CN-122024525-A - Unmanned aerial vehicle radio frequency individual hijacking recognition system integrating multisource federal learning and multi-mode sensing

CN122024525ACN 122024525 ACN122024525 ACN 122024525ACN-122024525-A

Abstract

The invention aims to provide an unmanned aerial vehicle radio frequency individual hijacking identification system integrating multi-source federal learning and multi-mode perception, wherein a multi-mode data acquisition and preprocessing module, a local feature extraction and initial discrimination module and a multi-mode fusion and hijacking risk assessment module form a front end detection link deployed at an unmanned aerial vehicle end and/or an edge node, a multi-source federal learning and global model aggregation module, a security and privacy protection module and an online sample acquisition and self-adaptive updating module form a rear end training link deployed at a cloud supervision center, and the two form an end-side-cloud collaborative unmanned aerial vehicle hijacking identification closed loop system. The method can form a federal learning system aiming at multi-mode data depth fusion such as flight control, radio frequency communication, airborne vision, ground track and the like.

Inventors

NIU NING
XIA YINGJUN
WANG GUANJUN
YANG QIAN
SHEN PENGPENG
LIU MALIANG

Assignees

海南大学

Dates

Publication Date: 20260512
Application Date: 20251230

Claims (10)

1. Unmanned aerial vehicle radio frequency individual hijacking recognition system integrating multisource federal learning and multi-mode perception is characterized by comprising: The multi-mode data acquisition and preprocessing module is used for synchronously acquiring flight control data, radio frequency and communication data, airborne vision data and ground monitoring track data at the unmanned plane end and the ground monitoring end through sliding time windows, and carrying out filtering, interpolation and alignment and normalization processing on various data to generate standardized multi-mode data fragments under a unified time reference; The local feature extraction and initial discrimination module is deployed at the unmanned aerial vehicle end and/or the edge computing node and is used for respectively constructing a flight control behavior coding sub-network, a radio frequency and instruction mode coding sub-network and a vision/environment coding sub-network aiming at the standardized multi-mode data fragment so as to extract a flight control feature vector, a radio frequency communication feature vector and a vision environment feature vector, and locally output corresponding mode level hijacking risk scores and confidence degrees; The multi-mode fusion and hijack risk assessment module is deployed at the unmanned plane end and/or the edge computing node and is used for executing mode alignment and interaction attention fusion on the flight control feature vector, the radio frequency communication feature vector and the visual environment feature vector based on a global multi-mode hijack recognition model issued from a cloud supervision center to obtain a fusion situation vector, and outputting a continuous hijack risk score of a 0-1 interval and a hierarchical hijack state judgment result at least comprising normal, suspicious and highly suspected hijack according to the fusion situation vector; The safety control and treatment decision module is used for selecting a corresponding treatment strategy from a preset safety strategy library according to the grading hijacking state judgment result, converting the treatment strategy into at least one control instruction comprising link switching, flight speed and height limitation, forced navigation, emergency hovering, task suspension and/or alarm information pushing to a ground supervision platform, and sending the control instruction to an unmanned plane flight control system and/or a ground supervision system after interface adaptation so as to realize automatic or semi-automatic safety treatment; The multi-source federal learning and global model aggregation module is deployed in the cloud supervision center and is used for receiving the multi-mode coding sub-network parameter increment uploaded by a plurality of unmanned aerial vehicles and/or edge nodes on the premise of not uploading original data, carrying out abnormal update detection and self-adapting adjustment on node aggregation weights according to the similarity between the parameter increment of each node and the average value of all the parameter increments in each federal training, updating the global multi-mode hijacking identification model parameters in a weighted summation mode and transmitting the parameters to each node; The security and privacy protection module is used for applying differential privacy noise to parameter increment uploaded by each node in the multi-source federal learning and global model aggregation process, carrying out encryption aggregation by adopting a secure multiparty calculation or secure aggregation protocol, and simultaneously recording a participating node identifier, an aggregation result hash value, an abnormal node list and a timestamp to form a traceable security audit log; The online sample acquisition and self-adaptive updating module is used for marking the flight fragment which is judged to be highly suspected hijacked, corresponding multi-modal data and treatment results as a novel risk sample and caching the novel risk sample in the system operation process, the novel risk sample is brought into a local training data set in the subsequent federal training round, and the global multi-modal hijacked identification model is updated through the multi-source federal learning and global model aggregation module, so that the system has continuous adaptive capacity to the novel hijacking mode; The multi-mode data acquisition and preprocessing module, the local feature extraction and initial discrimination module and the multi-mode fusion and hijack risk assessment module form a front end detection link deployed at an unmanned aerial vehicle end and/or an edge node, the multi-source federal learning and global model aggregation module, the safety and privacy protection module and the online sample acquisition and self-adaptive update module form a rear end training link deployed at a cloud supervision center, and the two form an unmanned aerial vehicle hijack identification closed loop system with end-side-cloud cooperation.
2. The unmanned aerial vehicle radio frequency individual hijacking recognition system integrating multi-source federal learning and multi-mode perception according to claim 1, wherein the multi-mode data acquisition and preprocessing module comprises: The flight control data sub-module is used for acquiring flight control parameters such as attitude angle, angular speed, linear speed, track height, acceleration, motor rotation speed and/or accelerator output and the like from an unmanned aerial vehicle flight control system, forming a fixed length time sequence with a first sampling frequency, carrying out band-pass filtering on each channel data, interpolating and supplementing missing samples, and carrying out normalization processing according to a preset range or statistic; The radio frequency and communication data sub-module is used for collecting radio frequency signals and communication logs of a remote control link and/or a graphic transmission link, performing short-time Fourier transform on the radio frequency signals in a sliding time window to generate a power spectral density spectrogram, performing logarithmic compression and spectral amplitude normalization on the spectrogram, aligning a control instruction sequence and handshake/reconnection events according to time stamps, and then encoding the aligned control instruction sequence and handshake/reconnection events into a fixed-length time sequence; the airborne visual data sub-module is used for acquiring video streams from forward-looking and/or downward-looking cameras of the unmanned aerial vehicle, extracting key frames according to inter-frame changes or fixed frame rates in a sliding time window, and performing target region cutting and size scaling on the key frames to obtain standardized images; The ground monitoring data sub-module is used for receiving the space position and identity information of the unmanned aerial vehicle output by the ground radar, the passive monitoring equipment or the low-altitude monitoring network, resampling, unifying coordinates, filtering and smoothing the track sequence, and normalizing the derived quantities of position, speed, acceleration and the like.
3. The unmanned aerial vehicle radio frequency individual hijacking recognition system integrating multi-source federal learning and multi-mode perception according to claim 1, wherein the local feature extraction and initial discrimination module comprises: The flight control behavior coding sub-network is used for extracting local time characteristics from the flight control data time sequence through a one-dimensional convolution layer, modeling the whole sequence through a cyclic neural network or a time sequence convolution network and outputting a fixed-length flight control characteristic vector; The radio frequency and instruction mode coding sub-network is used for extracting frequency-time local mode characteristics of the spectrogram through a two-dimensional convolutional neural network to obtain a radio frequency characteristic vector, extracting instruction mode characteristic vectors of a control instruction time sequence through a sequence model, and connecting or fusing the two to obtain a radio frequency communication characteristic vector; the visual/environment coding sub-network is used for extracting high-level visual features reflecting the situation information such as a no-fly zone, a sensitive target, surrounding aircrafts or strong light interference and the like from the key frame image through a target detection network and/or a semantic segmentation network, and fusing multi-frame results in a time dimension to form a visual environment feature vector; The modal risk estimation unit is used for outputting corresponding modal level hijacking risk scores and confidence degrees based on the local lightweight discrimination network input of the modal feature vectors, and adjusting the confidence degrees of the modalities according to the consistency among the multi-modal risk scores.
4. The unmanned aerial vehicle radio frequency individual hijacking identification system integrating multi-source federal learning and multi-modal awareness according to claim 1, wherein the multi-modal fusion and hijacking risk assessment module comprises: The modal alignment unit is used for carrying out time alignment on the flight control feature vector, the radio frequency communication feature vector and the visual environment feature vector in a unified time window, mapping each feature vector to a feature space with the same dimension through linear transformation, and obtaining an aligned multi-modal feature set; The interaction attention fusion unit is used for calculating basic weights based on the confidence coefficient and the data quality index of each mode, carrying out adjustment and normalization on the basic weights according to the similarity of the feature vectors among the modes to obtain attention weights alpha f 、α r 、α v , and carrying out weighted summation on the feature vectors of each mode under the constraint of meeting alpha f +α r +α v =1 to obtain fusion situation vectors; the hijacking risk scoring unit is used for inputting the fusion situation vector into a feedforward neural network, outputting a hijacking risk score R in a 0-1 interval through a Sigmoid activation function, and dividing the R into three states of normal, suspicious and highly suspected hijacking according to a preset first threshold R1 and a second threshold R2.
5. The unmanned aerial vehicle radio frequency individual hijacking identification system integrating multi-source federal learning and multi-mode perception according to claim 1, wherein the multi-source federal learning and global model aggregation module comprises: The federal task arrangement unit is used for issuing unified multi-mode coding network structures and training super parameters to each participating unmanned aerial vehicle end and/or edge node, and setting the local training round number and uploading frequency according to calculation power and bandwidth of each node; The local training and parameter uploading unit is deployed at each participating node and is used for executing multiple rounds of local model training based on local historical flight data and annotation hijacking events, calculating local model parameter increment and uploading to the cloud; The abnormal update detection unit is used for calculating cosine similarity between each node parameter increment and all parameter increment mean values in each round of federal training, and marking a node as a suspicious node and reducing aggregation weight or eliminating the suspicious node when the similarity of the node is lower than a preset threshold value; and the robust aggregation unit is used for carrying out weighted summation on the residual parameter increment according to the number of the effective training samples of each node after eliminating or reducing the weight of the suspicious nodes, generating the updating quantity of the global model parameters and updating the global multi-mode hijacking identification model.
6. The unmanned aerial vehicle radio frequency individual hijacking identification system integrating multi-source federal learning and multi-mode perception according to claim 1 is characterized in that the security and privacy protection module is specifically used for locally adding random noise conforming to Laplacian distribution or Gaussian distribution to parameter increments to be uploaded according to preset privacy budgets at each participating node so as to meet differential privacy constraint, aggregating the noisy parameter increments by adopting a secure multiparty computing protocol with homomorphic encryption or secret sharing at a cloud supervision center so that the cloud can not reversely push out real parameter increments of any node while acquiring an aggregation result, and forming a tamper-proof security audit log for the following malicious node tracking and model updating process audit by recording participation node identifiers, aggregation result hash values, marked suspicious node identifiers and similarity values, model version numbers and time stamps for each round of federal training.
7. The multi-source federal learning and multi-mode aware unmanned aerial vehicle radio frequency individual hijack recognition system of claim 1, wherein the security control and handling decision module wherein the policy matching unit selects an enhanced monitoring and manual validation policy when the hijack risk score R is greater than the first threshold R 1 and less than the second threshold R 2 , the control instruction generating unit generates one or more control instructions including at least one of increasing data acquisition frequency, switching backup control links, reducing unmanned aerial vehicle flight speed and/or upper altitude limit, and pushing pre-warning information to a ground supervision platform, and when the hijack risk score R is greater than or equal to the second threshold R 2 , the policy matching unit selects a forced security handling policy, the control instruction generating unit generates one or more control instructions including at least triggering a back-up procedure, performing emergency hover, clearing current automated task airlines, and/or cutting off control links with existing remote controls and accessing into a supervisory channel, and the interface adapting unit converts the control instructions into a control message supported by the unmanned aerial vehicle flight control system and/or a ground supervision platform format for transmitting the supervision information.
8. The unmanned aerial vehicle radio frequency individual hijacking recognition system integrating multi-source federal learning and multi-mode perception according to claim 1, wherein the online sample acquisition and self-adaptive updating module is further used for marking multi-mode data fragments in a corresponding time window as hijacking positive samples and marking normal flight fragments adjacent to the multi-mode data fragments as comparison samples after a manual supervision personnel confirms a certain high suspected hijacking event, and preferentially extracting training batches containing the new marked hijacking positive samples in subsequent federal training, so that the global multi-mode hijacking recognition model carries out self-adaptive adjustment on recognition thresholds and decision boundaries of the hijacking modes, and detection time of similar novel hijacking events is shortened and reporting omission probability is reduced.
9. The unmanned aerial vehicle radio frequency individual hijacking identification system integrating multi-source federal learning and multi-modal awareness is characterized by adopting an end-side-cloud layered deployment structure, wherein an unmanned aerial vehicle end deploys at least partial functions of a multi-modal data acquisition and preprocessing module, a local feature extraction and initial judgment module and a multi-modal fusion and hijacking risk assessment module to finish multi-modal awareness and rapid hijacking risk judgment on site, an edge node deploys enhanced versions of the multi-modal fusion and hijacking risk assessment module and a safety control and treatment decision module to carry out regional risk comprehensive assessment and treatment instruction unified issuing after receiving feature vectors or judgment results uploaded by a multi-frame unmanned aerial vehicle, and a cloud supervision center deploys the multi-source federal learning and global model aggregation module, a safety and privacy protection module and a global situation visualization training module to execute cross-regional model, strategy template updating and global hijacking event statistical analysis.
10. The method for using the unmanned aerial vehicle radio frequency individual hijack recognition system integrating multi-source federal learning and multi-mode perception according to the claims 1-9 is characterized by comprising the following steps of S1, acquiring flight control data, radio frequency and communication data, airborne visual data and ground monitoring track data at an unmanned aerial vehicle end and a ground monitoring end, filtering, interpolating and normalizing all mode data in a sliding time window through a multi-mode data acquisition and preprocessing module to form standardized multi-mode data fragments under a unified time window, S2, performing feature coding on the standardized multi-mode data fragments at each unmanned aerial vehicle end and/or edge node through a local feature extraction and initial discrimination module to obtain flight control feature vectors, radio frequency communication feature vectors and visual environment feature vectors, and outputting a mode hijack risk score and confidence, S3, executing mode alignment and interaction attention fusion on the feature vectors by a multi-mode fusion and risk assessment module in an online reasoning stage, generating a fusion situation risk R and a classification state judgment result, S4, performing a high-level hijack control strategy and a high-level security control strategy to be matched with a suspected user interface, performing a high-level security control strategy in a control platform or a real-time interface, performing a high-level security interface, and a warning command, and a user-level being triggered by a user interface, and a user interface being triggered by a control system, the method comprises the steps of calculating and uploading parameter increment, enabling a cloud supervision center to conduct abnormal update detection and robust weighted aggregation on each parameter increment through a multi-source federal learning and global model aggregation module, updating a global multi-mode hijacking identification model and transmitting the updated model to each node, and enabling an online sample acquisition and self-adaptive updating module to mark a flight segment which is judged to be highly suspected to be hijacked and a treatment result thereof as a novel risk sample in a system operation process, and adding the novel risk sample into a local training data set in subsequent federal training to enable the identification capability of the global multi-mode hijacking identification model to be continuously enhanced in a novel hijacking mode.

Description

Unmanned aerial vehicle radio frequency individual hijacking recognition system integrating multisource federal learning and multi-mode sensing Technical Field The invention belongs to the field of unmanned aerial vehicle hijacking recognition, and particularly relates to an unmanned aerial vehicle radio frequency individual hijacking recognition system integrating multi-source federal learning and multi-mode perception. Background With the massive application of various unmanned aerial vehicles such as multiple rotor wings and fixed wings in the fields of security patrol, aerial photography, logistics distribution and the like, the unmanned aerial vehicle autonomously flies in a complex airspace for a long time to become a normal state. Meanwhile, the security events such as illegal invasion, remote hijacking, navigation spoofing and the like of the unmanned aerial vehicle are gradually increased, namely, an attacker can cause the unmanned aerial vehicle to deviate from a preset route, fly into a no-fly zone and even strike a sensitive target by stealing a control key, forging a remote control signal and interfering or spoofing a navigation and communication link, so that risks are brought to public security and important infrastructure. Therefore, how to identify hijacking behavior in real time and take disposal measures in time in the unmanned aerial vehicle flight process becomes an important point of industry attention. In the prior art, common means for unmanned aerial vehicle safety mainly comprise the following categories: Link protection techniques based on encryption and authentication. By introducing encryption algorithms, authentication protocols and session key update mechanisms between the drone and the remote/ground station, control commands can be prevented to some extent from eavesdropping, tampering or falsification. However, when the key management is improper, an attacker grasps a legal control terminal or utilizes a system vulnerability, the problem of malicious control under legal identity or misuse of internal personnel still can occur only by depending on link encryption, and the unmanned aerial vehicle is difficult to discover in time that the unmanned aerial vehicle is in an unauthorized control state. Rule-based or single modality based anomaly detection methods. The partial scheme is characterized in that fixed threshold values and geofence (Geo-fence) rules are set on unmanned plane flight control data or navigation data, and an alarm or a return is triggered once the altitude, the speed overrun or the position overrun is detected. There are also solutions to build machine learning models using single modality data (such as gps+imu, flight trajectory or radio frequency signals) to identify GPS spoofing, link anomalies or suspicious behavior. The method can find partial abnormal flight, but has obvious limitations: The coverage of a single mode on a complex attack scene is limited, for example, when an attacker simultaneously controls a flight control instruction and a radio frequency link, a single-view track can still be in a reasonable range, a rule threshold and a geofence are difficult to cope with a dynamic environment and a novel attack method, false alarm or missing report easily occurs in a noise environment, the model is mostly trained on a single local or centralized server, and the data distribution difference and privacy compliance problems among different units and different models are not considered. Intelligent recognition scheme based on centralized data aggregation. Along with deep learning being applied to unmanned aerial vehicle anomaly detection, there are schemes to attempt to intensively upload a large number of flight logs, radio frequency data and image data to a cloud or a central server to train a unified anomaly identification model. While focused training is beneficial for fully mining data value, multiple constraints are faced in practical deployment: unmanned plane data of different operators or different areas relate to sensitive task information and privacy, are not suitable for centralized aggregation, large-scale original multi-mode data uploading can greatly increase communication bandwidth and storage cost and increase data leakage risk, centralized training can not fully reflect differences of each area and each model, and the model has limited recognition capability on specific small sample attack scenes. A distributed security scheme for federal learning and edge computing was initially introduced. It is proposed that federal learning is used to cooperatively train an intrusion detection or anomaly detection model among a plurality of devices, data privacy is protected by uploading local model parameters instead of original data, and an edge computing node is used to bear part of security recognition tasks, so that cloud pressure is relieved. However, the existing schemes are concentrated on limited modes such as network intrusion detection