A safe and reliable cooperative sensing method for unmanned aerial internet

Abstract

This invention relates to a secure and reliable collaborative perception method for UAV Internet, comprising: Step 1, constructing a joint optimization model for perception anomalies and transmission security issues based on a secure and reliable collaborative framework for UAV Internet; Step 2, dividing perception areas based on a balance between area and complexity, assigning roles according to UAV trust values, and completing collaborative cluster deployment through complementary capabilities game theory; Step 3, collecting perception data based on the deployed UAV cluster, and using the trust-enhanced Ridge Mahalanobis Distance (TERMD) algorithm according to the joint optimization model to detect perception data anomalies; Step 4, using the PFL-EPT algorithm to securely transmit perception results based on the perception data anomaly detection results; Step 5, updating UAV trust values ​​through a DTS mechanism based on task execution results, and feeding the trust values ​​back to Step 2, forming a trusted closed-loop iteration to continuously achieve the joint optimization goal.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) communication technology, and in particular to a secure and reliable collaborative perception method for UAV Internet. Background Technology

[0002] Traditional terrestrial communication is limited by network capacity and coverage, failing to meet the demands of wireless access services. This has led to the rise of highly mobile unmanned aerial vehicle (UAV)-assisted communication systems. Compared to terrestrial communication, air-to-ground communication can adapt to complex environments, expand coverage, and improve transmission efficiency, significantly optimizing communication system performance and service quality. However, the resulting Internet of Drones (IoD) faces two core challenges in regional perception tasks: abnormal perception data and vulnerable transmission security. Existing research in this area has significant limitations.

[0003] In terms of perception area division and task scheduling, existing research has proposed methods such as static priority division, dynamic area division, dynamic adjustment of energy perception, and deep reinforcement learning optimization. Although these methods have improved perception efficiency, communication connectivity, energy consumption balance, and intelligent scheduling, they have not fully matched the characteristics of the perception area with the performance of UAVs, nor have they utilized the complementary capabilities of swarms to improve perception accuracy. They have failed to achieve an effective balance between perception accuracy and resource scheduling efficiency.

[0004] In the field of anomaly detection in perceived data, distance-based detection methods are widely used due to their low complexity and lack of assumptions about data distribution. Among them, Mahalanobis distance, which can measure the correlation of feature dimensions, is the mainstream detection method. However, the covariance matrix of traditional Mahalanobis distance is easily affected by outliers, and the estimation is unstable in high-dimensional scenarios. Subsequent improvements such as MCD, regularization, ridge covariance matrix, and weighting mechanisms have gradually improved the detection stability in high-dimensional, small-sample scenarios, but they all ignore the performance factors of the UAV sensor itself, such as its accuracy and operating status, and cannot accurately assess and weight the data reliability of different UAVs.

[0005] Regarding the secure transmission of perception results, wireless transmission is vulnerable to malicious attacks that could lead to data tampering. Federated learning, by transmitting model parameters instead of the original data, has become the mainstream solution for secure transmission. However, classic algorithms such as FedAvg and FedProx suffer from convergence difficulties and insufficient model personalization in heterogeneous data scenarios. Improved solutions such as micro-wireless federated learning and multi-key homomorphic encryption only optimize global model performance and privacy protection; their rigid aggregation mechanisms cannot adapt to the diverse and personalized modeling needs of multiple drone swarms.

[0006] In terms of trust assessment of drones, the existing trust mechanism based on historical behavior, dynamic weighting, and multi-dimensional indicators can reduce the interference of bad nodes on perception data, but it has problems such as insensitivity to short-term behavior, insufficient assessment data, and single dimension. Moreover, it does not design differentiated assessment standards for different mission roles of drones, and cannot fairly and accurately reflect the true performance of each drone in specific missions.

[0007] In summary, current research on IoD communication security is fragmented, focusing on optimizing only a single aspect without building a secure and reliable collaborative framework covering the entire communication process. This makes it difficult to cope with overlapping security threats in complex and dynamic environments, ultimately leading to insufficient overall system resilience. Summary of the Invention

[0008] The purpose of this invention is to propose a secure and reliable collaborative perception method for the Internet of Unmanned Aerial Vehicles (UAVs) to solve the problems existing in the prior art.

[0009] To achieve the above objectives, the present invention provides the following solution: A secure and reliable collaborative perception method for UAV Internet includes: Step 1: Based on the pre-built UAV Internet security and trustworthy collaborative framework, and taking into account latency, energy consumption and trust scheduling constraints, construct a joint optimization model for perception anomaly and transmission security issues; Step 2: Divide the perception area based on a balance between area and complexity, assign roles according to the trust value of drones, and complete the collaborative cluster deployment through complementary capabilities game theory. Step 3: Based on the deployed drone cluster collecting perception data, and according to the joint optimization model, use the trust-enhanced Ridge Mahalanobis distance TERMD algorithm to detect anomalies in the perception data. Step 4: Based on the anomaly detection results of the sensing data, use the PFL-EPT algorithm to securely transmit the sensing results; Step 5: Update the UAV trust value through the DTS mechanism based on the task execution results, and feed the trust value back to Step 2 to form a trusted closed-loop iteration to continuously achieve the joint optimization goal.

[0010] Optionally, the pre-built UAV Internet security and trustworthy collaborative framework includes: an IoD system collaborative perception and communication scenario and a full-process security and trustworthy collaborative framework; the IoD system collaborative perception and communication scenario is based on a ground control station GCS, a heterogeneous UAV cluster, and a blockchain network; the IoD system collaborative perception and communication scenario includes a three-layer structure: the blockchain layer is responsible for storing and updating UAV trust values, the flight layer is where the heterogeneous UAV cluster performs perception tasks, and the ground layer integrates data through the ground control station GCS; The end-to-end secure and reliable collaborative framework includes: pre-communication perception area division and UAV task scheduling; during communication, perception anomaly detection and secure transmission of perception results; and trust closed-loop update after communication ends.

[0011] Optionally, the perception region can be divided based on a balance between region area and region complexity, including: Adopting the first One sensing area area Entropy of Land Feature Categories The proposed regional complexity modeling method integrates two dimensions: regional area and land feature category complexity. It comprehensively characterizes the perceptual difficulty of a region and divides the region into sub-regions with the closest complexity.

[0012] Optionally, roles can be assigned hierarchically based on drone trust values, and collaborative cluster deployment can be achieved through complementary capabilities and game theory. First, select the one with the highest trust value. UAVs are used as key UAVs in each sub-region. According to the region With the region The complexity and requirements of edge perception determine the characteristics of ordinary UAVs at the boundary of the region. Number From the remaining Select the UAV with the second highest trust value The frame serves as a regular UAV at the regional border. ,in The rest The UAV will serve as a regular UAV within the area. ; For ordinary UAVs in the region Adaptive deployment is used to construct a region selection mechanism based on game theory strategies; in the region selection mechanism, firstly from... shelf Among them, those ranked higher in trust were selected. A regular UAV The system selects a sensing area based on its compatibility with the sensing area, ensuring that each area has an initial ordinary UAV. Then, in the remaining A regular UAV In the middle, the one with the second highest trust value is selected again. The drone performs area selection; at this time, there are two ordinary UAVs in each area. This process is iterated round by round until all ordinary UAVs are involved. Region selection and deployment completed.

[0013] Optionally, the trust-enhanced Ridge Mahalanobis Distance (TERMD) algorithm is used for anomaly detection in perceptual data, including: Introducing a trust-driven weighted aggregation mechanism: This mechanism is used to aggregate data from various ordinary UAVs within a given region. When aggregating perception data, the first... A regular UAV In the Trust value after the round-robin communication task is completed As a regular UAV Weighting factors for perceived data; Ridge Covariance Matrix and Mahalanobis Distance Calculation: Based on the normalized mean of perceptual data, the concept of ridge regression is introduced to construct a positive definite and invertible ridge covariance matrix. The distance is then calculated based on this positive definite and invertible ridge covariance matrix. A regular UAV Mahalanobis distance; Dynamic threshold setting and trusted data aggregation: This involves obtaining all common UAVs... After calculating the spacetime Mahalanobis distance, all ordinary UAVs in this region will be included. The Mahalanobis distance is considered as the sample set, and the region is calculated. Mahalanobis distance mean of the perceived data and standard deviation And based on this, the judgment threshold for anomaly detection is set.

[0014] Optionally, using the PFL-EPT algorithm for secure transmission of sensing results includes: Personalized model updates and aggregation: in the obtained region Perception results Afterwards, the region Key UAV The initial global model will be received from the ground control station GCS. ,area Key UAV In local dataset Each node independently performs federated learning training tasks, generating personalized local model parameters. The training objective is to minimize the current global model. Lower region Dataset Local loss function ; The local model is uploaded to the ground control station GCS, and the ground control station GCS... The local model parameters are aggregated using similarity weighting to construct a personalized global model. The aggregated global model is then transmitted back in an encrypted manner. In the new round of local training and global synchronization, the entire federated learning process is conducted through key UAVs. The alternating optimization iterations between GCSs of ground control stations have been completed; Lightweight encrypted transmission mechanism: It integrates an end-to-end secure transmission protocol based on a physically non-clonable function and a lightweight key. This protocol ensures forward security and resistance to replay attacks while adapting to the resource constraints of UAVs.

[0015] Optionally, the DTS mechanism includes: Accuracy assessment is used to quantify the consistency between UAV output results and reality; ordinary UAV As data producers, their accuracy assessment depends primarily on the quality of the raw, perceived data provided, especially for key UAVs. As a regional manager, the core competency lies in the comprehensive judgment of information and the quality of decision-making; Reliability assessment is used to quantify the stability and behavioral consistency of UAVs in multiple consecutive missions; the reliability of ordinary UAVs considers their role suitability in the current mission and the stability of their historical behavior, while the reliability of critical UAVs focuses on the continuity of management responsibilities. Energy efficiency assessment is used to quantify the resource utilization efficiency of UAVs during task execution; the energy efficiency of a typical UAV is determined by hovering energy consumption. Sensing energy consumption Regarding communication energy consumption It consists of three parts, with the key UAV's energy efficiency including hovering energy consumption. Calculate energy consumption With communication energy consumption Three parts.

[0016] The beneficial effects of this invention are as follows: This invention proposes a comprehensive secure and trustworthy collaborative framework integrating "regional task scheduling – trust-enhanced anomaly detection – lightweight privacy transmission – trust update closed-loop feedback." This framework constructs an end-to-end collaborative protection mechanism that integrates "perception area division, node selection and complementary capability team formation, area perception, perception data anomaly removal, local model aggregation optimization, key encrypted transmission, and dynamic trust update." Specific innovations are as follows: 1) A method for regional partitioning and UAV collaborative task scheduling oriented towards the heterogeneity of sensing areas is proposed. Specifically, the strategy first dynamically partitions regions based on area and the complexity of land cover types; then, it adopts hierarchical UAV deployment, configuring high-performance nodes as key UAVs in each region to be responsible for data aggregation, while deploying suboptimal nodes as edge fusion units to ensure coverage continuity; finally, by designing a game mechanism based on capability complementarity and benefit optimization, it adaptively forms a group of ordinary UAVs in the region to collaboratively execute sensing tasks, thereby achieving optimized resource scheduling within the region.

[0017] 2) The TERMD algorithm is proposed to improve the accuracy of perception data in UAV group collaborative perception tasks. This algorithm innovates the traditional ridged Mahalanobis distance by introducing the historical trust value of UAVs to construct a trust-enhanced distance metric. Furthermore, by fusing the perception bias and trust score of each UAV, it enhances the influence of UAV perception data with high trust and reliable perception results in the aggregation process.

[0018] 3) The PFL-EPT algorithm is proposed to improve the security and privacy protection of UAV data transmission. This algorithm guides global aggregation by measuring the model similarity between regions, preserving the individual characteristics of each region while utilizing global consensus, thereby improving the performance of the model in heterogeneous environments. On this basis, the algorithm introduces a lightweight key mechanism to encrypt the transmitted parameters, achieving low-overhead end-to-end secure communication.

[0019] 4) A DTS mechanism is proposed to provide closed-loop feedback for task scheduling. This mechanism abandons the traditional single evaluation standard and instead comprehensively considers the accuracy, reliability and energy efficiency factors under different task attributes to conduct differentiated scoring, thereby dynamically and accurately measuring the credibility of UAV behavior under specific tasks. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of an IoD (In-place) area perception and communication scenario according to an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of the end-to-end secure and reliable collaborative framework according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the selection of a drone according to an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] This embodiment proposes a secure and reliable collaborative perception method for unmanned aerial vehicle (UAV) internet, mainly including: Step 1: Based on the pre-built UAV Internet security and trustworthy collaborative framework, and taking into account latency, energy consumption and trust scheduling constraints, construct a joint optimization model for perception anomaly and transmission security issues; Step 2: Divide the perception area based on a balance between area and complexity, assign roles according to the trust value of drones, and complete the collaborative cluster deployment through complementary capabilities game theory. Step 3: Based on the deployed drone cluster collecting perception data, and according to the joint optimization model, use the trust-enhanced Ridge Mahalanobis distance TERMD algorithm to detect anomalies in the perception data. Step 4: Based on the anomaly detection results of the sensing data, use the PFL-EPT algorithm to securely transmit the sensing results; Step 5: Update the UAV trust value through the DTS mechanism based on the task execution results, and feed the trust value back to Step 2 to form a trusted closed-loop iteration to continuously achieve the joint optimization goal.

[0026] The large-scale deployment of Unmanned Aerial Vehicles (UAVs) has given rise to the Internet of Drones (IoD), which faces two key challenges when performing area perception tasks: the inherent anomalies of perception data and the security vulnerabilities of perception results during transmission. Existing research often addresses these two issues in isolation, lacking a unified, secure, and reliable collaborative framework. Therefore, this embodiment, under the constraints of UAV energy consumption, latency, and trust scheduling, jointly optimizes the accuracy of perception data and the security of perception result transmission. Based on the established model, a secure and reliable collaborative framework is proposed that runs through the entire task execution process, encompassing "regional task scheduling—trust enhancement anomaly detection—lightweight privacy transmission—trust update closed loop." Before communication, the area is divided according to the perception area and the complexity of ground features. Combined with the performance indicators of each UAV, game theory optimization is used to complete the UAV task scheduling, achieving complementary and collaborative capabilities. In this embodiment, the trust value and perception bias of the UAV are incorporated into the communication process. A Trust-Enhanced Ridge-Based Mahalanobis Distance (TERMD) algorithm is proposed to improve the accuracy of anomaly detection. Furthermore, a Personalized Federated Learning with Encrypted Parameter Transmission (PFL-EPT) algorithm is proposed to achieve secure parameter aggregation while protecting privacy. After communication, a Differentiated Trust Score (DTS) mechanism is used to update the UAV trust value, forming a closed loop of trusted evidence storage and scheduling feedback. Simulation results show that the proposed framework achieves significant advantages in terms of anomaly data identification accuracy and data transmission security.

[0027] Furthermore, the pre-built UAV Internet security and trustworthy collaborative framework includes: IoD system collaborative perception and communication scenarios and a full-process security and trustworthy collaborative framework; The IoD system collaborative sensing and communication scenario is based on the ground control station GCS, heterogeneous UAV clusters, and a blockchain network. The IoD system collaborative sensing and communication scenario includes a three-layer structure: the blockchain layer is responsible for storing and updating UAV trust values, the flight layer is where the heterogeneous UAV cluster performs sensing tasks, and the ground layer integrates data through the ground control station GCS. The end-to-end secure and reliable collaborative framework includes: pre-communication perception area division and UAV task scheduling; during communication, perception anomaly detection and secure transmission of perception results; and trust closed-loop update after communication ends.

[0028] Furthermore, the perception region is divided based on a balance between region area and region complexity, including: Adopting the first One sensing area area Entropy of Land Feature Categories The proposed regional complexity modeling method integrates two dimensions: regional area and land feature category complexity. It comprehensively characterizes the perceptual difficulty of a region and divides the region into sub-regions with the closest complexity.

[0029] Furthermore, the collaborative cluster deployment is achieved by assigning roles based on drone trust values ​​and through complementary capabilities in a game-like process, including: First, select the one with the highest trust value. UAVs are used as key UAVs in each sub-region. According to the region With the region The complexity and requirements of edge perception determine the characteristics of ordinary UAVs at the boundary of the region. Number From the remaining Select the UAV with the second highest trust value The frame serves as a regular UAV at the regional border. ,in The rest The UAV will serve as a regular UAV within the area. ; For ordinary UAVs in the region Adaptive deployment is used to construct a region selection mechanism based on game theory strategies; in the region selection mechanism, firstly from... shelf Among them, those ranked higher in trust were selected. A regular UAV The system selects a sensing area based on its compatibility with the sensing area, ensuring that each area has an initial ordinary UAV. Then, in the remaining A regular UAV In the middle, the one with the second highest trust value is selected again. The drone performs area selection; at this time, there are two ordinary UAVs in each area. This process is iterated round by round until all ordinary UAVs are involved. Region selection and deployment completed.

[0030] Specifically, Step 1: This embodiment focuses on the collaborative sensing and communication scenario of an IoD system. Addressing the sensing anomalies and transmission security issues existing in the multi-UAV collaborative sensing process, a secure and reliable collaborative framework covering the entire process before, during, and after communication is constructed. This framework establishes a sensing anomaly risk model and an attack risk model, and comprehensively considers latency, energy consumption, and trust scheduling constraints. Based on this, the mathematical model for sensing anomalies and transmission security issues is finally constructed.

[0031] The application scenario constructed in this embodiment is as follows: Construct a collaborative sensing and communication scenario for an IoD system comprised of a ground control station (GCS), a heterogeneous UAV cluster, and a blockchain network. For example... Figure 1 As shown, this scenario comprises a three-layer structure: the blockchain layer is responsible for storing and updating UAV trust values; the flight layer consists of a heterogeneous UAV swarm performing sensing tasks; and the ground layer integrates data through the ground control station (GCS). In the blockchain layer, this embodiment defines the UAV set as... Each UAV corresponds to a trust value. Trust-based ranking The system selects the highest-ranked UAVs constitute a subset of perception tasks ,Right now At the flight level, the system will perceive the entire area. Divided into A subregion with balanced complexity Each sub-region Internally configured with a heterogeneous UAV group: multiple ordinary UAVs A key UAV The two work together to perform area perception tasks. Ordinary UAV Based on its deployment location, it can be further divided into: Ordinary UAVs in the area Responsible for sensing the types of ground features within the area; ordinary UAVs at the area boundaries. It focuses on sensing the types of ground features at the boundaries of the sensing area, and includes all ordinary UAVs. All are driven by trust, such as Figure 1 The task scheduling process shown (e.g.) Figure 1 (Steps 1-3). In the data aggregation phase, the key UAVs... Summary area ordinary UAV Perceived data Generate region aggregation results And further processed into data that can be used for transmission. The data is sent to the Ground Control Station (GCS). At the ground level, the GCS receives and integrates the key UAVs from each area. The aggregated perception results provide mobile users with a variety of services, including traffic monitoring, target tracking, and disaster relief.

[0032] Based on the aforementioned IoD area sensing and communication scenario, this embodiment further analyzes two key challenges faced in actual UAV communication: sensing data anomalies and data transmission security. During the sensing task execution phase, ordinary UAVs... Collected data It may be affected by both internal and external factors. From the perspective of internal factors, ordinary UAVs... Sensor components may experience decreased sensing accuracy due to physical damage or performance degradation, resulting in biased and abnormal sensing data. From an external threat perspective, malicious user MUs may acquire ordinary UAVs. To gain control, they deliberately uploaded incorrect perception data, thereby affecting the aggregated data. The accuracy of data transmission is crucial. During the data transmission phase, due to the broadcast and open nature of wireless channels, key UAVs... Transmitting sensing results to the ground control station GCS During the process, the mobile user may be subjected to eavesdropping, tampering, or replay attacks by malicious users (MU), who may send falsified perception results to the ground control station (GCS), leading to incorrect decisions by the GCS and causing significant harm to the mobile user.

[0033] End-to-end secure and trustworthy collaborative framework: To address the aforementioned challenges, this embodiment proposes the following... Figure 2 The illustrated end-to-end secure and reliable collaborative framework spans the entire lifecycle of communication tasks, constructing a complete assurance system from perception area segmentation to trust loop update. This includes perception area segmentation and UAV task scheduling before communication; perception anomaly detection and secure transmission of perception results during communication; and trust loop update after communication. It achieves efficient collaboration among multi-stage modules, effectively enhancing the system's perception accuracy while improving data security, ultimately constructing a structurally closed-loop and functionally complete IoD communication security assurance system.

[0034] 1) Before communication: Sensing area segmentation and UAV mission scheduling: To reduce the computational and communication load of a single UAV and improve the overall data perception accuracy of the system, this embodiment designs a regional perception scheduling model oriented towards regional heterogeneity, which mainly includes perception area division and performance-driven UAV scheduling.

[0035] In practical tasks, different regions exhibit significant differences in land cover composition and spatial scale. Using an equal-area division strategy would lead to uneven distribution of sensing difficulty across regions, causing UAV load imbalance and consequently affecting overall sensing performance. Therefore, this embodiment employs a method based on the [missing information - likely a specific method or strategy]. One sensing area area Entropy of Land Feature Categories The constructed region complexity modeling method aims to ensure a relatively balanced perceptual complexity for each sub-region. Specifically, it defines the region... Perceptual complexity as follows: (1); in, Indicates the region The total number of land cover types included. For the region Inner Area of ​​each land cover category within the region Percentage, expressed as This regional complexity modeling method effectively integrates two dimensions: regional area and land cover category complexity. It comprehensively characterizes the perceptual difficulty of a region. Ultimately, the goal is to divide the region into sub-regions with as similar a complexity as possible. .

[0036] This is followed by a performance-driven UAV scheduling phase. By comprehensively considering the trust level and performance indicators of the UAVs, highly reliable and high-performance UAVs are prioritized to participate in the communication task. The scheduling process of UAVs is as follows: Figure 3 As shown, when selecting UAVs to participate in the perception task, the system first selects the one with the highest trust value. UAVs are used as key UAVs in each sub-region. Subsequently, according to the region With the region The complexity and requirements of edge perception determine the characteristics of ordinary UAVs at the boundary of the region. Number The system from the remaining Select the UAV with the second highest trust value The frame serves as a regular UAV at the regional border. ,in The rest The UAV will serve as a regular UAV within the area. .

[0037] To enhance the intelligence and collaboration of regional task scheduling, this embodiment targets ordinary UAVs within the region. The adaptive deployment introduces a multi-agent game mechanism based on complementary capabilities and benefit optimization, which allows ordinary UAVs within the region to... By combining its own performance indicators with regional characteristics, it autonomously selects the optimal sub-region. Participating in perception tasks enables the overall optimization of performance matching, resource complementarity, and system benefits. This is to characterize the first... Ordinary UAVs in the frame area Multidimensional indicators, defining their attribute vectors as follows: (2); in, , They represent the first Ordinary UAVs in the frame area The computing power and remaining energy are the core indicators for subsequent team allocation and game optimization.

[0038] For ordinary UAVs To achieve adaptive deployment, this embodiment constructs a region selection mechanism based on a game-theoretic strategy. In this mechanism, the system first selects from... shelf Among them, those ranked higher in trust were selected. A regular UAV The system selects a sensing area based on its compatibility with the sensing area, ensuring that each area has an initial ordinary UAV. Then, in the remaining A regular UAV In the middle, the one with the second highest trust value is selected again. The drone performs area selection; at this time, there are two ordinary UAVs in each area. This process is iterated round by round until all ordinary UAVs are involved. Region selection and deployment are complete. During this process, each regular UAV... By selecting an independent strategy, its regional matching degree can reach a local optimum. When the system reaches Nash equilibrium, the following conditions are satisfied: (3); in, For the region Inner A regular UAV The region matching function is defined as follows: (4); in, and They are respectively regions Inner A regular UAV Sensing task matching degree and ordinary UAVs in the area The degree of collaborative matching, These are the weight parameters for the two, respectively. The first ordinary UAV to perform perception region selection in each region... Considering only its own performance and region Perceptual task matching degree Ordinary UAVs are deployed in each area. Based on this, a regular UAV was subsequently selected. Simultaneously consider the perceptual task matching degree and collaborative matching degree Two factors.

[0039] Perceived task matching degree Used to measure ordinary UAV The degree of adaptation between the UAV and the target perception area. The basis for this judgment is the same as that of a typical UAV. attribute vector The two key performance parameters are the adaptability to the region. If the computational power is good, the region with high land cover category information entropy is selected; conversely, if the remaining energy is high, the region with a larger area is selected for sensing. In other words, the region needs to be... area With the region Land feature category information entropy proportional allocation , and the A regular UAV Remaining energy Computing power proportional allocation The closest, based on this, the first is defined A regular UAV Performance and area Perception task matching degree function as follows: (5); As can be seen from formula (5), the task matching degree Perceived region features and the first A regular UAV The higher the attribute matching degree, The closer the value is to 1.

[0040] Collaborative matching degree Used to measure ordinary UAV The synergistic potential between groups. To quantify the heterogeneity and complementary synergistic potential of this cluster, the first... A regular UAV With this region Existing ordinary UAVs Collaborative matching degree function Using a regular UAV Attribute vector The variance between them is expressed as follows: (6); in, Indicates the region Ordinary UAVs within the inner area With the region Other ordinary UAVs attribute vector Calculate the overall variance across all dimensions. Use an exponential function to calculate the matching degree function. Controlled within (0, 1), and as The increase, A value close to 1 indicates a greater diversity of node attributes and more significant complementarity of capabilities within the region, which is conducive to the collaborative execution of complex perception tasks. Therefore, in the... A regular UAV When selecting a region, prioritize those with higher [percentages / areas]. Sub-regions are added to maximize overall perception performance and resource utilization efficiency.

[0041] After the game is completed, each sub-region The final complete UAV cluster is : (7); in, For the region Key UAV , and All belong to ordinary UAVs The former are regional boundary coordination nodes that enhance the continuity of regional perception, while the latter are nodes selected through game theory and distributed within the region.

[0042] 2) In communication: Sensing anomaly detection and secure transmission modeling After completing area division and UAV task scheduling, the system enters the core stage of task execution: data processing and transmission in communication. During this stage, ordinary UAVs... Perform area perception tasks and upload perception data to key UAVs. Key UAV It is responsible for two core tasks: first, improving the accuracy of sensing data through anomaly detection; and second, ensuring the privacy of sensing results through secure transmission mechanisms. This embodiment will first construct a mathematical model to evaluate the accuracy of sensing and the security of transmission, and further establish the system's latency and energy consumption constraints.

[0043] Perception accuracy modeling and anomaly risk: Based on the end-to-end secure and reliable collaborative framework proposed in this embodiment, in order to quantify the accuracy of the final perception results acquired by the ground control station's GCS, this embodiment introduces... Divergence, as an evaluation metric, is used to characterize the sensing results of GCS aggregation at ground control stations. Information on actual land cover categories The statistical deviation between them. Construct based on Divergence-based perceptual anomaly risk model It is expressed as follows: (8); in, and These are the actual land cover category information. Sensing results aggregated with ground control station GCS .

[0044] In the The perception results across multiple dimensions are analyzed by minimizing the perception anomaly risk model. This can maximize the improvement and This consistency enhances the overall perception accuracy of the system. Specifically, the perception results aggregated by the ground control station's GCS... Depend on Key UAVs in the region The obtained perception results are aggregated and represented as follows: (9); in, Indicates key UAV The weights of the perception results are used to measure the contribution of their aggregation results to global perception, and ; Indicates the region Key UAV The aggregated perception results.

[0045] Transmission security modeling and attack risk: To assess the privacy protection strength of the perceived results during transmission, this embodiment employs information theory to establish an attack risk model based on the difference between conditional entropy and information entropy. It is expressed as follows: (10); in, Aggregate perception results and data transmission The conditional entropy represents the condition when an attacker intercepts transmitted data. Subsequently, regarding the results of aggregated sensing... The remaining uncertainty; For aggregated sensing results The entropy is used to represent the result of aggregate perception when the attacker knows nothing about the system. Uncertainty. and The smaller the difference, the weaker the attacker's ability to infer plaintext information by intercepting the transmitted content, and the higher the system's encryption strength. and The calculation is expressed as follows: (11); in, Indicates aggregated perception results With transmitted data The joint probability distribution reflects the possibility that the two will occur simultaneously during the actual communication process of the system. To transmit data Under these conditions, aggregated sensing results The prior probability distribution represents the attacker's ability to infer plaintext information using intercepted information; For aggregated sensing results The marginal probability distribution reflects the natural distribution characteristics of the system for different land cover categories or perception states.

[0046] Latency and energy consumption constraint model: To ensure the timeliness and feasibility of system task execution, it is necessary to consider the critical UAVs. and ordinary UAV Strict constraints are established for latency and energy consumption in ordinary UAVs. As a data acquisition unit, its constraints primarily ensure the effective completion of the sensing task; key UAVs As a regional management node, its constraints need to meet more complex computing and communication requirements.

[0047] Ordinary UAV latency constraint: When performing perception tasks in a local area, the total latency of an ordinary UAV needs to meet a maximum latency limit to ensure the validity of the perceived data. Specifically, the latency of the first UAV... A regular UAV The task execution latency constraint is expressed as follows: (12); in, Indicates the system allows ordinary UAVs. The longest time limit for conducting area perception. Indicates the first After the first communication task is completed, the A regular UAV Total task execution latency, which includes data sensing time. With data upload time Specifically, it is represented as follows: (13); Energy consumption constraints for conventional UAVs: The total energy consumption during the execution of area sensing must be lower than its maximum energy budget. Specifically, the first A regular UAV The energy consumption constraint is expressed as follows: (14); in, Indicates the system allows ordinary UAVs. The maximum energy consumption for area sensing; Indicates the first After the first communication task is completed, the A regular UAV The total energy consumption. It consists of hovering energy consumption. Sensing energy consumption With communication energy consumption It consists of three parts. The specific formula is as follows: (15); Key UAV latency constraints: Key UAV The total latency when performing the aggregation task of region-aware data needs to meet the maximum latency limit to ensure the timeliness of data aggregation and data transmission. Specifically, the region... Key UAV The delay constraint is expressed as follows: (16); in, Indicates the key UAVs allowed by the system. The maximum time limit for aggregating regional sensing data and transmitting sensing results; Indicates the first After the second communication task is completed, the area Key UAV The total task execution latency includes data aggregation computation time. With result transmission time , means as follows: (17); Key UAV Energy Constraints: Key UAV The total energy consumption during the aggregation and transmission of regional sensing data must meet the maximum energy consumption limit to ensure sufficient power to support the aggregation of sensing data and the transmission of sensing results. Specifically, the regional... Key UAV The energy consumption constraint is expressed as follows: (18); in, Indicates the key UAVs allowed by the system. The maximum energy consumption for aggregating sensing data and transmitting sensing results. Indicates the first After the second communication task is completed, the area Key UAV Total energy consumption for task execution, including hovering energy consumption. Calculate energy consumption With communication energy consumption The specific formula is as follows: (19).

[0048] 3) After communication: Trust closed-loop update: The completion of the communication task marks the system entering a critical phase of self-optimization and trusted evolution. To achieve refined management of heterogeneous nodes and incentivize positive behavior, this embodiment proposes a role-driven differentiated trust update mechanism. This mechanism provides different trust updates for ordinary UAVs undertaking different responsibilities. and key UAV Trust assessment models with different focuses were established, taking into account the accuracy, reliability and energy efficiency of each task in the execution process, so as to accurately reflect their performance in different task attributes. The updated trust value will serve as the core basis for task scheduling and role allocation before the next round of communication, thus forming a reliable closed-loop feedback that drives the continuous optimization of the system.

[0049] Trust Update for Standard UAVs: Trust assessment for standard UAVs focuses on their individualized behavioral characteristics as data producers, reflecting their data quality, behavioral stability, and energy utilization efficiency during independent perception and data reporting. Specifically, the first... After the first communication task is completed, the A regular UAV Trust It is expressed as follows: (20); in, and The first After the next communication task is completed, the ordinary UAV The perception accuracy, historical reliability, and energy efficiency level together characterize the overall credibility of an individual task's execution; weighting coefficients Adaptive settings based on task scenarios to meet requirements This is to ensure the balance and interpretability among the multidimensional indicators.

[0050] Key UAV Trust Update: Unlike ordinary UAVs, key UAVs are responsible for regional management and data fusion. Their trust updates emphasize cross-node collaborative efficiency and decision robustness. They not only need to maintain their own high reliability but also ensure the fusion quality and transmission stability of the perception results from ordinary UAVs in the region. Therefore, in the first... After the next communication task is completed, the area Key UAV Trust The update is indicated as follows: (twenty one); in, and Measured separately as the first After the next communication task is completed, the key UAV Performance in terms of data fusion accuracy, collaborative reliability, and energy dispatch efficiency; weighting coefficients. satisfy .

[0051] By introducing role-differentiated weights during the trust evolution process and combining them with task-driven three-dimensional indicators, this model achieves personalized trust dynamic characterization of key UAVs and ordinary UAVs, ensuring that the system has structured trust adaptive capabilities in multi-level task collaboration.

[0052] To ensure the transparency, traceability, and tamper resistance of trust management, all historical behavior data of UAVs and the aforementioned trust update records are stored and consensus-based through a blockchain ledger. This not only effectively prevents trust manipulation and forgery but also provides the system with a globally verifiable and trustworthy data foundation, making the entire trust loop a robust and reliable cornerstone of system security.

[0053] To achieve a global balance between sensing accuracy and transmission security, this embodiment integrates models from the pre-communication, during-communication, and post-communication stages into a unified joint optimization problem. This problem aims to collaboratively minimize two core risks—sensing anomalies and data transmission attacks—under multiple constraints, including UAV task execution time, energy consumption, and trust-driven node selection. This objective is composed of the weighted sum of the optimized sensing anomaly risk and attack risk relative to their initial baseline values. Its complete mathematical model, the joint optimization model, is as follows: (twenty two); in, and These are the initial perceived anomaly risk and the initial attack risk, respectively, and the current perceived anomaly risk. and current attack risks Composition ratio and As two parts of the joint optimization objective of minimizing the probability of sensing-transmission risks, the parameters and These are used to characterize the weight proportions of perceived risk and transmission risk in the overall risk, respectively. Because the system operation satisfies... , Thus, the joint risk function is always in Within the interval; within the constraints 、 Key UAVs and ordinary UAV The time delay constraint ensures the effectiveness of data perception and perception result transmission; 、 Key UAVs and ordinary UAV Energy consumption constraints are in place to ensure the integrity of each UAV task execution. 、 Selecting constraints for trust-driven high-performance nodes during the first step During the second communication, only the current (number) communication is selected. Trust ranking (round) forward UAV as a key UAV Performs tasks such as perceptual data aggregation and transmission; trust value ranking. In the range They are classified as ordinary UAVs This allows for the execution of local sensing tasks. By selecting highly trusted UAVs to perform communication tasks, the accuracy of data sensing and the security of data transmission can be enhanced from the source. 。

[0054] The following challenges remain in solving the joint optimization problem in multi-UAV collaborative scenarios: 1) Insufficient data reliability in the perception phase: In complex airspace, data collected by multiple UAVs is greatly affected by their own performance, leading to biases in the perception results. Traditional methods struggle to extract reliable perception data from the perspective of UAV performance, thus reducing the accuracy of global perception.

[0055] 2) Security and privacy challenges during transmission: UAV networks exhibit significant regional heterogeneity, and direct global aggregation will weaken the personalized performance of the model. Simultaneously, sensing parameters face the risk of theft and tampering during transmission, while traditional encryption mechanisms are insufficient to meet the lightweight requirements of terminal UAVs.

[0056] 3) The challenge of accurately evaluating dynamic behavior in the feedback phase: UAVs exhibit diverse behavior patterns in different task scenarios. Traditional single evaluation indicators cannot accurately characterize their credibility in specific tasks, leading to imbalances in node selection and task allocation, which affects the overall system performance.

[0057] 4) The problem of full-process coupling in cross-stage collaboration: The three stages of perception, transmission and feedback influence and restrict each other. It is difficult to achieve global optimization by a single algorithm. It is necessary to deeply integrate trusted perception, secure transmission and trust assessment.

[0058] To address the aforementioned challenges, this embodiment proposes three algorithms: a trust-enhanced Ridge Mahalanobis distance (TERMD) algorithm to remove anomaly-aware data; a personalized federated learning (PFL-EPT) algorithm for encrypted parameter transmission to achieve privacy-preserving secure parameter aggregation; and a differentiated trust scoring (DTS) mechanism to update UAV trust values. Through the synergistic application of these three algorithms, a comprehensive security enhancement scheme covering the entire chain from data awareness to transmission scheduling is constructed.

[0059] Optimization and solution of mathematical model: Solving the joint optimization problem defined above, namely minimizing the sensing-transmission risk probability, is the objective of this embodiment. To achieve this goal, three core algorithms are designed, which run throughout the entire task execution process and form a collaborative optimization closed loop: During communication, the trust-enhanced Ridge Mahalanobis Distance (TERMD) algorithm is employed to directly minimize the risk probability of ordinary UAVs by improving the accuracy of the sensing data. Perceived abnormal risk Simultaneously, the personalized federated learning PFL-EPT algorithm, which employs parameter encryption transmission, ensures the security of the transmitted perception results, thereby directly minimizing the risk of malicious user MU attacks on the transmitted perception results. After the communication is completed, a Differentiated Trust Scoring (DTS) mechanism is used to dynamically update the trust value of the UAV after the communication is completed. This provides a precise basis for decision-making regarding task scheduling before the next round of communication, thereby optimizing system performance from the source.

[0060] These three algorithms do not operate in isolation, but rather form a closely collaborative feedback system: the performance of TERMD and PFL-EPT provides evaluation criteria for DTS, while the output of DTS, in turn, guides the weighting of TERMD and the node selection of PFL-EPT in the next round, jointly driving the continuous convergence of joint risk.

[0061] 1. Anomaly detection in perceptual data based on the TERMD algorithm: To address the challenge posed by anomalies in perceived data to the accuracy of aggregation results, this embodiment proposes the TERMD algorithm. The core innovation of this algorithm lies in the cross-domain fusion of behavioral trust models and physical perception data, enhancing the robustness and accuracy of anomaly detection through a triple collaborative mechanism: a trust-weighted mechanism is used to evaluate the performance of various ordinary UAVs within the region. When aggregating perceived data, its historical trust value is introduced as a weighting factor to enhance the influence of high-trust node data from the perspective of data value, thereby achieving dynamic suppression of malicious nodes and performance degradation nodes. The ridge correction mechanism performs ridge correction on the covariance matrix to solve the matrix ill-conditioned problem under high-dimensional small sample data from the perspective of numerical calculation, ensuring the stability of the anomaly discrimination basis. In the threshold setting, the spatiotemporal joint discrimination mechanism not only integrates the real-time perception deviation and trust score of the node, but also introduces a time smoothing term to capture the temporal continuity of perceived data, thereby achieving more accurate anomaly identification at the dynamic context level.

[0062] This design enables the TERMD algorithm to not only effectively identify outliers at the numerical level, but also to uncover deep anomalies from behavioral credibility and temporal evolution patterns, providing a reliable technical path for achieving low-perceived anomaly risk. The specific algorithm design is as follows.

[0063] Trust-weighted regional data aggregation: key UAVs Receiving ordinary UAV After uploading the perception data, the primary task is to perform high-quality aggregation of the regional data to form a reliable regional perception consensus. Traditional methods treat all data equally, making it difficult to defend against abnormal data injected by malicious nodes or low-quality data provided by nodes with degraded performance.

[0064] To address this issue, the primary innovation of the TERMD algorithm lies in the introduction of a trust-driven weighted aggregation mechanism. The core idea of ​​this mechanism is that the historical trustworthiness of a node should serve as a priori indicator of its current data reliability, and this should be reflected during the data fusion phase. Specifically, this embodiment introduces the... A regular UAV In the Trust value after the round-robin communication task is completed As a regular UAV Weighting factors for perceived data ensure that data from highly reliable nodes contributes more significantly. (Region) Key UAV The mean of the obtained area sensing data was calculated. It is expressed as follows: = (twenty three); in, For the first A regular UAV The perceived data, through trust weighting, ensures that the data of high-trust nodes in the mean calculation have a higher weight when forming regional consensus, thereby initially improving the overall accuracy and robustness at the source of data fusion. For the region Internal ordinary UAV The total number of ordinary UAVs at the boundaries of the two regions. With ordinary UAVs in the region The sum of the quantities, that is: (twenty four); To ensure a consistent sum of the means across all feature dimensions and to avoid imbalances in subsequent calculations caused by different dimensions, the above... After normalization, the mean of the updated sensing data is obtained. Represented as: = (25); in, for In the The values ​​of each dimension For realistic perception results In the The values ​​of each dimension This represents the percentage of the average value of each dimension of the perceived data within the total average value of the perceived data. This represents the sum of values ​​for each dimension of the actual perceived result.

[0065] Ridge covariance matrix and Mahalanobis distance calculation: based on normalized perceptual data mean Further calculation of the region Sample covariance matrix This is used to measure the collaborative change characteristics among perceived data in different dimensions, providing a metric basis for Mahalanobis distance calculation. Its expression is as follows: (26); However, in IoD scenarios with high-dimensional small samples or multicollinearity, the sample covariance matrix... Often exhibiting ill-conditioned or near-singular behavior, its inverse matrix becomes unstable, thus rendering Mahalanobis distance ineffective. Therefore, this embodiment further introduces the concept of ridge regression to construct a positive definite and invertible ridge-based covariance matrix. : (27); in, for 3D identity matrix The purpose of the ridge parameter is to make For all Both are positive definite and reversible.

[0066] Ridge parameters The selection of is crucial, as it controls the trade-off between the estimation's bias and variance. To find the optimal... This embodiment constructs an optimization objective aimed at minimizing the overall reconstruction error, which aims to preserve the general UAV to the greatest extent possible. While capturing the original features of the data, the ill-conditioned nature of the covariance matrix is ​​suppressed. Specifically, the square of the Frobenius norm is used as the optimization criterion, defined as follows: (28); in, Indicates the first A regular UAV Perceived data and regional mean The fitting error between the two is used to assess the degree of preservation of the original data structure; Then measure the covariance matrix To improve stability and suppress numerical instability.

[0067] After obtaining the ridge covariance matrix Then, further calculations were performed on the first... A regular UAV Mahalanobis distance Considering the same ordinary UAV In two adjacent time periods and Perception results for the same region within a given area typically exhibit temporal continuity. This embodiment introduces a temporal smoothing mechanism based on traditional Mahalanobis distance. Its specific form is as follows: (29); in, This is a time smoothing weighting coefficient used to balance the spatial consistency of the current moment with the temporal consistency of adjacent moments; and They are ordinary UAVs In time and The sensor data collected at any given moment; is the inverse of the covariance matrix.

[0068] This spatiotemporal Mahalanobis distance can not only capture instantaneous anomalies in the spatial distribution of nodes, but also effectively identify abrupt changes in their behavior over time, thereby significantly improving the detection capability against slow drifting or intermittent anomaly attacks.

[0069] Dynamic threshold setting and trusted data aggregation: This involves obtaining all common UAVs... spacetime Mahalanobis distance Subsequently, an adaptive discrimination mechanism is needed to define abnormal behavior. This embodiment abandons the fixed threshold method that requires prior knowledge and proposes a dynamic threshold setting method based on the regional data distribution characteristics. Specifically, it sets a threshold for all ordinary UAVs in this region. The Mahalanobis distance is considered as the sample set, and the region is calculated. Mahalanobis distance mean of the perceived data and standard deviation And based on this, set the threshold for anomaly detection. (Region) Mahalanobis distance threshold The definition is as follows: = + (30); in, This is the sensitivity coefficient, used to control the strictness of anomaly detection; and The specific calculation process is as follows: (31); If a regular UAV exists Mahalanobis distance Greater than this threshold, that is: when At that time, the corresponding perceived data is identified as an outlier and removed. This dynamic threshold can adapt to changes in data distribution in different regions and rounds, avoiding misjudgments or omissions caused by environmental differences.

[0070] After removing all data deemed abnormal, let the remaining number of valid sensory data be . This constitutes the region Trustworthy perception dataset Meanwhile, the region Key UAV According to region Trustworthy perception dataset Obtain the perception results that conform to this region This serves as the label for subsequent local model training in federated learning. Traditional simple averaging methods cannot withstand the data quality differences that still exist among the remaining nodes. Therefore, this embodiment proposes a dual-weight aggregation strategy that integrates trust and consistency to finely fuse trusted datasets. Its definition is as follows: (32); in, Indicates ordinary UAV The weight of perceived data is considered in conjunction with its trust value. and Mahal distance The dual benefits are defined as follows: (33); Numerator Trust Value As a positive incentive, nodes with reliable historical behavior are rewarded; the reciprocal of Mahalanobis distance. As a negative filter, nodes that deviate significantly from the group consensus are penalized, even if they are not judged as anomalous. The denominator ensures that all ownership values ​​are normalized.

[0071] This design allows nodes with high trust and low deviation to dominate the final aggregation, while the influence of nodes with low trust values ​​or large Mahalanobis distances is suppressed, thus achieving a secondary optimization of data quality.

[0072] In summary, the TERMD algorithm, through two core steps—"dynamic thresholding to eliminate dross and retain the best" and "dual-weighted selection of the best among the best"—ensures that the data input to the federated learning model possesses both high reliability and high accuracy. This directly improves the quality of region perception results, thereby effectively reducing the risk of perception anomalies in the system and laying a solid foundation for achieving the global optimization goal. The complete TERMD process is shown in Algorithm 1.

[0073] 2. Secure transmission of sensing results based on the PFL-EPT algorithm: To address the security and privacy threats faced by perception results during transmission, this embodiment proposes the PFL-EPT algorithm. This algorithm aims to collaboratively solve two core problems: balancing global consensus with local personalization needs in an IoD environment where data is not independently and identically distributed; and achieving end-to-end confidentiality and integrity of model parameter transmission under resource-constrained conditions.

[0074] The overall process of the algorithm is an encrypted iterative process between GCS and each key UAV, with its core consisting of two pillars: personalized model updates and secure parameter transmission.

[0075] 1) Personalized model updates and aggregation: In the area Perception results Afterwards, the region Key UAV The initial global model will be received from the ground control station GCS. Subsequently, the region Key UAV In local dataset Each node independently performs federated learning training tasks, generating personalized local model parameters. The training objective is to minimize the current global model. Lower region Dataset Local loss function Its definition is as follows: (34); in, Representing the global model In a single sample (ordinary UAV) (Regional sensing data) The loss value on; For global model For input samples The prediction.

[0076] The local model was then uploaded to the ground control station GCS, which... The local model parameters are aggregated using similarity weighting to construct a personalized global model. The aggregated global model is then transmitted back in an encrypted manner. In the new round of local training and global synchronization, the entire federated learning process is conducted through key UAVs. The alternating optimization iterations between ground control station GCSs have been completed.

[0077] Personalized local model updates: in the first During the local model update, the ground control station GCS will update the current personalized global model. Distribute to key UAVs Key UAV Based on region Dataset Perform a local model update. Specifically, update the region. In the Local model updated in rounds It is expressed as follows: (35); in, For the first Round global model The gradient of the loss function on local data; This is the learning rate.

[0078] To avoid "client drift" caused by simple local training on heterogeneous data, PFL-EPT further introduces a personalized fusion step. Specifically, in the first step... In the wheel, the area Key UAV Use the first Wheel Personalized Local Model and the local model A weighted fusion method is used to maintain the continuity and stability of local characteristics. Therefore, the region Key UAV In the Wheel Personalized Local Model The update process is represented as follows: = + (36); Among them, the fusion coefficient Used to control historical models Compared with the current model The proportion of this process in the personalization process effectively preserves the characteristic structure of local data, avoiding the loss of local information in traditional data fusion methods. This operation, by retaining historical model information during iteration, effectively enhances the model's adaptability to local data distribution, which is key to achieving personalization.

[0079] Personalized global model updates: regions Key UAV The encrypted personalized local model The results are uploaded to GCS for aggregation. GCS does not perform a simple averaging; instead, it employs a similarity-driven personalized aggregation strategy. First, it calculates the results for each key UAV. The model similarity between them is used to measure the individual differences of local models in the current iteration round. For regions Key UAV With the region Key UAV The similarity between their personalized local models It can be calculated using Euclidean distance, defined as follows: (37); in, For the first One sensing area In the A personalized local model for wheels.

[0080] Subsequently, to further guide the personalized aggregation of the global model, this embodiment introduces a similarity weighting mechanism, assigning it to other key UAVs. Personalized local models are closer to the key UAVs The aggregated data has higher parameter fusion weights. Specifically, a similarity weight calculation method based on the softmax normalization function is used for the region. Key UAV Assign corresponding similarity weights This weight reflects the region Key UAV Other key UAVs The degree of similarity between the personalized local model and the actual model is defined as follows: = softmax( ) = (38); Finally, the ground control station GCS performs weighted fusion of the personalized local models based on this similarity, thereby generating the first... Wheel's Personalized Global Model This method makes model updates more inclined to draw knowledge from nodes with similar learning objectives, naturally forming personalized model clusters. Specifically, the... Wheel's Personalized Global Model The update formula is as follows: (39); Repeat the above iterative process until the convergence condition is met or the preset number of training rounds is reached, at which point the process terminates to obtain the final local model. and global model This completes the federated modeling process, enabling the effective fusion of sensory data features from various regions and the secure sharing of models.

[0081] In summary, the personalized design of the PFL-EPT algorithm indirectly enhances the robustness and effectiveness of the global model by reducing model dependence on anomalous or dissimilarly distributed nodes. This not only improves the accuracy of perception tasks, but its paradigm of "replacing data flow with parameter transmission" fundamentally cuts off the path of original data leakage, providing core support for minimizing the risk of transmission attacks.

[0082] 2) Lightweight encrypted transmission mechanism: Although federated learning avoids the leakage of raw data by transmitting model parameters, the parameters themselves contain original information and may still expose privacy through reverse inference, and the transmission channel is subject to the risks of eavesdropping and tampering. Therefore, this embodiment integrates an end-to-end secure transmission protocol based on a physically unclonable function (PUF) and lightweight keys. This protocol ensures forward security and resistance to replay attacks while perfectly adapting to the resource constraints of UAVs. The specific algorithm design is as follows.

[0083] Equipment registration and PUF response generation: The ground control station GCS first registers the area... Key UAV Assign a unique identifier Ground control station GCS and area Key UAV All can utilize identity tokens And using physically non-cloning functions Unpredictable challenge-response pairs are generated, enabling the local generation and storage of key materials. The generation relationship is shown below: (40); in, In order to be with identity The bound local response value serves as the security foundation for subsequent key derivation and authentication processes. This method avoids the leakage risks of traditional key storage mechanisms and has advantages such as being lightweight, non-replicable, and resistant to physical attacks, making it suitable for resource-constrained UAV platforms.

[0084] Session key derivation: To achieve anonymity and traceability during communication, the GCS (Ground Control System) based on the aforementioned ground control station uses a region-based key derivation mechanism. Key UAV Identity identifier and The generated response value Through hash function Generation region Key UAV pseudonym This mechanism is used for anonymous authentication during this round of communication. While concealing the true identity, it ensures the traceability and timeliness of the communication process, preventing replay attacks and identity forgery. The calculation formula is as follows: (41); At this time, the ground control station GCS has Key UAV The pseudonym, the key UAV Each had its own pseudonym. Then, the ground control station GCS further... Key UAV The pseudonyms are combined and hashed to generate the session key value for this round of communication. This serves as the foundation for lightweight key derivation and perceptual data protection. The specific calculation is as follows: (42); For critical UAVs If the generated pseudonym is correct, it possesses the same key as the ground control station GCS, which is used for subsequent encryption and decryption processes. If an incorrect pseudonym is generated, it is determined to be a malicious UAV, and its key is set to 0. When an incorrect key is possessed, the ground control station GCS will not decrypt the transmitted content. This embodiment will use a regional... Key UAV key The generated representation is as follows: (43); in, For all critical UAVs The logic of pseudonyms, or, if honesty is key, then UAV. One of them can have the same key as the ground control station GCS.

[0085] Encrypted transmission and decryption in local models: in critical UAVs After configuring the communication key with the ground control station GCS, the transmission of all model parameters is protected by confidentiality and integrity. To enhance the security and anti-attack capabilities of transmitted data, the area... Key UAV The personalized local model parameters obtained in this round of training XOR the encrypted data with the remaining related content to form ciphertext, and then use this ciphertext to obtain the region. Key UAV Transmission data to the ground control station GCS It is represented as: (44); in, For the first The dynamic random number perturbation term increments automatically in each round, i.e.: This ensures that the encrypted content of each round of communication is not repeated, effectively preventing security risks caused by key reuse; For the region Key UAV The first Round timestamp, used for GCS inspection area Key UAV Whether the generated local model parameters were produced within the valid time period. The introduction of timestamps and random numbers effectively resists replay attacks and key reuse attacks.

[0086] After the ciphertext is generated, the area Key UAV Transmit data and timestamp The information is uploaded to the ground control station GCS. After receiving the information, the ground control station GCS first generates the first... Round timestamp Detection | - Whether it is within the valid time frame to defend against replay attacks and delay attacks. After successful verification, the ground control station GCS uses the same key and random number to perform inverse calculations to restore the original model parameters. : (45); Encryption and decryption of the global model: for the recovered key UAVs The original model parameters are aggregated by the ground control station's GCS (Global Control System) into a personalized global model, and the global model is then encrypted using the same encryption mechanism as the uplink transmission. Specifically, the ground control station's GCS will aggregate the personalized global model parameters... Key Random disturbance and timestamp Perform an XOR operation to generate the ground control station GCS, which is then transmitted to the critical UAV. Transmitted content It is represented as follows: (46); Next, the ground control station GCS needs to process the encrypted content. and timestamp Transmitted to the area Key UAV Subsequently, the key UAV Generate the first The time stamp of the start of the round And verify | Whether it's within the valid timeframe, then use the key synchronized in this round. With random numbers By performing the inverse operation, the personalized global model parameters were successfully decrypted. : (47); In summary, this lightweight encryption transmission mechanism utilizes physically unclonable functions. This implements the hardware's security cornerstone, providing anonymity through temporary pseudonyms and session keys. timestamp and random numbers The combined use of these technologies constructs a comprehensive security protection system covering authentication, confidentiality, integrity, and freshness. This mechanism ensures the security of model parameters during transmission with extremely low computational and communication overhead, directly and effectively minimizing the risk of transmission attacks in the joint optimization objective. The complete PFL-EPT process is shown in Algorithm 2.

[0087] 3. Trust closed-loop update based on DTS mechanism After a communication task is completed, the system performs a trust assessment based on the performance of each UAV in the task, providing an accurate and fair basis for decision-making in the next round of task scheduling and role allocation. To achieve this goal, this embodiment proposes a DTS (Distrusted Trust Assessment) mechanism. The core idea of ​​this mechanism is that the trust assessment criteria must be closely aligned with the role responsibilities of the nodes. Therefore, this embodiment provides a framework for ordinary UAVs undertaking different tasks. and key UAV Different accuracy, reliability, and energy efficiency assessment models with different focuses were designed.

[0088] 1) Accuracy assessment: In the process of UAVs performing tasks, accuracy assessment aims to quantify the consistency between the UAV's output and the actual situation, and is the core of measuring the quality of its task completion.

[0089] Accuracy of standard UAV: ​​Standard UAV As a data producer, the accuracy assessment of a node primarily depends on the quality of the raw sensing data provided. This embodiment uses the Mahalanobis distance obtained through the TERMD algorithm to evaluate the quality of the node's sensing data. Specifically, the larger the Mahalanobis distance, the more severely the node's sensing data deviates from the overall distribution, and the lower its accuracy. If the... A regular UAV Mahalanobis distance Exceeding the critical value If the perceived data is abnormal, its accuracy is set to 0. If the Mahalanobis distance is within the threshold, to reasonably reflect the impact of different degrees of deviation on accuracy, this embodiment uses an exponential decay function to model the perceived accuracy. Specifically, the first... After the first communication task is completed, the A regular UAV accuracy The calculation is expressed as follows: (48); Key UAV accuracy: Key UAV As a regional manager, their core competency lies in the quality of their comprehensive information assessment and decision-making. The evaluation of their accuracy should move beyond the precision of their own sensor data and instead focus on the reliability of their data fusion and decision-making. The accuracy after the completion of this communication task The calculation formula is as follows: (49); in, For its initial area perception results The final region perception conclusions after being integrated with federated learning Differences between Divergence, as a criterion for determining the region Key UAV Accuracy is a negative correlation metric, measuring the deviation between initial and final conclusions. The exponential function design reflects the non-linear decay relationship between difference and accuracy. When smaller, A value close to 1 indicates that the aggregation result of the key UAVs is highly consistent with the initial baseline, indicating high perception accuracy; while when... When it increases, A rapid decay to near zero indicates a significant increase in the degree to which the key UAV deviates from the true results during data aggregation. Specifically, the calculation formula is as follows: (50); in, and The initial region perception conclusions are respectively With the final area perception conclusion In the Perceptual data in multiple dimensions .

[0090] 2) Reliability assessment Reliability aims to quantify the stability and behavioral consistency of a UAV in a series of missions and is a key indicator for assessing its long-term trustworthiness.

[0091] Standard UAV Reliability: The reliability of a standard UAV considers its role suitability in the current task and the stability of its historical behavior. Specifically, the reliability of the UAV will be... After the first communication task is completed, the A regular UAV reliability It is expressed as follows: (51); in, They are ordinary UAVs Average matching degree and historical reliability The weight.

[0092] ordinary UAV Location Average regional matching degree It reflects its collaborative sensing capability with other ordinary UAVs in the area and is used to measure its overall matching degree in regional tasks, as shown below: (52); in, For the first A regular UAV The optimal region matching degree is given by formula (6); For the region The number of ordinary UAVs.

[0093] Historical reliability The design differs from indicators reflecting single-round task performance, focusing more on the behavioral stability and trustworthiness evolution characteristics of nodes over long-term operation. To achieve a more timely assessment that gives greater weight to recent performance, this embodiment introduces an exponentially weighted moving average model to calculate the historical trust performance of nodes. This model dynamically captures the trustworthiness trend of nodes by weighting and averaging past trust values. The specific calculation is as follows: (53); in, This is the current communication round; For UAV in the The trust value updated after the completion of the next communication task; The time weight corresponding to this historical trust value. This is the normalization coefficient, used to ensure a reasonable overall weight distribution.

[0094] Critical UAV Reliability: The reliability of critical UAVs focuses on the continuity of management responsibilities. As a regional manager, critical UAVs need to maintain stable trust performance and execution capabilities across multiple rounds of tasks; therefore, their reliability is directly determined by their historical trust performance. Specifically, the reliability of critical UAVs will be determined by their historical trust performance. After the second communication task is completed, the area Key UAV reliability It is expressed as follows: (54); This design emphasizes that key roles require a consistently high level of historical reliability, and any performance fluctuations or abnormal behavior should be reflected directly and quickly in their reliability.

[0095] 3) Energy efficiency assessment: Energy efficiency assessment aims to quantify the resource utilization efficiency of UAVs during mission execution and is a key indicator for measuring their mission sustainability and overall effectiveness. In resource-constrained IoD systems, energy-efficient nodes mean longer endurance and more stable service quality, which is crucial for the long-term robustness of the system. Therefore, this embodiment uses energy efficiency as one of the core indicators of trust assessment, and the assessment result is reflected as a normalized energy efficiency score, which is negatively correlated with energy consumption.

[0096] Energy efficiency of ordinary UAVs: The energy efficiency of ordinary UAVs mainly reflects the rationality of energy utilization and task execution efficiency in completing sensing and communication tasks. As data producers, ordinary UAVs need to balance sensing accuracy and communication quality under limited energy constraints. Therefore, their energy efficiency level not only reflects the total energy consumption but also the effective sensing results contributed per unit of energy consumption. For this reason, this embodiment defines ordinary UAVs from the perspective of energy use efficiency. energy efficiency Specifically, it is expressed as follows: (55); in, This represents the upper limit of energy consumption for a typical UAV under maximum workload, used for energy consumption normalization. For the first After the next communication task is completed, the UAV The total energy consumption is modeled in the same way as formula (16), that is: This expression reflects the general UAV The energy consumption structure characteristics of maintaining flight, information collection and data transmission are as follows, and the specific definitions and calculation methods of energy consumption in each stage are as follows.

[0097] Maintaining a stable flight attitude is a common feature of UAVs during mission execution. One of its basic functions, and its corresponding hovering energy consumption This constitutes an important component of the system's total energy consumption, as shown below: (56); in, Indicates ordinary UAV Power consumption during hovering To complete the first The total time required for this communication task is modeled in the same way as formula (13), that is: Data perception time Depends on UAV Size of perceived data for the region and its sensing rate Data transmission time The choice is based on the size of the perceived data. and transmission rate The decision is made jointly. The specific definitions of both are as follows: (57); in, Indicates the sensor's sampling frequency; The number of data bits generated in each sampling; For communication channel bandwidth; For transmission power; Channel gain; This represents noise power.

[0098] During the perception phase, ordinary UAVs The energy consumption for collecting information on land cover categories is determined by the sensing power. and perception of time Decision. Specifically, ordinary UAVs will be... Perceived energy consumption It is expressed as follows: (58); During the transmission phase, ordinary UAV Transmitting sensing data to key UAVs within the area, the energy consumption of this communication process is directly proportional to the amount of data and the energy consumption per unit of transmission. Specifically, this involves ordinary UAVs... The transmission energy consumed The definition is as follows: (59); in, To transmit the total amount of sensing data; The energy required to transmit each bit of data.

[0099] Key UAV Energy Efficiency: Key UAV The complexity and functional diversity of the tasks performed are significantly higher than those of ordinary UAVs. Therefore, its energy consumption structure is more complex. To measure the key UAV... To improve energy utilization efficiency, this embodiment introduces an energy efficiency function. The definition is as follows: (60); in, This represents the maximum energy consumption limit for critical UAVs during task execution. For the first After the next communication task is completed, the key UAV The total energy consumption is consistent with the modeling of formula (21), that is: This formula reflects the energy consumption sources of the key UAV in three main stages: flight hovering, data processing, and data transmission. The specific definitions and calculation methods of each component's energy consumption are as follows.

[0100] During mission execution, critical UAV Hovering energy consumption required to maintain stable flight attitude It is expressed as follows: (61); in, Indicates key UAV Power consumption during hovering; For its in the The total execution time of this communication task is modeled in the same way as formula (17), that is: Calculation time Depends on the size of the aggregated sensing data With processing rate Data transmission time The choice is based on the size of the transmitted data. and transmission rate The decision is made jointly. The specific definitions of both are as follows: (62); in, Indicates key UAV The processor's clock frequency; This represents the average number of computation cycles required per bit of data. for The transmission power.

[0101] During the data processing phase, key UAVs Data processing energy consumption The energy consumed in calculating Mahalanobis distance during outlier identification mainly comes from two parts: And the energy consumption of local model updates in collaboration with the ground control station GCS to complete the federated learning process. Specifically, it is expressed as follows: (63); First, the critical UAV uses the TERMD algorithm to remove anomalies from the sensing data of ordinary UAVs in the region. This process involves mean... Covariance matrix The calculation of, etc. Therefore, the energy consumption of anomaly detection data removal. It is expressed as follows: (64); in, , , , , and These correspond to the calculation of mean, covariance matrix, parameters, ridge covariance matrix, and various ordinary UAVs in removing outliers. The energy consumption required for the Mahalanobis distance and its threshold is modeled in this embodiment using the product of the number of floating-point operations and the energy consumption per operation. Specifically, it is represented as follows: (65); in, For ordinary UAVs to be identified The amount of perceived data; For feature dimensions; These represent the energy consumption of a single floating-point addition, division, and multiplication operation, respectively.

[0102] Subsequently, the key UAV The PFL-EPT algorithm is used to update the local parameters. This process is accumulated from the local model training energy consumption over multiple iterations, as shown below: (66); in, This indicates the number of rounds in which the model parameters are updated during the federated learning process; Size of the aggregated data; To train each The energy required to generate the data.

[0103] During the transmission phase, the key UAV The sensing results are transmitted to the ground control station GCS. Specifically, the transmission energy consumption... The definition is as follows: (67); in, The amount of sensing result data to be transmitted. The energy required to transmit each bit of data.

[0104] The complete DTS mechanism process is shown in Algorithm 3. In summary, the refined energy consumption model established in this embodiment not only provides a key energy efficiency dimension for trust assessment but also profoundly reveals the computational overhead of different algorithms (such as TERMD and PFL-EPT). This enables the DTS mechanism to effectively identify and incentivize high-quality nodes that can complete tasks with high quality while maintaining low energy consumption and long endurance, thereby significantly improving the long-term operational performance and stability of the entire IoD system.

[0105] This embodiment systematically studies the dual challenges posed by perceptual data anomalies and transmission security vulnerabilities in IoD systems. To address this complex problem, a joint optimization model aimed at minimizing both perceptual anomaly and attack risks is constructed, and based on this, a secure and reliable collaborative framework spanning the entire lifecycle of UAV missions is proposed. The core of this framework lies in its closed-loop design spanning the pre-, during-, and post-mission stages: In the pre-communication stage, dynamic partitioning of the perceptual region is achieved through region-based perceptual complexity modeling and multi-agent game optimization. Dynamic task scheduling is then performed based on the matching degree between each UAV and the perceptual region, as well as the performance matching degree between each UAV and the region's UAVs, achieving complementary collaboration and optimal deployment of heterogeneous resources. In the during-communication stage, the TERMD algorithm significantly improves perceptual accuracy from the data source through a collaborative mechanism of trust enhancement and ridge correction; the PFL-EPT algorithm, through personalized federated aggregation and lightweight encrypted transmission, constructs a robust privacy protection barrier while ensuring model performance. In the post-communication stage, the DTS mechanism, through role-driven differentiated evaluation, achieves accurate characterization of UAV node behavior and closed-loop updates of trust levels, providing a dynamic and reliable decision-making basis for continuous system optimization.

[0106] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for secure and trusted cooperative perception for UAV Internet of Things, characterized in that, include: Step 1: Based on the pre-built UAV Internet security and trustworthy collaborative framework, and taking into account latency, energy consumption and trust scheduling constraints, construct a joint optimization model for perception anomaly and transmission security issues; Step 2: Divide the perception area based on a balance between area and complexity, assign roles according to the trust value of drones, and complete the collaborative cluster deployment through complementary capabilities game theory. Step 3: Based on the deployed drone cluster collecting perception data, and according to the joint optimization model, use the trust-enhanced Ridge Mahalanobis distance TERMD algorithm to detect anomalies in the perception data. Step 4: Based on the anomaly detection results of the sensing data, use the PFL-EPT algorithm to securely transmit the sensing results; Step 5: Update the UAV trust value through the DTS mechanism based on the task execution results, and feed the trust value back to Step 2 to form a trusted closed-loop iteration to continuously achieve the joint optimization goal.

2. The method of claim 1, wherein, The pre-built UAV Internet security and trustworthy collaborative framework includes: an IoD system collaborative perception and communication scenario and a full-process security and trustworthy collaborative framework; wherein, the IoD system collaborative perception and communication scenario is based on the ground control station GCS, a heterogeneous UAV cluster, and a blockchain network; the IoD system collaborative perception and communication scenario includes a three-layer structure: the blockchain layer is responsible for storing and updating UAV trust values, the flight layer is where the heterogeneous UAV cluster performs perception tasks, and the ground layer integrates data through the ground control station GCS; The end-to-end secure and reliable collaborative framework includes: pre-communication perception area division and UAV task scheduling; during communication, perception anomaly detection and secure transmission of perception results; and trust closed-loop update after communication ends. 3.The UAV-Internet oriented secure and trusted collaborative perception method according to claim 1, characterized in that, The joint optimization model is as follows: in, and These are the initial perceived anomaly risk and the initial attack risk, respectively. and These represent the current perceived anomaly risk and the current attack risk, respectively, with parameters... and These are used to characterize the weight ratio of perceived risk and transmitted risk in the overall risk, respectively, and are subject to the constraints. 、 Key UAVs and ordinary UAV Delay constraints, 、 Key UAVs and ordinary UAV Energy consumption constraints 、 Select constraints for trust-driven high-performance nodes. For the region Key UAV, For the first A regular UAV, For key UAVs Regional aggregated data, For key UAV The results of the transmission of perception, For the first After the first communication task is completed, the A regular UAV Total task execution delay For ordinary UAVs allowed by the system The longest time limit for conducting area perception. For the first After the second communication task is completed, the area Key UAV Total task execution delay, For the critical UAVs allowed by the system The maximum time limit for aggregating regional sensing data and transmitting sensing results. For the first After the first communication task is completed, the A regular UAV Total energy consumption For ordinary UAVs allowed by the system Maximum energy consumption for area sensing. For the first After the second communication task is completed, the area Key UAV Total energy consumption for task execution For the critical UAVs allowed by the system The maximum energy consumption for aggregating sensing data and transmitting sensing results. In the first After the next communication task is completed, the area Key UAV Trust ranking In the first After the communication task is completed, the ordinary UAV Trust ranking To divide the area into numbers, The number of UAVs participating in the sensing.

4. The secure and reliable collaborative perception method for UAV Internet according to claim 1, characterized in that, The perception region is divided based on a balance between region area and region complexity, including: Adopting the first One sensing area area Entropy of Land Feature Categories The proposed regional complexity modeling method integrates two dimensions: regional area and land feature category complexity. It comprehensively characterizes the perceptual difficulty of a region and divides the region into sub-regions with the closest complexity.

5. The secure and reliable collaborative perception method for UAV Internet according to claim 3, characterized in that, The deployment of collaborative clusters is achieved through hierarchical role allocation based on drone trust values ​​and through complementary capabilities, including: First, select the one with the highest trust value. UAVs are used as key UAVs in each sub-region. According to the region With the region The complexity and requirements of edge perception determine the characteristics of ordinary UAVs at the boundary of the region. Number From the remaining Select the UAV with the second highest trust value The frame serves as a regular UAV at the regional border. ,in The rest The UAV will serve as a regular UAV within the area. ; For ordinary UAVs in the region Adaptive deployment is used to construct a region selection mechanism based on game theory strategies; in the region selection mechanism, firstly from... shelf Among them, those ranked higher in trust were selected. A regular UAV The system selects a sensing area based on its compatibility with the sensing area, ensuring that each area has an initial ordinary UAV. Then, in the remaining A regular UAV In the middle, the one with the second highest trust value is selected again. The drone performs area selection; at this time, there are two ordinary UAVs in each area. This process is iterated round by round until all ordinary UAVs are included. Region selection and deployment completed.

6. The secure and reliable collaborative perception method for UAV Internet according to claim 3, characterized in that, The TERMD algorithm, which employs trust-enhanced ridge Mahalanobis distance, is used for anomaly detection in perceptual data, including: Introducing a trust-driven weighted aggregation mechanism: This mechanism is used to aggregate data from various ordinary UAVs within a given region. When aggregating perception data, the first... A regular UAV In the Trust value after the round-robin communication task is completed As a regular UAV Weighting factors for perceived data; Ridge Covariance Matrix and Mahalanobis Distance Calculation: Based on the normalized mean of perceptual data, the concept of ridge regression is introduced to construct a positive definite and invertible ridge covariance matrix. The distance is then calculated based on this positive definite and invertible ridge covariance matrix. A regular UAV Mahalanobis distance; Dynamic threshold setting and trusted data aggregation: This involves obtaining all common UAVs... After calculating the spacetime Mahalanobis distance, all ordinary UAVs in this region will be included. The Mahalanobis distance is considered as the sample set, and the region is calculated. Mahalanobis distance mean of the perceived data and standard deviation And based on this, anomaly detection thresholds are set.

7. The secure and reliable collaborative perception method for UAV Internet according to claim 6, characterized in that, The ridge-modified covariance matrix is: in, For the region The sample covariance matrix, Positive definite invertible ridge-transformed covariance matrix for 3D identity matrix The purpose of the ridge parameter is to make For all Both are positive definite and reversible; Specifically, an optimization objective is constructed to minimize the overall reconstruction error, and the optimal ridge parameters are sought. The optimization objective aims to preserve the functionality of ordinary UAVs to the greatest extent possible. While capturing the original sensory data features, it suppresses the ill-conditioned nature of the covariance matrix.

8. The secure and reliable collaborative perception method for UAV Internet according to claim 1, characterized in that, Secure transmission of sensing results using the PFL-EPT algorithm includes: Personalized model updates and aggregation: in the obtained region Perception results Afterwards, the region Key UAV The initial global model will be received from the ground control station GCS. ,area Key UAV In local dataset Each node independently performs federated learning training tasks, generating personalized local model parameters. The training objective is to minimize the current global model. Lower region Dataset Local loss function ; The local model is uploaded to the ground control station GCS, and the ground control station GCS... The local model parameters are aggregated using similarity weighting to construct a personalized global model. The aggregated global model is then transmitted back in an encrypted manner. In the new round of local training and global synchronization, the entire federated learning process is conducted through key UAVs. The alternating optimization iterations between GCSs of ground control stations have been completed; Lightweight encrypted transmission mechanism: It integrates an end-to-end secure transmission protocol based on a physically non-clonable function and a lightweight key. This protocol ensures forward security and resistance to replay attacks while adapting to the resource constraints of UAVs.

9. The secure and reliable collaborative perception method for UAV Internet according to claim 8, characterized in that, The lightweight encrypted transmission mechanism includes: Equipment registration and PUF response generation: The ground control station GCS first registers the area... Key UAV Assign a unique identifier Ground control station GCS and area Key UAV All utilize identity verification and physically non-cloning functions Generate unpredictable challenge-response pairs, i.e., response values. This enables the local generation and storage of key materials. Session key derivation: based on response value Through hash function Generation region Key UAV pseudonym This is used for anonymous authentication during this round of communication. Ground control station GCS further Key UAV The pseudonyms are combined and hashed to generate the session key value for this round of communication. ; For critical UAVs If the pseudonym is generated correctly, it has the same key as the ground control station GCS, which is used for subsequent encryption and decryption processes. If the pseudonym is generated incorrectly, it is judged to be a malicious UAV, and its key is set to 0. When it has an incorrect key, the ground control station GCS will not decrypt the content transmitted by it. Encrypted transmission and decryption of local models: region Key UAV The personalized local model parameters obtained in this round of training XOR the encrypted data with the remaining related content to form ciphertext, and then use this ciphertext to obtain the region. Key UAV Transmission data to the ground control station GCS ; After the ciphertext is generated, the area Key UAV Transmit data and timestamp The information is uploaded to the ground control station GCS. After receiving the information, the ground control station GCS first generates the first... Round timestamp Detection | - If the verification is successful within the valid timeframe, the ground control station GCS uses the same key and random number to perform the inverse operation to restore the original model parameters. ; Encryption and decryption of the global model: for the recovered key UAVs The original model parameters are aggregated by the ground control station GCS into a personalized global model, and the global model is encrypted using the same encryption mechanism as the uplink transmission.

10. The secure and reliable collaborative perception method for UAV Internet according to claim 1, characterized in that, The DTS mechanism includes: Accuracy assessment is used to quantify the consistency between UAV output results and reality; Reliability assessment is used to quantify the stability and behavioral consistency of UAVs in consecutive multi-round tasks; Energy efficiency assessment is used to quantify the resource utilization efficiency of UAVs during task execution.

Images