An online adversarial learning test system combining spectrum and intelligence

Abstract

The present invention discloses an online adversarial learning and testing system combining spectrum and intelligence, including a hardware system and a software system. The hardware system includes: a spectrum transceiver module receives and transmits spectrum signals and transmits them to a software radio module, the software radio module performs preliminary processing on the spectrum signals and transmits them to an intelligent computing module, and the intelligent computing module completes the intelligent computing task of the spectrum signals. The software system includes: a system parameter configuration module completes parameter configuration, a spectrum processing module processes the received signals, and the processed data is input into an intelligent training module through a data management module for online training. After the training is completed, the model is subjected to an online intelligent evaluation test through an intelligent reasoning module. During the training and testing process, adversarial samples are introduced to enhance or verify the robustness of the model. The present invention can realize intelligent online training, testing, and adversarial processes based on real spectrum data, and improve the authenticity and timeliness of subsequent technology development.

Description

Technical Field

[0001] The present invention relates to the field of spectrum confrontation and artificial intelligence, and specifically proposes a spectrum intelligent online confrontation learning system. Background Art

[0002] With the research and development of wireless spectrum in applications such as 5G and 6G, spectrum application has received more and more attention. The application efficiency in the process of spectrum use, especially the frequency use strategy under interference confrontation conditions, is one of the key directions of future technology development. At the same time, deep learning artificial intelligence has made great technological breakthroughs, with the characteristics of efficient processing of massive information and rapid processing of complex problems, especially in the fields of computer vision and natural language processing, and has achieved results that surpass humans in terms of accuracy of processing results, processing speed and update frequency. Therefore, improving the efficiency of wireless spectrum use based on the characteristics of deep learning artificial intelligence technology is a promising technical direction.

[0003] However, while deep learning algorithms achieve high accuracy, they can be targeted by attacks. There are a large number of attack algorithms, such as the gradient-based PGD attack series, the optimization-based CW attack series, and the decision boundary-based deepfool series. By adding very small disturbances, the model can get wrong recognition results. Spectrum-based applications face strong security threats. In order to ensure the effective transmission of our spectrum information, we must consider the security and reliability of deep learning systems when facing attacks. Traditionally, many security-related contents have been considered in the spectrum field, but these mechanisms cannot be used to ensure the security of deep learning models. At present, the robustness of the model is mainly improved through adversarial training and data enhancement.

[0004] Currently, in the field of wireless spectrum, software systems represented by GNU Radio can well implement various spectrum process simulations at the hardware or software level. At the same time, in the field of deep learning, software systems represented by Pytorch can quickly and conveniently implement the training, testing, and tuning of various neural network models. However, a complete system that takes into account both spectrum and intelligence is a gap in current technical research, which has greatly restricted the development of spectrum technology combined with intelligence. Summary of the invention

[0005] The technical problem to be solved by the present invention is: the present invention discloses an online adversarial learning and testing system combining spectrum and intelligence, which can realize intelligent online training, testing and adversarial processes based on real spectrum data, without the need to build a large-scale data set in advance, thereby improving the authenticity and timeliness of subsequent technology development.

[0006] The technical solution adopted by the present invention is: an online adversarial learning and testing system combining spectrum and intelligence, including a hardware system and a software system, which can realize online adversarial learning and testing of intelligent algorithms based on real spectrum signals.

[0007] The hardware system includes a spectrum transceiver module, a software radio module and an intelligent computing module. The spectrum transceiver module uses a reconfigurable antenna to transmit the received spectrum signal to the software radio module, which performs preliminary processing on it and transmits it to the intelligent computing module, which completes the intelligent computing task of the spectrum signal.

[0008] The spectrum transceiver module uses a reconfigurable antenna as an antenna that can transmit and receive spectrum signals. It is required to be able to receive parameters such as spectrum and power and be controlled by the software system. The software radio module can use USRP devices, which are mainly responsible for the carrier filtering, modulation and demodulation of the antenna receiving data and other frequency processing processes. The intelligent processing module can be a high-performance computer, including a high-performance CPU and GPU, which can perform artificial intelligence algorithm calculations based on neural networks.

[0009] The software system is developed based on GNU Radio and Pytorch, including system parameter configuration module, spectrum processing module, data management module, adversarial sample generation module, intelligent training module, and intelligent reasoning module. The system parameter configuration module, data management module, adversarial sample generation module, intelligent training module, and intelligent reasoning module run on the intelligent computing module.

[0010] The software system working steps are as follows:

[0011] The parameter configuration module is used to configure the parameters required by the system.

[0012] The spectrum processing module completes the transmission, reception and processing of spectrum signals, mainly through the open source software GNU Radio.

[0013] The data management module processes the spectrum data from the spectrum processing module and manages the signals in batches.

[0014] The adversarial sample generation module mainly generates adversarial samples to counter the intelligent model, or introduces them into training to enhance the robustness of the intelligent model.

[0015] The intelligent training module sends the training set data input by the data management module and the adversarial sample generation module to the neural network of the intelligent training module in batches for training, and iterates until a converged model is obtained.

[0016] The intelligent reasoning module evaluates and tests the model performance under non-interference adversarial and interference adversarial conditions.

[0017] The system parameter configuration module mainly includes hardware parameter configuration and algorithm parameter configuration. The hardware parameter configuration includes the setting of spectrum sampling rate, center frequency, data type transmitted by hardware, channel gain, bandwidth, number of channels, etc. The algorithm configuration part configures various parameters of spectrum processing module, data management module, adversarial sample generation module, intelligent training module, and intelligent reasoning module, mainly including the configuration of parameters such as neural network structure, learning rate, number of iterations, adversarial perturbation size, model file save path, etc.

[0018] The spectrum processing module is responsible for applying parameter configuration to the hardware system and controlling the hardware system to complete the entire spectrum transmission and reception process. This includes antenna reception, signal detection, carrier frequency calibration, symbol phase synchronization, modulation and demodulation, encoding and decoding, etc. In order to ensure that the spectrum transmission and reception data is easy to confirm, a synchronization header can be added to the front end of the transmitted spectrum signal so that it can be aligned through the synchronization header correlator after reception.

[0019] The data management module takes into account the mismatch between the data sending and receiving rate of the hardware system and the processing rate of the software system. Usually, the data processing speed is far behind the data generation speed. A cache area is set up to solve the processing speed mismatch problem. The spectrum data is stored in the cache area in real time. The cache data follows the first-in-first-out principle. Each time, a batch of data sets containing all types are taken out from the cache area for data enhancement. The enhanced data is divided into three parts: training set, test set, and validation set according to the set ratio.

[0020] The adversarial sample generation module generates adversarial samples based on the adversarial attack algorithm of the neural network model, and sends the adversarial samples to the data management module and adds them to the training set, test set, and validation set for interference confrontation of subsequent intelligent models or enhancing model performance. If the original data and the generated adversarial samples are sent to the neural network model for training during the training process, a robust model can be obtained. Adversarial samples can be generated by any adversarial algorithm, including but not limited to white-box attacks, black-box attacks, gradient-based attacks, optimization-based attacks, decision-surface-based attacks, etc.

[0021] The intelligent training module calculates the model gradient information corresponding to the training set according to the training set divided by the data management module, and updates the neural network parameters to improve its accuracy on the training set. The cost function used in the training is the L2 regularized cross entropy function. The training algorithm uses the stochastic gradient descent (SGD) optimization algorithm to train the neural network weight parameters online. The recognition accuracy and loss function of the training set and the validation set are used to determine whether the network meets the end condition. When the accuracy of the training set is increasing and the accuracy of the validation set is decreasing, the training is stopped to obtain a trained network model.

[0022] The intelligent reasoning module is tested without attack, and the test set data obtained by the data management module is sent to the model trained by the intelligent training module in batches to obtain the model judgment result and statistically calculate the accuracy. In the attack test, the adversarial samples obtained by the adversarial sample generation module are sent to the trained model to obtain the judgment result, statistically calculate the accuracy; and calculate the bit error rate of the signal under no interference and interference.

[0023] The advantages of the present invention compared with the prior art are:

[0024] (1) The present invention can perform online training and testing on real spectrum signals, and can realize real-time spectrum confrontation without building a data set in advance, thus ensuring the timeliness of the entire process; the transmission and reception of real spectrum signals can be realized through spectrum transceiver devices such as USRP and other software radio devices and reconfigurable antennas. At the same time, the received and transmitted spectrum signals are transmitted in real channels and devices throughout the process, and all kinds of noise interference are real, so the test results are highly credible.

[0025] (2) The system of the present invention uses an intelligent model to perform intelligent computational processing of spectrum data, and can flexibly adjust the neural network architecture according to application needs. The neural network can be used to replace any link in the traditional spectrum process, so the spectrum process is definable and has good generalization ability.

[0026] (3) The spectrum intelligent online adversarial learning system proposed in the present invention can be applied to various wireless communication application scenarios involving interference or adversarial conditions, such as 6G; it can be used to test the frequency performance of spectrum algorithms under various channel conditions, test and then improve the anti-noise and anti-interference capabilities of spectrum algorithms, find the best strategy and improve training efficiency; it can be used for algorithm testing and development such as spectrum signal detection, and detect algorithms through online testing and training of real signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a system composition diagram of the present invention;

[0028] Figure 2 It is a flowchart of intelligent training;

[0029] Figure 3 This is a flowchart of intelligent reasoning. DETAILED DESCRIPTION

[0030] The present invention proposes a spectrum intelligent online adversarial learning system, including a hardware system and a software system. The hardware system includes a spectrum transceiver module, a software radio module and an intelligent computing module. The spectrum transceiver module uses a reconfigurable antenna to transmit the received spectrum signal to the software radio module, the software radio module preliminarily processes it and transmits it to the intelligent computing module, and the intelligent computing module completes the intelligent computing task of the spectrum signal. The software system is developed based on GNU Radio and Pytorch, including a system parameter configuration module, a spectrum processing module, a data management module, an adversarial sample generation module, an intelligent training module, and an intelligent reasoning module. The system parameter configuration module, the data management module, the adversarial sample generation module, the intelligent training module, and the intelligent reasoning module run on the intelligent computing module. The system parameter configuration module completes the parameter configuration, the spectrum processing module processes the received signal based on GNU Radio, the processed data is input into the intelligent training module through the data management module for online training, and after the training is completed, the model is subjected to online intelligent evaluation test through the intelligent reasoning module. During the training and testing process, adversarial samples are introduced to enhance or verify the robustness of the model. The present invention can realize intelligent online training, testing, and confrontation processes based on real spectrum data without the need to build a large-scale data set in advance, thereby improving the authenticity and timeliness of subsequent technology development.

[0031] The spectrum transceiver module uses a reconfigurable antenna as an antenna that can transmit and receive spectrum signals, and requires that its parameters such as the receivable spectrum and power can be controlled by the software system. The software radio module can use USRP devices, which are mainly responsible for the carrier filtering, modulation and demodulation of the antenna receiving data and other frequency processing processes. The intelligent processing module can be a high-performance computer, including a high-performance CPU and GPU, which can perform artificial intelligence algorithm calculations based on neural networks.

[0032] The parameter configuration module mainly includes hardware parameter configuration and algorithm parameter configuration. The hardware parameter configuration includes the settings of spectrum sampling rate, center frequency, data type transmitted by hardware, channel gain, bandwidth, number of channels, etc. The algorithm parameter configuration configures various parameters of the intelligent processing part of the system, mainly including the configuration of parameters such as neural network structure, learning rate, number of iterations, size of adversarial disturbance, and model file save path.

[0033] The spectrum processing module is responsible for applying parameter configuration to the hardware system and controlling the hardware system to complete the entire spectrum transmission and reception process. It includes antenna reception, signal detection, carrier frequency calibration, symbol phase synchronization, modulation and demodulation, encoding and decoding, etc. In order to ensure that the spectrum transmission and reception data is easy to confirm, a synchronization header can be added to the front end of the transmitted spectrum signal so that it can be aligned through the synchronization header correlator after reception. The specific work includes transmission control and reception control. For reception control, the spectrum processing module is based on GNU Radio, controls the spectrum transceiver module to receive the spectrum signal, and passes it to the software radio module for signal processing, such as gain, demodulation, synchronization, etc., and then passes the acquired processing data to the intelligent computing module to write into the data cache pool and connect the data management module. For transmission control, based on the data management module's requirements for the number and diversity of data sets, different spectrum signals are selected and corresponding synchronization headers are added, and then the spectrum data is transmitted through the software radio module and the spectrum transceiver module in turn.

[0034] The data management module takes into account the mismatch between the data transmission and reception rate of the hardware system and the processing rate of the software system. Usually, the data processing speed is far behind the data generation speed. A buffer area is set to solve the processing speed mismatch problem, manage the data transmission interval, and store the spectrum data in the buffer area in real time. The buffer area data follows the first-in-first-out principle. Each time, a batch of data sets containing all types are taken out from the buffer area for data enhancement. The enhanced data is divided into three parts: training set, test set, and verification set in a ratio of 7:2:1. The data management module includes three parts: data packaging, data enhancement, and data control, which jointly ensure data utilization and prevent data explosion in the buffer area. For data packaging, fixed batch packaging is performed according to the requirements of intelligent training or inference. For data enhancement, the data management module performs data enhancement methods such as translation, multi-angle rotation, and noise addition on the receiving end. For data control, it ensures that a batch of all label types required for training is taken from the buffer area each time, and the labels of different data are distinguished by setting different synchronization headers. In order to prevent the memory explosion caused by too much data in the buffer area, a critical value is set. When the data in the buffer area reaches the critical value, data transmission is automatically suspended. When the data volume decreases, data transmission is resumed.

[0035] The adversarial sample generation module generates adversarial samples based on the neural network model using the adversarial attack algorithm, and sends the adversarial samples to the data management module and adds them to the training set, test set, and validation set for interference confrontation of subsequent intelligent models or to enhance model performance. If the original data and the generated adversarial samples are sent to the neural network model for training during the training process, a robust model can be obtained. Adversarial attacks are mainly divided into white-box attacks and black-box attacks, and the attack methods can be divided into targeted attacks and non-targeted attacks. White-box attacks can obtain information such as model structure and gradients. First, the model gradient is obtained, and a suitable perturbation is selected and added to the data. The output adversarial samples make the model recognize incorrectly. Black-box attacks do not know the information of the model and need to obtain model information by continuously interacting with the attacked model and querying before attacking. Targeted attacks are to make the model recognize the specified category, while non-targeted attacks only require the model to recognize incorrectly. On the one hand, adversarial samples can enhance the enemy channel interference and cause the other party's intelligent model to recognize incorrectly. On the other hand, the model obtained after adding adversarial samples to intelligent training can reduce the probability of our signal being intercepted and recognized by the enemy and enhance the channel's anti-interference ability. Adversarial samples can be generated by any adversarial algorithm, including but not limited to white-box attack, black-box attack, gradient-based attack, optimization-based attack, decision-surface-based attack, etc. The present invention supports a variety of adversarial samples, including gradient-based attacks such as FGSM, PGD, MI-FGSM, etc., optimization-based CW, etc., and decision-surface-based deepfool, etc.

[0036] The intelligent training module calculates the model gradient information corresponding to the training set according to the training set divided by the data management module, and updates the neural network parameters to improve its accuracy on the training set. The cost function used in the training is the L2 regularized cross entropy function. The training algorithm uses the stochastic gradient descent (SGD) optimization algorithm to train the neural network weight parameters online. The recognition accuracy and loss function of the training set and the validation set are used to determine whether the network meets the end condition. When the accuracy of the training set is increasing and the accuracy of the validation set is decreasing, the training is stopped to obtain a trained network model. During the training process, the adversarial algorithm is applied to the current batch of data to generate adversarial samples. The adversarial algorithm can be expressed as first obtaining the gradient information of the model, and according to the set attack iteration steps and disturbance size, the adversarial algorithm is used to iteratively attack the model until the output adversarial sample causes the model to recognize errors, and then the adversarial sample at this time is returned and the adversarial sample and the original data are sent to the neural network model for training to obtain a robust model.

[0037] The intelligent reasoning module is tested without attack, and the test set data obtained by the data management module is sent to the model trained by the intelligent training module in batches to obtain the model judgment result and statistically calculate the accuracy. In the attack test, the adversarial samples obtained by the adversarial sample generation module are sent to the trained model to obtain the judgment result, statistically calculate the accuracy; and calculate the bit error rate of the signal under no interference and interference.

[0038] Example:

[0039] The present invention is a spectrum intelligent online confrontation learning system, which is composed of Figure 1 As shown, the system includes a hardware system and a software system, which can realize online adversarial learning and testing of intelligent algorithms based on real spectrum signals.

[0040] The hardware system includes a spectrum transceiver module, a software radio module and an intelligent computing module. The spectrum transceiver module uses a reconfigurable antenna to transmit the received spectrum signal to the software radio module, which performs preliminary processing on it and transmits it to the intelligent computing module, which completes the intelligent computing task of the spectrum signal.

[0041] The spectrum transceiver module uses a reconfigurable antenna, which is specifically a 1.4G omnidirectional antenna communication 1410-1470MHZ, the software radio module is specifically the ETTUS software radio USRP B210-Kit, and the intelligent computing module is specifically a high-performance computer, the processor chip is Intel(R) Core(TM) i7-9700K CPU@3.60GHz, and the graphics card is Geforce RTX 2070.

[0042] The software system is developed based on GNU Radio and Pytorch, including system parameter configuration module, spectrum processing module, data management module, adversarial sample generation module, intelligent training module, and intelligent reasoning module, and mainly runs on the intelligent processing module. The software system mainly runs on the intelligent computing module, that is, a high-performance computer. The computer operating system is Ubuntu 20.04LTS 64-bit, and the software used is Python version 3.6, pytorch version 1.6.0, and GNU Radio version 3.8.

[0043] The software system working steps are as follows:

[0044] Step 1: The parameter configuration module configures the parameters required by the system.

[0045] Step 2: The spectrum processing module completes the transmission, reception and processing of spectrum signals, mainly through the open source software GNURadio.

[0046] Step 3: The data management module processes the spectrum data from the spectrum processing module and manages the signals in batches.

[0047] Step 4: The adversarial sample generation module mainly generates adversarial samples to counter the intelligent model, or introduces them into training to enhance the robustness of the intelligent model.

[0048] Step 5: The intelligent training module sends the training set data obtained in steps 3-4 to the neural network of the intelligent training module in batches for training, and iterates until a converged model is obtained.

[0049] Step 6: The intelligent reasoning module evaluates the model performance under non-interference adversarial and interference adversarial conditions.

[0050] The parameter configuration module mainly includes hardware parameter configuration and algorithm parameter configuration. The hardware parameter configuration is as follows: the default sampling rate is set to 2.88GHz, the center frequency is 1.44GHz, the channel gain of the transmission and reception is 20dB, the bandwidth is 10MHz, and the number of channels is 1. The algorithm parameters are as follows: the initial learning rate of the neural network is 0.001, the batch size is 224, the momentum size is 0.9, the number of training rounds is 100, the learning rate strategy is cosine annealing, the default perturbation size of the adversarial attack is 0.01, and the number of iterations is 50.

[0051] The spectrum processing module is responsible for applying parameter configuration to the hardware system and controlling the hardware system to complete the entire spectrum transmission and reception process. It includes antenna reception, signal detection, carrier frequency calibration, symbol phase synchronization, modulation and demodulation, encoding and decoding, etc. In order to ensure that the spectrum transmission and reception data is easy to confirm, a synchronization header can be added to the front end of the transmitted spectrum signal so that it can be aligned through the synchronization header correlator after reception. The specific work includes transmission control and reception control. For reception control, the spectrum processing module is based on GNU Radio, controls the spectrum transceiver module to receive the spectrum signal, and passes it to the software radio module for signal processing. The gain is adjusted to 20dB, and the synchronization method uses the maximum likelihood estimation method to synchronize the data clock. The demodulation uses the modulation method of the transmission part for demodulation, and then passes the acquired processed data to the intelligent computing module to write it into the data cache pool, connecting the data management module. For the transmission control, based on the data management module's requirements for the number and diversity of data sets, the software radio platform of the transmission module generates 16 modulation signals of modulation modes using the open source software radio (GNU Radio) software platform according to the modulation signal principle: AM, DSB, SSB, 2ASK, 4ASK, 8ASK, 2FSK, 4FSK, 8FSK, MSK, 2PSK, 4PSK, 8PSK, 16QAM, 64QAM, 256QAM. The transmission module modulates the signals to the intermediate frequency of 1.44GHz, and then sends 16 modulation signals respectively. Different synchronization headers are set at the beginning of each signal, such as (4, 7, 8, 6, 11, 12, 5). Using the software radio platform and PC of the receiving module, the intermediate frequency is aligned to 1GHz. 100 groups of data of each modulation format are sent in a cycle. Each group of data is 2*1052, including a matrix of 1024-bit signal and 28-bit synchronization header. Then, the spectrum data is transmitted in sequence through the software radio module and the spectrum transceiver module.

[0052] The data management module takes into account the mismatch between the data transmission and reception rate of the hardware system and the processing rate of the software system. Usually, the data processing speed is far behind the data generation speed. A buffer area is set to solve the processing speed mismatch problem, manage the data transmission interval, and put the data in the spectrum processing module into the buffer module. All types of modulation modes are taken at one time. The data size taken out is 1600*2*1052. The data is shifted left and right, the phase is rotated by 30, 50, 70, 130, and various noises are added to obtain an enhanced data set with a size of 12800*2*1052. The enhanced data set is shuffled and divided into training set, validation set and test set according to 7:2:1, and the training set size is 8960*2*1052, the validation set size is 2560*2*1052, and the test set size is 1280*2*1052. The label of the data set is obtained according to the set synchronization header. In order to prevent the memory explosion caused by too much data in the buffer area, a critical value is set. When the data in the buffer area reaches the critical value, data transmission is automatically suspended. When the data volume decreases, data transmission is resumed.

[0053] The adversarial sample generation module generates adversarial samples based on the neural network model using the adversarial attack algorithm. The adversarial samples are sent to the data management module and added to the training set, test set, and validation set for interference confrontation with subsequent intelligent models or to enhance model performance. During the training process, if the original data and the generated adversarial samples are sent to the neural network model for training, a robust model can be obtained. Adversarial samples are generated in the following ways:

[0054] The PGD implementation formula is as follows:

[0055] x t =x0+noise,

[0056]

[0057] x t represents the generated adversarial sample, represents the gradient of the loss function L(θ,x,y) with respect to x, y is the label corresponding to the sample, θ is the model parameter, t=0,1,2,3,4...; x0 represents the initial sample, noise represents a noise with the same shape as the input signal, S represents the perturbation size, α represents the attack step size, and the attack process is represented as the initial sample is the noise plus the input signal. Each step of the adversarial sample is equal to the adversarial sample of the previous step plus the perturbation plus the sign of the adversarial sample gradient multiplied by the attack step size. After a round of calculation is completed, the current adversarial sample is clipped and iterated until the set number of iterations is reached.

[0058] FGSM, its implementation formula is as follows

[0059]

[0060] x' is the generated adversarial sample, ε represents the size of the perturbation, L(θ,x,y) is the loss function, θ is the model parameter, x represents the original sample, y represents the label corresponding to the sample, and sgn represents the sign operation. The initial sample plus the perturbation multiplied by the sign of the derivative of the loss function with respect to x, obtains the attack sample under the FGSM adversarial algorithm.

[0061] Deepfool is implemented by calculating the minimum distance between the original sample and the decision boundary and the optimal perturbation direction, and iteratively pushing the original sample to the decision boundary. Given a classifier, the adversarial sample calculation method is as follows:

[0062]

[0063] f y '=f y (x i )-f y0 (x i ),

[0064] The classifier is represented by f(x)=w T x+b, and its corresponding decision boundary plane Γ={x:w T x+b=0}, where w and b are model weight parameters. The DeepFool attack obtains a minimum perturbation distance f through each round of iteration. l ' and the optimal perturbation direction ω l ', x i+1 It refers to the adversarial sample generated in the i-th iteration, x0 represents the original sample, y0 represents the corresponding label, and y represents the set of attack labels. Indicates that x is under the y label i Find the gradient, f y (x i ) represents the sample x i The predicted value corresponding to the y label index on the classifier, l represents the adversarial sample label, ω y ' indicates the current disturbance direction, f y ' represents the current disturbance distance, where i = 0, 1, 2, 3, 4...

[0065] like Figure 2As shown, the intelligent training module calculates the model gradient information corresponding to the training set according to the training set divided by the data management module, and updates the neural network parameters to improve its accuracy on the training set. Specifically, the network uses resnet, the output layer size of the resnet network is changed to 224*16, the network weight is initialized using Heinitialization, and the cost function used in training is the L2 regularized cross entropy function, in which regularization technology and cross entropy loss function are used to prevent overfitting of training. According to the set parameters, the stochastic gradient descent (SGD) optimization algorithm is used to train the neural network weight parameters online, the initial learning rate is 0.01, and the cosine annealing strategy is used to adjust the learning rate. The total number of training times is 100. In the first round of iteration, take a batch of training data and change the dimension to 8960*1*1052*2, loop 40 times, each time the data size is 224*1*1052*2, send the data to the model for training, after training 40 times, update the gradient information and weight information of the model, send the validation set data to the model for verification, and then perform the next round of iterations. Continue to take the next batch of data from the data cache module to continue training. When the number of training rounds is reached, or when the accuracy of the training set increases and the accuracy of the validation set decreases, stop training and save the model. During the training process, if the adversarial training method is selected, the current batch of data is subjected to adversarial attacks. First, obtain the gradient information of the model, set the number of iterations and the size of the disturbance, use PGD, FGSM, and deepfool to iteratively attack the model until the output adversarial sample causes the model to recognize errors, return the adversarial sample at this time, and send the adversarial sample and the original data to the neural network model for training to obtain a robust model.

[0066] like Figure 3As shown in the figure, the intelligent reasoning module mainly evaluates the model modulation recognition and bit error rate analysis in the absence of interference and interference. The real-time spectrum intelligent online adversarial learning system is constructed by using GNU Radio combined with the deep learning algorithm of the host computer system (in this case, the Linux operating system of the PC). Under interference-free conditions, a batch of data (1600, 2, 1024) is obtained according to the data management module, and the data is expanded to (1, 1600, 2, 1024). The model trained by the intelligent training module is loaded, and the data model is tested to obtain the modulation recognition result of the model for the current signal. The model recognition accuracy is statistically analyzed. The receiving end adjusts the gain, maximum likelihood clock synchronization, and the modulation mode recognized by the model through GNU Radio to demodulate and process the received signal. According to the set synchronization header, the sending end data corresponding to the receiving end data is found, and the bit error rate and symbol error rate of 100,000 signal lengths are calculated. Under interference conditions, the data generated by the data management module generates adversarial samples such as pgd, deepfool, and FGSM under different disturbance sizes, and the same model is loaded for detection to verify the accuracy of the model under clean samples and adversarial samples and the robustness under different disturbance sizes. The interference data is sent to the receiver through GNU Radio to adjust the gain, maximum likelihood clock synchronization, and the modulation method identified by the model for demodulation. If the modulation recognition is wrong, the receiver cannot demodulate the original modulation method, and the bit error rate is extremely high, indicating that the interference is successful. If the modulation recognition is correct, the bit error rate of the signal is also calculated.

[0067] Parts of the present invention that are not described in detail belong to the well-known technology for those skilled in the art.

Claims

1. An online adversarial learning test system combining spectrum and intelligence, characterized in that: Including hardware system and software system; The hardware system includes spectrum transceiver module, software radio module and intelligent computing module; The spectrum transceiver module uses a reconfigurable antenna to receive and transmit spectrum signals and transmit them to the software radio module. The software radio module performs preliminary processing on the spectrum signals and transmits them to the intelligent computing module. The intelligent computing module completes the intelligent computing task of the spectrum signals. The software system includes a system parameter configuration module, a spectrum processing module, a data management module, an adversarial sample generation module, an intelligent training module, and an intelligent reasoning module. The system parameter configuration module, the data management module, the adversarial sample generation module, the intelligent training module, and the intelligent reasoning module run on the intelligent computing module; the system parameter configuration module completes the parameter configuration required for the online adversarial learning test system, and the spectrum processing module receives and transmits the received spectrum signals and processes them; the data management module batches and packages the radio frequency data processed by the spectrum processing module and inputs them into the intelligent training module; the adversarial sample generation module generates adversarial samples; The intelligent training module sends the training set data input by the data management module and the adversarial sample generation module into the neural network in batches for online training, and iterates until a converged model is obtained; the intelligent reasoning module evaluates and tests the model performance in non-interference confrontation and interference confrontation; the spectrum processing module applies the parameter configuration to the hardware system, and controls the hardware system to complete the entire spectrum reception and transmission process, including antenna reception, signal detection, carrier frequency calibration, symbol phase synchronization, modulation and demodulation, encoding and decoding; a synchronization header is added to the front end of the transmitted spectrum signal so that it can be aligned through the synchronization header correlator after reception; the specific work of the spectrum processing module includes transmission control and reception control; for reception control, the spectrum processing module controls the spectrum transceiver module to receive the spectrum signal, pass it to the software radio module for signal processing, pass the obtained processed data to the intelligent computing module to write it into the data cache pool, and connect the data management module; for transmission control, according to the data management module's requirements for the number and diversity of data sets, different spectrum signals are selected and corresponding synchronization headers are added, and spectrum data is transmitted through the software radio module and the spectrum transceiver module in turn.

2. According to claim 1, an online adversarial learning test system combining spectrum and intelligence is characterized in that: The reconfigurable antenna used by the spectrum transceiver module is an antenna that can transmit and receive spectrum signals, and its receiving spectrum and power parameters are required to be controllable by the software system.

3. The online adversarial learning test system combining spectrum and intelligence according to claim 1, characterized in that: The software radio module adopts USRP device to perform carrier filtering, modulation and demodulation processing on the data received by the antenna.

4. The online adversarial learning test system combining spectrum and intelligence according to claim 1, characterized in that: The system parameter configuration module includes hardware parameter configuration and algorithm parameter configuration. The hardware parameter configuration includes the setting of spectrum sampling rate, center frequency, data type of hardware transmission, channel gain of transmission and reception, bandwidth, and number of channels; the algorithm parameter configuration configures various parameters of the spectrum processing module, data management module, adversarial sample generation module, intelligent training module, and intelligent reasoning module, including the configuration of neural network structure, learning rate, number of iterations, adversarial perturbation size, and model file save path.

5. The online adversarial learning test system combining spectrum and intelligence according to claim 1 is characterized in that: The data management module sets a buffer area to manage the data transmission interval. The spectrum data is stored in the buffer area in real time. The data in the buffer area follows the first-in-first-out principle. Each time, a batch of data sets containing all types are taken out from the buffer area for data enhancement. The enhanced data is divided into three parts: training set, test set, and verification set according to the set ratio; The data management module includes data packaging, data enhancement, and data control. For data packaging, fixed batch packaging is performed according to the requirements of intelligent training or inference. For data enhancement, the data management module performs data enhancement on the data at the receiving end. For data control, it ensures that a batch of labels containing all the types of labels required for training is taken from the cache each time. The labels of different data are distinguished by setting different synchronization headers and setting critical values. When the data in the cache reaches the critical value, data sending is automatically suspended. When the data volume drops, data sending is resumed.

6. The online adversarial learning test system combining spectrum and intelligence according to claim 1, characterized in that: The adversarial sample generation module generates adversarial samples using an adversarial attack algorithm according to the neural network model, sends the obtained adversarial samples to the data management module and adds them to the training set, test set, and validation set, or during the training process, sends the original data and the generated adversarial samples together to the neural network model for training; Adversarial samples are generated based on adversarial algorithms, including white-box attacks, black-box attacks, gradient-based attacks, optimization-based attacks, and decision-surface-based attacks.

7. The online adversarial learning test system combining spectrum and intelligence according to claim 1, characterized in that: The intelligent training module calculates the model gradient information corresponding to the training set according to the training set divided by the data management module, and updates the neural network parameters; The cost function used in training is the L2 regularized cross entropy function. The training algorithm uses the stochastic gradient descent optimization algorithm to train the neural network weight parameters online. The recognition accuracy and loss function of the training set and the validation set are used to determine whether the network meets the end conditions. When the accuracy of the training set is increasing and the accuracy of the validation set is decreasing, the training is stopped to obtain a trained network model.

8. The online adversarial learning test system combining spectrum and intelligence according to claim 1, characterized in that: During the training process, the intelligent training module applies an adversarial algorithm to the current batch of data to generate adversarial samples, including: First, the gradient information of the model is obtained. According to the set attack iteration steps and disturbance size, the adversarial algorithm is used to iteratively attack the model until the output adversarial sample causes the model to recognize errors. The adversarial sample at this time is returned and the adversarial sample and the original data are sent to the neural network model for training to obtain a robust model.

9. The online adversarial learning test system combining spectrum and intelligence according to claim 1, characterized in that: When there is no attack test, the intelligent reasoning module sends the test set data obtained by the data management module to the model trained by the intelligent training module in batches to obtain the model judgment result and calculate the accuracy rate; when there is an attack test, the adversarial sample obtained by the adversarial sample generation module is sent to the trained model to obtain the judgment result and calculate the accuracy rate; Calculate the bit error rate of the signal without interference and with interference.

Images