Intelligent anti-interception transmission method based on full convolutional auto-encoder network

Abstract

The invention belongs to the technical field of wireless communication and deep learning, and particularly relates to an intelligent anti-interception transmission method based on a full convolutional auto-encoder network. According to the invention, a time-frequency domain combined signal processing scheme is designed through a pre-shared scrambling knowledge base and a pre-shared scrambling mode pattern: at a transmitting end, frequency domain notch, frequency domain scrambling and time domain phase rotation processing are performed on an original signal in sequence, and a transmission signal with time-frequency domain dual disturbance is constructed; and at a receiving end, the legal user sequentially executes time domain phase compensation and frequency domain notch processing based on the shared prior information, and intelligent reconstruction of the signal is realized by using a pre-trained full convolution auto-encoder network. According to the method, by constructing the asymmetric performance advantages of the legal receiver and the third party, it is ensured that the legal receiver can accurately recover the original signal, the third party cannot effectively analyze the signal due to lack of key disturbance parameters and a reconstruction algorithm, and the anti-interception performance of the system is remarkably improved while the communication quality is ensured.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communication and deep learning technology, and specifically relates to an intelligent anti-interception transmission method based on a fully convolutional autoencoder network. Background Art

[0002] With the rapid development of wireless communication technology and the exponential growth of intelligent terminal devices, such as smartphones and IoT devices, the security of communication systems faces unprecedented challenges. While traditional encryption schemes can ensure the confidentiality of data transmission, they still pose the risk of third-party interception and analysis during physical layer signal transmission, potentially exposing communication activity. In this context, improving the anti-interception capabilities of communication systems has become a key factor in ensuring communication quality.

[0003] The main goal of anti-interception technology is to change the time-frequency characteristics or expression of the signal through specific signal processing methods, eliminate its identifiable features and reconstruct the signal mapping relationship, making it difficult for unauthorized parties to use conventional detection methods to intercept and analyze the signal. The current mainstream anti-interception solutions mainly include the following types:

[0004] (1) Time-frequency domain processing technology: mainly includes two typical methods: direct sequence spread spectrum and frequency hopping. Direct sequence spread spectrum uses a pseudo-random code sequence to expand the signal spectrum, significantly reducing the power spectrum density of the signal, making it better hidden in the environmental noise; frequency hopping technology uses a preset frequency hopping sequence to achieve dynamic switching of the carrier frequency, effectively avoiding the synchronous capture of illegal receiving equipment.

[0005] (2) Transform Domain Processing: This technology dynamically generates a modulation waveform adapted to the current channel conditions through real-time perception of the channel environment and analysis of spectrum resources. This adaptively generated signal waveform can perfectly blend effective information into background noise, significantly improving the signal's concealment performance.

[0006] (3) Chaotic sequence technology: This technology utilizes the high randomness and initial value sensitivity of the pseudo-random sequence generated by the chaotic system, and uses it as a spread spectrum sequence or modulation parameter, etc., which can significantly enhance the anti-interception characteristics of the signal. This type of technology is usually used in combination with other anti-interception schemes to produce better protection effects.

[0007] (4) Signal masking technology: This technology completely hides the effective communication signal in the interference signal by designing a specific high-power interference signal. This technology can effectively circumvent conventional signal detection methods and achieve deep concealment of communication signals.

[0008] While existing anti-interception technologies can achieve a certain degree of signal concealment, the development of intelligent signal detection technology has challenged their effectiveness, necessitating the development of more differentiated and intelligent anti-interception solutions. Furthermore, while ensuring the performance of legitimate receivers, further enhancing resistance to intelligent third-party devices has become a new direction for the development of anti-interception technology. Summary of the Invention

[0009] To address these issues, the present invention proposes an intelligent, interception-resistant transmission solution. By jointly optimizing transmitter-side signal processing and receiver-side signal recovery methods, this solution creates an asymmetric performance advantage between legitimate receivers and third parties, reducing the legitimate receiver's bit error rate while improving the third party's bit error rate. The core of this solution is to make it difficult for third parties to effectively detect and analyze transmitted signals using traditional methods, while ensuring that legitimate receivers can accurately recover the original signal based on prior information. The following describes the intelligent, interception-resistant transmission technology proposed in this invention.

[0010] 1) Intelligent anti-interception transmission technology ideas

[0011] The intelligent anti-interception transmission technology proposed by the present invention is as follows: Figure 1 and Figure 2 As shown, Figure 1 It is the time-frequency domain scrambling module at the transmitter. Figure 2 This is the time-frequency domain descrambling module at the receiving end. The implementation of this technology requires that both the sender and receiver establish a shared scrambling knowledge base and scrambling pattern map in advance to define the time-varying perturbation position and phase change patterns. Based on this agreement, the system's specific signal processing flow can be divided into three stages: the transmitter, the legitimate receiver, and the third party.

[0012] like Figure 1 As shown, at the transmitter, the original signal is first determined. Then, a portion of the frequency band is notched based on the time-varying perturbation position change pattern agreed upon by the sender and receiver, resulting in a frequency-domain distorted signal. Subsequently, noise or interference of equal intensity to the original signal is added at the notched position, forming a frequency-domain scrambled signal with perturbations. Finally, the frequency-domain scrambled signal is phase-rotated based on the time-varying phase change pattern agreed upon by the sender and receiver, resulting in a transmitted signal with both frequency-domain and time-domain perturbations.

[0013] like Figure 2As shown, at the legitimate receiving end, the received signal is first determined. Then, the time-domain disturbance information is determined based on the pre-agreed time-varying phase variation pattern, and the time-domain phase is compensated to obtain the time-domain denoised signal. Subsequently, the frequency-domain disturbance information is determined based on the pre-agreed time-varying disturbance position variation pattern. The interference component is filtered out using a frequency-domain notch filter, resulting in a distorted signal with missing information in some frequency bands, i.e., the frequency-domain distorted signal. Finally, this frequency-domain distorted signal is input into a pre-trained fully convolutional autoencoder network. Leveraging the network's powerful feature extraction and signal reconstruction capabilities, the network achieves high-precision intelligent recovery of the original signal.

[0014] Compared to the legitimate receiving end, the third party at the third end will also receive a received signal with time-frequency domain disturbances. However, in general, since the third party cannot obtain the time-varying time-frequency domain disturbance information, it is difficult for it to accurately detect the useful signal from the disturbed signal. In addition, even if the third party knows the time-domain phase disturbance information and the frequency-domain disturbance position information, it is difficult to recover the useful signal from the distorted signal where some frequency band information is missing. This asymmetric design allows the legitimate receiver to maintain high-quality communication while significantly improving the third party's bit error rate, thereby achieving reliable anti-interception transmission.

[0015] This invention innovatively combines dynamic time-frequency domain scrambling with intelligent signal recovery, retaining the advantages of traditional anti-interception techniques while improving system performance through intelligent processing. By combining frequency-domain notch scrambling with phase rotation, combined with the signal reconstruction capabilities of a fully convolutional autoencoder network used in deep learning, it effectively protects against various third-party threats while ensuring the quality of legitimate communications.

[0016] 2) Analysis of the mechanism of intelligent anti-interception transmission technology

[0017] This section theoretically explains the working principle of intelligent anti-interception transmission technology, focusing on the signal transformation process and its mathematical relationship at the transmitter. Assume that the original signal of the sender is , which is expressed as

[0018]

[0019] in, is the original signal In the The sampling value at a time point, is the signal length, the superscript “ ” indicates a transpose operation.

[0020] Do the original signal Point discrete Fourier transform DFT to obtain frequency domain signal for

[0021]

[0022] in, is a frequency domain signal In the The spectrum value on the frequency component, is the DFT transformation matrix, that is

[0023]

[0024] ,

[0025] in, is a matrix The element in row k and column n, is the rotation factor, j is the imaginary unit, It is pi.

[0026] In order to prevent a third party from accurately recovering the original signal even if it intercepts the received signal , perform frequency domain zero-falling on the frequency domain signal. Assume that The frequency domain signal between is zero-pitched, then the frequency domain distortion signal after zero-pitching for

[0027]

[0028] In order to confuse the third party, an interference signal with the same strength as the original signal is generated, and its frequency band is limited to the frequency band between the zero-pitched frequencies of the original signal. for

[0029]

[0030] in, It is an interference signal In the The spectrum value of the frequency component.

[0031] The frequency domain distorted signal after zeroing and interference signals Phase superposition can obtain the frequency domain scrambled signal for

[0032]

[0033] Scramble the signal in the frequency domain Do The point discrete Fourier transform IDFT is used to obtain the time domain form of the frequency domain scrambled signal for

[0034]

[0035] in, It is the time domain form of the frequency domain scrambled signal In the The sampling value at each time point, with the superscript " ” denotes the conjugate operation.

[0036] Time domain form of frequency domain scrambled signal Perform time domain phase rotation Radians, get the time-frequency domain scrambled signal for

[0037]

[0038] in, is the time-frequency domain scrambled signal In the The sampling value at a time point.

[0039] Then, the time-frequency domain scrambled signal is used as the transmission signal and transmitted through the Gaussian channel.

[0040] The above analysis shows that the dual perturbation mechanism in the time and frequency domains employed by the transmitter effectively undermines the signal's recoverability. Legitimate receivers, leveraging pre-shared prior information and the robust recovery capabilities of deep neural networks, can achieve reliable recovery. However, third parties, lacking critical prior information and recovery algorithms, face significant information loss.

[0041] 3) Implementation steps of intelligent anti-interception transmission technology

[0042] In order to realize the intelligent anti-interception transmission technology proposed in the present invention, it is first necessary to establish a scrambling knowledge base and a scrambling pattern diagram to specify the time-varying disturbance position change law and the time-varying phase change law of the sender and receiver. Then, a simulation communication system with anti-interception capability is built. Subsequently, a data set is constructed using this system, and a fully convolutional autoencoder network is trained offline to enable it to restore the frequency domain distorted signal to the original signal. Finally, the scrambling knowledge base, the scrambling pattern diagram and the trained fully convolutional autoencoder network are combined to realize intelligent anti-interception transmission. The specific steps are:

[0043] S1. Establish a scrambling knowledge base and a scrambling pattern diagram: The schematic diagram of the scrambling knowledge base and the scrambling pattern diagram is as follows: Figure 3As shown. Specifically, the scrambling knowledge base consists of three key parameter libraries: a frequency domain notch parameter library, a frequency domain noise generation library, and a time domain phase rotation library. The frequency domain notch parameter library is used to specify key parameters in the frequency domain notch processing process, including the number of notch bands, the center frequency of each band, and the width of each notch band; the frequency domain noise generation library is used to specify key parameters in the frequency domain scrambling process, namely, the noise generation method, such as band-limited noise, random signals, and other interference modes; and the time domain phase rotation library is used to specify key parameters in the time domain phase rotation process, namely, the radian of phase rotation. These three parameter libraries together constitute N different scrambling parameter combinations, each of which corresponds to a specific frequency domain notch, frequency domain noise generation, and time domain phase rotation configuration. The scrambling pattern diagram, as a dynamic scheduling scheme for the scrambling knowledge base, has the core function of specifying the scrambling combination sequence used in each communication sequence. Through this time-sequential scheduling mechanism, the system can dynamically switch between different frequency-domain notching parameters, frequency-domain noise generation methods, and phase rotation radians, thereby achieving coordinated control of the time-varying perturbation position and phase variations. When both the sender and receiver synchronously share this scrambling knowledge base and scrambling pattern map, a completely consistent signal processing benchmark can be established: the transmitter dynamically scrambles the signal according to a predetermined pattern, while the legitimate receiver can accurately descramble and recover the signal based on the same timing rules.

[0044] S2. Build a simulation communication system with anti-interception capability: The overall framework of the simulation communication system in the present invention is as follows: Figure 4 As shown in Figure 1, it consists of three main components: transmitter, channel, and receiver. Specifically:

[0045] ① Transmitter: The transmitter first generates a random binary sequence as the transmission data, then performs QPSK modulation on this sequence to obtain the original signal. Subsequently, this signal is jointly scrambled in the time and frequency domains. This involves first notching a portion of the frequency band based on the time-varying perturbation position and injecting an interference signal of equal intensity to obtain a frequency-domain scrambled signal. This signal is then phase-rotated based on the time-varying phase pattern to obtain the time-domain scrambled signal, or the transmission signal. Finally, the transmission signal is upsampled and filtered to reduce effects such as inter-symbol interference.

[0046] ② Channel: The channel part adopts the Gaussian channel model, so that Gaussian noise is superimposed on the signal during transmission.

[0047] ③ Receiver: A legitimate receiver first obtains the received signal through matched filtering and downsampling. It then performs time-domain descrambling. This involves performing phase compensation based on the time-varying phase pattern to obtain a time-domain descrambled signal. Then, a frequency-domain notch filter is applied based on the position variation of the time-varying disturbance to remove interference, resulting in a frequency-domain distorted signal. This distorted signal is then fed into a fully convolutional autoencoder for frequency-domain compensation, yielding the time-domain compensated signal, or recovered signal. Finally, the recovered signal undergoes QPSK demodulation to obtain a received binary sequence, which is used as the received data.

[0048] S3. Construct a dataset for intelligent anti-interception transmission technology: First, based on the scrambling knowledge base and scrambling pattern diagram constructed in S1, pre-set parameters such as the frequency domain notch position and the time domain phase rotation radian, then generate a random binary sequence as the transmission data and let it pass through the simulated communication system constructed in S2. During the simulation process, the original signal of the transmitter and the frequency domain distorted signal obtained after time domain phase compensation and frequency domain notching at the receiver are synchronously recorded, and the frequency domain distorted signal is used as the input data in the dataset, and the original signal is used as the label in the dataset, so as to create a dataset for training deep neural networks. It should be noted that the frequency domain distorted signal and the original signal are both in complex form, and in the actual processing process, their real and imaginary parts need to be separated and processed independently to ensure that the input data and corresponding labels in the dataset are both in real form.

[0049] S4. Constructing a deep neural network: From the above analysis, it can be seen that the accuracy of the restored signal compared to the original signal is directly affected by the deep neural network. Therefore, in the processing of the deep neural network, the network structure and network parameters will directly affect the recovery accuracy of the restored signal. Considering that there is still a part of noise in the frequency domain distorted signal, the present invention uses a full convolutional autoencoder with denoising capability, that is, a signal generator network to characterize the mapping relationship between the frequency domain distorted signal and the original signal. Then, an efficient deep neural network training method is designed to optimize the parameters so that the deep neural network can generate a realistic original signal based on the frequency domain distorted signal.

[0050] Specifically, the signal generator network used in the present invention is as follows: Figure 5 As shown in Figure 2, the network has a symmetrical autoencoder structure. The encoder and decoder consist of four convolutional layers and four deconvolutional layers, respectively. Each layer uses a PReLU activation function for nonlinear processing. The output of the encoder's last layer is concatenated with random noise as the input to the decoder's first deconvolutional layer. A skip connection mechanism is introduced in each subsequent deconvolutional layer of the decoder. This means that the input to each subsequent deconvolutional layer is the output of the previous layer concatenated with the output of the corresponding symmetric convolutional layer.

[0051] This symmetric autoencoder structure offers significant technical advantages: the encoder effectively extracts deep signal features through layer-by-layer convolution, transforming the input data into a compact, low-dimensional representation; the decoder accurately reconstructs the signal using deconvolution and skip connections. In particular, the PReLU activation function used in the network is an improved ReLU activation function that allows for a slight slope on the negative axis. This helps alleviate the problem of the ReLU function's gradient being zero on the negative axis, leading to better model training.

[0052] The size of the frequency-domain distorted signal input and the restored signal output in the autoencoder network is (1 × 1024 × 1), where the first digit "1" represents the number of channels, the second digit "1024" represents the signal length, and the third digit "1" represents the signal width. During the network's forward propagation, the signal length is halved with each convolutional layer and doubled with each deconvolutional layer. This symmetrical structural design ensures precise matching of signal dimensions.

[0053] In order to assist the training of the signal generator network, the present invention also designs a discriminator network, whose structure is as follows: Figure 6 As shown in Figure 2, the discriminator network takes as input the concatenation of the distorted and original signals, or the concatenation of the distorted and restored signals. After extracting features from the input signal through three convolutional layers, the three-dimensional features are converted into one-dimensional vectors through a flattening layer. Finally, a fully connected layer with a sigmoid activation function outputs the discriminant result. The discriminator network's output value ranges from 0 to 1. Values ​​closer to 0 indicate that the discriminator network believes the input signal is more likely to be the concatenation of the distorted and restored signals; values ​​closer to 1 indicate that the discriminator network believes the input signal is more likely to be the concatenation of the distorted and original signals.

[0054] S5. Training deep neural networks: In order to enable the signal generator network to more accurately characterize the mapping relationship between the frequency domain distorted signal and the original signal, the present invention designs an efficient deep neural network training method to implement the training of the signal generator network. Specifically, the present invention adopts a two-stage training strategy that combines "generative adversarial training" and "direct supervision training". The first stage performs generative adversarial training based on the Generative Adversarial Networks (GAN) framework, and the second stage uses a direct supervision training method to further optimize network performance. The following describes the two training methods in detail, first introducing the direct supervision training method, and then introducing the generative adversarial training method based on GAN.

[0055] ① Directly supervised training

[0056] Assume that the frequency domain distorted signal is x and the true original signal is y, and denote the signal generator network as G. The training goal at this stage is to use the frequency domain distorted signal as the input of the signal generator network and the true original signal as the label, and optimize the parameters of the signal generator network through supervised learning. This paper adopts the Huber loss function as the loss function of the signal generator network, and its mathematical expression is as follows:

[0057]

[0058] in is the expectation operation, is the hyperparameter threshold. The Huber loss function combines the advantages of mean square error (MSE) and mean absolute error (MAE). When the prediction deviation is less than or equal to the threshold When the prediction error is greater than the threshold, the MSE loss function is used to ensure the gradient is smooth; When , MAE loss function is used to enhance the robustness to noise or outliers. In addition, the loss function is used to set the threshold , which can effectively constrain the impact of errors on the training process, thereby improving the stability of the model. It should be noted that the threshold It can be flexibly adjusted according to specific task requirements to adapt to different data distributions and training objectives.

[0059] ②Generative Adversarial Training Based on GAN

[0060] During the GAN-based generative adversarial training phase, adversarial training is used to make the samples generated by the signal generator network more closely resemble the real data distribution. This phase involves adversarial training of two networks: the signal generator G (gen) and the discriminator D (disc). The goal of the signal generator network is to generate a restored signal that is difficult for the discriminator network to distinguish and as close to the original signal as possible. The goal of the discriminator network is to accurately determine whether the input sample comes from real data or the generator network. Through alternating training, the two networks are continuously optimized, ultimately reaching a Nash equilibrium, where the discriminator cannot distinguish between generated samples and real samples. This is equivalent to the signal generator network being able to generate samples that conform to the real data distribution.

[0061] In this invention, the loss function of the signal generator consists of two parts. The first part is the adversarial (Adv) loss, which aims to make the discriminator identify the generated restored signal samples as the true original signal samples, that is,

[0062]

[0063] The second part is the Huber loss defined above, which ensures that the generated restored signal samples are as close as possible to the true original signal samples.

[0064] In order to balance the impact of Adv loss and Huber loss on the final loss of the signal generator network, a balance coefficient is set for each of the two losses. and , the two are weighted summed to form the final loss function of the signal generator network, that is,

[0065]

[0066] The loss function of the discriminator network also consists of two parts. The first part is the loss of the discriminant generated sample (Disc_gen), which aims to enable the discriminator network to correctly identify the restored signal samples generated by the signal generator and judge them as false, that is,

[0067]

[0068] The second part discriminates the loss of real samples (Disc_real), which aims to enable the discriminator network to correctly identify the real original signal and discriminate it as true, that is,

[0069]

[0070] In order to balance the impact of Disc_gen loss and Disc_real loss on the final loss of the discriminator network, a balance coefficient is set for each of the two losses. and , the two are weighted summed to form the final loss function of the discriminator network, that is,

[0071]

[0072] During training, the parameters of the signal generator network G and the discriminator network D are updated alternately. Each round of training begins by fixing the parameters of the signal generator network and updating those of the discriminator network. Subsequently, the parameters of the discriminator network are fixed and updated. This alternating training approach gradually improves the performance of both the signal generator and the discriminator, ultimately achieving a close match between the generated samples and the real-world data distribution.

[0073] In summary, by using a two-stage training strategy of "generative adversarial training" and "directly supervised training," the signal generator network is able to continuously optimize its restoration performance, minimizing the Huber loss between the generated restored signal and the true original signal. This training method not only improves the robustness of the network but also ensures that the generated restored signal is highly consistent with the true original signal, providing reliable technical support for the present invention's anti-interception transmission recovery of frequency-domain distorted signals.

[0074] S6. Perform intelligent anti-interception transmission: In the specific implementation process, the transmitter first generates a random binary sequence and processes it through the simulated communication system constructed in S2. The receiver performs time-domain phase compensation and frequency-domain notching on the received signal based on the pre-shared frequency-domain notch position information and time-domain phase rotation information, obtaining a frequency-domain distorted signal. This distorted signal is then input into the signal generator network trained in S5, which generates a recovered signal close to the original signal. Finally, the recovered signal is demodulated using QPSK to obtain a binary sequence that is highly consistent with the transmitter. In contrast, the third end, unable to obtain frequency-domain notch position information and time-domain phase rotation information, struggles to accurately perform time-frequency domain deinterference, resulting in a significant difference between its recovered binary sequence and the originally transmitted binary sequence. It can be seen that the receiver can obtain a relatively accurate transmitted sequence, while the third end cannot. This asymmetric performance advantage ensures the security of the communication process, thereby enabling the present invention to achieve the purpose of anti-interception transmission.

[0075] The beneficial effect of the present invention is that, through a pre-shared scrambling knowledge base and scrambling pattern diagram, the present invention designs a joint time-frequency domain signal processing scheme: at the transmitting end, the original signal is sequentially subjected to frequency domain notching, frequency domain scrambling, and time domain phase rotation processing to construct a transmission signal with dual perturbations in the time and frequency domains; at the receiving end, the legitimate user sequentially performs time domain phase compensation and frequency domain notching processing based on shared prior information, and utilizes a pre-trained fully convolutional autoencoder network to achieve intelligent reconstruction of the signal. This method ensures that the legitimate receiver can accurately recover the original signal by building an asymmetric performance advantage between the legitimate receiver and the third party, while the third party cannot effectively parse the signal due to the lack of key perturbation parameters and reconstruction algorithms. This significantly improves the system's anti-interception performance while ensuring communication quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 This is a schematic diagram of the structure of the time-frequency domain scrambling module at the transmitting end of the present invention.

[0077] Figure 2 This is a schematic diagram of the structure of the time-frequency domain descrambling module at the receiving end of the present invention.

[0078] Figure 3 This is a schematic diagram of the scrambling knowledge base and scrambling pattern diagram structure of the present invention.

[0079] Figure 4 It is a schematic diagram of the overall framework of the simulation communication system of the present invention.

[0080] Figure 5 Schematic diagram of the signal generator network structure of the present invention.

[0081] Figure 6 Schematic diagram of the discriminator network structure of the present invention.

[0082] Figure 7 This is a schematic diagram of the bit error rate performance of the technology proposed in this invention (only frequency domain scrambling and descrambling is performed).

[0083] Figure 8 This is a schematic diagram of the bit error rate performance of the technology proposed in this invention (joint time-frequency domain scrambling and descrambling). DETAILED DESCRIPTION

[0084] The practicality of the present invention is illustrated below with simulation examples and accompanying figures. First, system simulation parameters and deep neural network parameters are provided, followed by simulation results to evaluate the performance of the algorithm proposed in this invention. Specifically, the scrambling algorithm of this invention is used at the transmitter, the descrambling algorithm of this invention is used by the legitimate receiver at the receiver, and the third party at the third end does not perform any special processing on the received signal. The performance of the algorithm proposed in this invention is evaluated by comparing the bit error rates of the receiver and the third party.

[0085] The present invention considers a QPSK simulation communication system. Specifically, the communication signal in the simulation is a baseband QPSK signal, the number of symbols sent in a data is 1024, and the symbol rate is bps, the upsampling factor is 2, and the shaping filter uses the raised cosine function, where the roll-off coefficient , sampling frequency The signal bandwidth is 100 MHz. Furthermore, the frequency domain noise is generated as band-limited noise, with 8 phase rotation steps. The signal-to-noise ratio (Eb / N0) is considered to be 0 dB, 5 dB, 10 dB, 15 dB, and 20 dB. The frequency domain notch ratio is considered to be 0%, 10%, 20%, and 30%, with 1 notch segment and a notch center frequency of 50 MHz.

[0086] The design and parameter settings of the deep neural network are as follows. The signal generator network consists of four convolutional layers and four deconvolutional layers. Each layer has a convolution kernel size of (3×1) and a stride of (2×1), and each layer is followed by a Prelude-Reduced Unit (PReLU) activation function. The first convolutional layer has 1 input channel and 4 output channels; the second convolutional layer has 4 input channels and 16 output channels; the third convolutional layer has 16 input channels and 64 output channels; the fourth convolutional layer has 64 input channels and 256 output channels; the first deconvolutional layer has 512 input channels and 64 output channels; the second deconvolutional layer has 128 input channels and 16 output channels; the third deconvolutional layer has 32 input channels and 4 output channels; and the fourth deconvolutional layer has 8 input channels and 1 output channel.

[0087] The discriminator network consists of three convolutional layers, a flattening layer, and a fully connected layer. Each convolutional layer has a kernel size of (3×1) and a stride of (2×1), and each layer is followed by a LeakyReLU activation function. The first convolutional layer has 1 input channel and 4 output channels; the second convolutional layer has 4 input channels and 16 output channels; and the third convolutional layer has 16 input channels and 1 output channel. Following the convolutional layers is a flattening layer, which flattens the feature vectors obtained by the convolutional layers into one-dimensional vectors. Following the flattening layer is a fully connected layer with 128 input neurons and 1 output neuron, using the Sigmoid activation function.

[0088] In the first stage of training based on the “Generative Adversarial Training” method, the loss function of the signal generator network is , that is, the balance coefficient , where the threshold of Huber loss is set to ; The loss function of the discriminator network is , that is, the balance coefficient The training adopts an alternating iterative strategy, and each training round performs a parameter update of the discriminator network and a parameter update of the signal generator network in turn, for a total of 300 training rounds. The learning rate of the signal generator network is It follows the exponential decay rule, that is, , where q represents the current training round, is the initial learning rate of the signal generator network 0.02, is the final learning rate of the signal generator network 0.0002, The learning rate of the signal generator network is exponentially decayed at a rate of 0.023. The learning rate of the discriminator network is fixed at 0.0002. In addition, the optimizer used for both the signal generator network and the discriminator network is the Adam optimizer.

[0089] During the second stage of training based on the “directly supervised training” method, the loss of the signal generator network is , where the threshold of Huber loss is set to For each training round, the parameters of the signal generator network are updated once, and a total of 500 training rounds are performed. The learning rate of the signal generator network is fixed at 0.0002, and the optimizer used is the Adam optimizer.

[0090] For each fixed signal-to-noise ratio and frequency domain notch ratio, during the training phase, the present invention collected 10,000 data points for offline training. During the testing phase, the present invention evaluated the simulation performance of 5,000 data points.

[0091] Figure 7 The bit error rate performance of the legitimate receiver and the third party in the technology proposed in the present invention (only frequency domain scrambling and descrambling) is demonstrated. Through observation, it can be found that with the increase of Eb / N0, the bit error rates of the receiver and the third party both show a downward trend, but the downward trend of the receiver's bit error rate is more significant, and is always lower than the bit error rate of the third party. This shows that under different Eb / N0, the receiver always maintains a bit error rate performance advantage over the third party, and as Eb / N0 increases, the bit error rate performance advantage becomes more obvious. In addition, with the increase of the frequency domain notch ratio, the bit error rates of the receiver and the third party both show an upward trend. When the frequency domain notch ratio is 0%, the bit error rates of the receiver and the third party are relatively close, and when the frequency domain notch ratio is 10%, 20% and 30%, the bit error rate of the receiver is reduced by 0 to 4 orders of magnitude compared with the bit error rate of the third party. This shows that when the frequency domain notch ratio is 10%, 20% and 30%, the receiver has a greater bit error rate performance advantage than the third party. In summary, Figure 7 It shows that the technology proposed in this invention (only frequency domain scrambling and descrambling) has a strong anti-interception transmission capability in the Eb / N0 range of 0~20dB and the frequency domain notch ratio of 0%~30%.

[0092] Figure 8 The bit error rate performance of the legitimate receiver and the third party in the technology proposed in this invention (joint scrambling and descrambling in the time-frequency domain) is demonstrated. Through observation, it can be found that in the Eb / N0 range of 0~20dB and the frequency domain notch ratio of 0%~30%, the bit error rate of the third party is always maintained at around 0.5, which is equivalent to the level of random guessing; while the bit error rate of the receiver always maintains a significant performance advantage, and its bit error rate can be reduced by 0~5 orders of magnitude compared with the bit error rate of the third party. In particular, as Eb / N0 increases, the bit error rate advantage of the receiver becomes more obvious; at the same time, as the frequency domain notch ratio decreases, the bit error rate advantage of the receiver becomes more obvious. In summary, Figure 8 It shows that the technology proposed in this invention (joint scrambling and descrambling in the time-frequency domain) has a strong anti-interception transmission capability in the Eb / N0 range of 0~20dB and the frequency domain notch ratio of 0%~30%.

[0093] In addition, compared Figure 7 and Figure 8 It can be found that for the technology proposed in the present invention, when "only performing frequency domain scrambling and descrambling" is changed to "performing joint time-frequency domain scrambling and descrambling", that is, introducing time domain phase rotation and time domain phase compensation processing, the bit error rate of the receiver remains basically unchanged, while the bit error rate of the third party increases sharply. This shows that performing time domain scrambling, that is, adding phase rotation, can further enhance the anti-interception capability against third parties while ensuring the performance of the receiver.

Claims

1. An intelligent anti-interception transmission method based on a fully convolutional autoencoder network for a wireless communication system, wherein the system comprises a sender, a legitimate receiver, and a third party, and is characterized in that: The following steps are involved: S1. Establish a scrambling knowledge base and a scrambling pattern diagram, wherein the key parameter library of the scrambling knowledge base includes a frequency domain notch parameter library, a frequency domain noise generation library, and a time domain phase rotation library. These three parameter libraries together constitute N different scrambling parameter combinations, each of which corresponds to a frequency domain notch parameter, a frequency domain noise generation method, and a time domain phase rotation radian configuration; the scrambling pattern diagram serves as a dynamic scheduling scheme for the scrambling knowledge base, which specifies the scrambling combination sequence number used under each communication sequence; S2. Build a simulated communication system with anti-interception capabilities. The simulated communication system consists of three components: a transmitter, a channel, and a receiver. The transmitter and receiver have the scrambling knowledge base and scrambling pattern diagram established in S1. The working method of each component is as follows: Transmitter: First, a random binary sequence is generated and modulated to obtain the original signal. Then, the original signal is subjected to joint time-frequency domain scrambling, including frequency domain notching, frequency domain scrambling, and time domain phase rotation processing, to obtain a time-frequency domain scrambled signal. Finally, the time-frequency domain scrambled signal is upsampled and subjected to shaping filtering to obtain the transmitted signal. Channel: The transmitted signal is transmitted through a Gaussian channel; At the receiving end, the received signal is first processed through matched filtering and downsampling to obtain the received signal. Then, based on the shared time-varying disturbance position change law and the time-varying phase change law, the received signal is sequentially subjected to time-domain phase compensation and frequency-domain notching processing to obtain a frequency-domain distorted signal. Frequency-domain compensation is then implemented through a fully convolutional autoencoder, and the frequency-domain distorted signal is mapped to a recovered signal. Finally, the recovered signal is demodulated to obtain the received binary sequence. S3. Constructing a dataset for intelligent anti-interception transmission technology: Run a simulated communication system and synchronously record the original signal from the transmitter and the frequency-domain distorted signal from the receiver after time-domain phase compensation and frequency-domain notching. Use the frequency-domain distorted signal as input data and the original signal as a label to construct a training dataset. S4. Build a deep neural network: Use a fully convolutional autoencoder network as the signal generator network, and design a discriminator network to assist in the training of the signal generator network; S5. Training deep neural networks: A two-stage training strategy combining generative adversarial training and direct supervised training is adopted. The first stage is based on the generative adversarial network framework for generative adversarial training, and the second stage uses direct supervised training methods to further optimize network performance. S6. Perform intelligent anti-interception transmission: In the specific implementation process, the transmitter first generates a random binary sequence and processes it through the communication simulation system constructed in S2. The receiver performs time domain phase compensation and frequency domain notch processing on the received signal in sequence based on the pre-shared frequency domain notch position information and time domain phase rotation information to obtain a frequency domain distorted signal. Subsequently, the distorted signal is input into the trained signal generator network in S5, which generates a restored signal close to the original signal. Finally, the restored signal is demodulated to obtain a binary sequence that is highly consistent with the transmitter.

2. The intelligent anti-interception transmission method based on a fully convolutional autoencoder network according to claim 1 is characterized in that: In S2, the transmitting end performs null-stuck on the frequency domain signal of the original signal in a preset frequency band to obtain a frequency domain distorted signal; then, an interference signal is injected into the null-stuck frequency band to obtain a frequency domain scrambled signal; the preset frequency band is a continuous frequency band or a combination of multiple discrete sub-frequency bands; the intensity of the interference signal is the same as that of the original signal.

3. The intelligent anti-interception transmission method based on a fully convolutional autoencoder network according to claim 1 is characterized in that: In S2, the transmitter performs time-domain phase rotation processing on the frequency-domain scrambled signal, wherein the rotation radian is determined according to the time-varying phase variation law, to obtain a time-frequency domain scrambled signal.

4. The intelligent anti-interception transmission method based on a fully convolutional autoencoder network according to claim 3 is characterized in that: There are 8 gears for setting the phase rotation radian. The phase rotation gears of 1024 symbols in a data are the same, and the phase rotation gears between different data are independent.

5. The intelligent anti-interception transmission method based on a fully convolutional autoencoder network according to claim 1 is characterized in that: In S3, both the frequency domain distorted signal and the original signal are in complex form. In the actual processing process, their real and imaginary parts need to be separated and processed independently to ensure that the input data and corresponding labels in the dataset are in real form.

6. The intelligent anti-interception transmission method based on a fully convolutional autoencoder network according to claim 1 is characterized in that: In S4, the signal generator network has a symmetrical autoencoder structure. The encoder and decoder consist of 4 convolutional layers and 4 deconvolutional layers, respectively, and each layer is followed by a PReLU activation function for nonlinear processing. The output of the last layer of the encoder is concatenated with a random noise as the input of the first deconvolutional layer of the decoder. At the same time, a skip connection mechanism is introduced in each subsequent deconvolutional layer of the decoder, that is, the input of each subsequent deconvolutional layer is the output of the previous layer concatenated with the output of the symmetric convolutional layer of that layer.

7. The intelligent anti-interception transmission method based on a fully convolutional autoencoder network according to claim 6, characterized in that: In S4, the discriminator network takes the concatenation of the distorted signal and the original signal or the concatenation of the distorted signal and the restored signal as input. After the input signal passes through three convolutional layers to extract features, the three-dimensional features are converted into one-dimensional vectors through the flattening layer. Finally, the fully connected layer with Sigmoid as the activation function outputs the discrimination result; the output value of the discriminator network is between 0 and 1. The closer the value is to 0, the more likely the discriminator network believes that the input signal is to be the concatenation of the distorted signal and the restored signal; the closer the value is to 1, the more likely the discriminator network believes that the input signal is to be the concatenation of the distorted signal and the original signal.

8. The intelligent anti-interception transmission method based on a fully convolutional autoencoder network according to claim 1 is characterized in that: In S5, the first stage uses generative adversarial training, where the signal generator network and the discriminator network update parameters alternately; the second stage uses direct supervision training; Loss function used by the signal generator network for: ， ， ， in, and is the balance coefficient of the signal generator network, is the adversarial loss function, is the Huber loss function, is the expectation operation, is the discriminator network, x is the frequency domain distorted signal, y is the true original signal, is the signal generator network, is the hyperparameter threshold, and the Huber loss function indicates that when the prediction deviation is less than or equal to the threshold When the prediction error is greater than the threshold The mean absolute error is used; Loss function used by the discriminator network for: ， ， ， in, and is the balance coefficient of the discriminator network, Generate sample loss function for discrimination, is the loss function for discriminating true samples; The direct supervised training process adopts the Huber loss function as the loss function of the signal generator network.

9. The intelligent anti-interception transmission method based on a fully convolutional autoencoder network according to claim 8, characterized in that: During the first stage of generative adversarial training, the parameters of the signal generator network G and the discriminator network D are updated alternately. This means that in each round of training, the parameters of the signal generator network are first fixed and the parameters of the discriminator network are updated; then the parameters of the discriminator network are fixed and the parameters of the signal generator network are updated.

10. The intelligent anti-interception transmission method based on a fully convolutional autoencoder network according to claim 1, characterized in that: In S6, the trained signal generation network is trained based on the pre-shared frequency domain notch positions. In practice, it is necessary to establish a signal generator network model library and use different signal generator network models according to different spectrum notch positions.

Images