e uu ∈ε uu These represent links between sensor nodes and drones, base stations and drones, and drones and drones, respectively; e nm ∈ε represents a node and Is there a communicable link between them? nm =1 indicates the link is valid, e nm =0 indicates that there is no direct link; The time period is divided into T time slots, each with a length of τ, where T = {1, ..., t, ..., T} represents the set of time slots. At the beginning of time slot t, the demand r ∈ R is generated from the source sensing device. Transmitted to the destination base station Information is represented as Where the source node s r =i, destination node d r =b, the size of the demand r is Bit, The maximum latency tolerance for demand transmission; demand from the sensing device Upload to drone By drone Relay, from drone Download to base station This indicates that the demand r will originate from the source sensing device. Transmitted to the corresponding destination base station The complete route path, where R represents the set of requirements.

3. The blockchain-enabled zero-trust low-altitude intelligent network routing method according to claim 1, characterized in that, Step S1 further includes: The coordinates of sensor i and base station b are kept constant, and are denoted as Θ. i =(x i y i ,0) and Θ b =(x b y b ,0); In time slot t, the position of UAV u in the three-dimensional rectangular coordinate system is Θ u (t)=(x u (t), y u (t), z u (t)); In time slot t, the Euclidean distance d between node n∈I∪U and node m∈U∪B is... nm (t) is represented as: The distance between drones meets the following requirements: Where, d min This indicates the safe distance to avoid collisions between drones, (x n (t), y n (t), z n (t)) and (x m (t), y m (t), z m (t) represents the positions of node n and node m in the three-dimensional rectangular coordinate system at time t; the set of communicable UAVs of UAV u at time t is denoted as Γ. u (t), UAV u and UAV κ∈Γ u The distance d between (t) uκ (t) satisfies: Where, d u，max The maximum communication distance for the UAV u is given.

4. The blockchain-enabled zero-trust low-altitude intelligent network routing method according to claim 1, characterized in that, In step S2, the process of authenticating and authorizing drone joining or leaving requests using a zero-trust low-altitude intelligent network architecture includes: Before applying to join the cluster, the drone obtains unique identity information by registering at the ground control station. The drone packages the identity information, the ground control station's signature, and the identification document for joining the cluster into a transaction and broadcasts the transaction to the cluster leader drone in the low-altitude intelligent network for verification. After consensus is reached, the drone is allowed to join the cluster network. When a drone leaves the network, the drone in the cluster corresponding to the drone that left sends a corresponding transaction to the low-altitude intelligent network to dynamically update the status and changes of the link availability.

5. The blockchain-enabled zero-trust low-altitude intelligent network routing method according to claim 1, characterized in that, In step S3, an adaptive weighted trust model is used to update the reputation value of the drone. The reputation value of the drone includes direct trust considering the drone's demand forwarding rate, trust interaction degree, and probe packet reception rate, as well as indirect trust by aggregating the opinions of neighboring drones. The reputation value of the drone is: In the formula, This represents the initial reputation value of drone u. As a direct trust value, This is an indirect trust value; and They represent and The weight, and These are the reputation values ​​of drone u at time t and time t+1, respectively. The adaptive weighted trust model is as follows: In the formula, the auxiliary parameter β∈(0,1] is used to limit The range of values ​​for; The set of reputation values ​​for all drones was obtained through statistics. For drones, The value is in the range [0, 1]. Determine a given safety threshold based on safety requirements. Drones are categorized into two states: trusted and malicious; u (t) indicates whether the drone u is trustworthy: The set and number of malicious drones are respectively represented as:

6. The blockchain-enabled zero-trust low-altitude intelligent network routing method according to claim 1, characterized in that, Step S4 further includes: In a lightweight blockchain, drones are aggregated Divided into a collection of all drones and light drone collection Among them, the full-scale drone possesses the entire blockchain ledger and is responsible for blockchain consensus, including the broadcasting and verification of transactions; the lightweight drone only stores the block header and is responsible for generating local transactions and forwarding transactions. The drone with the highest reputation value and the strongest storage and computing power is designated as the master drone. The master drone acts as a blockchain client, collecting all transactions, verifying the signature and message authentication code of each transaction, and then compiling the collected transactions into a block. Other drones besides the master drone participate in the consensus as secondary drones. The consensus process is divided into four stages: collection, pre-preparation, preparation, and response. The total consensus latency of the lightweight blockchain consensus process is calculated. for: In the formula, and The total delays for the four stages of collection, pre-preparation, preparation, and response are respectively; K(t) and Let F(t) represent the total number of consensus drones and the number of malicious consensus drones at time slot t, respectively; F(t) = (K(t) - 1) / 3; ∈ s ,∈ v and ∈ m These represent the number of computation cycles for generating a signature, verifying a signature, and generating / verifying a message verification code, respectively. It refers to the CPU resources allocated by the main drone for block consensus; It refers to the CPU resources allocated by the replica drone for blockchain consensus; It is a collection of all secondary drones.

7. The blockchain-enabled zero-trust low-altitude intelligent network routing method according to claim 1, characterized in that, In step S5, a total transmission route delay model is constructed for multiple requirements from different source sensing devices to different destination base stations, and the nodes are included. and The path loss between them is represented by L. nm (t): Where λ represents the carrier frequency and c is the speed of light; and These represent the additional path loss under line-of-sight and non-line-of-sight propagation, respectively; when ω = 0, L nm (t) represents the path loss between drones; when ω=1, L nm (t) indicates from arrive Or from arrive Path loss; Pr nm (t) represents the probability of a line-of-sight link between the sensor node / base station and the drone: Where h nm (t) and χ nm (t) represents the difference between the height and horizontal distance between the two nodes, respectively; and It is a constant parameter; binary variables Indicates whether demand r passes through link e nm , binary variables The drone that represents the demand: A queue buffer is introduced for each node n. At time slot t-1, the demand received by node n is... At time slot t, C n (t), and Representing nodes respectively Queue length, maximum queue capacity, queue demand set, and number of transmission requests: The drone transmits requests in parallel, and all requests received from time slot t-1 are forwarded in time slot t. Based on demand and total bandwidth Adaptive allocation of channel bandwidth, total allocated bandwidth for: Based on Shannon's theory, at time slot t, the demand r from node arrive transmission rate for: in, and These are the transmission power and noise power between nodes n and m, respectively; the delay in transmitting the demand from node n to node m at time slot t is: Wherein, the delay of demand r from node n to node m. The routing path for transmitting demand r from source sensor node i to target base station b. End-to-end delay for: The number N of requests that failed to be transmitted fail (t) and total number of transmissions N total (t) = |R| Calculate the transmission success rate Among them, N fail (t)=N fail (t-1)+ψ(t); 8. The blockchain-enabled zero-trust low-altitude intelligent network routing method according to claim 1, characterized in that, In step S6, the optimization problem is: :s.t.d min ≤d nm (t) d uκ (t)≤d u，max in, Indicates whether requirement r is present. and Link e between nm ∈ε transmission; Indicates whether demand r is located at Above; from node The requirement r is only affected by another node take over.

9. The blockchain-enabled zero-trust low-altitude intelligent network routing method according to claim 1, characterized in that, Step S7 further includes: Optimization problem Rewritten as a distributed partially observable Markov decision process; in the distributed partially observable Markov decision process, each drone u is an independent agent and makes routing decisions, sending the demand to the next available drone; the set of agents is the set of drones u; in each time slot t, agent u observes the local state o from the environment. u (t), according to o u (t) Execute action a u (t), and receive a reward immediately. The environment shifts to the next observation state. u (t+1), the transition marker for agent u is f. u (t) represents; This represents the transformation of experience, where: In time slot t, state o u (t) is the drone n∈{u}∪Γ u The position of (t) Θ n (t), Queuing demand set and reputation value The set of: The joint state s(t) aggregates the observations of all agents, denoted as The state space is represented as The intelligent agent independently addresses the needs of each carrier. Make a decision Sub-action This indicates that the next-hop neighbor node selected by the drone u is used for relaying the request r: Where, in time slot t, z u (t) is related to drones A set of directly connected base station nodes; if a drone Connect to demand If the target base station is b, then the data is transmitted directly from drone u to base station b; otherwise, the data is relayed by a nearby drone. Let be the queuing demand set of UAV u in time slot t; in time slot t, Represents the set of actions for all agents. Represents the action space; Design a reward function with one hop as the minimum period: in It is the reward that agent u receives after requesting transmission r in time slot t: Among them, the auxiliary parameter ι is used to make and link e uk Delay They are of the same order of magnitude; It is a hyperparameter that balances the impact of latency and reputation score; Indicates link e uκ The probability of successful transmission; Indicates whether demand r is on link e uk Transmission; in time slot t, the total reward for all demands is: f(t)={f r (t)|r∈R} is the transformation flag set for all requirements, f r (t) indicates whether demand r has reached the target base station. Defined as: The cumulative reward set for all drones is γ u It is the cumulative reward for drone u; At time slot t, the strategy Drive the intelligent agent u to observe O u The demand r under (t) takes action The joint policy for demanding agent u is represented as: The strategy of all agents is: The goal is to find the optimal strategy. This represents the optimal demand joint strategy for agent u.

10. The blockchain-enabled zero-trust low-altitude intelligent network routing method according to claim 1, characterized in that, Step S8 further includes: An experience buffer for each agent is constructed by embedding distance into the reward: Where d uκ (t) and d κb (t) represents the distances between UAV κ and UAV u and the target base station b, respectively; Indicates whether demand r is on link e uκ transmission; For each agent u, the parameter is θ u The online network and parameters are The target network; historical experience is represented as Stored in the experience replay buffer of agent u; based on importance priority, from Extract historical experience transformation tuples from D, and then transform L(θ) into L(θ). u ) is calculated as the Q-value function Q(o) u (t), a u (t); θ u ) and Q objective function y u The mean square error between (t) is: in, Indicates from Samples v are obtained by sampling from the middle, and their corresponding loss values ​​are weighted and averaged. Tuple v is The transformed data, y u (t) is: Where, a′ u (t) represents the possible candidate actions of the drone u at time t. This means finding a′ that maximizes the value of the subsequent function. u (t), θ u These are online network parameters. These are the target network parameters, based on L(θ) u gradient of ) Update θ using gradient descent u , such that L(θ) u ) minimize: Where, α u It's the learning rate, a parameter. Updated periodically every W steps to match the online network parameter θ u μ is the soft update coefficient:

Images