An integrated framework and application of elastic federated learning and trust-aware task scheduling for heterogeneous UAV swarms

Abstract

This invention relates to an integrated framework for resilient federated learning and trust-aware task scheduling in heterogeneous UAV swarms. This framework aims to address the security challenges in collaborative learning among UAV swarms in adversarial environments, particularly the performance degradation caused by Byzantine attacks (i.e., malicious agents injecting malicious model updates). The core of this invention is the integration of multi-metric anomaly detection, dynamic trust management, and security-aware task scheduling to achieve joint optimization of learning robustness and operational efficiency. The technical solution encompasses three main layers: a trust-robust aggregation mechanism at the learning layer, a security-aware task scheduling algorithm at the operational layer, and cross-layer adaptive communication and computation optimization.

Description

Technical Field

[0001] This invention relates to the fields of edge computing, distributed machine learning and unmanned system cooperative control technology, specifically to an integrated framework and application of elastic federated learning and trust-aware task scheduling for heterogeneous unmanned aerial vehicle swarms. Background Technology

[0002] The proliferation of unmanned aerial vehicle (UAV) swarms in low-altitude urban airspace has fundamentally transformed the landscape of smart city infrastructure and services. Modern UAV swarms, equipped with advanced sensors and computing capabilities, serve as aerial data collection platforms and are widely used in areas such as traffic monitoring, emergency response, infrastructure inspection, and delivery services. As these autonomous systems become increasingly integrated into critical urban operations, the need for continuous learning and adaptation to dynamic environments becomes crucial. However, traditional centralized machine learning methods face significant limitations when applied to UAV swarms, including bandwidth constraints, privacy concerns, and the inherent latency of transmitting raw sensor data to cloud servers.

[0003] Federated learning, as a distributed machine learning paradigm, allows individual agents in a distributed UAV network to collaboratively train a model without sharing raw data, thus naturally aligning with the operational characteristics of UAV swarms. The application of federated learning in edge networks faces challenges such as device heterogeneity and resource constraints. However, existing research typically assumes benign actors and fails to consider adversarial scenarios in open edge networks. Furthermore, the operating environment of UAV swarms is physically accessible and vulnerable to attack vectors such as GPS spoofing, communication interference, and firmware manipulation. Among these, Byzantine attacks (i.e., malicious agents injecting carefully crafted malicious updates to degrade model performance) pose a serious risk to collaborative learning systems.

[0004] In practical applications of federated learning, existing technologies mainly revolve around Byzantine-robust aggregation methods, trust management systems, and task scheduling algorithms. Byzantine-robust aggregation methods such as Krum (selective update based on geometric distance), trimmed mean (coordinate-trimmed mean), and DETOX (combining filtering and damping mechanisms) provide theoretical guarantees, but they assume homogeneous clients and identically distributed (IID) data. These methods are less effective under non-IID data distributions and do not consider the heterogeneous resource constraints and operational tasks of UAV groups.

[0005] Trust management systems have been studied in distributed systems, such as EigenTrust (which calculates global trust values ​​based on eigenvector centrality) and CONFIDANT (which combines direct observation and indirect reputation information). However, these systems mainly focus on network layer behavior (such as communication security) and do not address learning layer attack detection. In UAV systems, trust mechanisms are mostly concentrated on network layer security, such as blockchain-based reputation systems or norm-based intrusion detection, lacking adaptation to collaborative learning scenarios.

[0006] In terms of task scheduling, existing methods mostly optimize traditional metrics such as energy efficiency, coverage, or collision avoidance, employing integer programming, auction mechanisms, or heuristic algorithms. Security considerations typically focus on communication security and trajectory privacy, such as anti-interference task offloading or GPS spoofing countermeasures, but trust scores or learning objectives are not integrated into the scheduling process. Existing frameworks often separate model training from operational tasks, neglecting the coupling between the two.

[0007] The existing technology has the following problems: 1. The existing Byzantine-robust aggregation method has poor adaptability to heterogeneous and non-IID data: Existing methods such as Krum and trimmed mean assume homogeneous clients and IID data. However, in real-world deployments, UAV swarms exhibit heterogeneous resources (e.g., computing power, battery capacity) and non-IID data distribution (due to each UAV observing different geographical areas). This leads to performance degradation in real-world scenarios, an inability to effectively filter adaptive Byzantine attacks, and slow model convergence or reduced accuracy. The shortcomings stem from invalid statistical assumptions, resulting in structural complexity and poor performance.

[0008] 2. Existing methods do not integrate operational constraints and task scheduling: The existing Byzantine-robust framework ignores the resource constraints (such as energy consumption and computational overhead) and task allocation requirements of UAV swarms. Task scheduling methods optimize traditional metrics (such as energy efficiency) without considering learning objectives or security factors, potentially leading to high-risk tasks being assigned to malicious agents, causing task failures and data integrity risks. This results in cumbersome processes and high costs, as additional mechanisms are needed to coordinate learning and operation.

[0009] 3. The trust management system cannot detect learning layer attacks: Existing trust systems (such as EigenTrust) focus on network layer behavior monitoring, but Byzantine attacks occur at the model update layer, where malicious agents may strategically alternate between honest and malicious behavior to evade detection. Static reputation systems cannot capture temporal dynamics, and intermittent connections exacerbate the difficulty of maintaining trust information. This results in poor performance and an inability to provide cross-layer security.

[0010] 4. Task scheduling methods ignore trust and learning objectives: Existing scheduling algorithms do not incorporate trust scores or learning objectives into their optimization, thus failing to achieve security-aware task allocation. For example, high-priority tasks may be assigned to low-trust agents, increasing the risk of task failure. Furthermore, resource allocation during model training impacts task performance, but existing methods do not jointly optimize this, resulting in structural complexity and inefficiency.

[0011] 5. The existing framework lacks overall integration: The independent operation of the learning, trust, and operational layers leads to security vulnerabilities and wasted resources. For example, Byzantine detection is disconnected from task scheduling, making it unable to dynamically adapt to threats; the resource configuration and learning parameter adjustments of heterogeneous UAVs are not coordinated, increasing costs and processing overhead. Summary of the Invention

[0012] To address the aforementioned deficiencies, the present invention aims to solve the following technical problems: 1. Address the issue that the Byzantine-robust method is not applicable to non-IID data in heterogeneous UAV clusters: Existing byzantine-robust aggregation methods (such as Krum and trimmed mean) assume homogeneous clients and identically distributed (IID) data. However, in actual UAV cluster deployments, each UAV observes different geographical areas, resulting in highly heterogeneous data distribution (non-IID). Specific drawbacks include: gradient statistical estimation is contaminated by outliers, making it impossible to distinguish between malicious updates and legitimate non-IID mutations; directional attacks (such as sign-flipping) are more difficult to detect with non-IID data, leading to decreased model accuracy.

[0013] To address this issue, this invention proposes a multi-index anomaly detection mechanism that combines gradient norm analysis, statistical residual checking, and orientation alignment evaluation to achieve high-precision Byzantine detection. The innovative design incorporates complementary multi-indexes to overcome the blind spots of single-index detection in non-IID scenarios; adaptive threshold adjustment dynamically optimizes detection sensitivity based on real-time data distribution.

[0014] 2. Achieve integrated optimization of learning objectives and operational tasks: The existing framework decouples federated learning from task scheduling, leading to wasted resources and the risk of task failure. Furthermore, malicious agents may be assigned to critical tasks, jeopardizing data integrity. The root of the problem lies in the scheduling algorithm's focus on optimizing traditional metrics (such as energy efficiency) without considering learning objectives (such as model convergence) and security constraints (such as trust scores).

[0015] Therefore, this invention innovatively designs a security-aware task scheduling algorithm that jointly optimizes learning objectives and operational tasks, achieving dynamic task allocation through multi-objective integer linear programming. It is the first to integrate trust scores and threat indicators into task scheduling, achieving cross-layer security optimization; simultaneously, it sets up a dynamic adaptation mechanism to adjust task allocation based on real-time trust updates.

[0016] 3. Provides cross-layer dynamic trust management: Existing trust systems (such as EigenTrust) focus on network layer behavior and cannot detect Byzantine attacks at the learning layer. Malicious agents may strategically alternate between honest and malicious behavior (such as intermittent attacks) to evade static reputation systems. Furthermore, intermittent connectivity in UAV networks delays the propagation of trust information, leading to detection lag.

[0017] Therefore, this invention innovatively introduces a dynamic trust management system, which updates trust scores based on long-term behavioral profiles to achieve fine-grained proxy discrimination. Specifically, a trust evolution model is designed, and behavior monitoring is performed using multi-indicator detection results: rewarding positive behavior, punishing malicious behavior, and allowing neutral behavior to be reversed. Finally, the trust score is used for both aggregated weighting and task scheduling, forming a closed-loop feedback loop and enabling cross-layer information sharing.

[0018] 4. Optimize task allocation to improve security and efficiency: Because task allocation ignores learning resource constraints, UAVs compete for resources between training and tasks. To address this issue, this invention develops an adaptive communication and computation mechanism that dynamically adjusts local training parameters based on threat level. Trust thresholds and threat levels are incorporated into the scheduling cost function, ensuring that high-trust agents are prioritized for high-priority tasks, thus reducing the risk of task failure.

[0019] 5. Construct an overall framework that couples the learning, trust, and operation layers: Existing methods suffer from security vulnerabilities and response delays due to inter-layer isolation (such as learning, trust, and operational independence). To address this issue, this invention proposes a holistic framework that achieves closed-loop defense by synchronously integrating threat assessment, trust updates, task scheduling, and aggregation. This closed-loop defense system enables Byzantine attack adaptation and adaptive resource configuration, reducing structural complexity and improving system robustness.

[0020] The technical solution adopted in this invention is an integrated framework for heterogeneous UAV swarm elastic federated learning and trust-aware task scheduling, comprising: a Byzantine fault-tolerant aggregation module, used to receive local model updates uploaded by each UAV, filter malicious updates based on a multi-index anomaly detection mechanism, and perform weighted aggregation based on the dynamic trust scores of each UAV to generate a global model update; a dynamic trust management module, connected to the Byzantine fault-tolerant aggregation module, used to update the trust scores of each UAV using projection gradient dynamics rules based on the results of the multi-index anomaly detection mechanism and the historical behavior of each UAV; a security-aware task scheduling module, connected to the dynamic trust management module, used to allocate suitable UAVs to tasks based on the attributes of the task to be executed, the resource status of each UAV, the dynamic trust score, and the threat level, by optimizing a cost function that includes trust deficit and threat level; and an adaptive communication and computation optimization module, connected to the security-aware task scheduling module and the Byzantine fault-tolerant aggregation module, used to dynamically adjust the local training parameters and communication strategies of each UAV based on the threat level faced by each UAV; wherein, the framework couples and optimizes model training, trust evaluation, and physical task scheduling through a synchronous closed-loop process.

[0021] Preferably, the multi-index anomaly detection mechanism updates the model for each received parameter. The triple complementary detection process is as follows: (1) Gradient norm analysis: Calculate the L2 norm of the updated gradient. The mean and standard deviation were estimated using robust pruning statistical methods, and the highest and lowest values ​​were excluded. The norm of the proportion is used to calculate the pruning mean. and pruning standard deviation , in, ; , An acceptable update set is defined as:

[0022] Among them, pruning ratio It needs to be greater than the Byzantine agency ratio. ,generally .

[0023] (2) Statistical residual check layer: Calculate residuals based on median absolute deviation (MAD). For each update, calculate the residual distance between it and the median coordinate.

[0024] The median is calculated along the coordinate dimensions, and MAD is defined as follows:

[0025] Identify updates that deviate from the distribution center trend, capture additive noise attacks or systematic biases, and accept the following update set: ; Among them, MAD is used as a robust scaling estimate. (3) Orientation alignment evaluation layer, calculates update and preliminary aggregation results. The cosine similarity, with a similarity threshold set to =0.85, Acceptable update sets are:

[0026] Ultimately, the updated set is the intersection of the three: .

[0027] Preferably, the dynamic trust management module updates the trust score using the following projection gradient dynamics rule. : in, The operator that projects the trust score onto the interval [0,1], with an update value. The determination rule is as follows: If a local update is marked as malicious, then Among them, the penalty rate If the local update is accepted and the contribution is positive, then Reward rate If it is a neutral behavior, then .

[0028] Preferably, the security-aware task scheduling module models the task allocation problem as an integer linear programming problem. , in, The total cost is the comprehensive cost of assigning drone i to perform task k, and the comprehensive cost function is:

[0029] in, For task urgency, a parameter taking values ​​in the range [0, 1] represents the urgency of task k; For energy efficiency; A trust deficit; Threat level; These are the weighting coefficients.

[0030] Preferably, the adaptive communication and computation optimization module adjusts the response based on the threat level of the unmanned aerial vehicle (UAV). The following adaptive adjustments will be made: High threat ( Training rounds Quantization bit depth Inject differential privacy noise, standard deviation , Low threat ( Training rounds Quantization bit depth No noise.

[0031] Preferably, the adaptive communication and computation optimization module is further used to update the gradient before the UAV uploads the update. Perform uniform quantization: Where q is the number of quantization bits. Top-k sparsity processing for gradient transmission only: in, For binary masks, k can be set to 10%-20% of the gradient dimension.

[0032] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the functions of the modules of an integrated framework.

[0033] The beneficial effects of this invention are: 1. Innovation and advantages of the multi-index anomaly detection mechanism, (1) Significantly improved detection accuracy (see Table 1): The three mechanisms complement each other, and a single attack is difficult to avoid all detection dimensions at the same time. Optimized computational efficiency: Each layer of detection can be computed in parallel, and the overall complexity remains O(nlogn), which is suitable for resource-constrained UAV environments. (2) Enhanced adaptability: It does not rely on the IID assumption and adapts to heterogeneous data distribution through local statistical features. A single attack is difficult to avoid all detection dimensions at the same time. Experimental data show that the detection rate reaches 52.8%, while the baseline method is almost 0. As shown in Table 1 below, the detection rates of baseline methods such as FedAvg and Trimmed mean are both 0.000, the detection rate of Krum is also 0.000, while the Malicious detection rate of the Proposed method is 0.528.

[0034] II. Innovation and Advantages of Dynamic Trust Management System (1) Rapid Response Capability: The proportional damping design ensures smooth changes in trust values ​​while maintaining sensitivity to malicious behavior. Experiments show that significant separation is achieved within 10 rounds. (2) Anti-Manipulation: Parameter calibration resists periodic attack strategies, and malicious agents cannot maintain high trust through alternating behaviors. (3) Cross-Layer Security Assurance: Trust scores are used for both weight aggregation and task scheduling to form a unified security metric.

[0035] III. Innovation and Advantages of Trust-Weighted Aggregation Mechanism: (1) Improved Model Quality: High-quality updates from high-trust agents receive greater weight, accelerating model convergence. Experiments show that the verification accuracy is stable at around 60%. (2) Enhanced Security: Even if malicious updates pass detection, their impact is limited by low trust scores. (3) Resource Efficiency: Avoids completely discarding suspicious updates and makes full use of all available information.

[0036] IV. Innovation and advantages of security-aware task scheduling algorithm: (1) Maximize operation efficiency: dynamically balance security and efficiency, and the task completion rate is less than 5% of the baseline. (2) Optimize resources: consider battery status β_i and communication bandwidth b_i to extend system battery life. (3) Risk control: high-threat tasks are automatically assigned to high-trust agents to reduce the risk of task failure.

[0037] V. Innovation and advantages of adaptive communication and computing mechanisms: (1) Improved energy efficiency: Adaptive adjustment reduces unnecessary computing and communication overhead, with energy consumption difference from the baseline of <3%. (2) Guaranteed real-time performance: Sparsification and quantization reduce communication volume by more than 50%, suitable for bandwidth-constrained environments. (3) Security adaptation: Automatically enhances protection when the threat level is high, balancing security and efficiency.

[0038] VI. Overall Framework Integration Innovation and Advantages: (1) Comprehensive Improvement in System Performance: Detection rate 52.8% → 50+ percentage points higher than baseline; Task completion rate >95% → Approaching optimal operating efficiency. Energy consumption increase <3% → Resource overhead is controllable. (2) Enhanced Engineering Practicality: Modular design facilitates deployment and maintenance; adjustable parameters adapt to different application scenarios; supports large-scale UAV group expansion. (3) Theoretical Solidity: Convergence mathematical proof; trust separation theory guarantee; optimal scheduling algorithm guarantee.

[0039] This invention, through the synergistic effect of the aforementioned innovations, achieves significant improvements in security and performance while maintaining low resource consumption. The tight integration of the various technical components forms a virtuous cycle of "detection-trust-scheduling-training," providing a practical security solution for mission-critical UAV applications. Experimental data fully validates the practical effects of the theoretical advantages, demonstrating that this invention significantly outperforms existing technologies in terms of detection accuracy, operational efficiency, resource utilization, and system adaptability. This comprehensive advantage makes this invention particularly suitable for deployment in smart city UAV applications with stringent security requirements.

[0040] Table 1. Average performance over 36 rounds at 20% Byzantine nodes. Security perception (method described in this paper) 0.596 2.369 0.528 0.489 0.030 0.436 Safety perception (fine-tuning) 0.590 2.248 0.528 0.522 0.031 0.462 Trimming the mean 0.533 2.399 0.403 0.544 0.032 0.524 Krum 0.610 2.376 0.000 0.544 0.031 0.556 Naive Mean 0.384 3.803 0.000 0.544 0.031 0.556 Attached Figure Description

[0041] Figure 1 This is a diagram illustrating the malicious detection rate for each aggregation strategy in each round.

[0042] Figure 2 for Figure 2 A diagram illustrating the task completion rate per round by the scheduler.

[0043] Figure 3 This is a schematic diagram illustrating the evolution of the average trust score for participants in UAVs.

[0044] Figure 4 This is a diagram illustrating the verification accuracy during 36 rounds of collaborative learning.

[0045] Figure 5 Architecture diagram of the integrated framework for elastic federated learning and trust-aware scheduling of heterogeneous drone swarms Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0047] This invention relates to an integrated framework for resilient federated learning and trust-aware task scheduling in heterogeneous UAV swarms. This framework aims to address the security challenges in collaborative learning among UAV swarms in adversarial environments, particularly the performance degradation caused by Byzantine attacks (i.e., malicious agents injecting malicious model updates). The core of this invention is the integration of multi-metric anomaly detection, dynamic trust management, and security-aware task scheduling to achieve joint optimization of learning robustness and operational efficiency. The technical solution encompasses three main layers: a trust-robust aggregation mechanism at the learning layer, a security-aware task scheduling algorithm at the operational layer, and cross-layer adaptive communication and computation optimization. The technical solutions for each improvement of the invention will be described in detail below.

[0048] The trust-robust aggregation mechanism of this invention employs a multi-layered defense strategy, combining statistical analysis, geometric characteristics, and behavioral patterns to accurately identify and filter malicious model updates while ensuring model convergence. This mechanism comprises three sub-modules: multi-indicator anomaly detection, trust-weighted aggregation, and dynamic trust updates.

[0049] 1. Multi-indicator anomaly detection: This module updates each received model. (from) A triple-complementary detection method is used to capture different forms of adversarial behavior. The detection process is as follows: (1) Gradient norm analysis (first detection layer): Technical feature: Calculating the L2 norm of the updated gradient The mean and standard deviation were estimated using robust pruning statistics. Specifically, the highest and lowest values ​​were excluded. Proportion (e.g.) Calculate the trimming mean using the norm value of (=0.2). and pruning standard deviation .

[0050] in, ; .

[0051] Working principle: Filters updates with extreme amplitudes (such as additive noise attacks) to prevent outliers from contaminating statistical estimates. The acceptable update set is defined as: Among them, the pruning ratio It needs to be greater than the Byzantine agency ratio. (generally ,like =0.2), to ensure the robustness of the estimation.

[0052] (2) Statistical residual check (second detection layer): Technical features: Residuals are calculated based on median absolute deviation (MAD). For each update, the residual distance to the median coordinates is calculated. The median is calculated along the coordinate dimensions. MAD is defined as: .

[0053] Operating principle: Identifies updates that deviate from the distribution center trend, capturing additive noise attacks or systematic biases. Acceptable update set: Among them, MAD, as a robust scaling estimate, is insensitive to outliers and is suitable for non-IID data.

[0054] (3) Orientation alignment assessment (third detection layer): Technical features: computational updates and preliminary aggregation results The cosine similarity (calculated from the updates after the first two layers of filtering). The similarity threshold is set to... =0.85.

[0055] How it works: It ensures that the update direction aligns with the optimization goal, effectively combating symbol flip attacks. Acceptable update sets are:

[0056] Preliminary aggregation It needs to be calculated based on a trusted updated subset to avoid initial contamination.

[0057] Ultimately, the updated set is the intersection of the three: This multi-index design improves detection accuracy through a complementary mechanism and maintains robustness under non-IID data.

[0058] 2. Trust-weighted aggregation: After filtering, a trust-weighted scheme is used to calculate the global update to strengthen the influence of reliable agents.

[0059] Technical features: Global update Calculate using the following formula: ;in, ,yes

[0060] Trust score in round t.

[0061] How it works: Agents with higher trust scores contribute more to model evolution, while the influence of newly added or previously labeled agents is limited, thus reducing the impact of malicious updates. Trust scores need to be updated dynamically, and the initial score can be set to a uniform value (e.g., 0.5).

[0062] 3. Trust Update Mechanism: Trust scores are dynamically adjusted based on behavioral analysis, forming a closed-loop feedback loop. (1) Technical features: Trust score ( Follow the projection gradient dynamics update rule: .

[0063] in, UAV trust score; The operator represents the projection of the trust score onto the interval [0,1]; update amount Determined by the proportional damping rule, a differentiated update strategy is adopted based on the following three behavioral patterns: a. If the update is marked as malicious: Penalty rate Provide sufficient deterrence; If the contribution model is updated and improved: Reward rate Ensure a steady increase in trust; c. Neutral behavior: Allows the system to recover from temporary anomalies, avoiding excessive penalties.

[0064] (2) Working principle: The proportional design prevents the trust value boundary from oscillating and ensures that the honest and malicious proxy scores are asymptotically separated.

[0065] (3) Implementation conditions: parameters and Calibration is required to balance sensitivity and stability.

[0066] (4) Separation characteristic: This update mechanism can ensure that the trust scores of honest agents and malicious agents are asymptotically separated. a. The expected trust score of the honest agent converges to: ; b. The expected trust score of a malicious proxy converges to: ; c. Among them ;ensure .

[0067] The effect of multi-index detection is as follows Figure 1 As shown, the method of the present invention has a detection rate of over 0.5 in the first 8 rounds and maintains a high level, while the detection rate of the baseline method (such as FedAvg) is zero, demonstrating its superiority.

[0068] I. Security-Aware Task Scheduling Algorithm This invention models task allocation as a multi-objective optimization problem, integrating trust scores, resource constraints, and threat indicators to achieve secure and efficient task scheduling. The algorithm includes optimization formulas, a greedy heuristic, and the optimal Hungarian method.

[0069] 1. Multi-objective optimization formula (1) Technical characteristics: The task allocation problem is formalized as an integer linear programming problem: ;in, Let $\frac{i}{k}$ be the total cost, representing the comprehensive cost of assigning drone $i$ to perform task $k$, with the goal of minimizing it.

[0070] (2) The constraints are as follows: a. Each task can be assigned a maximum of one UAV: ; b. Each UAV can be assigned a maximum of one task: ; c. Trust threshold: ,like ;in, This represents the minimum trust threshold for the task.

[0071] Energy constraints: If the battery level .in, The drone battery level indicates the remaining battery power of drone i at the current moment; Task energy consumption represents the energy that task k is expected to consume during execution.

[0072] (3) Cost function: .

[0073] in, For task urgency, a parameter taking values ​​in the range [0, 1] represents the urgency of task k; For energy efficiency; A trust deficit; Threat level; These are the weighting coefficients.

[0074] (4) Working principle: Balance task urgency, energy efficiency, trust and threat mitigation to ensure that high-risk tasks are assigned to high-trust agents.

[0075] 2. Scheduling Algorithm Implementation (1) Greedy heuristic: 1) Core Algorithm Ideas and Design Principles Greedy heuristics employ a strategy of immediate optimal local decision-making, consistently selecting the allocation scheme that appears best at the moment during each round of scheduling. This design is based on the important consideration that, in large-scale drone swarm scenarios, the computational cost of finding the globally optimal solution is often too high, while greedy algorithms can provide high-quality feasible solutions within an acceptable timeframe.

[0076] The core of the algorithm lies in the single-task allocation guarantee mechanism, which ensures that each task is assigned to at most one drone, and each drone undertakes at most one task. This one-to-one allocation mode avoids resource conflicts and simplifies the complexity of system management.

[0077] 2) Detailed step-by-step analysis of the process flow a. Input parameter definition: UAV sets can be used.

[0078] A set of tasks to be assigned.

[0079] : Execute the task The cost function.

[0080] : This represents the minimum trust threshold for the task.

[0081] b. Detailed execution process Phase 1: Feasible Solution Space Construction and Preprocessing The algorithm first performs a feasibility screening of all tasks and drones. For each task to be assigned, the system iterates through all available drones, checking whether they meet the basic requirements of the task. These requirements include two aspects: trust threshold and energy constraints. The drone's current trust score must meet or exceed the minimum trust standard specified by the task; at the same time, the drone's remaining battery power must be sufficient to complete the energy consumption required to complete the task.

[0082] The selected drones are included in the candidate set for the task and then sorted in ascending order according to their cost function values. This sorting process ensures that the system can quickly locate the candidate drone with the best cost in the subsequent allocation stage.

[0083] Phase Two: Priority-Based Sequential Allocation The system processes tasks sequentially according to their priority. For each current task, the algorithm selects the lowest-cost, unallocated, valid drone from its candidate drone list. This selection mechanism ensures that high-priority tasks receive the optimal drone resources first.

[0084] During the allocation process, the system updates the drone's status information in real time. Once a drone is successfully assigned a task, its status is immediately marked as "assigned," thus avoiding resource conflicts in subsequent task allocations. This real-time status update mechanism ensures the consistency of the allocation scheme.

[0085] Phase 3: Handling and Optimizing Unassigned Tasks For tasks that cannot be assigned due to insufficient resources or unmet constraints, the system will reserve them for the next scheduling cycle. Simultaneously, the algorithm will record the allocation results of this round, and this data will be used for subsequent trust assessment and system optimization.

[0086] 3) Time complexity analysis and performance characteristics The time complexity of this algorithm is primarily determined by the sorting operation. For a system with n tasks, each task needs to sort its set of candidate drones. Assuming an average of m candidate drones per task, the time complexity of sorting is linear-logarithmic. Therefore, the overall time complexity is linearly related to the product of the number of tasks and the number of drones, while also being affected by the sorting operation.

[0087] In practical engineering implementations, by employing efficient sorting algorithms and reasonable data structures, the algorithm can maintain good real-time performance even in large-scale scenarios. This computational efficiency makes the algorithm particularly suitable for applications with high response time requirements.

[0088] 4) Analysis of Applicable Scenarios and Limitations Greedy heuristics are best suited for scenarios involving drones and a large number of tasks, typically showing significant advantages when the number of nodes exceeds 50. Their main advantages lie in high computational efficiency, simple implementation, and relatively low resource consumption. However, this algorithm cannot guarantee a globally optimal solution and may get stuck in local optima under certain circumstances.

[0089] 5) Time complexity

[0090] (2) Hungarian method: 1) Mathematical Foundations and Theoretical Framework of Algorithms The Hungarian method, based on bipartite graph matching theory in combinatorial optimization, transforms the task assignment problem into a mathematical problem of finding the minimum weight perfect matching. The core of this method lies in establishing a complete bipartite graph model, where the set of drones constitutes the left-hand nodes, the set of tasks constitutes the right-hand nodes, and the edges represent feasible assignment relationships.

[0091] The algorithm is based on the Kuhn-Munkres theorem, which guarantees the existence of a polynomial-time algorithm for finding the minimum weight matching in a complete bipartite graph. By maintaining a set of feasible vertex labels, the algorithm can systematically search for the optimal matching scheme.

[0092] a. Bipartite graph construction: Left-hand node set X: UAV nodes, |X|=m The right-hand side node set Y: task nodes, |Y| = n Edge set E: Feasible edge allocation

[0093] l Edge weight w(i,k): Allocation cost

[0094] 2) Step-by-step analysis of the complete algorithm process ① Initialization phase: Problem modeling and parameter setting First, a cost matrix is ​​constructed, where each element represents the cost of a specific drone performing a specific task. For infeasible allocation combinations (such as insufficient drone trust score or insufficient energy), the corresponding position in the matrix is ​​set to infinity, thereby automatically excluding these invalid allocations during the calculation process.

[0095] Simultaneously, two sets of label variables are initialized: one set corresponds to the drone node, and the other set corresponds to the task node. The initial label settings follow specific rules to ensure the validity of subsequent calculations.

[0096] ② Iterative optimization phase: Augmented path search and matching expansion The algorithm iteratively improves the matching scheme. In each iteration, the system attempts to find an augmenting path—a special path whose start and end nodes are unmatched, and whose edges alternate between inside and outside the current match.

[0097] Once an augmenting path is found, the number of matches or the quality of matches can be increased or improved by reversing the matching state of the path (changing a matching edge to a non-matching edge, and vice versa). This process is repeated until no more augmenting paths can be found.

[0098] ③ Labeling adjustment stage: Maintenance and optimization of feasible solutions When an augmenting path cannot be found directly, the algorithm expands the search space by adjusting vertex labels. The amount of label adjustment needs to be precisely calculated to ensure that existing matching properties are not violated, while creating conditions for discovering new augmenting paths.

[0099] 3) Algorithm characteristics and performance analysis The time complexity of the Hungarian method is on the order of the cube of the number of nodes, meaning that the computation time increases significantly as the problem size grows. However, the algorithm's greatest advantage is that it guarantees a globally optimal solution, that is, finding the allocation scheme with the minimum total cost among all possible allocation schemes.

[0100] Regarding memory usage, the algorithm needs to store the complete cost matrix, which can lead to significant memory overhead for large-scale problems. Therefore, this method is more suitable for small to medium-sized scheduling scenarios.

[0101] 4) Constraint handling mechanism The algorithm handles various constraints in practical applications through multiple mechanisms. Hard constraints (such as minimum trust requirements and energy requirements) are reflected in the existence of edges during the bipartite graph construction phase, ensuring that only allocations that satisfy all basic requirements are considered.

[0102] 5) Time complexity Each UAV requires O(n) label updates, and each label update requires O(n) path searches, resulting in a total complexity of O(n).

[0103] The task completion rate of the scheduler is shown in the figure. The method of the present invention has a difference of less than 5% from the baseline within 36 rounds, indicating that it maintains operational efficiency while ensuring safety.

[0104] II. Adaptive Communication and Computation To adapt to resource constraints and threat environments, this invention introduces threat-adaptive parameter selection, quantization, and sparsification techniques to dynamically adjust local training and communication parameters.

[0105] 1. Threat Adaptive Parameter Selection (1) Technical characteristics: Each UAV is based on the threat level Select training parameters: High threat ( Training rounds Quantization bit depth Inject differential privacy noise (standard deviation) ).

[0106] low threat ( Training rounds Quantization bit depth No noise.

[0107] (2) Working principle: minimize computational footprint and information leakage under high threat, and maximize contribution under low threat.

[0108] (3) Implementation conditions: The threat level is updated in real time by the environmental monitoring module, for example, based on GPS spoofing or communication interference indicators.

[0109] 2. Quantization and Sparsification (1) Quantization technique: Apply uniform quantization to gradient updates: Where q is the number of quantization bits (e.g., 8 or 16), reducing communication load.

[0110] (2) Sparsification technique: Only transmit the top-k largest components of the gradient: .

[0111] in, For binary masks, k can be set to 10%-20% of the gradient dimension.

[0112] (3) Working principle: It preserves the information gradient direction and reduces communication overhead by more than 50%, making it suitable for bandwidth-limited UAV networks.

[0113] (4) Implementation conditions: Quantization and sparsification need to be applied after local training and reconstructed before global aggregation.

[0114] III. Framework Integration and Workflow The framework of this invention operates in synchronous rounds, each round containing four phases: (1) Threat assessment and trust update: updating threat levels based on environmental data and historical behavior. and trust score .

[0115] (2) Safety-aware task allocation: Use scheduling algorithms to allocate tasks to ensure that constraints are met.

[0116] (3) Adaptive local training: UAV selects to perform local model training based on parameters.

[0117] (4) Byzantine fault-tolerant aggregation: The central node performs multi-index detection and trust-weighted aggregation.

[0118] The entire process forms a closed loop, ensuring that learning and operational goals are aligned. The dynamic evolution of trust scores is as follows: Figure 3 As shown, the trust value of malicious agents continuously decreases, while that of honest agents remains stable, demonstrating effectiveness. The model convergence is as follows: Figure 4 As shown, the accuracy of this invention is stable near 0.6 in the presence of Byzantine faults, which is close to the baseline.

[0119] Verification of the technical solution: 1. Multi-index Anomaly Detection Mechanism (1) The implementation method completely covers three detection layers: norm filtering, residual check, and direction alignment (2) Parameter selection (β = 0.2, = 0.85) ensures the balance between detection accuracy and false alarms (3) The experimental results show a detection rate of 52.8%, supporting "high-precision detection" Dynamic Trust Management System (1) The implementation includes the complete processes of trust initialization, update rules, and participation control (2) Proportional damping design ( = 0.08, = 0.12) ensures stability and responsiveness (3) The experimental results of trust separation support "effectively distinguishing honest and malicious agents" Security-Aware Task Scheduling (1) Provide two algorithm implementations: greedy heuristic and Hungarian optimal (2) The cost function integrates four key factors: urgency, energy, trust, and threat (3) The task completion rate > 95% verifies "optimized task allocation" Adaptive Communication Mechanism (1) Implement a quantization scheme: .

[0120] (2) Sparsification scheme: Top-k selection (k = 0.1d) (3) The communication overhead is reduced by more than 50%, supporting "efficient communication" Verification of Consistency with Technical Solutions 1. Performance Consistency: (1) Theoretical convergence is guaranteed through the learning rate control (η ≤ 1 / L) in the implementation (2) The trust separation theorem is instantiated through the proportional update rule (3) The experimental results show that all performance indicators meet or exceed the theoretical expectations 2. Security Consistency: (1) The implementation meets the Byzantine resilience requirements (tolerating f < N / 2 malicious nodes) (2) Multi-level defense covers all known attack types (sign flipping, additive noise, zero gradient) (3) The dynamic trust mechanism provides continuous security monitoring and adaptation 3. Resource Consistency: (1) Computational complexity control: Aggregation algorithm O(NlogN), scheduling algorithm O(N^3) or O(N^2logN) (2) Memory usage optimization: sparse matrix storage, incremental update (3) Energy efficiency: Dynamically adjust training intensity to match UAV power status.

[0121] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. An integrated framework for heterogeneous unmanned aerial vehicle (UAV) swarm elastic federated learning and trust-aware task scheduling, characterized in that, include: The Zhanting fault-tolerant aggregation module is used to receive local model updates uploaded by each drone, filter malicious updates based on a multi-indicator anomaly detection mechanism, and perform weighted aggregation based on the dynamic trust scores of each drone to generate a global model update. The dynamic trust management module, connected to the Byzantine fault-tolerant aggregation module, is used to update the trust score of each UAV based on the results of the multi-index anomaly detection mechanism and the historical behavior of each UAV using the projection gradient dynamics rule. The security awareness task scheduling module, connected to the dynamic trust management module, is used to allocate appropriate drones to tasks based on the attributes of the tasks to be executed, the resource status of each drone, the dynamic trust score, and the threat level, by optimizing a cost function that includes the trust deficit and the threat level. An adaptive communication and computation optimization module, connected to the security awareness task scheduling module and the Byzantine fault-tolerant aggregation module, is used to dynamically adjust the local training parameters and communication strategies of each UAV according to the threat level faced by each UAV. The framework optimizes model training, trust evaluation, and physical task scheduling through a synchronous closed-loop process.

2. The heterogeneous UAV swarm elastic federated learning and trust-aware task scheduling integration framework according to claim 1, characterized in that, The aforementioned multi-index anomaly detection mechanism updates the model for each received parameter. The triple complementary detection process is as follows: (1) Gradient norm analysis: Calculate the L2 norm of the updated gradient. The mean and standard deviation were estimated using robust pruning statistical methods, and the highest and lowest values ​​were excluded. The norm of the proportion is used to calculate the pruning mean. and pruning standard deviation , in, ; , An acceptable update set is defined as: Among them, pruning ratio It needs to be greater than the Byzantine agency ratio. ,generally ; (2) Statistical residual check layer: Calculate residuals based on median absolute deviation (MAD). For each update, calculate the residual distance between it and the median of the coordinates. The median is calculated along the coordinate dimensions, and MAD is defined as follows: Identify updates that deviate from the distribution center trend, capture additive noise attacks or systematic biases, and accept the following update set: ； Among them, MAD serves as a robust scaling estimate; (3) Orientation alignment evaluation layer, calculate update and preliminary aggregation results The cosine similarity, with a similarity threshold set to =0.85, Acceptable update sets are: Ultimately, the updated set is the intersection of the three: .

3. The heterogeneous UAV swarm elastic federated learning and trust-aware task scheduling integration framework according to claim 2, characterized in that, The dynamic trust management module updates the trust score using the following projection gradient dynamics rule. : ， in, This is an operator that projects the trust score onto the interval [0,1]. Update volume The determination rule is: if a local update is marked as malicious, then Among them, the penalty rate If the local update is accepted and the contribution is positive, then Reward rate ; If it is a neutral behavior, then .

4. The integrated framework for heterogeneous UAV swarm elastic federated learning and trust-aware task scheduling according to claim 3, characterized in that, The security-aware task scheduling module models the task allocation problem as an integer linear programming problem. ， in, The total cost is the comprehensive cost of assigning drone i to perform task k, and the comprehensive cost function is: ， in, For task urgency, a parameter taking values ​​in the range [0, 1] represents the urgency of task k; For energy efficiency; A trust deficit; Threat level; These are the weighting coefficients.

5. The integrated framework for heterogeneous UAV swarm elastic federated learning and trust-aware task scheduling according to claim 4, characterized in that, The adaptive communication and computation optimization module adjusts the response based on the threat level of the unmanned aerial vehicle (UAV). The following adaptive adjustments will be made: High threat ( Training rounds Quantization bit depth Inject differential privacy noise, standard deviation , Low threat ( Training rounds Quantization bit depth No noise.

6. The integrated framework for heterogeneous UAV swarm elastic federated learning and trust-aware task scheduling according to claim 5, characterized in that, The adaptive communication and computation optimization module is also used to update the gradient before the UAV uploads the update. Perform uniform quantization: Where q is the number of quantization bits. Top-k sparsity processing for gradient transmission only: in, For binary masks, k can be set to 10%-20% of the gradient dimension.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the functions of each module of the integrated framework as described in any one of claims 1-6.

Images