Iot-based ai physical layer anti-interception secure communication method and system

Abstract

This invention relates to the field of secure communication technology at the physical layer of the Internet of Things (IoT), specifically to an AI-based physical layer anti-interception secure communication method and system for IoT. The method generates an original communication signal at the transmitting end, then calls an improved stochastic resonance algorithm that adaptively adjusts nonlinear system parameters based on the current channel noise statistical characteristics to hide the original signal, generating a modulated concealed signal and an unrelated artificial noise signal. These two signals are weighted and mixed to form a synthesized transmitted signal, which is then transmitted. At the receiving end, the noisy received signal is received, and a trained AI signal separation network is used to separate the noise-suppressed signal and the remaining noise signal. Finally, the improved stochastic resonance algorithm is used in reverse from the noise-suppressed signal to recover the estimated signal of the original signal. This invention achieves adaptive signal concealment under dynamic channels and effective signal separation by the receiver under unknown interference, improving the system's robustness and communication reliability under non-ideal conditions.

Description

Technical Field

[0001] This invention relates to the field of IoT physical layer secure communication technology, and in particular to an AI physical layer anti-interception secure communication method and system based on IoT. Background Technology

[0002] Internet of Things (IoT) devices transmit data in open wireless environments, making the signals highly susceptible to interception and analysis by unauthorized receivers. Physical layer security is therefore an important research area.

[0003] In existing technologies, one approach is to interfere with the eavesdropper by superimposing artificial noise at the transmitting end. The legitimate receiver needs to use channel information or a preset key to eliminate the noise's influence. However, if the artificial noise is unrelated to the signal, its interference effect is limited; if it is related to the signal, the design is complex and the signal characteristics are easily exposed.

[0004] Another approach is to utilize the stochastic resonance effect to enhance weak signals in strong noise. Traditional stochastic resonance methods are mostly designed for known, fixed noise and employ preset nonlinear system parameters. However, the wireless channel environment of the Internet of Things (IoT) is dynamically changing, and the statistical characteristics of background noise are time-varying. A stochastic resonance system with fixed parameters cannot maintain the optimal concealment of the signal in a changing channel, leading to an increased probability that the signal can be detected in noise.

[0005] At the receiver, traditional physical layer security schemes rely on precise prior knowledge of the transmitter's signal processing methods, such as the accurate structure of artificial noise or a perfect estimate of the channel state, to extract the original signal from the mixed signal and strong noise. In real dynamic channels, such ideal conditions are difficult to meet, resulting in decreased received signal separation performance and a significant increase in the bit error rate.

[0006] In summary, existing solutions struggle to simultaneously achieve high signal concealment for eavesdroppers and reliable signal recovery for legitimate receivers in dynamic wireless environments. This invention aims to address the issues of adaptive signal concealment in dynamic channels and effective signal separation by the receiver under unknown interference. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing technologies by proposing an AI physical layer anti-interception secure communication method and system based on the Internet of Things. This method enables adaptive concealment of signals under dynamic channels and effective signal separation by the receiver under unknown interference, thereby improving the robustness and communication reliability of the system under non-ideal conditions.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: IoT-based AI physical layer anti-interception secure communication methods and systems, including: Generate the raw communication signal to be transmitted at the IoT transmitter; An improved stochastic resonance algorithm is invoked to perform signal hiding processing on the original communication signal. The improved stochastic resonance algorithm adaptively adjusts the nonlinear system parameters based on the statistical characteristics of the current channel noise to generate a modulated hidden signal. An artificial noise signal unrelated to the original communication signal is generated at the IoT transmitter. The modulated covert signal and the artificial noise signal are weighted and mixed at the physical layer to generate a synthetic transmission signal; The synthesized transmission signal is transmitted via a wireless channel; The IoT receiver receives noisy signals transmitted via wireless channels. At the IoT receiver, a trained AI signal separation network is invoked to separate the noise-suppressed signal and the remaining noise signal from the noisy received signal; The improved stochastic resonance algorithm is invoked in reverse from the noise-suppressed signal to recover the estimated signal of the original communication signal.

[0009] As a further aspect of the present invention, the step of invoking an improved stochastic resonance algorithm to perform signal hiding processing on the original communication signal, wherein the improved stochastic resonance algorithm adaptively adjusts the nonlinear system parameters based on the statistical characteristics of the current channel noise to generate a modulated hidden signal, includes: Extract the feature vector of the original communication signal, the feature vector including signal power spectral density, zero-crossing rate and time-domain envelope shape; Real-time monitoring of background noise in the wireless channel to obtain the statistical characteristics of the current channel noise, including noise variance, power spectrum and higher-order cumulative quantities; The original communication signal is input as a periodic weak driving signal into a bistable state trap nonlinear system, and the equation of motion of the bistable state trap nonlinear system is determined by the system potential function parameters and the damping coefficient. Based on the statistical characteristics of the current channel noise, the system potential function parameters and the damping coefficient are dynamically adjusted through a parameter adapter, so that the current channel noise and the original communication signal resonate within the bistable potential trap nonlinear system. In the resonant state, the signal energy output by the bistable state trap nonlinear system shifts from the frequency of the original communication signal to the vicinity of the system characteristic frequency, forming the modulated covert signal with altered spectral characteristics that matches the statistical characteristics of the background noise.

[0010] As a further aspect of the present invention, based on the statistical characteristics of the current channel noise, the system potential function parameters and the damping coefficient are dynamically adjusted through a parameter adapter, including: Based on the noise variance, the critical barrier height reference value of the bistable potential trap nonlinear system is calculated, and the critical barrier height reference value is proportional to the square root of the noise variance. Based on the power spectrum of the noise, the concentrated frequency band of the noise energy is identified, and the adjustment amount of the system characteristic frequency is calculated so that the system characteristic frequency deviates from the concentrated frequency band of the noise energy. The non-Gaussian intensity of the noise is estimated using the higher-order cumulative amount of the noise, and the weight of the nonlinear term of the damping coefficient is adjusted accordingly to match the statistical distribution of the noise. Based on the critical barrier height reference value, the adjustment amount of the system characteristic frequency, and the weight of the nonlinear term, the real-time configuration parameters of the system potential function parameters and the damping coefficient are generated and injected into the bistable potential trap nonlinear system.

[0011] As a further aspect of the present invention, the generation of an artificial noise signal unrelated to the original communication signal at the Internet of Things transmitter includes: Obtain the physical fingerprint characteristics of the IoT transmitter hardware, including the nonlinear characteristics of the radio frequency front-end power amplifier and the phase noise spectrum of the local oscillator; Based on the physical fingerprint features, a baseband noise sequence with a specific statistical distribution is generated by a pseudo-random sequence generator. The baseband noise sequence is uncorrelated with the original communication signal in both the time and frequency domains. The baseband noise sequence is shaped and filtered, and the frequency response characteristics of the filter are designed to be complementary to the multipath delay power spectrum of the current wireless channel. The filtered baseband noise sequence is upconverted to the transmit carrier frequency to generate the artificial noise signal. The power level of the artificial noise signal is set according to a preset signal-to-interference-plus-noise ratio target.

[0012] As a further aspect of the present invention, the modulated covert signal and the artificial noise signal are weighted and mixed at the physical layer to generate a synthetic transmission signal, including: Based on the preset concealment index and the signal-to-noise ratio requirement of the receiver, the power mixing ratio of the modulated concealed signal and the artificial noise signal is calculated. The modulated covert signal and the artificial noise signal are power scaled according to the power mixing ratio; The modulated covert signal with power scaling and the artificial noise signal with power scaling are added in phase in the complex domain to form a hybrid baseband signal; The hybrid baseband signal is subjected to digital-to-analog conversion and radio frequency modulation to generate the final synthesized transmission signal.

[0013] As a further aspect of the present invention, at the IoT receiver, a trained AI signal separation network is invoked to separate the noise-suppressed signal and the residual noise signal from the noisy received signal, including: The AI ​​signal separation network includes an encoder network and a decoder network. The encoder network is used to extract multi-scale depth features of the noisy received signal, and the decoder network is used to reconstruct signal components and noise components from the depth features. The temporal sampling sequence of the noisy received signal is input into the encoder network, which extracts multi-level feature maps from fine-grained to coarse-grained through multi-layer one-dimensional convolution and downsampling operations. A feature attention fusion layer is set between the encoder network and the decoder network to perform weighted fusion of feature maps from different scales, highlighting feature channels related to the modulated hidden signal; The decoder network outputs two reconstructed sequences in parallel through multi-layer deconvolution and upsampling operations, and by fusing skip connection features from the encoder: a noise-suppressed signal sequence and a residual noise signal sequence.

[0014] As a further aspect of the present invention, the trained AI signal separation network is obtained through the following method: A training dataset is constructed, and the training data samples include: sample cover signals, sample artificial noise signals, sample channel noise, and sample received signals generated by mixing them. The sample cover signals are generated by calling the improved stochastic resonance algorithm on the original sample signals. The received sample signal is used as the input to the AI ​​signal separation network, and the corresponding sample hidden signal and sample artificial noise signal are used as the desired output. Define a loss function that includes the mean square error between the reconstructed hidden signal and the sample hidden signal, and the mean square error between the reconstructed artificial noise signal and the sample artificial noise signal, and the weighted sum of the two errors; The parameters of the AI ​​signal separation network are iteratively updated using the backpropagation algorithm and optimizer until the loss function converges to a preset threshold, thus obtaining the trained AI signal separation network.

[0015] As a further aspect of the present invention, the construction of the training dataset includes: A large number of typical IoT communication signals were collected as a sample raw signal library; Simulate various typical wireless channel environments to generate sample channel models with different multipath, Doppler, and path loss characteristics; For each pair of original sample signals and sample channel models, the following steps are performed: the improved stochastic resonance algorithm is invoked to process the original sample signals and generate a sample cover signal; a corresponding sample artificial noise signal is generated based on the characteristics of the sample channel model; the sample cover signal and the sample artificial noise signal are transmitted through the sample channel model, and Gaussian white noise is added to obtain the sample received signal; The original sample signal, the corresponding hidden sample signal, the sample artificial noise signal, and the sample received signal are combined into a four-tuple training sample, and all four-tuples constitute the training dataset.

[0016] As a further aspect of the present invention, the improved stochastic resonance algorithm is invoked in reverse from the noise-suppressed signal to recover the estimated signal of the original communication signal, including: During the processing at the transmitting end of the improved stochastic resonance algorithm, inverse system parameters for signal concealment are obtained and recorded. These inverse system parameters are independently generated by the transmitting and receiving ends using the same channel noise estimation algorithm based on a pre-shared key, without explicit transmission. The inverse system parameters include inverse potential function parameters and inverse damping coefficients. A nonlinear inverse bistable potential trap system is constructed at the receiving end, and its equation of motion is determined by the inverse potential function parameters and the inverse damping coefficient. The noise suppression signal is used as the input driving signal for the inverse bistable state trap nonlinear system. In the inverse bistable state trap nonlinear system, signal energy is transferred back from the resonant frequency to near the baseband frequency of the original communication signal; The output of the inverse bistable state trap nonlinear system is low-pass filtered and sampled for decision-making to recover the estimated signal of the original communication signal.

[0017] As a further aspect of the present invention, the present invention also includes an IoT-based AI physical layer anti-interception secure communication system, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the above-described IoT-based AI physical layer anti-interception secure communication method.

[0018] Compared with the prior art, the advantages and positive effects of the present invention are as follows: 1) This invention designs an improved stochastic resonance algorithm. Based on the real-time channel noise statistical characteristics, it adaptively adjusts the barrier height and damping coefficient of its nonlinear system, so that the system can dynamically match the power spectrum distribution and intensity of the current channel noise. This allows the energy of the signal to be transmitted to be more effectively modulated into the specific statistical characteristics of the noise background. When the signal is transmitted in the wireless channel, its waveform characteristics are highly integrated with the statistical characteristics of the natural channel noise. It is difficult for the eavesdropper to distinguish the signal components from the background noise through conventional energy detection or feature analysis. The signal concealment does not depend on the absolute intensity of the noise, and low-probability detection can be achieved in various noise environments. 2) This invention designs a trained AI signal separation network to separate noise-suppressed signals and residual noise signals from noisy received signals. This network learns the internal structure mapping relationship of the synthesized transmitted signal under complex channel distortion through a large number of samples. It can identify the signal characteristic pattern after adaptive random resonance modulation without knowing the precise form of artificial noise or the specific state information of the channel in advance, and separate it from the mixed artificial noise and channel noise. This blind separation capability reduces the receiver's dependence on the transmitter parameters and channel state information. Even when the channel estimation is inaccurate or the characteristics of artificial noise are not fully known, it can still recover the signal components from severe mixed interference and directly output the denoised signal, thereby improving the robustness and communication reliability of the system under non-ideal conditions. Attached Figure Description

[0019] Figure 1 This is a state diagram of the IoT-based AI physical layer anti-interception secure communication method described in this invention. Figure 2 A flowchart for the dynamic adjustment of parameters in a bistable state trap nonlinear system; Figure 3 A flowchart for generating artificial noise signals. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] like Figure 1 As shown, this invention provides an AI physical layer anti-interception secure communication method based on the Internet of Things, the specific method including: At the IoT transmitter, the original communication signal to be transmitted is generated, and then an improved stochastic resonance algorithm is invoked to perform signal hiding processing on the original communication signal. The core of this improved stochastic resonance algorithm lies in adaptively adjusting the parameters of its internal nonlinear system based on the statistical characteristics of the current channel noise monitored in real time, thereby generating a modulated covert signal. Simultaneously, the transmitter independently generates an artificial noise signal unrelated to the original communication signal. At the physical layer, the modulated covert signal and the artificial noise signal are weighted and mixed according to a specific power mixing ratio, thereby generating the final synthetic transmission signal and radiating it through the wireless channel. At the IoT receiver, the device receives a noisy received signal that has been transmitted through the wireless channel and is mixed with noise. The receiver invokes a pre-trained AI signal separation network to process this noisy received signal. This AI signal separation network can separate the noise-suppressed signal and the residual noise signal from the complex mixed signal. The separated noise-suppressed signal is essentially an estimate of the modulated covert signal. The receiving end reverses the process of the improved stochastic resonance algorithm from the noise-suppressed signal to recover the estimated signal of the original communication signal. The inverse system parameters are distributed through a pre-shared key or secure control channel between legitimate devices. The eavesdropping party cannot perform the reverse recovery due to the lack of these parameters, thus completing secure communication.

[0022] In one embodiment of the present invention, this embodiment describes the specific process of performing signal hiding processing on the original communication signal using the improved stochastic resonance algorithm.

[0023] The process begins with the extraction of feature vectors from the original communication signal, encompassing the signal power spectral density, zero-crossing rate, and temporal envelope shape. Simultaneously, the system monitors the background noise of the wireless channel in real time to obtain the statistical characteristics of the current channel noise, including noise variance, power spectrum, and higher-order cumulants. The original communication signal is input as a periodic weakly driven signal into a bistable state-well nonlinear system. The dynamic behavior of this system is described by its equations of motion, which are determined by the system potential function parameters and damping coefficients. To achieve signal concealment, based on the acquired statistical characteristics of the current channel noise, the system potential function parameters and damping coefficients are dynamically adjusted via a parameter adapter. The goal of this adjustment is to induce a cooperative stochastic resonance effect between the current channel noise and the original communication signal within the bistable state-well nonlinear system.

[0024] The adjustment logic of the parameter adapter specifically includes several steps, such as... Figure 2As shown, based on the monitored noise variance, a reference value for the critical barrier height required by the bistable potential trap nonlinear system is calculated. This reference value is proportional to the square root of the noise variance. According to the noise power spectrum, the frequency band where noise energy is concentrated is identified, and the required adjustment amount for the system's characteristic frequency is calculated, aiming to deviate the system's characteristic frequency from this noise energy concentration band. Using the higher-order cumulative amount of noise, the non-Gaussian intensity of the noise is estimated, and the weight of the nonlinear term in the damping coefficient is adjusted accordingly to match the statistical distribution of the noise. Based on the calculated critical barrier height reference value, the system characteristic frequency adjustment amount, and the nonlinear term weight, a set of real-time system potential function parameters and damping coefficient configuration parameters are generated and injected into the bistable potential trap nonlinear system. After the system parameters are adaptively adjusted as described above, in a resonant state, the energy distribution of the signal output by the bistable potential trap nonlinear system changes, shifting from the baseband frequency of the original communication signal to near the system's characteristic frequency, thereby generating a modulated covert signal with altered spectral characteristics that matches the statistical characteristics of the background noise.

[0025] In specific implementation, the improved stochastic resonance algorithm performs signal hiding processing on the original communication signal. The processing begins with the extraction of the feature vector of the original communication signal. The feature vector includes the signal power spectral density, zero-crossing rate, and time-domain envelope shape. At the same time, the system monitors the background noise of the wireless channel in real time to obtain the statistical characteristics of the current channel noise. The statistical characteristics include noise variance, power spectrum, and higher-order cumulative quantity.

[0026] In some embodiments, feature vector extraction is achieved by a digital signal processing module, which performs time-frequency analysis on the original communication signal to calculate the power spectral density, obtains the zero-crossing rate by counting signal zero-crossing events, and obtains the time-domain envelope shape using an envelope detection circuit or algorithm.

[0027] In practical implementation, the original communication signal is input as a periodic weak driving signal into a bistable state trap nonlinear system. The equation of motion of the bistable state trap nonlinear system is determined by the system potential function parameters and the damping coefficient. The equation of motion of the bistable state trap nonlinear system is expressed as: in: These are the state variables of a bistable state-well nonlinear system. It is time. It is the damping coefficient. and These are the system potential function parameters. It is the original communication signal. This represents the current channel noise. It can be understood that this equation describes the cooperative dynamics of the original communication signal and the current channel noise within a nonlinear potential well.

[0028] In practical implementation, the system potential function parameters and damping coefficients are dynamically adjusted through a parameter adapter based on the statistical characteristics of the current channel noise. The adjustment logic of the parameter adapter includes calculating the critical barrier height reference value of the bistable potential well nonlinear system based on the noise variance. The critical barrier height reference value is proportional to the square root of the noise variance. The noise energy concentration band is identified based on the power spectrum of the noise, and the adjustment amount of the system characteristic frequency is calculated to make the system characteristic frequency deviate from the noise energy concentration band. The non-Gaussian intensity of the noise is estimated using the higher-order cumulative amount of the noise, and the weight of the nonlinear term of the damping coefficient is adjusted accordingly to match the statistical distribution of the noise.

[0029] Optionally, the critical barrier height reference value is calculated using a lookup table that stores the mapping relationship between noise variance and barrier height. In some embodiments, the adjustment amount of the system characteristic frequency is generated by a frequency shift algorithm, which dynamically adjusts the system potential function parameters based on the peak frequency of the noise power spectrum.

[0030] In practical implementation, real-time configuration parameters for the system potential function and damping coefficient are generated based on the critical barrier height reference value, the adjustment amount of the system characteristic frequency, and the weight of the nonlinear term. These parameters are then injected into the bistable potential trap nonlinear system. Under resonant conditions, the signal energy output by the bistable potential trap nonlinear system shifts from the frequency of the original communication signal to the vicinity of the system characteristic frequency, forming a modulated covert signal with altered spectral characteristics that matches the statistical characteristics of the background noise. It can be understood that the energy transfer process allows the modulated covert signal to incorporate background noise in the frequency domain, thereby enhancing its anti-interception capability.

[0031] Optionally, the dynamic adjustment of the parameter adapter is performed in real time by a microcontroller, which periodically updates the system parameters based on the output of the noise monitoring module. In a specific implementation, the generation of the modulated covert signal is achieved through a digital signal processor, which executes a numerical integration algorithm for the bistable state-well nonlinear system to output a signal sequence.

[0032] In one embodiment of the present invention, such as Figure 3As shown, the generation of the artificial noise signal begins with acquiring the physical fingerprint characteristics of the IoT transmitter hardware. These characteristics include the nonlinear characteristics of the RF front-end power amplifier and the phase noise spectrum of the local oscillator. Based on these physical fingerprint characteristics, a baseband noise sequence with a specific statistical distribution is generated through a pseudo-random sequence generator. This baseband noise sequence is designed to be uncorrelated with the original communication signal in both the time and frequency domains. The baseband noise sequence is then subjected to shaping filtering, with the frequency response characteristics of the filter specifically designed to complement the multipath delay power spectrum of the current wireless channel. The filtered baseband noise sequence is up-converted to the transmit carrier frequency to generate the final artificial noise signal, the power level of which is set according to a preset signal-to-interference-plus-noise ratio (SNR) target. After generating the modulated covert signal and the artificial noise signal, a weighted mixing stage is entered. This stage first calculates an optimal power mixing ratio based on preset communication covertness indicators and receiver SNR requirements. Based on this power mixing ratio, the modulated covert signal and the artificial noise signal are then subjected to corresponding power scaling. The modulated covert signal with power scaling and the artificial noise signal with power scaling are added in phase in the complex domain to form a hybrid baseband signal. This hybrid baseband signal is then subjected to digital-to-analog conversion and radio frequency modulation to generate a synthetic transmit signal that can be transmitted by the antenna.

[0033] In practical implementation, the generation of artificial noise signals and their weighted mixing with covert signals begin with acquiring the physical fingerprint characteristics of the IoT transmitter hardware. These physical fingerprint characteristics include the nonlinear characteristics of the RF front-end power amplifier and the phase noise spectrum of the local oscillator. Specifically, the nonlinear characteristics of the RF front-end power amplifier are characterized by measuring the input-output amplitude response and extracting the third-order cross-tuning point value. The phase noise spectrum of the local oscillator is measured using a spectrum analyzer and described as noise power density values ​​at a set of discrete frequency points.

[0034] In some embodiments, a baseband noise sequence with a specific statistical distribution is generated based on physical fingerprint features using a pseudo-random sequence generator. The initialization seed of the pseudo-random sequence generator is derived from the hash value of the physical fingerprint features. The generated baseband noise sequence is uncorrelated with the original communication signal in both the time and frequency domains. In a specific implementation, the baseband noise sequence is shaped and filtered. The frequency response characteristics of the shaped filter are designed to be complementary to the multipath delay power spectrum of the current wireless channel. This complementarity is reflected in the fact that the shaped filter has a lower gain in the high-power band of the channel's multipath delay power spectrum, and the impulse response of the shaped filter... satisfy: in: It is the frequency response of the shaped filter. It is a digital angular frequency. It is a positive real-valued scaling factor less than 1 used to control the strength of complementarity. It is the normalized multipath delay power spectral density function of the current wireless channel. It can be understood that this design aims to enhance the transmit power of artificial noise signals at frequencies where the channel is fading, thereby worsening the reception conditions for potential interceptors at these frequencies.

[0035] In specific implementations, the filtered baseband noise sequence is up-converted to the transmit carrier frequency to generate an artificial noise signal. The power level of the artificial noise signal is set according to a preset signal-to-interference-plus-noise ratio (SINR) target. This preset SINR target is a system configuration parameter used to balance communication concealment and receiver demodulation reliability. In some embodiments, the power level setting is achieved through an automatic gain control loop, which adjusts the digital gain of the baseband noise sequence based on real-time power measurements of the synthesized transmit signal.

[0036] In practical implementation, the weighted mixing stage calculates the power mixing ratio of the modulated covert signal and the artificial noise signal based on a preset concealment index and the receiver's signal-to-noise ratio (SNR) requirement. The preset concealment index can be a waveform similarity threshold, and the receiver's SNR requirement is determined by the receiver's demodulation threshold. Power mixing ratio The calculation can be expressed as the ratio of the modulated covert signal power to the artificial noise signal power. Its specific value is obtained by solving an optimization problem constrained by covertness requirements and receiver signal-to-noise ratio requirements. In practice, the modulated covert signal and artificial noise signal are power scaled according to the power mixing ratio. This power scaling is achieved through a digital multiplier, which multiplies the modulated covert signal sequence by a coefficient. Multiply the artificial noise signal sequence by a coefficient Optionally, the power mixing ratio can be a time-varying parameter that is adaptively adjusted based on the channel estimation results.

[0037] In practical implementation, the modulated covert signal with power scaling and the artificial noise signal with power scaling are added in phase in the complex domain to form a hybrid baseband signal. The addition operation is performed in the adder of the digital baseband processing unit. In another implementation, the hybrid baseband signal undergoes digital-to-analog conversion and radio frequency modulation to generate the final synthesized transmit signal. The digital-to-analog converter converts the digital hybrid baseband signal into an analog signal, and the radio frequency modulator up-converts the analog baseband signal to the target radio frequency band and amplifies it. It can be understood that in-phase addition ensures that the modulated covert signal and the artificial noise signal achieve waveform superposition at the transmit antenna. Optionally, additional hardware nonlinearity may be introduced in the radio frequency modulation stage. This nonlinearity has been pre-included in the physical fingerprint characteristics and is therefore considered during the generation of the artificial noise signal.

[0038] In one embodiment of the present invention, the AI ​​signal separation network adopts an encoder-decoder architecture, where the encoder network is responsible for extracting multi-scale depth features of the noisy received signal, and the decoder network reconstructs the signal and noise components based on these depth features. During processing, the temporal sampling sequence of the noisy received signal is input into the encoder network. The encoder network consists of multiple one-dimensional convolutional layers and interleaved downsampling layers, extracting multi-level feature maps from fine-grained to coarse-grained levels through layer-by-layer processing. A feature attention fusion layer is set between the encoder and decoder networks. This layer adaptively weights and fuses feature maps from different scales of the encoder, enhancing the weights of feature channels related to the modulated hidden signal and suppressing irrelevant features. The decoder network receives the attention-fused depth features and gradually restores the spatial resolution of the signal through multiple deconvolutions and upsampling operations. During decoding, the decoder fuses feature information from corresponding levels of the encoder through skip connections to supplement details. The decoder network outputs two reconstructed sequences in parallel at its end, corresponding to a noise-suppressed signal sequence and a residual noise signal sequence, respectively.

[0039] In its implementation, the AI ​​signal separation network, which separates signal components from noisy received signals, comprises an encoder network and a decoder network. The encoder network extracts multi-scale depth features from the noisy received signal, while the decoder network reconstructs the signal and noise components from these depth features. Specifically, the temporal sampling sequence of the noisy received signal is input into the encoder network. The encoder network extracts multi-level feature maps from fine-grained to coarse-grained levels through multiple layers of one-dimensional convolution and downsampling operations. Each layer of the encoder network consists of a one-dimensional convolutional layer, an activation function layer, and a pooling layer in sequence. The one-dimensional convolutional layers use kernels of different widths to capture features at different time scales, and the pooling layers perform downsampling operations to expand the receptive field of the feature maps. In essence, the multi-level feature maps represent the noisy received signal at different levels of abstraction and different temporal resolutions.

[0040] In some embodiments, the kernel width of the one-dimensional convolutional layer is set to [7,5,3,3], and the corresponding downsampling factor is set to [2,2,2,1], thereby forming a four-layer coding structure. For the specific configuration parameters of the network layers, please refer to Table 1.

[0041] Table 1: Partial Hierarchical Configuration Table of AI Signal Separation Network Encoder In practical implementation, a feature attention fusion layer is set between the encoder network and the decoder network to perform weighted fusion of feature maps from different scales. The feature attention fusion layer receives the feature maps output from each layer of the encoder network, first applies global average pooling to each scale of the feature map to compress it into channel description vectors, and then generates attention weights for each channel through a shared multilayer perceptron network. The calculation can be expressed as: in: Indicates the index at the feature layer level. Indicates the channel index. It is the first Channel description vectors of the layer feature map, , and , These are the weights and bias parameters of the multilayer perceptron network. It is the ReLU activation function. It is the sigmoid activation function. Weighted fusion is achieved by multiplying the original feature map with the corresponding channel attention weights channel by channel. This operation can be understood as highlighting feature channels relevant to the modulated hidden signal and suppressing irrelevant or interfering feature channels.

[0042] In some embodiments, the feature attention fusion layer further includes a cross-scale feature stitching operation, which upsamples the weighted multi-scale feature maps to the same temporal resolution and then stitches them together along the channel dimension.

[0043] In specific implementations, the decoder network outputs two reconstructed sequences in parallel through multi-layer deconvolution and upsampling operations, fusing skip connection features from the encoder. Each layer of the decoder network contains an upsampling operation, a deconvolution layer, and a skip connection corresponding to the feature map of the encoder layer. The skip connection concatenates the feature map from the encoder network with the upsampled feature map from the decoder network along the channel dimension, providing detailed information to the decoder. Optionally, the upsampling operation can be implemented using nearest neighbor interpolation or transposed convolution. In specific implementations, the decoder network outputs two reconstructed sequences in parallel at the end: a noise-suppressed signal sequence and a residual noise signal sequence. The output layer consists of two independent one-dimensional convolutional layers, each with a kernel width of 1, mapping the high-dimensional features to a single-channel one-dimensional time series. These two one-dimensional time series correspond to the estimates of the modulated hidden signal and the artificial noise signal, respectively. Optionally, during the training phase, constraints can be imposed on the noise-suppressed signal sequence and the residual noise signal sequence to make their sum approximate the noisy input signal as closely as possible; this serves as an additional reconstruction loss term.

[0044] In one embodiment of the present invention, the construction of the training dataset is a prerequisite for network training. The construction process includes multiple steps: collecting a large number of typical IoT communication signals to form a rich sample raw signal library; simulating various typical wireless channel environments to generate sample channel models with different multipath, Doppler shift, and path loss characteristics; for each pair consisting of a sample raw signal and a sample channel model, the following operations are performed to generate training samples: calling an improved stochastic resonance algorithm to process the sample raw signal and generate the corresponding sample cover signal; and generating a corresponding sample artificial noise signal based on the characteristics of the current sample channel model. Statistical features are intentionally introduced to confuse the samples during the construction of the training dataset. When generating the sample artificial noise signal, its power spectral density and variance distribution are forced to be consistent with or highly overlap with the sample channel noise. Simultaneously, based on the characteristics of the sample channel model, a transmitter-specific nonlinear transformation marker is introduced into the sample artificial noise signal, derived from a nonlinear mapping of physical fingerprint features, ensuring structural distinguishability even with similar statistical distributions. The generated sample cover signal and sample artificial noise signal are transmitted through the sample channel model for simulation, and Gaussian white noise is added to obtain the simulated received sample signal. A training sample is formed by combining the original sample signal, its corresponding hidden sample signal, the sample artificial noise signal, and the sample received signal. A large number of such training samples constitute the complete training dataset. After obtaining the training dataset, the training process of the AI ​​signal separation network is initiated, using the sample received signal as the input and the corresponding hidden sample signal and sample artificial noise signal as the desired output. The network's loss function is defined, comprising two parts: the mean squared error between the network-reconstructed hidden signal and the sample hidden signal, and the mean squared error between the network-reconstructed artificial noise signal and the sample artificial noise signal. The total loss is the weighted sum of these two errors. Using the backpropagation algorithm and optimizer, all trainable parameters in the AI ​​signal separation network are iteratively updated until the value of the loss function converges to below a preset threshold. At this point, the desired trained AI signal separation network is obtained.

[0045] At the IoT receiver, the residual noise signal sequence output by the AI ​​signal separation network is acquired. The statistics of this sequence are calculated as channel state feedback features. Specifically, it includes: in, Represents variance. Indicates kurtosis, This indicates the peak-to-average power ratio.

[0046] The receiving end transmits the data through a low-rate control channel. The data is then transmitted back to the transmitter. The transmitter deploys a reinforcement learning agent whose action space involves adjusting the power mixing ratio. Define the reward function. To balance concealment and reliability: in, As a weighting factor, Based on the current Estimated bit error rate.

[0047] The launcher is rewarded Update power mixing ratio (i.e., the weighted proportion of artificial noise): in For learning rate, For the reward function with respect to The gradient is calculated. Through iterative updates, the system converges in real time to the optimal covert communication equilibrium point.

[0048] In practical implementation, the trained AI signal separation network is obtained through specific methods. Constructing a training dataset is a prerequisite for network training. The training data samples include sample covert signals, sample artificial noise signals, sample channel noise, and sample received signals generated by mixing them. The sample covert signals are generated by applying an improved stochastic resonance algorithm to the original sample signals. In practical implementation, a large number of typical IoT communication signals are collected as the original sample signal library. The original sample signal library covers signals with various modulation methods and data rates, such as signals with binary phase shift keying, quadrature phase shift keying, and Gaussian frequency shift keying modulation. The signal data rates cover typical IoT application scenarios from low speed to high speed. In practical implementation, various typical wireless channel environments are simulated to generate sample channel models with different multipath, Doppler, and path loss characteristics. Multipath characteristics are described by extended typical urban channel models or rural area channel models. Doppler frequency shift simulates the impact of terminal movement speed, and path loss follows a log-normal shadowing fading model.

[0049] In some embodiments, the sample channel model is implemented as a tapped delay line model, where each tap has a specific delay, average power, and Doppler spectrum. The channel model parameter set is shown in Table 2.

[0050] Table 2: Sample Channel Model Parameter Configuration Table In specific implementations, for each pair of original sample signals and sample channel models, operations are performed to generate training samples. These operations include calling an improved stochastic resonance algorithm to process the original sample signals to generate a hidden sample signal, generating a corresponding artificial noise signal based on the characteristics of the sample channel model, transmitting the hidden sample signal and the artificial noise signal through the sample channel model, and adding Gaussian white noise to obtain the received sample signal. In some embodiments, the generation of the artificial noise signal needs to consider the characteristics of the corresponding sample channel model, and the complementary frequency response design of the shaping filter will be dynamically calculated based on the multipath delay power spectrum of the current sample channel model. In specific implementations, the original sample signal, the corresponding hidden sample signal, the artificial noise signal, and the received sample signal are combined into a quadruplet training sample. All quadruplets constitute the training dataset, and each quadruplet is stored in the dataset in a time-aligned manner to ensure a one-to-one correspondence between the sampling points of the signal sequence. Optionally, the training dataset can be further divided into a training set, a validation set, and a test set for model training, hyperparameter tuning, and performance evaluation.

[0051] In practical implementation, the received sample signal is used as the input to the AI ​​signal separation network, and the corresponding sample hidden signal and sample artificial noise signal are used as the expected output. A loss function is defined to guide network training. The loss function includes the mean square error between the reconstructed hidden signal and the sample hidden signal, and the mean square error between the reconstructed artificial noise signal and the sample artificial noise signal. The weighted sum of these two errors constitutes the total loss. Loss Function The specific form is as follows: in: This represents the length of the signal sequence in a training batch. It is the first signal reconstructed by the AI ​​signal separation network. Hidden signals at each sampling point It is the first of the corresponding sample hidden signals One sampling point, It is the first signal reconstructed by the AI ​​signal separation network. Artificial noise signal at each sampling point It is the first of the corresponding sample artificial noise signals One sampling point, The weighting coefficients, ranging from 0 to 1, are used to balance the importance of the two errors. This loss function directly constrains the network output to approximate the true signal components.

[0052] In practice, the parameters of the AI ​​signal separation network are iteratively updated using a backpropagation algorithm and an optimizer. The backpropagation algorithm calculates the gradient of the loss function relative to the parameters of each layer of the network, and the optimizer uses an adaptive moment estimation algorithm to update the parameters based on the gradient. This iterative update continues until the average value of the loss function on the validation set converges to a preset threshold. Optionally, the preset threshold can be set to a small positive number, and the training process can be terminated early when the validation set loss no longer decreases significantly over multiple consecutive training epochs. In practice, after obtaining the trained AI signal separation network, its network parameters are fixed and deployed in the signal processing unit of the IoT receiver.

[0053] In one embodiment of the present invention, this embodiment describes the inverse processing procedure for recovering the original communication signal estimate from a noise-suppressed signal. This process begins with key parameters acquired and recorded during the signal concealment process at the transmitting end. These parameters are system parameters used by the improved stochastic resonance algorithm when generating the modulated concealed signal, and are used as inverse system parameters at the receiving end, specifically including the inverse potential function parameters and the inverse damping coefficient. At the receiving end, an inverse bistable state-well nonlinear system is constructed based on these inverse system parameters. The equation of motion of this system is determined by the inverse potential function parameters and the inverse damping coefficient. The noise-suppressed signal output from the AI ​​signal separation network is used as a driving signal input to this inverse bistable state-well nonlinear system. In this inverse system, the signal energy undergoes a reverse transfer, shifting from the resonance frequency back to the vicinity of the baseband frequency of the original communication signal. The output of the inverse bistable state-well nonlinear system is low-pass filtered to remove out-of-band noise and high-frequency components. The filtered signal is sampled and decided upon to recover the estimated signal of the original communication signal, completing the entire secure communication decoding process.

[0054] In practice, the process of reverse-invoking the improved stochastic resonance algorithm begins with acquiring and recording the inverse system parameters used for signal concealment from the processing of the improved stochastic resonance algorithm at the transmitting end. These inverse system parameters include the inverse potential function parameters and the inverse damping coefficient. In practice, the inverse potential function parameters and the inverse damping coefficient are complementary parameters that the transmitting end synchronously calculates and stores when adaptively adjusting the nonlinear system parameters based on the statistical characteristics of the current channel noise. These parameters are transmitted to the legitimate IoT receiver via a secure control channel or in a reserved field in the communication protocol.

[0055] In some embodiments, the inverse potential function parameters and inverse damping coefficients can be independently regenerated at the receiver using a pre-shared key and the same channel noise estimation algorithm, thus eliminating the need for explicit transmission in the channel. In a specific implementation, an inverse bistable state-well nonlinear system is constructed at the receiver, and the equations of motion for this system are determined by the inverse potential function parameters and the inverse damping coefficients.

[0056] In specific implementations, the noise-suppressed signal is used as the input driving signal for the inverse bistable state-well nonlinear system. The noise-suppressed signal is an estimated sequence of modulated hidden signals extracted from the noisy received signal by the AI ​​signal separation network. In the inverse bistable state-well nonlinear system, signal energy is transferred back from the resonant frequency to near the baseband frequency of the original communication signal. This process is the reverse of the random resonant energy transfer at the transmitter. Its physical essence is to use nonlinear system parameter adjustments to place the system operating point in another steady state, causing a reverse shift in the spectrum of the driving signal. In some embodiments, the numerical solution of the inverse bistable state-well nonlinear system is implemented in real time on a digital signal processor using the fourth-order Runge-Kutta method. Optionally, to ensure numerical stability, the solution steps need to be adaptively adjusted according to the highest frequency component of the input noise-suppressed signal.

[0057] In practical implementation, the output of the inverse bistable state-well nonlinear system undergoes low-pass filtering and sampling decision-making to recover the estimated signal of the original communication signal. The purpose of low-pass filtering is to remove high-frequency noise components generated during the inverse processing and any possible residual out-of-band interference. The cutoff frequency of the low-pass filter is set to 1.1 to 1.3 times the baseband bandwidth of the original communication signal to ensure distortion-free signal transmission and effective suppression of out-of-band noise. In practical implementation, the sampling decision-making operation is performed on the filtered analog signal or high-sampling-rate digital signal. The decision-maker determines the signal amplitude, phase, or frequency at the optimal sampling time based on the modulation scheme of the original communication signal, and finally outputs the bitstream estimate of the original communication signal. It can be understood that sampling decision-making is a standard step in communication demodulation, and its specific implementation depends on the modulation and coding scheme of the original signal. Optionally, for digitally modulated signals, timing synchronization and carrier phase recovery can be performed before sampling decision-making. In practical implementation, the entire inverse processing process—including inverse system construction, driving, filtering, and decision-making—is completed in a dedicated signal processing chip or programmable gate array at the IoT receiver.

[0058] In one embodiment of the present invention, an IoT-based AI physical layer anti-interception secure communication system is also included. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the above-described IoT-based AI physical layer anti-interception secure communication method.

[0059] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A secure communication method for AI physical layer anti-interception based on the Internet of Things, characterized in that, The method includes: Generate the raw communication signal to be transmitted at the IoT transmitter; An improved stochastic resonance algorithm is invoked to perform signal hiding processing on the original communication signal. The improved stochastic resonance algorithm adaptively adjusts the nonlinear system parameters based on the statistical characteristics of the current channel noise to generate a modulated hidden signal. An artificial noise signal unrelated to the original communication signal is generated at the IoT transmitter. The modulated covert signal and the artificial noise signal are weighted and mixed at the physical layer to generate a synthetic transmission signal; The synthesized transmission signal is transmitted via a wireless channel; The IoT receiver receives noisy signals transmitted via wireless channels. At the IoT receiver, a trained AI signal separation network is invoked to separate the noise-suppressed signal and the remaining noise signal from the noisy received signal; The improved stochastic resonance algorithm is invoked in reverse from the noise-suppressed signal to recover the estimated signal of the original communication signal.

2. The IoT-based AI physical layer anti-interception secure communication method according to claim 1, characterized in that, The improved stochastic resonance algorithm is invoked to perform signal hiding processing on the original communication signal. The improved stochastic resonance algorithm adaptively adjusts the nonlinear system parameters based on the statistical characteristics of the current channel noise to generate a modulated hidden signal, including: Extract the feature vector of the original communication signal, the feature vector including signal power spectral density, zero-crossing rate and time-domain envelope shape; Real-time monitoring of background noise in the wireless channel to obtain the statistical characteristics of the current channel noise, including noise variance, power spectrum and higher-order cumulative quantities; The original communication signal is input as a periodic weak driving signal into a bistable state trap nonlinear system, and the equation of motion of the bistable state trap nonlinear system is determined by the system potential function parameters and the damping coefficient. Based on the statistical characteristics of the current channel noise, the system potential function parameters and the damping coefficient are dynamically adjusted through a parameter adapter, so that the current channel noise and the original communication signal resonate within the bistable potential trap nonlinear system. In the resonant state, the signal energy output by the bistable state trap nonlinear system shifts from the frequency of the original communication signal to the vicinity of the system characteristic frequency, forming the modulated covert signal with altered spectral characteristics that matches the statistical characteristics of the background noise.

3. The IoT-based AI physical layer anti-interception secure communication method according to claim 2, characterized in that, Based on the statistical characteristics of the current channel noise, the system potential function parameters and the damping coefficient are dynamically adjusted through a parameter adapter, including: Based on the noise variance, the critical barrier height reference value of the bistable potential trap nonlinear system is calculated, and the critical barrier height reference value is proportional to the square root of the noise variance. Based on the power spectrum of the noise, the concentrated frequency band of the noise energy is identified, and the adjustment amount of the system characteristic frequency is calculated so that the system characteristic frequency deviates from the concentrated frequency band of the noise energy. The non-Gaussian intensity of the noise is estimated using the higher-order cumulative amount of the noise, and the weight of the nonlinear term of the damping coefficient is adjusted accordingly to match the statistical distribution of the noise. Based on the critical barrier height reference value, the adjustment amount of the system characteristic frequency, and the weight of the nonlinear term, the real-time configuration parameters of the system potential function parameters and the damping coefficient are generated and injected into the bistable potential trap nonlinear system.

4. The IoT-based AI physical layer anti-interception secure communication method according to claim 1, characterized in that, The generation of artificial noise signals unrelated to the original communication signals at the IoT transmitter includes: Obtain the physical fingerprint characteristics of the IoT transmitter hardware, including the nonlinear characteristics of the radio frequency front-end power amplifier and the phase noise spectrum of the local oscillator; Based on the physical fingerprint features, a baseband noise sequence with a specific statistical distribution is generated by a pseudo-random sequence generator. The baseband noise sequence is uncorrelated with the original communication signal in both the time and frequency domains. The baseband noise sequence is shaped and filtered, and the frequency response characteristics of the filter are designed to be complementary to the multipath delay power spectrum of the current wireless channel. The filtered baseband noise sequence is upconverted to the transmit carrier frequency to generate the artificial noise signal. The power level of the artificial noise signal is set according to a preset signal-to-interference-plus-noise ratio target.

5. The IoT-based AI physical layer anti-interception secure communication method according to claim 1, characterized in that, The modulated covert signal and the artificial noise signal are weighted and mixed at the physical layer to generate a synthetic transmission signal, including: Based on the preset concealment index and the signal-to-noise ratio requirement of the receiver, the power mixing ratio of the modulated concealed signal and the artificial noise signal is calculated. The modulated covert signal and the artificial noise signal are power scaled according to the power mixing ratio; The modulated covert signal with power scaling and the artificial noise signal with power scaling are added in phase in the complex domain to form a hybrid baseband signal; The hybrid baseband signal is subjected to digital-to-analog conversion and radio frequency modulation to generate the final synthesized transmission signal.

6. The IoT-based AI physical layer anti-interception secure communication method according to claim 1, characterized in that, At the IoT receiver, a trained AI signal separation network is invoked to separate the noise-suppressed signal and the residual noise signal from the noisy received signal, including: The AI ​​signal separation network includes an encoder network and a decoder network. The encoder network is used to extract multi-scale depth features of the noisy received signal, and the decoder network is used to reconstruct signal components and noise components from the depth features. The temporal sampling sequence of the noisy received signal is input into the encoder network, which extracts multi-level feature maps from fine-grained to coarse-grained through multi-layer one-dimensional convolution and downsampling operations. A feature attention fusion layer is set between the encoder network and the decoder network to perform weighted fusion of feature maps from different scales, highlighting feature channels related to the modulated hidden signal; The decoder network outputs two reconstructed sequences in parallel through multi-layer deconvolution and upsampling operations, and by fusing skip connection features from the encoder: a noise-suppressed signal sequence and a residual noise signal sequence.

7. The IoT-based AI physical layer anti-interception secure communication method according to claim 6, characterized in that, The trained AI signal separation network was obtained through the following method: A training dataset is constructed, and the training data samples include: sample cover signals, sample artificial noise signals, sample channel noise, and sample received signals generated by mixing them. The sample cover signals are generated by calling the improved stochastic resonance algorithm on the original sample signals. The received sample signal is used as the input to the AI ​​signal separation network, and the corresponding sample hidden signal and sample artificial noise signal are used as the desired output. Define a loss function that includes the mean square error between the reconstructed hidden signal and the sample hidden signal, and the mean square error between the reconstructed artificial noise signal and the sample artificial noise signal, and the weighted sum of the two errors; The parameters of the AI ​​signal separation network are iteratively updated using the backpropagation algorithm and optimizer until the loss function converges to a preset threshold, thus obtaining the trained AI signal separation network.

8. The IoT-based AI physical layer anti-interception secure communication method according to claim 7, characterized in that, The construction of the training dataset includes: A large number of typical IoT communication signals were collected as a sample raw signal library; Simulate various typical wireless channel environments to generate sample channel models with different multipath, Doppler, and path loss characteristics; For each pair of original sample signals and sample channel models, the following steps are performed: the improved stochastic resonance algorithm is invoked to process the original sample signals and generate a sample cover signal; a corresponding sample artificial noise signal is generated based on the characteristics of the sample channel model; the sample cover signal and the sample artificial noise signal are transmitted through the sample channel model, and Gaussian white noise is added to obtain the sample received signal; The original sample signal, the corresponding hidden sample signal, the sample artificial noise signal, and the sample received signal are combined into a four-tuple training sample, and all four-tuples constitute the training dataset.

9. The IoT-based AI physical layer anti-interception secure communication method according to claim 1, characterized in that, The improved stochastic resonance algorithm is invoked in reverse from the noise-suppressed signal to recover the estimated signal of the original communication signal, including: During the processing at the transmitting end of the improved stochastic resonance algorithm, inverse system parameters for signal concealment are obtained and recorded. These inverse system parameters are independently generated by the transmitting and receiving ends using the same channel noise estimation algorithm based on a pre-shared key, without explicit transmission. The inverse system parameters include inverse potential function parameters and inverse damping coefficients. A nonlinear inverse bistable potential trap system is constructed at the receiving end, and its equation of motion is determined by the inverse potential function parameters and the inverse damping coefficient. The noise suppression signal is used as the input driving signal for the inverse bistable state trap nonlinear system. In the inverse bistable state trap nonlinear system, signal energy is transferred back from the resonant frequency to near the baseband frequency of the original communication signal; The output of the inverse bistable state trap nonlinear system is low-pass filtered and sampled for decision-making to recover the estimated signal of the original communication signal.

10. An AI physical layer anti-interception secure communication system based on the Internet of Things, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the IoT-based AI physical layer anti-interception secure communication method according to any one of claims 1 to 9.

Images