A method and apparatus for radio frequency fingerprint recognition based on few-sample open-set incremental learning using meta-learning
Patent Information
- Application Number
- CN202410334573.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-22
- Publication Date
- 2026-09-01
- Estimated Expiration
- 2044-03-22
AI Technical Summary
然而，现有的DL技术基类分类器有以下问题：可以准确地对旧类进行分类，但不能对新类进行分类；通过直接训练分类器，模型偏向于拟合新类，忘记了旧类的知识；由于小样本限制，难以确定新类的开集识别阈值
[0071] This invention proposes a radio frequency fingerprint recognition scheme for small samples. It extracts pseudo-tasks from the training set to realize environmental simulation, which is suitable for continuous incremental processes. Compared with traditional incremental learning methods, it achieves better recognition results.
Smart Images
Figure CN118245872B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of radio frequency fingerprint recognition technology, and in particular to a radio frequency fingerprint recognition method and apparatus based on meta-learning few-sample open set incremental learning. Background Technology
[0002] In recent years, deep learning (DL) technology has been widely applied to radio frequency fingerprinting and has generated numerous public wireless security applications, such as modulation identification, IoT device authentication, spectrum monitoring, and aviation management. Utilizing the powerful feature extraction capabilities of DL models to manage intelligent communication devices holds significant promise. Previous research has explored various DL-based RFID fingerprinting techniques to improve the identification efficiency of wireless devices. However, existing DL base classifiers suffer from the following problems: they can accurately classify old classes but not new ones; by directly training the classifier, the model tends to fit the new class, forgetting knowledge of the old classes; and due to small sample size limitations, it is difficult to determine the open set recognition threshold for new classes.
[0003] Establishing a generalized evolutionary model that combines few-shot learning (FSL), open-set recognition (OSR), and multi-stage continuous increments requires overcoming the following challenges: catastrophic forgetting and overfitting, unknown class distribution shifts, and unknown open-set thresholds. Existing few-shot class increment (FSCIL) algorithms appear to address these challenges, but still struggle to determine the open-set boundary for small-shot samples. The lack of samples prevents us from seeing the true class distribution and understanding the sample's position within it, hindering accurate estimation of the open-set boundary. Furthermore, the multi-stage continuous increment process exacerbates these challenges, leading to a significant decrease in the efficiency of existing increment algorithms such as iCARL, OSR, and FSCIL.
[0004] Therefore, existing technologies have not achieved accurate prediction results in radio frequency fingerprint recognition based on small sample open set incremental learning. In order to solve these problems, there is an urgent need to provide an effective radio frequency fingerprint recognition method in the current complex signal environment. Summary of the Invention
[0005] To overcome the shortcomings of existing detection methods, this invention proposes a radio frequency fingerprint recognition method and apparatus based on meta-learning and few-sample open-set incremental learning.
[0006] The technical solution adopted by this invention to solve its technical problem is:
[0007] The first invention relates to a radio frequency fingerprinting method based on meta-learning and few-sample open-set incremental learning, the method comprising the following steps:
[0008] 1) Feature pre-training: Extract more signal modal information, obtain features using basic session data, and embed them into network and classifier weights;
[0009] 2) Meta-task sampling: The base class dataset is divided into support set, multi-sample query set, and open set. A large number of simulated small sample open set incremental recognition tasks are extracted from the basic session data to simulate the process of real small sample open set incremental recognition tasks.
[0010] 3) Meta-incremental training: The network is generalized between new and old class samples, and incremental learning is optimized by using a similarity-based cross-entropy loss function. Based on the idea of meta-learning, the incremental learning task of a large number of simulated samples is optimized.
[0011] 4) Optimization of small sample open set recognition: Simulate open set scenarios in the meta-task to optimize open set recognition capabilities, and use clustering samples to gradually reduce intra-class discrepancies;
[0012] 5) Adaptive RF fingerprint open set recognition mechanism: Employs a multi-head self-attention mechanism to explore the relationships between prototype points, uses attention blocks to generate inverted points, and then uses an optimization function to optimize the meta-learning task.
[0013] Furthermore, in step 1), the feature pre-training specifically includes the following process:
[0014] Step 101, Feature Extraction: Extract more signal modal information, mainly the instantaneous amplitude, instantaneous phase, instantaneous frequency, and center instantaneous frequency of the signal. The calculation formulas for each feature are as follows:
[0015]
[0016]
[0017]
[0018]
[0019] Furthermore, the symbol x(n) is defined as:
[0020] Step 102, Model pre-training: Train the feature embedding network and classifier in base class training. The weights of the classifier are represented by the average embedding of the class prototype or the most representative feature in the class.
[0021] class prototype w j It can be calculated using the following formula:
[0022]
[0023] Where I(·) represents the index function.
[0024] The probability p of each base class training dialogue j The calculation formula is as follows:
[0025]
[0026] Here, sim() represents the cosine similarity function.
[0027] Finally, the cross-loss function is used as the loss function for training the dialogue in each base class section.
[0028] Furthermore, in step 2), the meta-task sampling includes the following process;
[0029] Step 201: Divide the base class dataset into three non-overlapping datasets.
[0030] T = {S, Q} * Q O |C S C * C O}
[0031] in To support the set, N classes are randomly sampled each time, with K samples from each class.
[0032] in For a multi-sample query set, each time |Y0| classes are randomly sampled, and each class has K samples, where Y0 is the label space of the support set S.
[0033] Q O For an open set, N classes are randomly sampled each time, with K samples from each class.
[0034] Step 202 involves extracting a large number of simulated FSOSIL tasks from the base session data to simulate the process of real FSOSIL tasks. The sampled simulated FSOSIL tasks aim to provide a method for neural networks to learn generalizable embeddings.
[0035] In step 3), the meta-incremental training includes the following process:
[0036] Step 301: During the incremental session, the classification model will continuously receive new session tasks, requiring the network to generalize between new and old class samples. To simulate the incremental session, the new class incremental samples and all class samples are represented using the support set and the all class query set from the base session, respectively. In this process, the network first generates new prototypes for the small support set in the incoming task using an equation, and then directly uses the pre-trained old class weights to identify the old class. Therefore, the classifier weights for the task... The update will be performed in the following ways:
[0037] Step 302, incremental learning is optimized using a similarity-based cross-entropy loss function, as follows:
[0038]
[0039] Step 303: Finally, based on the idea of meta-learning, it is necessary to simultaneously optimize the incremental learning task on a large number of simulated samples. This allows the neural network to learn how to adapt to this environment. The meta-learning optimization loss is shown below:
[0040]
[0041] In step 4), the optimization of small sample open set identification includes the following process:
[0042] Step 401: Simulate open set scenarios in the meta-task to optimize open set recognition capabilities. To fairly control the distance between class prototypes, an orthogonalization loss is introduced in the inner product space. In the orthogonal space, all prototypes are orthogonalized vectors with a cosine similarity of 0. To ensure that the prototype points of the new class are orthogonal to those of the old class, an orthogonalization loss l is applied in the meta-task. or ,Right now
[0043]
[0044]
[0045] Where M is the mask matrix, and in each meta-task, the updated prototype points... The prototype points will be orthogonalized by the meta-calibrator T, which guarantees that the similarity between prototype points of different classes is 0.
[0046] Step 402: After controlling for inter-class distance, it is necessary to further control for intra-class distance. Clustered samples are used to progressively reduce the disparity. This process can be achieved using the following formula.
[0047]
[0048]
[0049]
[0050] l e l represents Euclidean distance. c This indicates cosine similarity.
[0051] Step 403: Calculate the element-wise loss using the following formula.
[0052]
[0053] Where α and β represent hyperparameters.
[0054] In step 5), adaptive RFF open set identification includes the following process:
[0055] Step 501: To adapt to the continuous incremental few-shot learning task, we added a generative network G. θ We train the dataset using numerous meta-tasks to learn how to synthesize inverted points R. For multi-sample query sets, we extract the corresponding prototype point P of the i-th class. i and using the generator network G θ Generate inverted point R i The generator network structure primarily employs the standard Multi-Head Self-Attention (MHSA) mechanism, which can effectively uncover the relationships between prototype points. We apply P... i The attention block between W and the inverted point R is generated as follows:
[0056]
[0057]
[0058]
[0059] Among them G q G k G v R represents the parameters of MHSA. i This represents the inverted point of the i-th class.
[0060] Step 502: For a multi-sample query set, after obtaining the inverse points of the target class, we can use the binary cross-entropy function to optimize it:
[0061]
[0062] For the constructed open set samples, we expect them to be related to the non-target class prototype point P. i The similarity is smaller, while the reciprocal point R is smaller. i The similarity is higher, which can be expressed as:
[0063]
[0064] Therefore, for general tasks, the loss function is defined as:
[0065]
[0066] Step 503, the final FSOCIL meta-learning task loss can be calculated using the following formula:
[0067]
[0068] This allows for the optimization of the FSOCIL meta-learning task.
[0069] A second aspect of the present invention relates to a radio frequency fingerprint recognition device, comprising a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement a radio frequency fingerprint recognition method based on meta-learning and few-sample open-set incremental learning according to the present invention.
[0070] A third aspect of the present invention relates to a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements a radio frequency fingerprinting method based on meta-learning and few-sample open-set incremental learning.
[0071] This invention proposes a radio frequency fingerprint recognition scheme for small samples. It extracts pseudo-tasks from the training set to realize environmental simulation, which is suitable for continuous incremental processes. Compared with traditional incremental learning methods, it achieves better recognition results.
[0072] The technical concept of this invention is to propose a signal enhancement scheme for small sample sizes. Based on the idea of meta-learning, a FSOSIL meta-task is defined, and a large number of pseudo-tasks are extracted from the training set to achieve environmental simulation. Furthermore, multi-task training is performed using meta-learning techniques, enabling the neural network to adapt to such an environment. Finally, to solve the FSOSR problem, an open loss is added to the meta-task, and an open-set thresholdless mechanism is proposed. This mechanism is applicable to continuous incremental processes and automatically generates open-set thresholds for small sample classes.
[0073] The main advantages of this invention are as follows: it establishes an FSOSIL framework for RFF recognition, implements a generalized RFF continuous incremental learning system, defines the FSOSIL meta-task, and designs a multi-task training mechanism based on a prototype network. It proposes a soft orthogonalization loss and an open loss to achieve automatic calibration of prototype points and OSR, achieving superior recognition performance compared to traditional incremental learning methods. Attached Figure Description
[0074] Figure 1 This is a flowchart of the method of the present invention.
[0075] Figure 2 This is a structural diagram of an adaptive radio frequency fingerprint open set recognition framework. Detailed Implementation
[0076] The invention will be further described below with reference to the accompanying drawings.
[0077] Example 1
[0078] Reference Figure 1 A radio frequency fingerprint recognition method based on few-sample open-set incremental learning using meta-learning, the method comprising the following steps:
[0079] 1) Feature pre-training: Extract more signal modal information, obtain features using basic session data, and embed them into network and classifier weights;
[0080] 2) Meta-task sampling: The base class dataset is divided into support set, multi-sample query set, and open set. A large number of simulated small sample open set incremental recognition tasks are extracted from the basic session data to simulate the process of real small sample open set incremental recognition tasks.
[0081] 3) Meta-incremental training: The network is generalized between new and old class samples, and incremental learning is optimized by using a similarity-based cross-entropy loss function. Based on the idea of meta-learning, the incremental learning task of a large number of simulated samples is optimized.
[0082] 4) Optimization of small sample open set recognition: Simulate open set scenarios in the meta-task to optimize open set recognition capabilities, and use clustering samples to gradually reduce intra-class discrepancies;
[0083] 5) Adaptive RF fingerprint open set recognition mechanism: Employs a multi-head self-attention mechanism to explore the relationships between prototype points, uses attention blocks to generate inverted points, and then uses an optimization function to optimize the meta-learning task.
[0084] Furthermore, in step 1), the feature pre-training specifically includes the following process:
[0085] Step 101, Feature Extraction: Extract more signal modal information, mainly the instantaneous amplitude, instantaneous phase, instantaneous frequency, and center instantaneous frequency of the signal. The calculation formulas for each feature are as follows:
[0086]
[0087]
[0088]
[0089]
[0090] Furthermore, the symbol x(n) is defined as:
[0091] Step 102, Model pre-training: Train the feature embedding network and classifier in base class training. The weights of the classifier are represented by the average embedding of the class prototype or the most representative feature in the class.
[0092] class prototype w j It can be calculated using the following formula:
[0093]
[0094] Where I(·) represents the index function.
[0095] The probability p of each base class training dialogue j The calculation formula is as follows:
[0096]
[0097] Here, sim() represents the cosine similarity function.
[0098] Finally, the cross-loss function is used as the loss function for training the dialogue in each base class section.
[0099] Furthermore, in step 2), the meta-task sampling includes the following process;
[0100] Step 201: Divide the base class dataset into three non-overlapping datasets.
[0101] T = {S, Q} * Q O |C S C * C O}
[0102] in To support the set, N classes are randomly sampled each time, with K samples from each class.
[0103] in For a multi-sample query set, each time |Y0| classes are randomly sampled, and each class has K samples, where Y0 is the label space of the support set S.
[0104] Q O For an open set, N classes are randomly sampled each time, with K samples from each class.
[0105] Step 202 involves extracting a large number of simulated FSOSIL tasks from the base session data to simulate the process of real FSOSIL tasks. The sampled simulated FSOSIL tasks aim to provide a method for neural networks to learn generalizable embeddings.
[0106] In step 3), the meta-incremental training includes the following process:
[0107] Step 301: During the incremental session, the classification model will continuously receive new session tasks, requiring the network to generalize between new and old class samples. To simulate the incremental session, the new class incremental samples and all class samples are represented using the support set and the all class query set from the base session, respectively. In this process, the network first generates new prototypes for the small support set in the incoming task using an equation, and then directly uses the pre-trained old class weights to identify the old class. Therefore, the classifier weights for the task... The update will be performed in the following ways:
[0108]
[0109] Step 302, incremental learning is optimized using a similarity-based cross-entropy loss function, as follows:
[0110]
[0111] Step 303: Finally, based on the idea of meta-learning, it is necessary to simultaneously optimize the incremental learning task on a large number of simulated samples. This allows the neural network to learn how to adapt to this environment. The meta-learning optimization loss is shown below:
[0112]
[0113] In step 4), the optimization of small sample open set identification includes the following process:
[0114] Step 401: Simulate open set scenarios in the meta-task to optimize open set recognition capabilities. To fairly control the distance between class prototypes, an orthogonalization loss is introduced in the inner product space. In the orthogonal space, all prototypes are orthogonalized vectors with a cosine similarity of 0. To ensure that the prototype points of the new class are orthogonal to those of the old class, an orthogonalization loss l is applied in the meta-task. or ,Right now
[0115]
[0116]
[0117] Where M is the mask matrix, and in each meta-task, the updated prototype points... The prototype points will be orthogonalized by the meta-calibrator T, which guarantees that the similarity between prototype points of different classes is 0.
[0118] Step 402: After controlling for inter-class distance, it is necessary to further control for intra-class distance. Clustered samples are used to progressively reduce the disparity. This process can be achieved using the following formula.
[0119]
[0120]
[0121]
[0122] l e l represents Euclidean distance. c This indicates cosine similarity.
[0123] Step 403: Calculate the element-wise loss using the following formula.
[0124]
[0125] Where α and β represent hyperparameters.
[0126] In step 5), adaptive RFF open set identification includes the following process:
[0127] Step 501: To adapt to the continuous incremental few-shot learning task, we added a generative network G. θ We train the dataset using numerous meta-tasks to learn how to synthesize inverted points R. For multi-sample query sets, we extract the corresponding prototype point P of the i-th class. i and using the generator network G θ Generate inverted point R i The generator network structure primarily employs the standard Multi-Head Self-Attention (MHSA) mechanism, which can effectively uncover the relationships between prototype points. We apply P... i The attention block between W and the inverted point R is generated as follows:
[0128]
[0129]
[0130]
[0131] Among them G q G k G v R represents the parameters of MHSA. i This represents the inverted point of the i-th class.
[0132] Step 502: For a multi-sample query set, after obtaining the inverse points of the target class, we can use the binary cross-entropy function to optimize it:
[0133]
[0134] For the constructed open set samples, we expect them to be related to the non-target class prototype point P. i The similarity is smaller, while the reciprocal point R is smaller. i The similarity is higher, which can be expressed as
[0135]
[0136] Therefore, for general tasks, the loss function is defined as:
[0137]
[0138] Step 503, the final FSOCIL meta-learning task loss can be calculated using the following formula:
[0139]
[0140] This allows for the optimization of the FSOCIL meta-learning task.
[0141] Example 2
[0142] This embodiment relates to a radio frequency fingerprint recognition device, including a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, they are used to implement a radio frequency fingerprint recognition method based on meta-learning and few-sample open set incremental learning according to Embodiment 1.
[0143] Example 3
[0144] This embodiment relates to a computer-readable storage medium storing a program that, when executed by a processor, implements a radio frequency fingerprinting method based on meta-learning and few-sample open-set incremental learning, as described in Embodiment 1.
[0145] This invention proposes a radio frequency fingerprint recognition scheme for small sample sizes. It extracts pseudo-tasks from the training set to simulate the environment, making it suitable for continuous incremental processes. Compared to traditional incremental learning methods, it achieves superior recognition performance. This method belongs to the field of radio frequency fingerprint recognition technology and successfully achieves accurate recognition of small sample radio frequency fingerprints.
[0146] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.
Claims
1. A radio frequency fingerprint recognition method based on few-sample open-set incremental learning of meta-learning, characterized in that: The specific steps are as follows: 1) Feature pre-training: Extract more signal modal information, obtain features using basic session data, and embed them into network and classifier weights; 2) Meta-task sampling: The base class dataset is divided into support set, multi-sample query set, and open set. A large number of simulated small sample open set incremental recognition tasks are extracted from the basic session data to simulate the process of real small sample open set incremental recognition tasks. 3) Training meta-increment: The network is generalized between new and old class samples, and the incremental learning is optimized by the cross-entropy loss function based on similarity. Based on the idea of meta-learning, the incremental learning task of a large number of simulated samples is optimized. 4) Optimize small sample open set recognition: Simulate open set scenarios in the meta-task to optimize open set recognition capabilities, and use clustering sample methods to gradually reduce intra-class discrepancies; 5) Adaptive RF fingerprint open set recognition mechanism: Employs a multi-head self-attention mechanism to mine the relationships between prototype points, uses attention blocks to generate inverted points, and then uses an optimization function to optimize the meta-learning task; specifically including: Step 501, add a generator network It is trained using a large number of meta-tasks to learn how to synthesize inverted points R; for multi-sample query sets, the prototype points of the corresponding i-th class are extracted. and using a generative network Generate inverted point The generator network structure employs the standard multi-head self-attention (MHSA) mechanism, which can effectively uncover the relationships between prototype points; Applications The attention block between W and the inverted point R is generated as follows: , = , = , （15） ， （16） = + ， （17） in Indicates the parameters of MHSA, Represents the inverted point of the i-th class; Step 502: For a multi-sample query set, after obtaining the inverse points of the target class, optimize it using the binary cross-entropy function: ))+ ))（18） For the constructed open set samples, the expectation is that they are similar to the prototype points of the non-target class. The similarity is smaller, while the reciprocal point is smaller. The similarity is higher, which can be expressed as ))+ ))（19） Therefore, for general tasks, the loss function is defined as: Step 503, the final FSOCIL meta-learning task loss can be calculated using the following formula: This optimizes the FSOCIL meta-learning task; among which, ( The cross-entropy loss is based on similarity. The prototype orthogonalization loss is ) ( ( ), The loss is the intra-class distance constraint loss. , , This is a hyperparameter.
2. The radio frequency fingerprint recognition method based on few-sample open-set incremental learning according to meta-learning as described in claim 1, characterized in that: Step 1) Feature pre-training specifically includes: Step 101, Feature Extraction: Extract more signal modal information, including the instantaneous amplitude, instantaneous phase, instantaneous frequency, and center instantaneous frequency. The calculation formulas for each feature are as follows: （1） （2） (n-1),n=1,2,...,N-1 (3) （4） Furthermore, the sample symbol x(n) is defined as: , } Step 102, Model pre-training: Train the feature embedding network and classifier in base class training; the weights of the classifier are represented by the average embedding of the class prototype or the most representative feature in the class. in Indicates the index function; The probability of each base class training dialogue The calculation formula is as follows: （6） Finally, the cross-loss function is used as the loss function for training the dialogue in each base class section.
3. The radio frequency fingerprint recognition method based on meta-learning and few-sample open-set incremental learning as described in claim 1, characterized in that: The meta-task sampling in step 2) specifically includes: Step 201: Divide the base class dataset into three non-overlapping datasets. in To form a support set, N classes are randomly sampled each time, with K samples from each class; in For a multi-sample query set, random sampling is performed each time. | classes, each class has K samples, where Y0 is the label space of the support set S; The set is open, and N classes are randomly sampled each time, with K samples from each class; Step 202 involves extracting a large number of simulated FSOSIL tasks from the base session data to simulate the process of real FSOSIL tasks; the sampled simulated FSOSIL tasks are intended to provide a method for neural networks to learn generalizable embeddings.
4. The radio frequency fingerprint recognition method based on meta-learning and few-sample open-set incremental learning as described in claim 2, characterized in that: Step 3) specifically includes the following training meta-increments: Step 301: During the incremental session, the classification model will continuously receive new session tasks, which requires the network to generalize between new class samples and old class samples. To simulate the incremental session, the new class incremental samples and all class samples are represented by the support set and the query set of all classes in the base session, respectively. In this process, the network first generates a new prototype for a small number of support sets in the input task according to formula (5), and then directly uses the pre-trained old class weights to identify the old class. Therefore, the classifier weights of the task The update will be performed in the following ways: = （7） Step 302, incremental learning is optimized using a similarity-based cross-entropy loss function, as follows: （8） Step 303: Finally, based on the idea of meta-learning, it is necessary to simultaneously optimize the incremental learning task on a large number of simulated samples. This allows the neural network to learn how to adapt to this environment. The meta-learning optimization loss is shown below: （9）。 5. The radio frequency fingerprint recognition method based on meta-learning and few-sample open-set incremental learning as described in claim 1, characterized in that: Step 4) specifically includes optimizing small sample open set recognition: Step 401: Simulate an open set scenario in the meta-task to optimize open set recognition capability; to fairly control the distance between class prototypes, an orthogonalization loss is introduced in the inner product space; in the orthogonal space, all prototypes are orthogonalized vectors with a cosine similarity of 0; to ensure that the prototype points of the new class are orthogonal to the old class, an orthogonalization loss is applied in the meta-task. ,Right now ( )=M , （10） = Where M is the mask matrix, and in each meta-task, the updated prototype points... The prototype points will be orthogonalized by the meta-calibrator T, which guarantees that the similarity between prototype points of different classes is 0. Step 402: After controlling for inter-class distance, it is necessary to further control for intra-class distance; clustered samples are used to gradually reduce the differences; this process can be achieved using the following formula. Represents Euclidean distance. Indicates cosine similarity; Step 403: Calculate the element-wise loss using the following formula. Where α and β represent hyperparameters.
6. A radio frequency fingerprint recognition device, characterized in that, The device includes a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the radio frequency fingerprint recognition method based on few-sample open-set incremental learning according to any one of claims 1-5.
7. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, implements a radio frequency fingerprint recognition method based on few-sample open-set incremental learning according to any one of claims 1-5.