A network intrusion detection method and system for a UAV cluster

Abstract

The application belongs to the field of unmanned aerial vehicle network security and intrusion detection, and provides a network intrusion detection method and system for unmanned aerial vehicle clusters, to synchronously collect cluster node network communication and flight control telemetry data at fixed time steps, form unified samples through space-time alignment of sliding windows, generate pretreatment features through standardization processing, freeze rule parameters and generate unique feature signatures; then, replay fingerprints, injected semantics and cross-modal residual scores are calculated, which are integrated with four types of state basic probabilities output by a lightweight model into a comprehensive evidence vector, which is mapped into a D-S evidence theory basic probability distribution; finally, the distribution is exchanged between nodes, and after trusted discount processing, multi-source evidence is iteratively fused, and a detection result is output to judge intrusion in combination with double belief thresholds. The application constructs a multi-modal collaborative detection framework, accurately identifies injected and replay attacks and stably alarms, and improves system engineering maintainability and long-term operation robustness through versioning and simulation closed loop.

Description

Technical Field

[0001] This invention belongs to the field of UAV network security and intrusion detection technology, and provides a network intrusion detection method and system for UAV swarms. Background Technology

[0002] The drone swarm is a distributed collaborative communication network built on the basis of a wireless ad hoc network. Each drone node in the swarm is a peer-to-peer communication entity. The nodes achieve overall operation scheduling of the swarm through real-time task command distribution, flight control status telemetry feedback, and formation collaborative information exchange. The status perception, behavior decision-making, and action execution of each node are interconnected and mutually restrictive. The task execution efficiency and continuity of the swarm depend entirely on the reliable data exchange and collaborative consistency between nodes. At the same time, this type of communication network has inherent characteristics such as high-speed node movement, dynamic network topology reconstruction, wireless link quality being susceptible to environmental interference fluctuations, and limited onboard computing / communication / storage resources of drones. The communication connection between nodes is adjusted in real time according to the formation position.

[0003] The network characteristics and application scenarios of drone swarms make them vulnerable to network attacks and various network intrusion behaviors. On the one hand, drone swarms use wireless self-organizing networks for data transmission, and the wireless communication links are not physically isolated, making them easy for attackers to intercept, eavesdrop on, or even tamper with. Moreover, the difficulty of node identity authentication and link validity verification under dynamic topology is greatly increased, providing opportunities for attacks. On the other hand, swarms have extremely high requirements for the real-time, validity, and timing of data. Attackers do not need to damage network hardware; they can interfere with the coordination and consistency between nodes simply by forging and repeatedly sending data in a legitimate format. Among these, fake data injection attacks and replay attacks are the main types of intrusions due to their strong concealment and low implementation cost. These two types of attacks can disrupt the data semantics and timing consistency between nodes without significantly changing the traditional network traffic statistics. The resource-constrained nature of drone swarms makes it difficult to deploy complex front-end security protection mechanisms, further increasing the risk of intrusion. Once an intrusion occurs, it will directly lead to swarm formation instability, mission command execution errors, and even cause overall swarm failure and complete mission interruption.

[0004] Current technologies for network intrusion detection in drone swarms primarily employ centralized deep learning-based detection schemes, supplemented by some general network security technologies. The core detection logic of this mainstream approach is as follows: using a ground control station or centralized server as the core node, network communication data and some flight control status data from all nodes in the drone swarm are collected uniformly. After feature alignment and format conversion of the collected heterogeneous data, a training dataset with a fixed field structure and fixed dimensions is constructed. A deep learning or ensemble learning model is trained on this dataset in an offline environment to classify and identify attack types such as distributed denial-of-service attacks, spoofed data injection, and replay attacks. The trained detection model is then fully distributed to each drone's onboard terminal or edge processing node. Online detection is performed on the real-time collected network data through model inference, and finally, the onboard terminal or ground station outputs intrusion alerts based on the inference results. While existing technologies can achieve basic network attack identification and alerting, the false positive and false negative rates are relatively high, and the model lacks sufficient engineering deployability, real-time performance, and long-term robustness. Specifically:

[0005] It is incompatible with heterogeneous data from different drone models and is difficult to adapt to data distribution drift caused by mission phase switching.

[0006] It lacks the ability to identify covert intrusion behaviors and cannot effectively identify fake data injection and replay attacks that are legally formatted but semantically or temporally abnormal.

[0007] Without considering the conflicts between nodes and the uncertainties in the reasoning process, when the evidence is contradictory or the detection results are in a boundary state, it is easy to make incorrect judgments due to the one-sidedness of a single piece of evidence, which greatly increases the risk of misjudgment in local detection and reduces the reliability of detection reasoning.

[0008] The single-node independent detection mode lacks a collaborative detection mechanism among UAV cluster nodes. The observation results of a single node are easily affected by factors such as link obstruction, environmental disturbances, and local perspective deviations, resulting in deviations. Relying solely on single-node observations cannot reflect the overall status of the cluster, which can easily lead to missed or false alarms. Summary of the Invention

[0009] To address the aforementioned technical issues, this invention provides a network intrusion detection method and system for unmanned aerial vehicle (UAV) swarms. Under the conditions of dynamic topology and resource constraints in UAV swarm systems, it establishes a portable, deployable, and sustainably evolving multimodal collaborative intrusion detection framework, enabling accurate identification and stable alarms for injection and replay attacks. Furthermore, it enhances the system's engineering maintainability and long-term operational robustness through model versioning and simulation closed-loop.

[0010] The technical solution of the present invention includes: Network communication data and flight control telemetry data of each node in the UAV cluster are collected synchronously at fixed time steps. The network communication data and flight control telemetry data are spatiotemporally aligned to form a unified sample. A preprocessing mapping function is constructed based on field mapping, encoding, and normalization to convert the unified sample into standardized preprocessing features.

[0011] To address replay attacks, the repetition rate, backtracking rate, out-of-order arrival ratio, and arrival interval drift are calculated within a sliding window and weighted to obtain a replay fingerprint score. To address injection attacks, the violation degree of field legal domain constraints, the violation degree of task phase constraints, and the abrupt change degree of short-term command distribution are calculated and weighted to obtain an injection semantic score. A cross-modal prediction function is constructed based on the coupling relationship between the UAV network state and the flight control state. The cross-modal prediction function calculates the residual vector between the actual network observations and the predicted values, and the residual vector is used to calculate the cross-modal residual score through the feature weight matrix.

[0012] The preprocessed features are input into the lightweight detection model, and the output node has a basic probability of being in one of four states: normal, replay, injection, and composite. The replay fingerprint score, injection semantic score, cross-modal residual score and the basic probability of the four states are integrated into a comprehensive evidence vector. The uncertainty score of the comprehensive evidence vector is calculated, and the comprehensive evidence vector is mapped to the basic probability allocation of the DS evidence theory with the uncertainty score correction term.

[0013] Each node shares a basic probability allocation with its neighboring nodes. The basic probability allocation of the neighboring nodes is discounted to obtain a discounted probability allocation. Each node iteratively merges its own basic probability allocation with the discounted probability allocation of all neighboring nodes to obtain multi-source evidence. The belief value of the state with the highest probability among the four states is calculated using the multi-source evidence. The belief value is compared with two preset belief thresholds, and three detection results are output for each state: normal, suspicious, and confirmed. Based on the detection results, it is determined whether an intrusion has occurred.

[0014] Furthermore, all parameters of the preprocessing mapping function are processed to generate a unique feature signature. The preprocessing mapping function parameters include a fixed field mapping table, an encoding dictionary, a normalized numerical range, and a missing value handling strategy. The feature signature is a unique hash value of the parameter, used for consistency verification during model training, inference, updating, and rollback processes.

[0015] Furthermore, the arrival interval drift is the degree of deviation of the mean and variance of the current window message arrival interval from the normal baseline statistics, which is calculated by weighting the mean drift and variance drift; the short-term instruction distribution abruptness is calculated using the Jensen-Shannon distance to determine the difference between the current window control instruction category distribution and the historical normal baseline distribution; the cross-modal prediction function is constructed by a lightweight regressor, which can be a linear regression model, a tree model, or a lightweight neural network.

[0016] Furthermore, the uncertainty score is obtained based on the degree of conflict between the replay fingerprint score, the injected semantic score, the cross-modal residual score, and the dispersion of the basic probability distribution; when the evidence conflict is obvious or the probability distribution is uniform, the uncertainty score is increased to reduce the risk of misjudgment.

[0017] Furthermore, the iterative fusion adopts the Dempster combination rule, which incorporates the evidence conflict coefficient into the uncertainty quality during the fusion process to avoid distortion of the fusion results caused by strong conflicts; the credibility discount coefficient is obtained by normalizing the link quality / reputation index and the node evidence conflict index using the Sigmoid function.

[0018] Furthermore, when a composite intrusion is identified, the process also includes composite attack deconstruction and attack source localization, including: calculating the relative magnitudes of replay contribution and injection contribution to determine the dominant type of composite attack; calculating replay contribution and injection contribution by combining corresponding scores and short-window trend terms; calculating the risk value of each node in the cluster using a node risk formula, generating a Top-K list of suspicious nodes by sorting by risk value; the node risk formula integrates the node's evidence score, fusion conflict degree, and cross-modal residual; combining neighborhood fusion conflict characteristics and cross-modal residual propagation characteristics to identify suspicious link sets, and using the UAV cluster communication topology as a graph structure, expanding outward from high-risk nodes to determine the attack impact range.

[0019] Furthermore, it also includes dynamic model updates, including: constructing a drift statistic based on the proportion of nodes with abnormal scores exceeding the standard in the neighborhood and the average conflict degree of neighborhood fusion, where the average conflict degree of neighborhood fusion is the average conflict coefficient when a node merges evidence with all its neighboring nodes; when the drift statistic continuously exceeds a preset threshold for L time windows, federated incremental training is initiated, the onboard end performs incremental training based on new local data and reports parameter increments and training summaries, and the aggregation end weights and aggregates the parameter increments according to node credibility and data quality to form a global model, and sends the update to all nodes in the cluster.

[0020] Furthermore, during dynamic model updates, version management is included before global model deployment. This includes binding the preprocessing mapping function, feature signature, and current model parameters into version triples, recording version change descriptions and regression verification results. The regression verification results are quantitative evaluation data of core performance indicators such as detection accuracy, false positive rate, false negative rate, confidence stability, and evidence fusion conflict level, based on simulation / test data covering multiple flight conditions, topology changes, and attack modes, as well as compliance judgment conclusions for version deployment.

[0021] Furthermore, before loading the new version of the global model on the airborne end, the consistency between the local feature signature and the new version is verified. If the verification fails, loading is refused. If the new version experiences performance abnormalities during online operation, the model, preprocessing mapping function, and feature signature are synchronously rolled back to the version triplet of the previous stable version according to the version pointer.

[0022] This invention also provides a network intrusion detection system for drone swarms, used to implement the aforementioned network intrusion detection method for drone swarms, comprising: The data acquisition and preprocessing module is used to synchronously acquire network communication data and flight control telemetry data of each node in the UAV cluster at a fixed time step, form a unified sample through spatiotemporal alignment of sliding window, generate preprocessed features based on frozen preprocessing mapping function, and complete feature signature consistency verification.

[0023] The attack feature quantization module is used to calculate the replay fingerprint score, injection semantic score, and cross-modal residual score, thereby enabling the quantitative characterization of replay attacks, injection attacks, and cross-modal state mismatches.

[0024] The local evidence processing module is used to output the basic probabilities of four states through a lightweight detection model, integrate the rule-based scores and basic probabilities into a comprehensive evidence vector, and map it to the basic probability allocation of DS evidence theory.

[0025] The cluster collaborative detection module is used to realize the basic probability allocation and exchange between nodes, process neighborhood evidence based on the credibility discount coefficient and complete the iterative fusion of multi-source evidence, and determine the intrusion detection result through dual belief threshold.

[0026] The attack tracing module is used to break down the dominant attack type during complex intrusions, generate a Top-K list of suspicious nodes, identify suspicious link sets, and determine the scope of attack impact.

[0027] The model update and version management module is used to build drift statistics and trigger federated incremental training, realize the binding, storage, verification and rollback of model version triples, and complete the regression verification and visual operation and maintenance of the new version.

[0028] The technical solution provided by this invention has the following advantages compared with the prior art: 1. By synchronously collecting two core types of data—network communication and flight control telemetry—and combining them with a sliding window to achieve spatiotemporal alignment, and then using a preprocessing mapping function to convert the unified samples of each node into unified standardized preprocessed features, the data processing methods throughout the model training and inference process are consistent. This effectively avoids detection failures caused by differences in data formats of different UAV models and improves the adaptability of the detection model to heterogeneous UAV swarm scenarios.

[0029] 2. Abandoning the limitations of traditional single-modal detection, this method constructs multi-dimensional quantifiable indicators and generates corresponding scores for the temporal anomalies of replay attacks and the semantic anomalies of injection attacks. At the same time, it designs cross-modal residual scores by utilizing the coupling relationship between the network and flight control status, achieving a comprehensive and accurate characterization of covert attacks that are "formatted legally but semantically / temporally abnormal". This fills the blind spot of traditional detection methods in the identification of covert attacks and improves the accuracy of attack detection.

[0030] 3. By integrating the rule-based attack score with the basic probability predicted by the lightweight model into a comprehensive evidence vector, and introducing uncertainty score and mapping it to the basic probability allocation of DS evidence theory, the organic integration of interpretable rule evidence and intelligent model reasoning evidence is achieved. This effectively absorbs evidence conflicts and boundary sample noise, reduces the risk of misjudgment and missed judgment caused by single evidence reasoning, and improves the reliability of local detection reasoning.

[0031] 4. By exchanging neighborhood evidence between nodes and constructing a credibility discount coefficient based on link quality and node reputation, evidence from low-credibility nodes is suppressed. Then, cluster-level collaborative verification is achieved through iterative fusion of multi-source evidence. Combined with dual-belief thresholds, hierarchical judgment is completed. This effectively avoids observation bias caused by link obstruction and local disturbances of single nodes, improves the stability and reliability of intrusion detection results, and adapts to the dynamic topology and multi-node collaborative operation characteristics of UAV clusters.

[0032] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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.

[0034] Figure 1 Workflow diagram for a drone swarm collaborative intrusion detection system. Detailed Implementation

[0035] The following detailed description of a specific embodiment of the present invention is provided in conjunction with the accompanying drawings. However, it should be understood that the scope of protection of the present invention is not limited to the specific embodiment.

[0036] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the technical solution of this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0037] In the description of the embodiments of the present invention, unless otherwise stated, "a plurality of" means two or more.

[0038] like Figure 1 As shown, the present invention provides a network intrusion detection method and system for unmanned aerial vehicle (UAV) swarms, comprising: Step 1: Synchronously collect network communication data and flight control telemetry data from each node of the UAV swarm at fixed time steps. Use a sliding window to spatiotemporally align the two types of data to form a unified sample. Construct a preprocessing mapping function based on field mapping, encoding, and normalization to convert the unified sample into standardized preprocessed features. Simultaneously, freeze all parameters of the preprocessing mapping function and generate a unique feature signature. The frozen parameters include a fixed field mapping table, encoding dictionary, normalized numerical range, and missing value handling strategy. The feature signature is a unique hash value of the parameters, used for consistency verification during model training, inference, updating, and rollback.

[0039] Step 2: For replay attacks, calculate the sequence repetition rate, sequence backtracking rate, out-of-order arrival ratio, and arrival interval drift within a sliding window, and weight them to obtain a replay fingerprint score. For injection attacks, calculate the field legal domain constraint violation degree, task phase constraint violation degree, and short-term command distribution mutation degree, and weight them to obtain an injection semantic score. Construct a cross-modal prediction function based on the coupling relationship between the UAV network state and flight control state, calculate the residual vector between the actual network observations and the predicted values, and combine it with the feature weight matrix to obtain the cross-modal residual score. The arrival interval drift is the degree of deviation of the mean and variance of the current window message arrival interval from the normal baseline statistics, and is calculated by weighting the mean drift and variance drift. The short-term command distribution mutation degree uses the Jensen-Shannon distance to calculate the difference between the current window control command category distribution and the historical normal baseline distribution. The cross-modal prediction function is constructed by a lightweight regressor, which can be a linear regression model, a tree model, or a lightweight neural network.

[0040] Step 3: Input the preprocessed features into the lightweight detection model, and output the basic probabilities of nodes in four states: normal, replay, injection, and composite. Integrate the replay fingerprint score, injection semantic score, cross-modal residual score, and the four basic probabilities into a comprehensive evidence vector. Map the comprehensive evidence vector to the basic probability allocation of DS evidence theory through the uncertainty score. The uncertainty score is obtained based on the degree of conflict of the replay fingerprint score, injection semantic score, cross-modal residual score, and the dispersion of the basic probability distribution. When the evidence conflict is obvious or the probability distribution is uniform, increase the uncertainty score to reduce the risk of misjudgment.

[0041] Step 4: Each node exchanges basic probability allocations with neighboring nodes. A credibility discount coefficient is constructed based on link quality, historical consistency, and node reputation. The basic probability allocations of neighboring nodes are discounted to obtain discounted probability allocations. The local basic probability allocation and the discounted probability allocations of all neighboring nodes are iteratively fused to obtain multi-source evidence. The belief value of the state with the highest probability among the four states is calculated. This belief value is compared with two preset belief thresholds, and three detection results—normal, suspicious, and confirmed—are output to determine whether an intrusion has occurred. The iterative fusion uses an improved Dempster combination rule. During the fusion process, the evidence conflict coefficient is absorbed into the uncertainty quality to avoid distortion of the fusion result caused by strong conflicts. The credibility discount coefficient is obtained by normalizing the link quality / reputation index and the node evidence conflict index using the Sigmoid function.

[0042] Specifically, at the data level, this invention organizes the drone swarm operation information at a fixed time step: Let the set of drone nodes in the swarm be... For each node The system at discrete time On the one hand, collect network communication data This includes message arrival time, sequence number or message counter, message type, source / destination identifier, transmission direction, and link quality indicators. On the other hand, it collects flight control telemetry data. This includes information such as position, velocity, attitude, mission phase, and formation relationships, used to characterize the UAV's motion state and cooperative behavior at the current moment. A sliding window alignment method is used, with a window length of [missing information]. Then the uniform sample is represented as: ,in This represents the statistical aggregation of network-side window data (throughput, retransmission rate, packet loss rate, RTT, etc.). This allows subsequent discrimination to utilize both communication layer information and flight status information, thus avoiding missed detections caused by single-modal features at the source.

[0043] To address the migration difficulties arising from variations in different drone models, protocol stacks, and data field dimensions, and to generate statistically applicable network state features, this invention defines a preprocessing mapping function. as follows: Preprocessing mapping function Original unified sample Converted into standardized preprocessing features ,Right now The function executes the following three steps in sequence: The first step is field mapping: Based on the field mapping table C, the heterogeneous field names of different drone models are uniformly mapped to standard field names.

[0044] The second step, feature encoding and normalization: for continuous features... Using standardized formulas Normalization is performed on discrete features. Use fixed dictionary encoding Convert to numerical value.

[0045] The third step is missing value handling: If missing values ​​are encountered during the above operations, they are filled or marked according to strategy S.

[0046] The preprocessor frozen state is defined as follows: And generate a signature hash for the feature set and mapping relationship: All subsequent training, inference, updates, and rollbacks will use the same method. and Consistency verification is performed. This mechanism ensures that the model training and deployment phases use completely consistent input representations, avoiding implicit distribution drift caused by field changes, and making the system reproducible and transferable across scenarios.

[0047] A typical characteristic of replay attacks is that the attacker repeatedly sends historically valid messages, thereby disrupting message timing consistency and affecting the real-time performance of the control loop. Since such attacks often do not change the message format and basic statistics, relying solely on ordinary traffic classifiers can easily result in missed detections. Therefore, this invention uses "timing consistency fingerprints" as one of the core pieces of evidence for replay detection.

[0048] Let the message sequence within the window be Then define: Sequence repetition rate: Sequence backtracking rate: Arrival interval sequence: Arrival interval drift (characterized by distribution distance, which can be expressed as mean-variance shift): in , This is the normal baseline statistic. , The weights are directly fixed based on expert experience. The resulting replay fingerprint score is then calculated.

[0049] Sequence repetition rate indicates how many message sequence numbers are repeated within a sliding window (containing n messages), numerator The value k represents the number of messages that meet the condition that there exists a position j before k and the sequence number of position j is equal to the sequence number of position k. In other words, the sequence number of message k does not appear for the first time in the window but is a repeat of the sequence number of a previous message. The denominator n is the total number of messages in the window. Therefore, the higher this ratio is, the greater the possibility of a replay attack, because in normal communication, the sequence number should be incremental and unique.

[0050] Sequence regression rate represents the proportion of sequence numbers that have regressed. The count is determined by how many packets k have a sequence number less than the sequence number of the previous packet. The denominator is n-1 because only n-1 pairs of packets can be compared. When an attacker inserts a historical packet, the sequence number may suddenly become smaller. This is called the rollback phenomenon. In a normal network, even if there are out-of-order arrivals, the overall trend of the sequence number should be increasing. Frequent rollbacks are a typical characteristic of replay attacks.

[0051] The arrival interval sequence represents a sequence of time intervals between message arrivals. This represents the time difference between the k-th message and the (k-1)-th message. k iterates through all message pairs within the window from 2 to n. This sequence records the time pattern of message arrival. Under normal communication conditions, the interval sequence should conform to a certain stable statistical distribution.

[0052] Interval of arrival drift measures the degree to which the statistical characteristics of the interval of arrival in the current window deviate from the normal baseline. The first term... It is mean shift, which calculates the absolute difference between the current window interval mean and the baseline mean. (The second term...) It is variance drift, which calculates the absolute difference between the current window interval variance and the baseline variance. α and β are two weighting coefficients used to balance the relative importance of mean drift and variance drift. When an attacker replays packets in batches, a large number of packets will be generated in a short period of time, causing the mean of the arrival interval to decrease and the variance to change abruptly. This drift can capture this anomaly.

[0053] Based on the above definition, the system analyzes the sequence number and arrival time series within a sliding window to form a set of features that can stably characterize the replay behavior, including sequence repetition rate, sequence backtracking rate, out-of-order arrival ratio, and the degree of drift in the arrival interval distribution.

[0054] The sequence repetition rate measures the density of repeated packets within a window, the sequence backoff rate measures the extent to which historical packets are inserted into the current stream, the out-of-order ratio is used to capture inconsistencies between the arrival order and the sending logic, and the arrival interval drift is used to characterize the abrupt changes in the interval statistical distribution caused by burst replay.

[0055] The system combines the above features into a replay fingerprint score, making the replay judgment interpretable and quantifiable, and maintaining stable differentiation ability under the background of link fluctuations and normal disturbances.

[0056] The key challenge of injection attacks lies in "legitimate format but abnormal semantics": attackers can construct seemingly normal control commands or telemetry data, making it difficult for traditional detection methods based on syntax or simple statistical features to identify them.

[0057] This invention decomposes semantic consistency into field validity domains, task phase constraints, and instruction distribution mutations. Let the node... At any moment The task phase is The control message category is The field vector is

[0058] Field validity constraints can be written as the degree of violation of interval and combination constraints: Task phase constraint definition allows for the set of categories Its degree of violation is: Short-time instruction distribution abruptness: Let the category distribution within the window be... The baseline distribution is Using Jensen-Shannon distance: Field valid domain constraint violation degree This indicator is used to detect whether the field values ​​of control or telemetry messages exceed a reasonable range. The formula is as follows: This represents a field vector of node i at time t. This vector contains multiple fields such as velocity, altitude, and heading angle. This represents the specific value of the m-th field. This represents the predefined range of valid values ​​for the m-th field, with the lower bound being... The upper boundary is Indicator function Its function is to determine whether the m-th field is within the valid range. If it is not, it returns 1 to indicate a constraint violation; if it is, it returns 0 to indicate a valid range. It sums the values ​​of all m fields. This gives you the total number of fields that violated the value range constraints. The larger this number, the more likely the message is to be forged and injected, because normal control commands or telemetry data should meet physical and logical value restrictions. For example, the speed of a drone cannot be negative or exceed physical limits. When an attacker injects fake data, they may construct unreasonable combinations of fields.

[0059] Task phase constraint violation This metric checks whether the type of control message received currently matches the logic of the current task stage. This indicates the type of control message received by node i at time t, such as takeoff command, hovering command, landing command, etc. This indicates the current task phase of node i, such as takeoff, cruise, or landing. It is a predefined set of allowed categories representing the current task phase. Which types of control messages are valid? Indicator function Determine the type of message received. If it is not in the allowed set, return 1 to indicate a violation of the constraint. For example, if a drone is in the landing phase but receives a take-off command, this is obviously abnormal. Under normal circumstances, control messages should match the mission phase. When an attacker injects false commands, they may not consider the context of the current mission state, resulting in inappropriate control types.

[0060] Short-time instruction distribution abrupt change This indicator is used to capture sudden changes in the distribution of control command types over a short period of time. This indicates the probability distribution of various control messages within the current time window; for example, takeoff commands account for 10%, hovering commands account for 60%, and maneuvering commands account for 30%. This indicates that the baseline distribution probability obtained from historical normal data is used as a reference standard. The Jensen-Shannon divergence measures the distance or difference between two probability distributions. A larger divergence value indicates a more significant deviation between the current window's instruction distribution and the normal baseline. During normal flight, the instruction type distribution should be relatively stable and consistent with mission characteristics. However, when an attacker injects a batch of false instructions of a specific type, the proportion of that type of instruction can suddenly increase, thus... Distribution mutation and The indicator that produces significant differences is used to identify covert injection attack patterns through statistical distribution comparisons.

[0061] Therefore, the injection semantic score is defined as: Based on the above design, the system establishes legal domain constraints for the control and telemetry fields, specifying the range of values, field combination relationships, and mutual exclusion relationships. Exceeding the constraints constitutes semantic abnormality evidence.

[0062] Simultaneously, a task phase consistency constraint is introduced, which binds the set of allowed control command types and command frequency ranges to the task phase, thereby identifying "control types that should not occur in the current phase" or "abnormally high-frequency control behaviors".

[0063] Examples of field combination relationships Example 1: Constraints of Speed ​​and Throttle Opening in Drone Flight. There is a physical coupling relationship between the current flight speed v and the throttle opening. Under normal circumstances, if the speed v < 5 m / s (low-speed flight), then the throttle opening should be between 20% and 40%. If the detected speed is only 3 m / s but the throttle opening reaches 90%, although each field is within its own legal range when viewed individually (speed 0-50 m / s is legal, throttle 0-100% is legal), their combination violates the laws of physics. This constitutes a violation of the field combination relationship, which is very likely due to an attacker injecting an unreasonable combination of control parameters.

[0064] Example 2: The combination of altitude and air pressure constraints. There is a fixed physical correspondence between the flight altitude of a UAV and the air pressure sensor reading. For every 100 meters increase in altitude, the air pressure decreases by about 12 hPa. If the system detects an altitude of 500 meters (relative to the takeoff point) but the air pressure value shows 101.3 kPa (close to sea level air pressure), this indicates that the combination of these two fields does not conform to the laws of atmospheric physics. Under normal circumstances, the air pressure corresponding to an altitude of 500 meters should be about 95 kPa. This abnormal combination may mean that the telemetry data has been tampered with and injected.

[0065] Examples of field mutual exclusion relationships: Example 1: Takeoff Mode and Landing Mode: Mutually Exclusive In the flight mode field of a drone, takeoff mode (takeoff_mode=1) and landing mode (landing_mode=1) are mutually exclusive because a drone cannot be in both takeoff and landing states at the same time. If the system detects that both fields are set to 1 in a certain control message, that is, takeoff_mode=1 and landing_mode=1 are both true, this violates the mutual exclusion constraint. In normal control logic, only one of these two flag bits can be 1 or both can be 0. This mutual exclusion violation is a clear data forgery feature.

[0066] Example 2: Manual control and auto cruise are mutually exclusive: In the control permission field, manual_control (manual control flag) and auto_cruise (auto cruise flag) are mutually exclusive. When manual_control=true, it means that the pilot is manually operating the drone. At this time, auto_cruise must be false, and vice versa. If the control command injected by the attacker contains both manual_control=true and auto_cruise=true, this violates the mutual exclusion constraint of control permissions. This is because the control of the drone cannot belong to both the manual and automatic systems at the same time. This contradictory combination of fields is typical evidence of injection attacks.

[0067] Examples of task phase consistency constraints: Mission phase consistency constraints are essentially rule recognition, but expressed and calculated using formulas. Specifically, the system pre-stores a rule table that tells the system rules such as "only takeoff commands and altitude adjustment commands can appear during the takeoff phase," "only route tracking commands and formation coordination commands can appear during the cruise phase," and "only descent commands and landing confirmation commands can appear during the landing phase." These rules are pre-defined manually based on flight common sense and mission logic. When the system detects that the UAV is in the landing phase but has received a takeoff command, it checks the rule table and finds that "the takeoff command is not in the allowed list for the landing phase," and then uses a formula... Record the violation as 1, and the absence of a violation as 0. Therefore, this formula simply expresses the action of "looking up a table to determine if a violation has occurred" using mathematical symbols. That represents that rule table. The symbols represent the process of looking up a table and making a judgment. Therefore, it is not a mathematical model formula learned through training, but a transformation of manually defined if-then rules into a numerical calculation form, which makes it convenient to perform weighted summation with other scores later.

[0068] Example 1: Formation adjustment commands should not be issued during the landing phase: When the drone is in the landing phase ( The set of control command types allowed to appear at that time. It should only include commands directly related to landing, such as descent speed control, attitude stabilization, and landing confirmation. It should not include formation adjustment commands (formation_adjust), because landing is a solo operation phase requiring focused attention on the landing maneuver. If the system detects the drone is landing but receives a formation adjustment command, it will be problematic. ,because This triggers a task phase constraint violation, which is likely due to the attacker injecting inappropriate control commands without understanding the task context.

[0069] Example 2: High-frequency maneuver commands should not be issued during the hovering phase: When the drone is in the hovering phase ( During hovering, under normal circumstances, the system should primarily receive position-keeping and fine-tuning commands. The frequency of maneuver commands should be very low, for example, no more than once every 10 seconds. If the system detects that it has received 5 maneuver commands consecutively within a 1-second window during the hovering phase, this abnormally high frequency of control behavior violates the command frequency range constraint of the mission phase. This is because the essence of hovering is to maintain stability and should not involve frequent maneuvers. This abnormal frequency is likely due to an attacker injecting false maneuver commands in bulk in an attempt to interfere with the stable hovering of the drone.

[0070] In addition, the system performs statistical analysis on the distribution of short-term instruction types and calculates the degree of abrupt changes in the distribution in order to capture sudden changes in instruction patterns caused by injection attacks.

[0071] The aforementioned semantic verification output injection score enables the system to form a stable response to covert injections and avoids misjudging short-term fluctuations caused by normal maneuvers as attacks.

[0072] Relying solely on evidence from the network or flight control side may still leave blind spots, especially when attackers deliberately control the insignificance of a single modality. The system needs to further utilize the "coupling relationship between network behavior and flight status" to establish strong constraints.

[0073] This invention introduces cross-modal consistency residuals to measure whether changes in network-side observations can be explained by changes in flight state, thereby identifying coupling disruptions caused by injection or replay. Let the cross-modal prediction function be:

[0074] in It is a lightweight regressor (linear regression / tree model / lightweight neural network can all be implemented and solidified).

[0075] The residual vector and residual score are defined as follows: in This is a feature weight matrix used to emphasize network metrics that are sensitive to attacks.

[0076] After spatiotemporal alignment, the system estimates a reasonable network state range using flight control and topology states. This includes factors such as the relationship between distance, relative motion, and link quality, or using lightweight regression models to predict reasonable ranges for packet loss, RTT, and retransmission under current maneuvering conditions. The system then calculates the residual between actual network observations and predictions, generating a cross-modal residual score. When injection attacks forge control or telemetry, causing state inconsistencies, or replay attacks cause control lag leading to asynchronous responses, the cross-modal residual continuously increases within the window, becoming a crucial basis for countering covert attacks.

[0077] After obtaining the replay fingerprint score, the injected semantic score, and the cross-modal residual score, the airborne side combines the output of the lightweight detection model to form a local evidence vector.

[0078] A lightweight detection model is used to supplement abnormal patterns that rules cannot cover. Its inputs are a unified feature vector and evidence scores, and its output is the base probability for four states: "normal," "replay," "injection," and "composite." Let the classification framework be:

[0079] These represent normal (N), replay (R), injection (I), and composite (C), respectively.

[0080] The lightweight detection model outputs the posterior probability: The three scores (replay fingerprint score, injection semantic score, and cross-modal residual score) are expert evidence calculated based on rules. They are obtained through sliding window statistics, semantic constraint checks, and cross-modal prediction residual calculations, respectively. These three scores are manually designed interpretable indicators, and the higher the score value, the greater the probability of the corresponding attack type.

[0081] The posterior probability is the output of the lightweight detection model. The model takes a preprocessed feature vector as input and outputs the probability distribution for four states (normal N, replay R, injection I, and composite C), such as P(N|feature)=0.7, P(R|feature)=0.1, P(I|feature)=0.1, and P(C|feature)=0.1. It is called "posterior" because it is a probability estimate of the state after observing the data features. This model is trained by machine learning and is used to capture complex abnormal patterns that cannot be covered by rules.

[0082] The base probability is actually another name for the "posterior probability." In the documentation, they refer to the same thing: the four-state probabilities output by a lightweight model. They are just different terms used in different contexts. When it is emphasized that "this is the output of the model," it is called the posterior probability. When it is emphasized that "this is a component to be integrated with other evidence," it is called the base probability.

[0083] The logical relationship between them is as follows: The system simultaneously obtains evidence from two paths. The first path is a rule-based evidence chain, which calculates three scores through a manually designed algorithm. These scores directly reflect the feature strength of a specific attack pattern. The second path is a model inference chain, which inputs preprocessed features into a lightweight neural network or machine learning model, and the model outputs the posterior probabilities (base probabilities) of four states. Then, the system integrates the results of these two paths to form a comprehensive evidence vector. This vector contains seven elements: [replay score, injection score, cross-modal score, P(N), P(R), P(I), P(C)]. In other words, the three rule-based scores and the four model output probabilities are equally placed in a single vector as comprehensive evidence.

[0084] The comprehensive evidence vector is: To achieve neighborhood-based collaborative verification, this invention further maps local evidence to the basic probability allocation (BBA) of DS evidence theory and explicitly introduces uncertainty mass to absorb noise and boundary samples. Mapping is based on probability assignment .

[0085] Define a linear scoring function: An uncertainty score is also set: in This indicates the degree of conflict of evidence or instability at the boundaries (e.g., increased uncertainty when the scores of the three categories are close).

[0086] The final BBA definition is as follows: When replaying fingerprints is significant and injected semantic evidence is weak, local evidence tends to favor replay; when injected semantic evidence is significant and replay evidence is weak, local evidence tends to favor injection; when both types of evidence are significant or cross-modal residuals are significant, local evidence tends to favor the composite state; when all evidence is weak and the model output is normal, local evidence tends to favor the normal state; when evidence is conflicting or near a threshold, the system improves uncertainty quality to avoid prematurely drawing erroneous conclusions. Through this mapping process, this invention achieves a unified expression of "interpretable evidence + learned model probability," providing consistent input for subsequent fusion.

[0087] In UAV swarm networks, single-node observations can be affected by link obstruction, local viewpoint bias, or short-term disturbances, easily leading to false alarms and missed alarms. This invention performs evidence fusion verification within the neighborhood, improving decision stability through group consistency. Each node generates BBA (local evidence) locally, exchanges evidence with neighboring nodes, and performs DS combination of the received multi-source evidence locally. Considering the possibility of node failures, abnormal nodes, or even adversarial nodes in the UAV swarm environment, the system introduces a credibility discount coefficient on neighboring evidence before fusion. This discount coefficient is jointly determined by link quality, historical consistency, and reputation, thereby reducing the interference of low-credibility nodes on the fusion result.

[0088] Let the bee colony communication diagram be as follows: ,node The neighborhood of .node Collect neighborhood evidence A discounting process is applied to suppress unreliable evidence. The discount factor is defined as:

[0089] in For link quality / reputation metrics, This serves as an indicator of evidence conflict at this node. It is Sigmoid.

[0090] BBA after discount: The two pieces of evidence were successively fused using the standard Dempster combination rule. , : The conflict coefficient is calculated as follows: The neighborhood evidence is iteratively synthesized to obtain The output category is the one with the highest belief value:

[0091] The state is then determined using a dual threshold method: When the conflict level increases significantly, the system incorporates the conflict into the uncertainty quality to prevent distortion in the fusion process under strong conflict. After fusion, the system uses the maximum belief category as the output category and employs a dual-threshold mechanism to distinguish between "suspicious" and "confirmed" attacks: when the maximum belief exceeds the high threshold, it is judged as a confirmed attack; when it falls between the low and high thresholds, it is judged as a suspicious attack; otherwise, it is judged as normal. This mechanism makes the detection results stable and progressive, reducing the propagation of false alarms under dynamic topology conditions of UAV swarms.

[0092] In the embodiments provided by this invention, when a composite intrusion is determined, the method further includes composite attack disassembly and attack source localization, including: calculating the relative magnitude of replay contribution and injection contribution to determine the dominant type of composite attack; the replay contribution and injection contribution are calculated by combining the corresponding scores and short-window trend terms; calculating the risk value of each node in the cluster using the node risk formula, generating a Top-K list of suspicious nodes by sorting by risk value; the node risk formula integrates the node's evidence score, fusion conflict degree, and cross-modal residual; combining the neighborhood fusion conflict characteristics and cross-modal residual propagation characteristics, identifying a set of suspicious links, and using the UAV cluster communication topology as a graph structure, expanding outward from high-risk nodes to determine the attack impact range.

[0093] Upon confirmation of an attack, this invention not only issues an alert but also outputs interpretable diagnostic information. For combined attacks or attacks where injection and replay are both significant (specifically, calculating the relative magnitudes of replay and injection contributions, comparing these two contributions to determine which attack type is dominant, e.g., a "combined attack dominated by replay and supplemented by injection" or "combined attack dominated by injection and supplemented by replay"), such detailed analysis provides more precise guidance for subsequent emergency response. This is because different handling strategies may be required for combined attacks with different dominant types; for example, a replay-dominated combined attack may require prioritizing time-series verification and sequence number monitoring, while an injection-dominated combined attack requires prioritizing semantic consistency checks and task-stage constraint verification), the system decomposes the evidence based on contribution, defining the contribution as follows:

[0094] in This represents the short-window trend term (such as the first difference or regression slope).

[0095] After calculating the relative magnitudes of replay contribution and injection contribution, the system provides a conclusion of "replay-dominated composite attack" or "injection-dominated composite attack", thus providing a clear direction for subsequent handling strategies.

[0096] Simultaneously, the system sorts node-level risks, forming a Top-K list of suspicious nodes. Node risk is defined as follows:

[0097] Based on this, the node ranking is obtained, and link anomalies are evaluated by combining neighborhood fusion conflict and cross-modal residual propagation characteristics, outputting a set of suspicious links and an estimated impact range. The impact range is determined by using the UAV swarm communication topology as a graph structure, and the affected subnets are identified by expanding outward from high-risk nodes, enabling the system to have the capability from "detection" to "location", thus enhancing its engineering application value.

[0098] In the embodiments provided by this invention, dynamic model updating is also included, specifically: a drift statistic is constructed based on the proportion of nodes with abnormal scores exceeding the standard in the neighborhood and the average conflict degree of neighborhood fusion. The average conflict degree of neighborhood fusion is the average conflict coefficient when a node fuses evidence with all its neighboring nodes. When the drift statistic continuously exceeds a preset threshold for L time windows, federated incremental training is initiated. The airborne end performs incremental training based on new local data and reports parameter increments and training summaries. The aggregation end weights and aggregates the parameter increments according to node credibility and data quality to form a global model and issues updates to all nodes in the cluster. The switching of UAV cluster mission phases, environmental changes, and changes in adversarial strategies can cause data distribution drift. To adapt to data distribution drift, this invention uses evidence trend-triggered federated updates.

[0099] Define drift statistics: in The average conflict degree of neighborhood fusion.

[0100] When the following conditions are met: For L time windows Then, a federated update is initiated. The federated update performs local incremental training on the onboard end and reports parameter increments and training summaries. The aggregation end weights and aggregates data according to node credibility and data quality to form a global model, and then distributes the new model parameters, corresponding thresholds, and fusion configurations to all nodes. This update method matches model evolution with the risk situation, avoids the communication waste caused by fixed-period updates, and also improves the system's adaptability to new injection and replay variants.

[0101] Drift statistics The calculation formula is: This formula consists of two parts. The first part performs statistics on all nodes i within the neighborhood V, checking whether the cross-modal residual score and replay fingerprint score of each node exceed their respective thresholds. and If the number of nodes exceeds the limit, the function returns 1; otherwise, it returns 0. Then, the sum of all nodes is calculated and divided by the total number of neighboring nodes. This yields the percentage of nodes in the neighborhood that have abnormally exceeded their scores; Part Two It is the average conflict degree of neighborhood integration. This represents the average degree of evidence conflict among neighboring nodes during the DS evidence fusion process. The degree of conflict increases when there is a large contradiction between the BBA of different nodes. ρ is a weighting coefficient used to control the contribution ratio of the degree of conflict in the drift statistic.

[0102] The meaning of evidence trend is to observe the drift statistic. The system uses patterns of change over a time window to determine whether a persistent shift in data distribution is occurring. Specifically, the system monitors... Time series, if If the score exceeds the preset threshold for L consecutive time windows (e.g., 10 consecutive time windows), a clear "upward trend" is formed. This indicates that more and more nodes in the neighborhood are abnormally exceeding the threshold or that the evidence conflict between nodes is continuously intensifying. This continuous abnormal increase is the manifestation of the "evidence trend". It indicates that the current model may no longer be adapted to the new data distribution. This may be due to changes in normal behavior patterns caused by task phase switching, the emergence of new attack patterns that the original model cannot accurately identify, or environmental changes that cause a systematic shift in network features. Therefore, federated incremental training needs to be triggered to update the model to adapt to the new data distribution.

[0103] Simply put: Evidence trend = through monitoring drift statistics Time series changes are used to discover persistent shifts in data distribution. When L consecutive windows exceed the threshold, it is considered that a drift trend requiring model updates has emerged, rather than an occasional short-term fluctuation.

[0104] Average Conflict Degree of Neighborhood Integration The evidence fusion process, derived from Section 5.6 "Neighborhood DS Evidence Fusion and Cluster Confirmation Mechanism," involves the following specific calculation steps: Step 1: Calculation of the conflict coefficient for a single fusion: When node i wants to fuse two BBAs (e.g., its own), and neighbor node j When calculating the conflict coefficient, use the formula provided in the document. The meaning of K here is the sum of the mass products of the two BBAs in mutually exclusive states. For example, if The "normal" state was assigned a mass of 0.8. The "Replay" state was assigned a quality of 0.7, since "Normal" and "Replay" are mutually exclusive. If so, this term contributes 0.8 × 0.7 = 0.56 to the conflict coefficient K. The total conflict coefficient K is obtained by summing the products of all mutually exclusive state pairs. The larger K is, the more contradictory the judgments of the two pieces of evidence are.

[0105] Step 2: Multiple fusions within the neighborhood generate multiple conflict coefficients: The neighborhood V of node i contains several neighboring nodes, let the size of the neighborhood be... Node i needs to merge its own BBA with the BBA of each neighboring node. Assume there are 3 neighboring nodes in the neighborhood. This will result in three fusion calculations, yielding three conflict coefficients. Each K value reflects the degree of evidence conflict between node i and its corresponding neighbor.

[0106] Step 3: Calculate the average conflict degree: Average conflict degree of neighborhood fusion This involves averaging the conflict coefficients obtained when node i is merged with all its neighboring nodes, using the following formula: In other words, summing all the K values ​​and dividing by the number of neighbors gives the average value, which reflects the overall level of evidence conflict within node i's neighborhood. The high value indicates that node i and its neighbors generally disagree on the current state. A very low value indicates that the judgments of the nodes in the neighborhood are relatively consistent.

[0107] When the model is dynamically updated, version management is included before the global model is deployed. This includes binding the preprocessing mapping function, feature signature, and current model parameters into a version triplet, recording version change descriptions and regression verification results. The regression verification results are quantitative evaluation data of core performance indicators such as detection accuracy, false positive rate, false negative rate, confidence stability, and evidence fusion conflict level, based on simulation / test data covering multiple flight conditions, topology changes, and attack modes, as well as compliance judgment conclusions for version deployment. To achieve traceable and rollback-capable engineering operation, this invention defines the model version as a triplet: preprocessing frozen state, feature signature, and model parameter set. After each federated update, the system generates a new triplet version and stores it in the database, while simultaneously recording change descriptions and regression verification results.

[0108] Before loading the new version of the global model on the onboard device, the consistency between the local feature signature and the new version is verified. If the verification fails, loading is rejected. If the new version experiences performance anomalies during online operation, the model, preprocessing mapping function, and feature signature are synchronously rolled back to the version triplet of the previous stable version according to the version pointer. In other words, the onboard device first verifies the consistency between the feature signature and the local input mapping before loading the new version to prevent implicit errors caused by input structure mismatch; after loading, an online switch is performed and the switch time is recorded.

[0109] If the false alarm rate increases significantly, the detection performance decreases, or the operation and maintenance are deemed abnormal during online operation, the system will roll back to the previous stable version according to the version pointer, and at the same time restore the corresponding preprocessing frozen state and feature signature to ensure end-to-end consistency recovery.

[0110] By using triplet binding, the system avoids engineering risks such as "the model changes but the preprocessing doesn't / the features change but the model isn't synchronized," ensuring a reliable closed-loop operation.

[0111] To ensure the reliability of version evolution, this invention incorporates simulation regression verification and visualized operation and maintenance into the closed loop of system operation.

[0112] The simulation module continuously generates data covering different flight conditions, different topology changes, and different injection / replay modes, which are used to supplement training data, verify version regression, and reproduce online issues.

[0113] Before each new version is released, regression testing is performed based on a unified preprocessing frozen state to evaluate false positives and false negatives, confidence stability and fusion conflict level under multiple operating conditions, and the verification results are used as a necessary condition for the version to be included in the repository.

[0114] The visual console displays the real-time attack situation, risk trends, ranking of suspicious nodes and links, scope of impact, and reasons for federal update triggers and version change records of drone swarms. This enables the detection results to be interpretable, the model evolution to be traceable, and the operation and maintenance decisions to be supported, ultimately forming a long-term operational closed-loop system for drone swarm intrusion detection.

[0115] This invention also provides a network intrusion detection system for drone swarms, used to implement the aforementioned network intrusion detection method for drone swarms, comprising: The data acquisition and preprocessing module is used to synchronously acquire network communication data and flight control telemetry data of each node in the UAV cluster at a fixed time step, form a unified sample through spatiotemporal alignment of sliding window, generate preprocessed features based on frozen preprocessing mapping function, and complete feature signature consistency verification.

[0116] The attack feature quantization module is used to calculate the replay fingerprint score, injection semantic score, and cross-modal residual score, thereby enabling the quantitative characterization of replay attacks, injection attacks, and cross-modal state mismatches.

[0117] The local evidence processing module is used to output the basic probabilities of four states through a lightweight detection model, integrate the rule-based scores and basic probabilities into a comprehensive evidence vector, and map it to the basic probability allocation of DS evidence theory.

[0118] The cluster collaborative detection module is used to realize the basic probability allocation and exchange between nodes, process neighborhood evidence based on the credibility discount coefficient and complete the iterative fusion of multi-source evidence, and determine the intrusion detection result through dual belief threshold.

[0119] The attack tracing module is used to dissect the dominant attack type during complex intrusions, generate a Top-K list of suspicious nodes, identify suspicious link sets, and determine the scope of attack impact.

[0120] The model update and version management module is used to build drift statistics and trigger federated incremental training, realize the binding, storage, verification and rollback of model version triples, and complete the regression verification and visual operation and maintenance of the new version.

[0121] It should be noted that any parts not disclosed or specifically described in this invention are existing technology or conventional configurations, and their specific structures and working principles will not be elaborated further. In this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0122] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Other modifications can be readily implemented by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and examples shown and described herein.

Claims

1. A network intrusion detection method for unmanned aerial vehicle (UAV) swarms, characterized in that, include: The network communication data and flight control telemetry data of each node in the UAV cluster are collected synchronously by a sliding window and spatiotemporally aligned to form a unified sample. The samples of each node are then uniformly converted into standardized preprocessed features through a preprocessing mapping function. To address replay attacks, the repetition rate, backtracking rate, out-of-order arrival ratio, and arrival interval drift are calculated within a sliding window and weighted to obtain a replay fingerprint score. To address injection attacks, the violation degree of field legal domain constraints, the violation degree of task phase constraints, and the abrupt change degree of short-term command distribution are calculated and weighted to obtain an injection semantic score. A cross-modal prediction function is constructed based on the coupling relationship between the UAV network state and the flight control state. The cross-modal prediction function calculates the residual vector between the actual network observations and the predicted values, and the residual vector is used to calculate the cross-modal residual score through the feature weight matrix. The preprocessed features are input into the lightweight detection model, and the basic probabilities of the nodes in four states are output: normal, replay, injection, and composite. The replay fingerprint score, injection semantic score, cross-modal residual score and the basic probabilities of the four states are integrated into a comprehensive evidence vector. The uncertainty score of the comprehensive evidence vector is calculated, and the comprehensive evidence vector is mapped to the basic probability allocation of the DS evidence theory with the uncertainty score as the correction term. Each node shares a basic probability allocation with its neighboring nodes. The basic probability allocation of the neighboring nodes is discounted to obtain a discounted probability allocation. Each node iteratively merges its own basic probability allocation with the discounted probability allocation of all neighboring nodes to obtain multi-source evidence. The belief value of the state with the highest probability among the four states is calculated using the multi-source evidence. The belief value is compared with two preset belief thresholds, and three detection results are output for each state: normal, suspicious, and confirmed. Based on the detection results, it is determined whether an intrusion has occurred.

2. The network intrusion detection method for unmanned aerial vehicle (UAV) swarms according to claim 1, characterized in that, The preprocessing mapping function's parameters are used to generate a unique feature signature. The preprocessing mapping function's parameters include a fixed field mapping table, an encoding dictionary, a normalized numerical range, and a missing value handling strategy. The signature is a unique hash value for all parameters, used for consistency verification during model training, inference, updating, and rollback.

3. The network intrusion detection method for unmanned aerial vehicle (UAV) swarms according to claim 1, characterized in that, The arrival interval drift is the degree of deviation of the mean and variance of the current window message arrival interval from the normal baseline statistics, and is calculated by weighting the mean drift and the variance drift. The short-term instruction distribution abruptness is calculated using the Jensen-Shannon distance to determine the difference between the current window control instruction category distribution and the historical normal baseline distribution. The cross-modal prediction function includes a lightweight regressor, which is one of a linear regression model, a tree model, or a lightweight neural network.

4. The network intrusion detection method for unmanned aerial vehicle (UAV) swarms according to claim 1, characterized in that, The uncertainty score is obtained based on the degree of conflict of the replay fingerprint score, the injected semantic score, the cross-modal residual score, and the degree of dispersion of the distribution of the basic probability. When there is obvious conflict of evidence or the probability distribution is uniform, the uncertainty score is increased to reduce the risk of misjudgment.

5. A network intrusion detection method for unmanned aerial vehicle (UAV) swarms according to claim 1, characterized in that, The iterative fusion adopts the Dempster combination rule, and the evidence conflict coefficient is absorbed into the uncertainty quality during the fusion process to avoid distortion of the fusion result caused by strong conflicts; The credibility discount coefficient is obtained by normalizing the link quality / reputation index and the node evidence conflict index using the Sigmoid function.

6. The network intrusion detection method for unmanned aerial vehicle (UAV) swarms according to claim 1, characterized in that, When a complex intrusion is identified, the process also includes dismantling the complex attack and locating the attack source, including: The relative magnitudes of replay contribution and injection contribution are calculated to determine the dominant type of composite attack. The replay contribution and injection contribution are calculated by combining the corresponding scores and short window trend terms. The risk value of each node in the cluster is calculated by the node risk formula, and a Top-K list of suspicious nodes is generated by sorting the risk values. The node risk formula integrates the evidence score of the node, the degree of conflict, and the cross-modal residual. By combining the characteristics of neighborhood fusion conflict and cross-modal residual propagation, suspicious link sets are identified, and the attack impact range is determined by expanding outward from high-risk nodes using the communication topology of UAV swarms as a graph structure.

7. A network intrusion detection method for unmanned aerial vehicle (UAV) swarms according to claim 1, characterized in that, It also includes dynamic model updates, including: A drift statistic is constructed based on the proportion of nodes with abnormal scores exceeding the standard in the neighborhood and the average conflict degree of neighborhood fusion. The average conflict degree of neighborhood fusion is the average conflict coefficient when a node merges evidence with all its neighboring nodes. When the drift statistics continuously exceed the preset threshold for L time windows, federated incremental training is initiated. The onboard unit performs incremental training based on local new data and reports the parameter increments and training summary. The aggregation unit weights and aggregates the parameter increments according to node credibility and data quality to form a global model and sends the update to all nodes in the cluster.

8. A network intrusion detection method for unmanned aerial vehicle (UAV) swarms according to claim 7, characterized in that, When the model is dynamically updated, version management is also included before the global model is deployed, including: The preprocessing mapping function, feature signature, and current model parameters are bound to a version triplet, and the version change description and regression validation results are recorded. The regression verification results are quantitative evaluation data of core performance indicators such as detection accuracy, false alarm rate, false negative rate, confidence stability, and evidence fusion conflict level, based on simulation / test data covering multiple flight conditions, topology changes, and attack modes, as well as compliance judgment conclusions for version deployment.

9. A network intrusion detection method for unmanned aerial vehicle (UAV) swarms according to claim 8, characterized in that, Before loading the new version of the global model on the airborne end, verify the consistency between the local feature signature and the new version. If the verification fails, refuse to load. If a performance anomaly occurs during the online operation of the new version, the model, preprocessing mapping function, and feature signature will be synchronously rolled back to the version triplet of the previous stable version according to the version pointer.

10. A network intrusion detection system for drone swarms, used to implement the network intrusion detection method for drone swarms as described in any one of claims 1 to 9, characterized in that, include: The data acquisition and preprocessing module is used to synchronously acquire network communication data and flight control telemetry data of each node of the UAV cluster at a fixed time step, form a unified sample through spatiotemporal alignment of sliding window, generate preprocessing features based on frozen preprocessing mapping function and complete feature signature consistency verification. The attack feature quantification module is used to calculate the replay fingerprint score, injection semantic score, and cross-modal residual score, thereby realizing the quantitative characterization of replay attacks, injection attacks, and cross-modal state mismatch. The local evidence processing module is used to output the basic probabilities of four states through a lightweight detection model, integrate the rule-based scores and basic probabilities into a comprehensive evidence vector, and map it to the basic probability allocation of DS evidence theory. The cluster collaborative detection module is used to realize the basic probability allocation and exchange between nodes, process neighborhood evidence based on the credibility discount coefficient and complete the iterative fusion of multi-source evidence, and determine the intrusion detection result through the dual belief threshold. The attack tracing module is used to disassemble the dominant attack type during complex intrusions, generate a Top-K list of suspicious nodes, identify suspicious link sets, and determine the scope of attack impact. The model update and version management module is used to construct drift statistics and trigger federated incremental training, realize the binding, storage, verification and rollback of model version triples, and complete the regression verification and visual operation and maintenance of the new version.

Images