A radio frequency fingerprint recognition method and system for distributed Internet of Things devices
By building a layered distributed network topology and federated learning, the problem of RF fingerprint recognition in distributed scenarios in the Internet of Things is solved, and high accuracy and robust RF fingerprint recognition is achieved. It is suitable for multi-access scenarios and reduces the Privacy risk.
Patent Information
- Application Number
- CN202411080323.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-08
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-08-08
AI Technical Summary
The existing technology is difficult to effectively perform RF fingerprint recognition in distributed scenarios, especially in scenarios where large-scale communication devices are accessed in the Internet of Things, and the existing methods are difficult to adapt in small sample scenarios and difficult to process complex signals. feature.
By building a layered distributed network topology, using RF edge nodes to collect time-domain RF signals and perform data augmentation and preprocessing, converting them into spectrum diagrams, combining federated learning for local and global training, building a RF fingerprint recognition model, and realizing a distributed model Collaborative training to improve recognition accuracy and robustness.
It realizes high-accuracy RF fingerprint recognition in distributed Internet of Things, reduces the risk of privacy leakage, is suitable for multiple access scenarios, and improves system robustness and model generalization capabilities.
Smart Images

Figure CN118606827B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of signal recognition, and in particular to a radio frequency fingerprint recognition method and system for distributed Internet of Things devices. Background Art
[0002] With the rapid development of the Internet of Things, the number of devices connected to a single cellular network has increased dramatically. The traditional key management system (PKC) needs to bear more overhead to store and manage keys, which brings challenges to the fast and lightweight identity recognition needs of the Internet of Things. Therefore, RF fingerprint recognition based on physical hardware damage of the transmitter has become a new paradigm. Compared with traditional methods, RF fingerprint recognition does not require key management and is suitable for large-scale IoT device access scenarios. In addition, the deviations caused by firmware damage are small and not easily tampered with.
[0003] Existing technical solutions related to RFID include:
[0004] For example, the small sample RF fingerprint recognition method based on triplet loss with patent publication number CN116127298A uses an energy detection algorithm to detect and collect burst signals, obtains a signal sample set by preprocessing the signal samples, and trains the embedding network under the constraint of the triplet loss function, which can realize the feature extraction of RF fingerprints. At the same time, the principal component analysis (PCA) dimensionality reduction technology is combined to further reduce the fingerprint dimension to obtain a good RF fingerprint classification capability, and realize RF fingerprint recognition in small sample scenarios. However, the above scheme only targets RF fingerprint recognition in small sample scenarios, which is inconsistent with the scenario of multi-sample big data in the Internet of Things. At the same time, the application scenario of single-point training is not suitable for the multi-access scenario of the development of the Internet of Things.
[0005] For example, the patent publication number CN116522228A is a radio frequency fingerprint recognition method based on feature imitation federated learning. This method aims to solve the privacy and security issues in heterogeneous distributed scenarios of edge data. Based on federated learning, the KL distance between the pseudo-feature prediction values of the local classifier and the federated average global classifier is used to optimize the global classifier to obtain a global classifier to adapt to heterogeneous data distribution scenarios.
[0006] For example, the patent publication number CN114598517A is a method for radio frequency distribution identification of mobile devices based on federated learning. This method uses a federated learning algorithm to achieve distributed radio frequency fingerprint identification, targeting the characteristics of distributed storage of radio frequency signal data and difficulty in centralized training models. At the same time, an additional dynamic sample selection algorithm is added to dynamically allocate tasks according to the data processing capabilities of heterogeneous devices, reducing calculation time.
[0007] The problem with the above scheme is that it is difficult to directly learn the features of large quantities of signal data using deep learning methods, because the features of signal data are complex and difficult to summarize. At present, there is little research on RF fingerprint recognition in distributed scenarios, and the network topology of the existing Internet of Things is distributed, and RF fingerprint recognition technology for distributed scenarios is urgently needed. Summary of the invention
[0008] In order to overcome the defects and shortcomings of the prior art, the present invention provides a radio frequency fingerprint recognition method and system for distributed Internet of Things devices. The present invention deploys multiple radio frequency edge nodes. Since the radio frequency devices may move, all edge nodes are required to be able to identify the radio frequency devices in the system, so that the radio frequency signals collected by each edge node obey independent and identical distribution; at the same time, in order to reduce the risk of data leakage and improve the robustness of the system, a federated learning method is adopted to coordinate the training of radio frequency fingerprint recognition models with each edge node, and the radio frequency signal is converted into a spectrum diagram to extract the radio frequency fingerprint. Based on the radio frequency fingerprint recognition, the presence of radio frequency devices in the detection area can be detected, so that fingerprint recognition with higher accuracy can be achieved, which has the advantages of anti-interference, strong robustness, high recognition accuracy, and the significance of safe deployment in distributed Internet of Things, and realizes the characteristics of data not leaving the domain and collaborative training between nodes.
[0009] In order to achieve the above object, the present invention adopts the following technical solutions:
[0010] The present invention provides a radio frequency fingerprint recognition method for distributed Internet of Things devices, comprising the following steps:
[0011] Construct a hierarchical distributed network topology, including a core layer consisting of central servers, a convergence layer consisting of edge nodes, and an access layer consisting of radio frequency devices. Each edge node is used as a local node, and the central server is used as a global node.
[0012] The edge node receives the time-domain RF signal emitted by the RF device within the scope within the set time, obtains the orthogonal amplitude modulation signal sample, and performs data enhancement and data preprocessing;
[0013] Based on short-time Fourier transform, the orthogonal amplitude modulation signal samples after data enhancement and data preprocessing are processed into a spectrum diagram;
[0014] Build a radio frequency fingerprint recognition model and iteratively perform local and global training based on a federated learning loop;
[0015] The spectrum graph is used as a training set for local training of the RF fingerprint recognition model, multiple local nodes are selected as nodes participating in the current round of training, the RF fingerprint recognition model is trained based on the training set, and the model parameters are sent to the global node;
[0016] The global node aggregates the model parameters based on the federated averaging algorithm and sends the updated model parameters to the local nodes;
[0017] The RF fingerprint recognition model trained synchronously by each local node converts the RF signal into RF fingerprint, identifies the identity of the RF device, and monitors the situation in the domain.
[0018] As a preferred technical solution, the edge node receives the time domain radio frequency signal emitted by the radio frequency device within the scope within a set time, t The signal received at the moment is expressed as:
[0019] ;
[0020] in, Expressed as the time-varying channel impulse response, is additive white Gaussian noise, Indicates hardware damage to the receiver. Represents a convolution operation.
[0021] As a preferred technical solution, the data enhancement specifically includes:
[0022] Multipath and Doppler frequency shift fading are added to the signal received by the edge node, and a channel with multipath and Doppler effects is simulated by increasing the number of signal data through replication, and the replicated and increased signal data passes through the channel, and artificial Gaussian white noise is added to the signal to simulate different signal-to-noise ratio levels.
[0023] As a preferred technical solution, the data preprocessing specifically includes:
[0024] Execute the data packet detection algorithm to determine whether there is a valid wireless data packet in the orthogonal amplitude modulation signal sample. If there is a data packet, execute the synchronization algorithm to locate the starting point of the received data packet and perform carrier frequency offset compensation.
[0025] As a preferred technical solution, the orthogonal amplitude modulation signal samples after data enhancement and data preprocessing are processed into a spectrum diagram based on short-time Fourier transform, specifically including:
[0026] The long-time signal is divided into multiple overlapping short time periods, each time period is used as a window, and the Fourier transform is applied to the signal in each window to obtain the spectrum information. The short-time Fourier transform is specifically expressed as:
[0027] ;
[0028] in, j and k It represents the number of rows. Krepresents the discrete short-time Fourier complex matrix S the number of rows, is the discrete short-time Fourier complex matrix S , c represents the discrete short-time Fourier complex matrix S The number of columns;
[0029] Analyzing Discrete Short-Time Fourier Complex Matrix S The amplitude of the signal is obtained, and the spectrum diagram of the signal is expressed as:
[0030] ;
[0031] in, represents the discrete short-time Fourier complex matrix S The amplitude of .
[0032] As a preferred technical solution, the discrete short-time Fourier complex matrix S It is expressed as:
[0033] ;
[0034] in, The elements in the matrix represent the channel gain components of K rows and C columns, is the component of the original signal with K rows and C columns, Signal Hardware damage disturbance.
[0035] As a preferred technical solution, the RF fingerprint recognition model is based on VGG11 as the extractor of RF fingerprint, is stacked by 8 convolutional layers, and uses Batch pooling and ReLU activation function to extract RF fingerprints from the spectrum map, and uses two layers of linear layers to classify the RF fingerprints. In the last linear layer, the softmax function is used to find the category with the highest probability among multiple RF transmitters.
[0036] As a preferred technical solution, the local training specifically includes:
[0037] Select N local nodes as the nodes participating in the training in the current round. The local model of the nodes participating in the training in each round is expressed as , the model parameters are expressed as ;
[0038] Each local node trains the RF fingerprint recognition model based on the training set and obtains the local loss, which is expressed as:
[0039] ;
[0040] Where n represents the node number participating in the current round of training, e represents the number of rounds of the current iteration, Indicates the number of training sets for the nth local node;
[0041] The objective function of each task is expressed as:
[0042] ;
[0043] Back-propagate the local loss to update the local model parameters of the current round, expressed as:
[0044] ;
[0045] in, represents the learning rate of the task;
[0046] Train all N local nodes participating in the current round to obtain the model parameter set Send to the global node.
[0047] As a preferred technical solution, the global node aggregates the model parameters based on the federated average algorithm and sends the updated model parameters to the local node. The updated model parameters are specifically expressed as:
[0048] ;
[0049] in, represents the number of n-th local node training sets, and d represents the total amount of data in the local node training sets participating in the current round of training.
[0050] The present invention also provides a radio frequency fingerprint recognition system for distributed Internet of Things devices, which is used to implement the radio frequency fingerprint recognition method for the above-mentioned distributed Internet of Things devices, and the system includes: a hierarchical distributed network, a signal acquisition module, a data enhancement module, a data preprocessing module, a short-time Fourier transform module, a radio frequency fingerprint recognition model construction module, a model training module, and a radio frequency fingerprint recognition module;
[0051] The hierarchical distributed network includes a core layer composed of a central server, a convergence layer composed of edge nodes, and an access layer composed of radio frequency devices, with each edge node being used as a local node and the central server being used as a global node;
[0052] The signal acquisition module is used to receive the time domain radio frequency signal emitted by the radio frequency device within the scope within a set time, and obtain the orthogonal amplitude modulation signal sample;
[0053] The data enhancement module is used to perform data enhancement processing on the acquired signal;
[0054] The data preprocessing module is used to perform data preprocessing on the acquired signal;
[0055] The short-time Fourier transform module is used to process the orthogonal amplitude modulation signal samples after data enhancement and data preprocessing into a frequency spectrum based on short-time Fourier transform;
[0056] The radio frequency fingerprint recognition model construction module is used to construct a radio frequency fingerprint recognition model;
[0057] The model training module is used to iteratively perform local training and global training based on a federated learning cycle;
[0058] The spectrum graph is used as a training set for local training of the RF fingerprint recognition model, multiple local nodes are selected as nodes participating in the current round of training, the RF fingerprint recognition model is trained based on the training set, and the model parameters are sent to the global node;
[0059] The global node aggregates the model parameters based on the federated averaging algorithm and sends the updated model parameters to the local nodes;
[0060] The RF fingerprint recognition module is used to convert the RF signal into a RF fingerprint based on the trained RF fingerprint recognition model, identify the identity of the RF device, and monitor the situation in the domain.
[0061] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0062] (1) The present invention constructs a hierarchical distributed network topology in a distributed Internet of Things. Through the cooperation of radio frequency devices, edge nodes (responsible for the collection, monitoring and positioning of radio frequency signals) and central servers (core business processing), it is more suitable for the current scenario of large-scale multi-access communication devices in the Internet of Things. It solves the model training problem in the distributed Internet of Things and can be applied to the regional detection of radio frequency devices. It effectively disperses the system load, reduces the risk of single point failure, and improves the robustness of the system.
[0063] (2) The present invention uses short-time Fourier transform to process the original RF signal into a spectrum diagram, and uses the RF fingerprint recognition model to extract the RF fingerprint, which solves the problem that complex signals are difficult to process. Different from the existing solution that directly uses RF complex signals for training, the present invention has the advantage that the CV deep learning model can be used to extract the RF fingerprint, which can improve the accuracy of RF fingerprint recognition.
[0064] (3) The present invention adopts federated learning to achieve collaborative training between edge nodes without leaving the scope of privacy radio frequency signals. By utilizing the collaborative integration training of multiple data nodes, the generalization ability of the model is improved. The advantage is that the risk of privacy leakage is reduced and it is suitable for a wide range of wireless identity authentication scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0065] Figure 1It is a flow chart of the radio frequency fingerprint identification method of the distributed Internet of Things device of the present invention;
[0066] Figure 2 A transmission schematic diagram of a hierarchical distributed network topology structure of the present invention;
[0067] Figure 3 It is a schematic diagram of the classification results of radio frequency fingerprint recognition of the radio frequency fingerprint recognition method of the distributed Internet of Things device of the present invention. DETAILED DESCRIPTION
[0068] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0069] Example 1
[0070] like Figure 1 As shown, this embodiment provides a radio frequency fingerprint recognition method for distributed Internet of Things devices, and the method specifically includes the following steps:
[0071] S1: Build a hierarchical distributed network topology consisting of a central server, edge nodes, and radio frequency devices, divided into: access layer, aggregation layer, and core layer;
[0072] like Figure 2 As shown, the core layer consists of central servers (global nodes), which are used for training global models in federated learning and processing core services.
[0073] The convergence layer consists of edge nodes (local nodes), which are used for training local models in federated learning, collecting RF signals within the node scope (monitoring domain), and monitoring and locating RF devices in the implementation phase;
[0074] The access layer is composed of radio frequency devices, which mainly send radio frequency signals and business data to the edge nodes of the aggregation layer. Radio frequency devices are mobile, which means that the signals of the same radio frequency device can be collected by different edge nodes. Once the radio frequency device moves beyond the monitoring range of the previous edge node, it will switch to the monitoring domain of the new edge node.
[0075] Each edge node can jointly train the RF fingerprint recognition model in a collaborative manner under the premise of ensuring that the RF signal within the scope does not leave the scope. Local training and global training are the process of federated learning. The advantage is that the data does not leave the scope, and only the training parameters are exchanged to jointly train a global security recognition model with other nodes. This embodies collaboration and security. Federated learning (i.e. local and global training) completes collaborative training and privacy protection (privacy data is secure without leaving the scope). VGG11 enables RF identification energy efficiency and can distinguish which device the RF signal comes from.
[0076] S2: M edge nodes collect time-domain RF signals emitted by RF devices within the scope within a set time, and perform data enhancement and preprocessing;
[0077] In this embodiment, the number of edge nodes is M, which is expressed as: , where the edge node It is possible to collect the original RF data of legitimate devices within the scope of the received signal. t The moment can be expressed as:
[0078] ;
[0079] in, Expressed as the time-varying channel impulse response, is additive white Gaussian noise, Indicates hardware damage to the receiver. Represents the convolution operation;
[0080] In this embodiment, 30 LoRa development boards are used as RF transmitters and USRP N210 as receivers, where M=27, meaning there are 27 distributed edge nodes in total. Considering the homogeneous distributed scenario, the RF signal collected by each edge node may be sent by 1 to 30 LoRas.
[0081] Each distributed edge node collects 1,000 RF signals, which may come from 1 to 30 RF devices. Each edge node performs data enhancement and preprocessing on the collected RF signals.
[0082] In order to enhance the robustness of the recognition model to noise, it is necessary to perform data enhancement on the received signal y(t). The channel effects of data enhancement include multipath and Doppler shift. That is, multipath and Doppler shift fading are added to the received signal y(t) to make the model more robust. Specifically, the following are included:
[0083] (1) The training data should be collected in a stationary scenario with line-of-sight wireless transmission so that it can be assumed to experience frequency-flat slow fading;
[0084] (2) Increasing the number of training samples by replication. In this embodiment, the training set is doubled. More replication may improve system performance, but it will significantly increase the requirements for disk and memory storage.
[0085] (3) Use Matlab to simulate channels with multipath and Doppler effects;
[0086] (4) The samples generated in step (2) are passed through the channel in step (3), and artificial Gaussian white noise is added to the signal to simulate different signal-to-noise ratio (SNR) levels;
[0087] Based on the original data, three pre-processing steps are required to improve data quality and the reliability of RF signals. The specific steps include packet detection, synchronization, and carrier frequency offset (CFO) compensation. First, a packet detection algorithm is performed at the receiver to determine whether there is a valid wireless data packet in the captured orthogonal amplitude modulation signal (IQ) sample; if there is a data packet, a synchronization algorithm is performed to accurately locate the starting point of the received data packet; finally, since the oscillator frequency is sensitive to temperature changes, the CFO of the received signal should be estimated and compensated.
[0088] After collection and preprocessing, the edge nodes The RF raw data collected within the scope can be expressed as .
[0089] S3: Processing the RF signal of the IQ sample obtained in step S2 into a spectrum diagram using a short-time Fourier transform (STFT);
[0090] In this embodiment, the complex IQ radio frequency signal is processed into a spectrum diagram according to short-time Fourier transform, and the specific steps are as follows:
[0091] Short-time Fourier transform divides a long-time signal into multiple short time periods, and then performs Fourier transform on the signal in each time period to obtain the spectrum information in that time period.
[0092] The calculation process of STFT is as follows:
[0093] Divide the long-term signal into multiple overlapping short time periods, each of which is called a window;
[0094] Apply Fourier transform to the signal in each window to obtain spectrum information;
[0095] Repeat the above steps until the entire signal is covered.
[0096] The discrete short-time Fourier transform can be expressed mathematically as:
[0097] ;
[0098] Among them, j and k represent the components of the number of rows, j, k = 1, 2, ..., K, K represents the number of rows in the matrix S, is the component of the discrete short-time Fourier complex matrix S, and c represents the number of columns of the matrix, which can be expressed as:
[0099] ;
[0100] Wherein, K is the number of rows of S and the length of the window function w[k], R is the number of hops, in this embodiment, SF=7, B=125kHz, sampling rate Ts=1μs, K is 256, R is 128, and C is calculated to be 63;
[0101] The processing can obtain the short-time Fourier complex matrix S with 102 rows and 63 columns:
[0102] ;
[0103] in, The elements in the matrix represent the channel gain components of K rows and C columns, is the component of the original signal with K rows and C columns, Signal By analyzing the amplitude of the short-time Fourier transform matrix S, we can get the spectrum of the signal and analyze the spectrum characteristics of the RF signal:
[0104] ;
[0105] S4: Iteratively perform local training and global training based on the federated learning cycle;
[0106] In this embodiment, all nodes are divided into two categories: local nodes and global nodes. Each edge node is used as a local node, the central server is used as a global node, and the spectrum graph is used as a training set for the local node to train the radio frequency fingerprint recognition model;
[0107] In this embodiment, the role of the local node is to use the spectrum graph within the scope to train the fingerprint recognition model and send the updated model parameters to the global node. The global node aggregates all submitted model parameters to obtain a fingerprint recognition network applicable to the spectrum graphs within the domain of all local nodes.
[0108] In this embodiment, each edge node processes the radio frequency signals collected within the scope into a spectrum graph based on steps S2 and S3 and encapsulates it into a training set. The training set can be expressed as To prevent overfitting of model training, 10% of the total data needs to be used as a validation set, where each The number of images is 900 and the dimension is 102×63. The training set of the entire model is defined as .
[0109] S5: When starting iterative training, select N local nodes to participate in each round;
[0110] In this embodiment, for model training, a predetermined number of iterations E needs to be set in advance to determine whether the loop reaches the training condition; for each round of iteration, N nodes need to be selected as the nodes participating in the training of the current round. The selection criteria can be selected based on the communication environment conditions. If the communication conditions are close or similar, N nodes can be randomly selected. These selected nodes can be represented as ; For the first iteration, each local node and the global node will initialize the same model parameters .
[0111] In this embodiment, the number of iterations E is preferably E=400, and N is preferably 3. Usually, the nodes participating in the training in each round of iteration will select 10% of the total distributed nodes, because too large a participation ratio will increase the cost of communication and synchronization. These selected nodes can be expressed as .
[0112] In this embodiment, in the first iteration, each local node and the global node will initialize the same parameters of the RF fingerprint recognition model. . Each local node will train the RF fingerprint recognition model, and the global node will aggregate the RF fingerprint recognition models trained by all local nodes. It should be noted that what the local node communicates with the global node is the model parameters rather than the data, so as to achieve the privacy protection of the spectrum graph data and realize the collaborative training of local nodes. The RF fingerprint recognition model is based on VGG11 as the extractor of RF fingerprint, which has powerful image processing capabilities. It is stacked with 8 convolutional layers, each of which uses a smaller 3x3 convolution kernel, and uses Batch pooling and ReLU activation functions. The pooling layer uses a 2x2 maximum pooling operation and a 1x1 average pooling. After feature extraction by VGG11, the original spectrum graph will become a 512×3 RF fingerprint, and then the RF fingerprint is classified using two linear layers. In the last linear layer, the softmax function is used to find the category with the highest probability among the 30 RF transmitters.
[0113] S6: The N local nodes participating in the current round train local models based on the local training set and send network parameters to the global node;
[0114] In this embodiment, the specific steps of local training include:
[0115] The local model of the node participating in each round of training can be expressed as , and its model parameters can be expressed as , where e represents the number of rounds of the current iteration.
[0116] For each local node, the local model is trained based on the local training set using the model parameters of the previous round, and the local loss is obtained, which can be expressed as:
[0117] ;
[0118] Among them, n represents the node number participating in the current round of training , the goal of each task is achieved by minimizing the loss function based on supervised learning. The objective function of each task can be expressed as:
[0119] ;
[0120] Back-propagate the local loss to update the local model parameters of the current round:
[0121] ;
[0122] in, represents the learning rate of the task;
[0123] All N local nodes participating in the current round are trained to obtain the model parameter set Send to the global node.
[0124] S7: The global node aggregates N network parameters to update the global model and sends it to all local nodes; at the same time, it determines whether the current number of iterations has reached the predetermined requirement. If the upper limit of the number of iterations has been reached, the training is terminated. Otherwise, steps S5, S6, and S7 are repeated.
[0125] In this embodiment, the global training specifically includes the following steps:
[0126] The global node receives the parameter set sent by the local node in this round, aggregates it using the federated averaging algorithm FedAvg, and obtains the updated model parameters. It is expressed as:
[0127] ;
[0128] in, represents the number of n-th local node training sets, d represents the total amount of data in the local node training sets participating in the current round of training, ;
[0129] Since the present invention considers a homogeneous distributed scenario, the amount of data of each distributed edge node is the same, which means that the weights of the update parameters of all local nodes are the same, and the aggregated global parameters are Distribute to all local nodes.
[0130] Determine whether the current round e is greater than the specified iteration number E. If so, end the loop training; otherwise, continue to loop through S5, S6, and S7 until e>E.
[0131] S8: After each edge node synchronizes the RF fingerprint recognition model, the edge node identifies the RF devices in the area based on the RF fingerprint recognition and monitors the situation in the area;
[0132] In this embodiment, the trained RF fingerprint recognition model is distributed to all edge nodes, and the RF devices within the scope are broadcast to start communication and execute business logic. The edge node can establish a connection with the RF devices within the scope, and convert the RF signal into a RF fingerprint based on the RF fingerprint recognition model and identify the identity of the RF device. The edge node monitors the survival of the RF devices in the domain and feeds back to the central server. From the perspective of the central server, it can be located in which edge node's scope a certain RF device is active, and it can be monitored whether there is a lost RF device.
[0133] The test data of the present invention uses IQ samples and device labels from the same 30 RF transmitters but not present in the training to test the trained RF fingerprint recognition model. Figure 3 As shown, the classification result is a confusion matrix, which is usually used to visualize classification results. The overall accuracy rate reaches 95.78%, which means that the trained RF fingerprint recognition model can classify RF transmitters 1-30 with high accuracy.
[0134] The present invention can be deployed in a hierarchical star topology consisting of multiple radio frequency devices, multiple edge devices and core servers. The radio frequency signal is expanded into a spectrum graph through data preprocessing and short-time Fourier transform. Based on federated learning, the radio frequency signal is used as privacy data to realize the collaborative training of the radio frequency fingerprint recognition model through each distributed edge node without leaving the domain, thereby realizing high-accuracy radio frequency recognition within the system. In addition, even when the mobile radio frequency device moves across nodes, it can be recognized by nodes in other systems.
[0135] Example 2
[0136] This embodiment provides a radio frequency fingerprint recognition system for distributed Internet of Things devices, which is used to implement the radio frequency fingerprint recognition method for distributed Internet of Things devices in the above-mentioned embodiment 1. The system includes: a hierarchical distributed network, a signal acquisition module, a data enhancement module, a data preprocessing module, a short-time Fourier transform module, a radio frequency fingerprint recognition model construction module, a model training module, and a radio frequency fingerprint recognition module;
[0137] In this embodiment, the hierarchical distributed network includes a core layer composed of a central server, a convergence layer composed of edge nodes, and an access layer composed of radio frequency devices, each edge node is used as a local node, and the central server is used as a global node;
[0138] In this embodiment, the signal acquisition module is used to receive the time domain radio frequency signal emitted by the radio frequency device within the scope within a set time, and obtain the orthogonal amplitude modulation signal sample;
[0139] In this embodiment, the data enhancement module is used to perform data enhancement processing on the acquired signal;
[0140] In this embodiment, the data preprocessing module is used to perform data preprocessing on the acquired signal;
[0141] In this embodiment, the short-time Fourier transform module is used to process the orthogonal amplitude modulation signal samples after data enhancement and data preprocessing into a spectrum diagram based on short-time Fourier transform;
[0142] In this embodiment, the radio frequency fingerprint recognition model construction module is used to construct a radio frequency fingerprint recognition model;
[0143] In this embodiment, the model training module is used to iteratively perform local training and global training based on a federated learning cycle;
[0144] The spectrum graph is used as a training set for local training of the RF fingerprint recognition model, multiple local nodes are selected as nodes participating in the current round of training, the RF fingerprint recognition model is trained based on the training set, and the model parameters are sent to the global node;
[0145] The global node aggregates the model parameters based on the federated averaging algorithm and sends the updated model parameters to the local nodes;
[0146] In this embodiment, the RF fingerprint recognition module is used to convert the RF signal into a RF fingerprint based on the trained RF fingerprint recognition model, identify the identity of the RF device, and monitor the situation in the domain.
[0147] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.
Claims
1. A radio frequency fingerprint recognition method for distributed IoT devices. It is characterized in that The steps include: Construct a hierarchical distributed network topology, including a core layer consisting of central servers, a convergence layer consisting of edge nodes, and an access layer consisting of radio frequency devices. Each edge node is used as a local node, and the central server is used as a global node. The edge node receives the time-domain RF signal emitted by the RF device within the scope within the set time, obtains the orthogonal amplitude modulation signal sample, and performs data enhancement and data preprocessing; Based on short-time Fourier transform, the orthogonal amplitude modulation signal samples after data enhancement and data preprocessing are processed into a spectrum diagram; Build a radio frequency fingerprint recognition model and iteratively perform local and global training based on a federated learning loop; The spectrum graph is used as a training set for local training of the RF fingerprint recognition model, multiple local nodes are selected as nodes participating in the current round of training, the RF fingerprint recognition model is trained based on the training set, and the model parameters are sent to the global node; The global node aggregates the model parameters based on the federated averaging algorithm and sends the updated model parameters to the local nodes; The RF fingerprint recognition model trained synchronously by each local node converts the RF signal into RF fingerprint, identifies the identity of the RF device, and monitors the situation in the domain.
2. The radio frequency fingerprint recognition method for distributed Internet of Things devices according to claim 1, It is characterized in that The edge node receives the time domain radio frequency signal emitted by the radio frequency device within the scope within the set time, and t The signal received at the moment is expressed as: ; in, Expressed as the time-varying channel impulse response, is additive white Gaussian noise, Indicates hardware damage of the receiver. Represents a convolution operation.
3. The radio frequency fingerprint recognition method for distributed Internet of Things devices according to claim 1, It is characterized in that The data enhancement specifically includes: Multipath and Doppler frequency shift fading are added to the signal received by the edge node, and a channel with multipath and Doppler effects is simulated by increasing the number of signal data through replication, and the replicated and increased signal data passes through the channel, and artificial Gaussian white noise is added to the signal to simulate different signal-to-noise ratio levels.
4. The radio frequency fingerprint recognition method for distributed Internet of Things devices according to claim 1, It is characterized in that The data preprocessing specifically includes: Execute the data packet detection algorithm to determine whether there is a valid wireless data packet in the orthogonal amplitude modulation signal sample. If there is a data packet, execute the synchronization algorithm to locate the starting point of the received data packet and perform carrier frequency offset compensation.
5. The radio frequency fingerprint recognition method for distributed Internet of Things devices according to claim 1, It is characterized in that Based on short-time Fourier transform, the orthogonal amplitude modulation signal samples after data enhancement and data preprocessing are processed into a spectrum diagram, specifically including: The long-time signal is divided into multiple overlapping short time periods, each time period is used as a window, and the Fourier transform is applied to the signal in each window to obtain the spectrum information. The short-time Fourier transform is specifically expressed as: ; in, j and k It represents the number of rows. K represents the discrete short-time Fourier complex matrix S the number of rows, is the discrete short-time Fourier complex matrix S , c represents the discrete short-time Fourier complex matrix S The number of columns; Analyzing Discrete Short-Time Fourier Complex Matrix S The amplitude of the signal is obtained, and the spectrum diagram of the signal is expressed as: ; in, represents the discrete short-time Fourier complex matrix S The amplitude of .
6. The radio frequency fingerprint recognition method for distributed Internet of Things devices according to claim 5, It is characterized in that Discrete Short-Time Fourier Complex Matrix S It is expressed as: ; in, The elements in the matrix represent the channel gain components of K rows and C columns, is the component of the original signal with K rows and C columns, Signal Hardware damage disturbance.
7. The radio frequency fingerprint recognition method for distributed Internet of Things devices according to claim 1, It is characterized in that The RF fingerprint recognition model is based on VGG11 as the extractor of RF fingerprint. It is composed of 8 stacked convolutional layers, and uses Batch pooling and ReLU activation function to extract RF fingerprints from the spectrum map. The RF fingerprints are classified using two linear layers. In the last linear layer, the softmax function is used to find the category with the highest probability among multiple RF transmitters.
8. The radio frequency fingerprint recognition method for distributed Internet of Things devices according to claim 1, It is characterized in that The local training specifically includes: Select N local nodes as the nodes participating in the training in the current round. The local model of the nodes participating in the training in each round is expressed as , the model parameters are expressed as ; Each local node trains the RF fingerprint recognition model based on the training set and obtains the local loss, which is expressed as: ; Where n represents the node number participating in the current round of training, e represents the number of rounds of the current iteration, Indicates the number of training sets for the nth local node; The objective function of each task is expressed as: ; Back-propagate the local loss to update the local model parameters of the current round, expressed as: ; in, represents the learning rate of the task; Train all N local nodes participating in the current round to obtain the model parameter set Send to the global node.
9. The radio frequency fingerprint recognition method for distributed Internet of Things devices according to claim 8, It is characterized in that The global node aggregates the model parameters based on the federated average algorithm and sends the updated model parameters to the local node. The updated model parameters are specifically expressed as: ; in, represents the number of n-th local node training sets, and d represents the total amount of data in the local node training sets participating in the current round of training.
10. A radio frequency fingerprint recognition system for distributed IoT devices. It is characterized in that The radio frequency fingerprint recognition method for implementing the distributed Internet of Things device described in any one of claims 1 to 9, the system comprising: a hierarchical distributed network, a signal acquisition module, a data enhancement module, a data preprocessing module, a short-time Fourier transform module, a radio frequency fingerprint recognition model building module, a model training module, and a radio frequency fingerprint recognition module; The hierarchical distributed network includes a core layer composed of a central server, a convergence layer composed of edge nodes, and an access layer composed of radio frequency devices, with each edge node being used as a local node and the central server being used as a global node; The signal acquisition module is used to receive the time domain radio frequency signal emitted by the radio frequency device within the scope within a set time, and obtain the orthogonal amplitude modulation signal sample; The data enhancement module is used to perform data enhancement processing on the acquired signal; The data preprocessing module is used to perform data preprocessing on the acquired signal; The short-time Fourier transform module is used to process the orthogonal amplitude modulation signal samples after data enhancement and data preprocessing into a frequency spectrum based on short-time Fourier transform; The radio frequency fingerprint recognition model construction module is used to construct a radio frequency fingerprint recognition model; The model training module is used to iteratively perform local training and global training based on a federated learning cycle; The spectrum graph is used as a training set for local training of the RF fingerprint recognition model, multiple local nodes are selected as nodes participating in the current round of training, the RF fingerprint recognition model is trained based on the training set, and the model parameters are sent to the global node; The global node aggregates the model parameters based on the federated averaging algorithm and sends the updated model parameters to the local nodes; The RF fingerprint recognition module is used to convert the RF signal into a RF fingerprint based on the trained RF fingerprint recognition model, identify the identity of the RF device, and monitor the situation in the domain.
