Unmanned aerial vehicle image transmission signal safety identification method and system based on ST-LFNet and active honeypot defense

Abstract

The invention discloses an unmanned aerial vehicle image transmission signal safety identification method and system based on ST-LFNet and active honeypot defense, and is applied to the cross technical field of wireless communication signal processing and next-generation brain-like artificial intelligence. Mapping to a time-frequency domain through short-time Fourier transform; a dynamic computing resource scheduling mechanism based on threat scores is adopted, in a low threat mode, only a main processor is started to run an ST-LFNet network to carry out space-time fluid feature extraction, and in a high threat mode, a security coprocessor is activated to carry out active honeypot camouflage based on the game theory; the voltage drop and the temperature gradient in the chip are monitored in real time, the KL divergence of real-time observation signal distribution and the reference normal template is calculated, and when the KL divergence suddenly increases and exceeds a preset threshold value, a trapping mode is triggered immediately. According to the invention, the unification of high-precision signal supervision and high-level physical security protection is realized.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of wireless communication signal processing, next-generation neuromorphic AI, hardware security, and embedded high-performance computing, and more specifically to a method and system for secure identification of drone image transmission signals based on ST-LFNet and proactive honeypot defense. Background Technology

[0002] With the rapid development of the low-altitude economy, the widespread application of unmanned aerial vehicles (UAVs) has brought serious challenges to public safety and privacy protection. Effective regulation of "unauthorized" UAVs relies heavily on the accurate interception and identification of their downlink video transmission signals. These signals carry video streams and telemetry data, and their physical layer waveforms contain unique "radio frequency fingerprints" that can identify device models and even individual identities.

[0003] However, existing physical layer recognition technologies face the fundamental obstacle of signal non-stationarity. First, the high-dynamic Doppler effect is significant. The high-speed maneuvers of racing or military drones cause severe Doppler shifts, disrupting the subcarrier orthogonality of OFDM signals and leading to spectral distortion, broadening, and aliasing. Second, multipath fading exhibits rapidly time-varying characteristics. In urban canyons or complex terrain, the channel impulse response (CIR) fluctuates dramatically over time. Limitations of existing models: Traditional convolutional neural networks (CNNs) or long short-term memory networks (LSTMs) are essentially static mapping models based on discrete time steps. They implicitly assume that signal features are stationary within the sampling interval. This "discretization assumption" has an inherent contradiction with the "continuous dynamic signal," making it difficult for the model to capture rapidly changing dynamic features. When drones perform high-maneuvering maneuvers such as sharp turns or dives, traditional models are prone to a sharp drop in recognition rate due to feature tracking failure.

[0004] To overcome the limitations of discrete models, institutions such as MIT proposed the Liquid Neural Network (LNN). Inspired by biological neural systems (such as Caenorhabditis elegans), LNNs directly model the continuous state evolution of neurons using ordinary differential equations (ODEs). Unlike traditional networks, LNNs have two core advantages: first, causal reasoning ability, enabling them to understand the physical processes of data changes; and second, adaptability, with their time constant varying with the input, adapting to the container (data) like a "liquid."

[0005] However, the application of LNNs to high-frequency radio frequency signal processing is currently unexplored. More challenging is the fact that LNNs involve complex numerical integration operations (such as the Runge-Kutta method). On resource-constrained edge devices, achieving efficient computation while ensuring numerical stability is a pressing problem. Existing general-purpose processors (CPUs) or graphics processing units (GPUs) are inefficient and consume excessive power when handling such fine-grained serial iterative calculations, making them unsuitable for deployment on portable detection terminals.

[0006] Furthermore, when edge devices carrying high-value AI recognition models are deployed in physically uncontrollable outdoor environments, the devices themselves become the primary target of attackers. Side-channel attacks (SCA) are a highly threatening non-intrusive attack method. Attackers do not need to crack encryption algorithms; they only need to analyze the physical information unintentionally leaked during device operation (such as power consumption fluctuations, electromagnetic radiation, and execution time) to steal model parameters or keys.

[0007] Especially for ODE-based LNN models, their numerical integration process exhibits highly structured physical fingerprints. For example, the four-step iterative computation of the RK4 solver generates periodic characteristic power ripples, which attackers can easily isolate using differential energy analysis (DPA) or correlated electromagnetic analysis (CEMA). Among existing defense methods, random masking significantly increases computational latency, failing to meet the demands of real-time signal processing; while passive physical shielding is ineffective against low-frequency magnetic fields and increases equipment size and cost.

[0008] In summary, the current field of drone monitoring faces a dual technological paradox of algorithms and hardware: on the one hand, in order to deal with the non-stationary signals of high-speed moving targets, there is an urgent need to introduce continuous-time neuromorphic computing models based on ODE (such as LNN), but the high computational complexity of such models makes it difficult to deploy them efficiently on edge devices; on the other hand, the numerical iteration process unique to LNN models exposes significant side-channel characteristics, making edge devices extremely vulnerable to physical attacks, while traditional passive defense methods further deteriorate the real-time performance of the system.

[0009] Therefore, how to provide a new technological paradigm that can accurately capture continuous dynamic signal characteristics by using differential equation modeling, and mask the physical fingerprint of numerical calculations through active hardware mechanisms, so as to achieve high-precision signal recognition and high-level physical security protection without sacrificing real-time performance, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0010] In view of this, this invention provides a method and system for secure identification of UAV image transmission signals based on ST-LFNet and proactive honeypot defense. It aims to overcome the "double paradox" faced in the existing fields of radio regulation and edge computing security: the "mismatch between discrete static models and continuous dynamic signals" at the algorithm performance level, and the "incompatibility between high-level protection requirements and limited computing resources at the edge" at the physical security level. It creatively proposes a new technical paradigm that deeply integrates "algorithmic fluidization" and "defensive honeypotting." Through paradigm shift at the mathematical modeling level and game-theoretic design at the hardware architecture level, it achieves ultimate capture of non-stationary signals and proactive immunity to physical attacks, thereby constructing a next-generation cognitive radio system with both high-precision perception capabilities and strong self-protection capabilities.

[0011] To achieve the above objectives, the present invention adopts the following technical solution: A method for secure identification of UAV image transmission signals based on ST-LFNet and active honeypot defense includes: Step 1: Perform high-fidelity physical reconstruction on the collected UAV downlink signal and map it to the time-frequency domain through short-time Fourier transform to generate a time-frequency tensor; Step 2: Employ a dynamic computing resource scheduling mechanism based on threat scoring. In low-threat mode, only the main processor is enabled to run the ST-LFNet network for spatiotemporal fluid feature extraction. In high-threat mode, the security coprocessor is activated to perform proactive honeypot camouflage based on game theory. Step 3: Monitor the voltage drop and temperature gradient inside the chip in real time, calculate the KL divergence between the real-time observed signal distribution and the reference normal template, and immediately trigger the trapping mode when the KL divergence suddenly increases beyond the preset threshold.

[0012] Optionally, in step 1, the collected UAV downlink signal is subjected to high-fidelity physical reconstruction, specifically as follows: Receive signal The model is a non-stationary random process superimposed with Ricean fading and Doppler shift, and the mathematical expression is as follows:

[0013] in, This represents the number of multipath paths. For the first The time-varying fading coefficient of the path; For time-varying delay; For Doppler frequency shift; It is additive white Gaussian noise.

[0014] Optionally, in step 1, after performing high-fidelity physical reconstruction on the collected UAV downlink signal, the method further includes: adaptively estimating and compensating for gain mismatch using a blind source separation algorithm based on second-order statistics. and phase mismatch And DC bias removal and automatic gain control preprocessing.

[0015] Optionally, in step 2, the dynamic computing resource scheduling mechanism based on threat scoring is as follows: Threat Score It integrates side-channel monitoring data, historical attack records, and sensitivity to the current geographical location; when When, it is judged to be in low threat mode, when At that time, it was determined to be a high-threat mode; among them, The threshold for determining low-threat mode; This is the threshold for determining a high-threat mode.

[0016] Optionally, in step 2, when activating the security coprocessor to perform game theory-based active honeypot spoofing in high-threat mode, it also includes: calling a pre-trained lightweight generative adversarial network to generate realistic fake data to drive the honeypot, so that the power consumption pattern generated by the honeypot has statistical distribution characteristics that are extremely similar to real data, thereby confusing the attacker.

[0017] Optionally, in step 2, the ST-LFNet network consists of three cascaded modules: a multi-scale temporal convolution module, a spatial fluid neural module, and a model training module. Multi-scale temporal convolution module: This is a multi-branch parallel convolution structure. Let the input be... The m-th branch uses a kernel size of The output of the filter is calculated as follows:

[0018] in, This represents the batch normalization operation; it extracts transient burst features of the signal in parallel using convolutional kernels with different receptive fields, and finally fuses the features of each branch through a concatenation operation to generate a robust temporal feature vector. ; Spatial Fluid Liquid Neural Module: Based on the theory of liquid time constant, defining states Continuous derivative:

[0019] Expanded to:

[0020] in, This is the adaptive damping term, which determines the viscosity or inertia of the system. This is a nonlinear driving term, representing the continuous injection and modulation of the neural state by external signals; Using a fourth-order Runge-Kutta solver, the iterative formula is as follows:

[0021] Model training module: Employs adjoint sensitivity analysis, calculating the loss function with respect to parameters by solving the adjoint ODE in reverse time. gradient:

[0022] Among them, the accompanying state Following the reverse evolution process:

[0023] Finally, an adaptive gating mechanism is used to fuse temporal features with fluid neural features:

[0024] in, Indicates a splicing operation; It is a learnable gating coefficient that dynamically determines whether the final classification decision depends more on transient features or long-term fluid features.

[0025] Optionally, in step 2, the game theory-based active honeypot camouflage includes two strategies: complementary energy masking and twin shadow networks with minimizing mutual information. Complementary energy masking: The system establishes an instruction-level power consumption model. If the main core is about to execute a high-power floating-point multiply-accumulate operation, the honeypot core is scheduled to execute a low-power no-operation or wait for instructions, and vice versa; thus reducing the total power consumption. Approaching the preset stable trajectory :

[0026] in, Injected random Gaussian noise; Twin Shadow Network and Minimizing Mutual Information: The co-kernel runs a shadow network with the exact same structure as ST-LFNet, but with specially initialized weights; the optimization objective is to minimize the observed signal. Compared with actual parameters Mutual messaging, while maximizing interaction with shadow parameters. Mutual information:

[0027] in, , These are the weighting coefficients.

[0028] Optionally, in step 3, the KL divergence between the real-time observed signal distribution and the baseline normal template is calculated as follows:

[0029] in, For signal distribution; This is the baseline normal template.

[0030] Optionally, in step 3, the trapping mode is as follows: Entrapment Response: Once an attack is confirmed, the system will not immediately issue an alarm, but will instead enter entrapment mode; Logical trapping: The real recognition result is suppressed, and the main processor outputs a preset false result; Physical trapping: The honeypot path generates a specific power consumption pattern to mislead attackers into thinking they have found the key. At the same time, the system records the attacker's behavioral characteristics for subsequent tracing and evidence collection.

[0031] This invention also provides a UAV image transmission signal security identification system based on ST-LFNet and active honeypot defense, utilizing a method for secure identification of UAV image transmission signals based on ST-LFNet and active honeypot defense, comprising: Signal preprocessing module: used to perform high-fidelity physical reconstruction of the acquired UAV downlink signal and map it to the time-frequency domain through short-time Fourier transform to generate a time-frequency tensor; Resource scheduling module: Used to employ a dynamic computing resource scheduling mechanism based on threat scoring. In low-threat mode, only the main processor is enabled to run the ST-LFNet network for spatiotemporal fluid feature extraction. In high-threat mode, the security coprocessor is activated to perform proactive honeypot camouflage based on game theory. Monitoring and trapping module: used to monitor the voltage drop and temperature gradient inside the chip in real time, calculate the KL divergence between the real-time observed signal distribution and the reference normal template, and immediately trigger the trapping mode when the KL divergence suddenly increases beyond the preset threshold.

[0032] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method and system for secure identification of UAV image transmission signals based on ST-LFNet and active honeypot defense, achieving the following beneficial effects: (1) A qualitative leap has been achieved in the accuracy of non-stationary signal recognition (algorithm robustness): By introducing ST-LFNet, this invention fundamentally solves the "sampling bias" problem of discrete models when processing continuous dynamic signals. Experiments demonstrate that when UAVs perform highly maneuverable actions such as sharp turns and dives, causing drastic changes in channel characteristics, this invention can maintain the continuity and integrity of feature extraction. Especially in low signal-to-noise ratio (SNR < 0dB) and complex electromagnetic environments (such as urban canyons), its recognition accuracy and generalization ability are significantly better than traditional CNN or LSTM models, greatly reducing the false alarm rate and false alarm rate.

[0033] (2) A physical-level "iron wall" (physical security) against side-channel attacks has been constructed: This invention transforms "passive defense" into "active deception." By generating deceptive noise through hardware honeypots, this invention not only masks the side-channel characteristics of the real model but also generates false parameter distributions to lure attackers. This exponentially increases the sample size required for differential energy analysis (DPA) and correlated electromagnetic analysis (CEMA) targeting this device, theoretically reducing the success rate of side-channel attacks to near zero. This provides defense-grade physical security for edge intelligent devices deployed in unattended environments.

[0034] (3) A dynamic balance between high-performance computing and high-level security has been achieved (system performance): This invention solves the long-standing zero-sum game between "security" and "efficiency." By offloading the defense load from the main processor through a heterogeneous architecture, this invention avoids the 2-5 times performance loss typically associated with traditional masking techniques. This enables complex ODE fluid network algorithms to run at full speed on resource-constrained embedded devices while enjoying the highest level of physical protection, significantly improving the system's performance per watt and practical value, and providing a completely new technical path for the design of next-generation secure edge AI terminals. Attached Figure Description

[0035] 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0036] Figure 1 This is a schematic diagram of the method flow provided by the present invention.

[0037] Figure 2 This is a schematic diagram of the ST-LFNet network architecture provided by the present invention.

[0038] Figure 3 This is a schematic diagram of the active honeypot defense waveform principle provided by the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Example 1: Embodiment 1 of this invention discloses a method for secure identification of UAV image transmission signals based on ST-LFNet and active honeypot defense, such as... Figure 1 As shown, it includes: Step 1: Perform high-fidelity physical reconstruction on the collected UAV downlink signal and map it to the time-frequency domain through short-time Fourier transform to generate a time-frequency tensor.

[0041] First, the system acquires UAV downlink signals covering the 2.4GHz / 5.8GHz frequency band using a broadband RF frontend. This frontend typically includes a low-noise amplifier (LNA), a mixer, and a high-sampling-rate analog-to-digital converter (ADC). Due to the complexity of the actual physical environment, the received signals are often severely distorted, containing various noises and interferences. To extract effective features, the acquired UAV downlink signals undergo high-fidelity physical reconstruction, specifically as follows: Considering the real physical environment, the received signal The model is a non-stationary stochastic process superimposed with Rician fading and Doppler shift, and the mathematical expression is as follows:

[0042] in, This represents the number of multipath paths. For the first The time-varying fading coefficient of the path; For time-varying delay; For Doppler frequency shift; The noise used is additive white Gaussian noise. This model not only considers the line-of-sight (LoS) component but also includes multipath scattering components, accurately reflecting the statistical characteristics of UAV image transmission signals. Compared to traditional additive white Gaussian noise (AWGN) channel models, this model more closely approximates the real-world outdoor channel environment.

[0043] To address the common IQ amplitude-phase imbalance problem in zero-IF receivers, after high-fidelity physical reconstruction of the acquired UAV downlink signal, a blind source separation algorithm based on second-order statistics is also included. This algorithm can adaptively estimate and compensate for gain mismatch without relying on pilot signals. and phase mismatch This restores the orthogonality of the signals and eliminates image interference. In addition, it includes standard preprocessing steps such as removing DC offset and automatic gain control (AGC).

[0044] The corrected time-domain complex signal is mapped to the time-frequency domain through a short-time Fourier transform (STFT) to generate a time-frequency tensor. The time-frequency tensor is used to simultaneously characterize the burst characteristics of the signal in the time domain and the multicarrier characteristics in the frequency domain; the window function length and overlap rate of the short-time Fourier transform are adaptively adjusted according to the symbol period of the signal. This step transforms the one-dimensional time series into a two-dimensional time-frequency image, preserving the spectral characteristics of the signal changing over time (such as frequency hopping patterns, OFDM subcarrier spacing, etc.), providing rich input information for subsequent fluid networks. Preferably, the window function of the STFT adopts a Hamming window, and the overlap rate is set to 50% to 75% to balance time resolution and frequency resolution.

[0045] Step 2: Employing a dynamic computing resource scheduling mechanism based on threat scoring, in low-threat mode, only the main processor (AP) is activated to run the ST-LFNet network for spatiotemporal fluid feature extraction, aiming for the lowest power consumption and highest energy efficiency. During this time, the coprocessor is in sleep or low-power standby mode. In high-threat mode, the security coprocessor (SCP) is activated to perform proactive honeypot camouflage based on game theory (the SCP does not participate in actual signal identification).

[0046] To balance security and energy efficiency with limited edge computing resources, the system introduces a threat scoring-based approach. The dynamic computing resource scheduling mechanism is as follows: Threat Score It integrates side-channel monitoring data, historical attack records, and sensitivity to the current geographical location; when When the system determines the environment to be safe (e.g., in a controlled laboratory environment or a non-hostile area), it is classified as a low-threat mode. When a potential threat is detected (e.g., abnormal electromagnetic scanning or being in an unknown field environment)); [the following is a possible interpretation:] This mode is classified as high-threat (when a potential threat is detected (e.g., abnormal electromagnetic scanning or being in an unknown field environment)). The threshold for determining low-threat mode; This is the threshold for determining a high-threat mode.

[0047] To enhance deception, in high-threat mode, when activating the security coprocessor for game theory-based active honeypot spoofing, it also includes: calling a pre-trained lightweight generative adversarial network (GAN) to generate realistic fake data to drive the honeypot, so that the power consumption pattern generated by the honeypot has statistical distribution characteristics that are extremely similar to real data, thereby deceiving attackers.

[0048] The ST-LFNet network is inspired by fluid dynamics and aims to solve the problem that discrete networks cannot handle continuous dynamic signals. Figure 2 As shown, the network consists of three cascaded modules: a multi-scale temporal convolution module, a spatial fluid neural module, and a model training module. Multi-scale temporal convolutional module (MS-1D-CNN): This is a multi-branch parallel convolutional structure. UAV signals typically contain extremely short frequency-hopping pulses and long frame synchronization sequences. To capture features at both scales simultaneously, this invention designs a multi-branch parallel convolutional structure, assuming the input is... The m-th branch uses a kernel size of The output of the filter is calculated as follows:

[0049] in, This indicates batch normalization; using convolution kernels with different receptive fields (e.g., ... The transient burst features of the signal are extracted in parallel, and finally the features of each branch are fused through a concatenation operation to generate a robust time-domain feature vector. ; Fluid Liquid Modeling: This is the theoretical core of this invention. This module treats the feature vector as a continuous fluid. Unlike the discrete state updates of RNNs, this module is based on the theory of liquid time constant (LTC) and defines the state... Continuous derivative:

[0050] More specifically, to enhance the non-linear expressive power, it can be expanded as follows:

[0051] In-depth analysis of the physical mechanism: The key to this dynamic equation lies in the introduction of an adaptive mechanism.

[0052] in, The adaptive damping term determines the viscosity or inertia of the system when the input signal... When the Doppler effect changes drastically, When the term approaches 1, the effective damping increases, and the time constant decreases. This means the network "responds faster" and can sensitively track high-frequency jitter. Conversely, when the signal is stable, the time constant increases, and the network tends to retain memory. This mechanism allows the network to adaptively adapt to the dynamic changes of non-stationary signals, much like a fluid adapting to a container. As a nonlinear driving term, it represents the continuous injection and modulation of the neural state by external signals, ensuring a continuous inflow of information.

[0053] To accurately execute the above differential equations on digital hardware, this invention employs a fourth-order Runge-Kutta solver (RK4), with the following iterative formula:

[0054] RK4 achieves this in a time step. By performing four intermediate state samplings (discretizing a continuous time step into four intermediate slopes) and weighted summation, the true trajectory of a continuous system can be approximated with fourth-order accuracy. Although computationally intensive, this high-order approximation greatly ensures the fidelity of feature extraction and effectively overcomes the challenge of nonstationarity.

[0055] Model Training Module: Since the RK4 solver involves multiple iterations, direct backpropagation would lead to an excessively deep computational graph, causing memory explosions. Therefore, this invention employs the Adjoint Sensitivity Method, which calculates the loss function with respect to parameters by solving the adjoint ODE in the backpropagation time. The gradient is obtained, thus avoiding the need to store all intermediate states of the numerical integrator during backpropagation, significantly reducing the memory overhead of edge devices:

[0056] Among them, the accompanying state Following the reverse evolution process:

[0057] This approach enables online fine-tuning on resource-constrained edge devices, and the memory overhead is independent of the number of integration steps, only a fraction of the total memory required. .

[0058] Spatiotemporal fusion: Finally, an adaptive gating mechanism is used to fuse temporal features with fluid neural features.

[0059] in, Indicates a splicing operation; It is a learnable gating coefficient that dynamically determines whether the final classification decision depends more on transient features or long-term fluid features.

[0060] While ST-LFNet's RK4 solution process boasts high accuracy, its "four-step iteration" generates highly characteristic "four-peak" power ripple, making it easily identifiable by attackers. To address this, the coprocessor concurrently executes a game-theoretic-based proactive honeypot camouflage. This part represents the essence of the hardware-software co-design, encompassing complementary energy masking and two strategies: twin shadow networks and minimizing mutual information. Complementary energy masking: The goal of this strategy is to "smooth out peak loads and fill valleys." The system establishes an instruction-level power model. If the main core is about to execute a high-power floating-point multiply-accumulate (FMA) operation, the honeypot core is scheduled to execute a low-power no-op (NOP) or wait for instructions, and vice versa. Mathematically, this reduces the total power consumption. Approaching the preset stable trajectory :

[0061] in, The injected random Gaussian noise is used to eliminate the correlation of energy fluctuations in the time domain through this complementary mechanism, thus rendering differential energy analysis (DPA) attacks ineffective.

[0062] Twin Shadow Network and Minimizing Mutual Information: The goal of this strategy is to "confuse the audience." A co-kernel runs a shadow network with an identical (isomorphic) structure to ST-LFNet, but with specially initialized weights; the optimization objective is to minimize the observed signal. Compared with actual parameters Mutual messaging, while maximizing interaction with shadow parameters. Mutual information to resist correlation electromagnetic analysis (CEMA):

[0063] in, , These are the weighting coefficients. Through this game theory, even if an attacker successfully separates the signal, they will obtain incorrect honeypot parameters, thus protecting the security of the real model.

[0064] The active honeypot defense waveform principle, such as Figure 3 As shown.

[0065] Step 3: Monitor the voltage drop and temperature gradient inside the chip in real time, calculate the KL divergence between the real-time observed signal distribution and the reference normal template, and immediately trigger the trapping mode when the KL divergence suddenly increases beyond the preset threshold.

[0066] To achieve proactive defense, the system must possess sensing capabilities. This system integrates a highly sensitive side-channel monitoring unit (SCMU).

[0067] The SCMU uses an on-chip distributed ring oscillator (RO) array to monitor the voltage drop and temperature gradient inside the chip in real time, and calculates the KL divergence between the real-time observed signal distribution and the reference normal template (this is used as a statistic to determine whether there is an anomalous attack), as follows:

[0068] in, For signal distribution; This is the baseline normal template.

[0069] If the KL divergence suddenly increases beyond the preset threshold, it means that there is a change in load impedance caused by the access of an external probe, or frequency drift caused by electromagnetic injection.

[0070] The trapping method is as follows: Entrapment Response: Once an attack is confirmed, the system will not immediately issue an alarm, but will instead enter entrapment mode; Logical decoy: The real identification result is suppressed, and the main processor outputs a preset false result (e.g., marking the enemy reconnaissance drone as a civilian whitelist device). Physical trapping: The honeypot path generates specific power consumption patterns (e.g., simulating the power consumption characteristics of the standard AES encryption algorithm) to mislead attackers into believing they have found the key, thus wasting their time and resources. The system simultaneously records the attacker's behavioral characteristics (such as attack time, frequency, and waveform features) for subsequent tracing and forensics.

[0071] The technical solution of this invention is based on the following core theoretical concepts and architectural design: (1) Fluidized Reconstruction at the Algorithm Level: From “Static Slices” to “Continuous Fluid” This invention breaks through the traditional stacking paradigm of discrete layers in artificial neural networks (ANNs) and proposes a spatiotemporal liquid fluid network (ST-LFNet) architecture.

[0072] Dynamic Modeling: This concept views the feature evolution of image transmission signals as a continuously flowing "fluid" in a high-dimensional feature space. It abandons the traditional static graph mapping mechanism and instead utilizes a fluid neural module based on ordinary differential equations (ODEs) to describe the continuous rate of change of neuronal states.

[0073] Adaptive Solving: By introducing an embedded high-order numerical integrator (such as the Runge-Kutta Solver), the network can adaptively adjust the "integration step size" and "time constant" according to the drastic fluctuations in the input signal (such as sudden Doppler shifts). This gives the model fluid-like physical adaptability, enabling it to accurately approximate the true dynamic trajectory of the signal in the continuous time domain, thereby achieving endogenous compensation for channel distortion and non-stationary characteristics.

[0074] (2) Active honeypots at the hardware level: From "passive defense" to "game-theoretic deception": This invention introduces the concept of non-cooperative game theory to construct a hardware-level active side-channel honeypot.

[0075] Heterogeneous Parallel Phantom Tasks: Utilizing a heterogeneous multi-core architecture, while high-performance computing cores execute real inference tasks, independent coprocessor cores execute meticulously designed "phantom tasks" in parallel. These tasks are not simply random noise generation, but rather computational loads that complement or are structurally twins of the real computing tasks in terms of power consumption ripple and electromagnetic radiation patterns.

[0076] Maximizing Information Entropy: The core of this concept lies in constructing a "side-channel mirage." Through dynamic scheduling based on the principle of minimizing mutual information, highly deceptive physical fingerprints are actively synthesized. This fundamentally disrupts the statistical correlation between the physically leaked information observed by the attacker and the real data inside the device, maximizing the attacker's decoding entropy. This not only prevents the attacker from extracting the key but may even mislead them to the wrong attack path.

[0077] (3) System-level resource decoupling: from "resource contention" to "orthogonal isolation": This invention employs a system design that combines physical isolation with logical coupling. The high-performance computing core and the security coprocessing core are hardware isolated through independent power and clock domains. This design ensures that the honeypot defense mechanism does not occupy the main business's computing cycle or introduce additional processing latency, thus guaranteeing the orthogonality of "defense" and "computation" at the physical level and solving the performance regression problem caused by traditional software defense methods.

[0078] It should be understood that the ST-LFNet algorithm architecture and proactive honeypot defense mechanism proposed in this invention have broad applicability and scalability. For those skilled in the art, various modifications, combinations, and evolutions can be made to this invention without departing from its core principles. For example: At the algorithm level, the types of ordinary differential equation (ODE) solvers (such as Runge-Kutta), the specific activation functions of liquid neurons, or the network hierarchical structure mentioned in the paper can all be adaptively adjusted or replaced according to actual computing resources. At the hardware level, the heterogeneous computing resources are not limited to a specific combination of CPU and coprocessor, but can also utilize FPGA, ASIC, DSP or cloud-edge collaborative architecture to achieve equivalent logical functions. At the application level, the signal recognition and defense logic of this invention is also applicable to other non-stationary wireless communication scenarios besides drone image transmission.

[0079] Example 2: Embodiment 2 of this invention discloses a UAV image transmission signal security identification system based on ST-LFNet and active honeypot defense, comprising: Signal preprocessing module: used to perform high-fidelity physical reconstruction of the acquired UAV downlink signal and map it to the time-frequency domain through short-time Fourier transform to generate a time-frequency tensor; Resource scheduling module: Used to employ a dynamic computing resource scheduling mechanism based on threat scoring. In low-threat mode, only the main processor is enabled to run the ST-LFNet network for spatiotemporal fluid feature extraction. In high-threat mode, the security coprocessor is activated to perform proactive honeypot camouflage based on game theory. Monitoring and trapping module: used to monitor the voltage drop and temperature gradient inside the chip in real time, calculate the KL divergence between the real-time observed signal distribution and the reference normal template, and immediately trigger the trapping mode when the KL divergence suddenly increases beyond the preset threshold.

[0080] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0081] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for secure identification of UAV image transmission signals based on ST-LFNet and active honeypot defense, characterized in that, include: Step 1: Perform high-fidelity physical reconstruction on the collected UAV downlink signal and map it to the time-frequency domain through short-time Fourier transform to generate a time-frequency tensor; Step 2: Employ a dynamic computing resource scheduling mechanism based on threat scoring. In low-threat mode, only the main processor is enabled to run the ST-LFNet network for spatiotemporal fluid feature extraction. In high-threat mode, the security coprocessor is activated to perform proactive honeypot camouflage based on game theory. Step 3: Monitor the voltage drop and temperature gradient inside the chip in real time, calculate the KL divergence between the real-time observed signal distribution and the reference normal template, and immediately trigger the trapping mode when the KL divergence suddenly increases beyond the preset threshold.

2. The method for secure identification of UAV image transmission signals based on ST-LFNet and active honeypot defense according to claim 1, characterized in that, In step 1, the collected UAV downlink signal is reconstructed using high-fidelity physical reconstruction, specifically as follows: Receive signal The model is a non-stationary random process superimposed with Ricean fading and Doppler shift, and the mathematical expression is as follows: in, This represents the number of multipath paths. For the first The time-varying fading coefficient of the path; For time-varying delay; For Doppler frequency shift; It is additive white Gaussian noise.

3. The method for secure identification of UAV image transmission signals based on ST-LFNet and active honeypot defense according to claim 1, characterized in that, In step 1, after performing high-fidelity physical reconstruction on the collected UAV downlink signal, the method further includes: using a blind source separation algorithm based on second-order statistics to adaptively estimate and compensate for gain mismatch. and phase mismatch And DC bias removal and automatic gain control preprocessing.

4. The method for secure identification of UAV image transmission signals based on ST-LFNet and active honeypot defense according to claim 1, characterized in that, In step 2, the dynamic computing resource scheduling mechanism based on threat scoring is as follows: Threat score It integrates side-channel monitoring data, historical attack records, and the sensitivity of the current geographical location; when When, it is judged to be in low threat mode, when At that time, it was determined to be a high-threat mode; among them, The threshold for determining the low-threat mode; The threshold for determining the high-threat mode.

5. The method for secure identification of UAV image transmission signals based on ST-LFNet and active honeypot defense according to claim 1, characterized in that, In step 2, when activating the security coprocessor to perform game theory-based active honeypot spoofing in high-threat mode, it also includes: calling a pre-trained lightweight generative adversarial network to generate realistic fake data to drive the honeypot, so that the power consumption pattern generated by the honeypot has statistical distribution characteristics that are extremely similar to real data, thereby deceiving attackers.

6. The method for secure identification of UAV image transmission signals based on ST-LFNet and active honeypot defense according to claim 1, characterized in that, In step 2, the ST-LFNet network consists of three cascaded modules: a multi-scale temporal convolution module, a spatial fluid neural network module, and a model training module. The multi-scale temporal convolution module is a multi-branch parallel convolution structure, with the input being... The m-th branch uses a kernel size of The output of the filter is calculated as follows: in, This represents the batch normalization operation; it extracts transient burst features of the signal in parallel using convolutional kernels with different receptive fields, and finally fuses the features of each branch through a concatenation operation to generate a robust temporal feature vector. ; The spatial fluid liquid neural module: based on the theory of liquid time constant, defines states. Continuous derivatives: Expanded to: in, This is the adaptive damping term, which determines the viscosity or inertia of the system. This is a nonlinear driving term, representing the continuous injection and modulation of the neural state by external signals; Using a fourth-order Runge-Kutta solver, the iterative formula is as follows: The model training module employs adjoint sensitivity analysis, calculating the loss function with respect to parameters by solving the adjoint ODE in the reverse time. gradient: Among them, the accompanying state Following the reverse evolution process: Finally, an adaptive gating mechanism is used to fuse temporal features with fluid neural features: in, Indicates a splicing operation; It is a learnable gating coefficient that dynamically determines whether the final classification decision depends more on transient features or long-term fluid features.

7. The method for secure identification of UAV image transmission signals based on ST-LFNet and active honeypot defense according to claim 1, characterized in that, In step 2, the game theory-based active honeypot camouflage includes two strategies: complementary energy masking and twin shadow networks with minimizing mutual information. The complementary energy masking involves establishing an instruction-level power consumption model. If the main core is about to execute a high-power floating-point multiply-accumulate operation, the honeypot core is scheduled to execute a low-power no-operation or wait for instructions, and vice versa; thus reducing the total power consumption. Approaching the preset stable trajectory : in, Injected random Gaussian noise; The twin shadow network and minimizing mutual information: the co-core runs a shadow network with the exact same structure as ST-LFNet, but with specially initialized weights; the optimization objective is to minimize the observed signal. Compared with actual parameters Mutual messaging, while maximizing interaction with shadow parameters. Mutual information: in, , These are the weighting coefficients.

8. The method for secure identification of UAV image transmission signals based on ST-LFNet and active honeypot defense according to claim 1, characterized in that, In step 3, the KL divergence between the real-time observed signal distribution and the baseline normal template is calculated as follows: in, For signal distribution; This is the baseline normal template.

9. A method for secure identification of UAV image transmission signals based on ST-LFNet and active honeypot defense according to claim 1, characterized in that, In step 3, the trapping mode is specifically as follows: Entrapment response: Once an attack is confirmed, the system does not immediately issue an alarm, but instead enters entrapment mode; Logical trapping: The real recognition result is suppressed, and the main processor outputs a preset false result; Physical trapping: The honeypot path generates a specific power consumption pattern to mislead attackers into thinking they have found the key. At the same time, the system records the attacker's behavioral characteristics for subsequent tracing and evidence collection.

10. A drone image transmission signal security identification system based on ST-LFNet and active honeypot defense, utilizing the drone image transmission signal security identification method based on ST-LFNet and active honeypot defense as described in any one of claims 1-9, characterized in that, include: Signal preprocessing module: used to perform high-fidelity physical reconstruction of the acquired UAV downlink signal and map it to the time-frequency domain through short-time Fourier transform to generate a time-frequency tensor; Resource scheduling module: Used to employ a dynamic computing resource scheduling mechanism based on threat scoring. In low-threat mode, only the main processor is enabled to run the ST-LFNet network for spatiotemporal fluid feature extraction. In high-threat mode, the security coprocessor is activated to perform proactive honeypot camouflage based on game theory. Monitoring and trapping module: used to monitor the voltage drop and temperature gradient inside the chip in real time, calculate the KL divergence between the real-time observed signal distribution and the reference normal template, and immediately trigger the trapping mode when the KL divergence suddenly increases beyond the preset threshold.

Images