Small sample incremental emitter individual identification method based on knowledge distillation and graph model

Abstract

The present application relates to the technical field of radiation source individual identification, and in order to propose a radiation source individual identification method in a complex scene, the technical scheme adopted by the present application is a small sample incremental radiation source individual identification method based on knowledge distillation and a graph model, a backbone network is first obtained by training using known radiation source data; then, pseudo-class incremental training is performed: two classifiers are respectively connected to the same adaptive graph model classifier, and the adaptive graph model classifier is optimized through training; then, an incremental training stage is entered, on the one hand, the teacher network is used to perform knowledge distillation on the student network; on the other hand, the adaptive graph model classifier that has been trained will adaptively connect the context information of the new small amount of radiation source individuals; finally, the trained backbone network and the adaptive graph model classifier connected thereto are used to realize small sample class incremental radiation source individual identification. The present application is mainly applied to radiation source individual identification occasions.

Description

Technical Field

[0001] This invention relates to the field of radiation source individual identification technology, specifically to a small-sample incremental radiation source individual identification method based on knowledge distillation and graph models. Background Technology

[0002] The rapid development of the Internet of Things (IoT) and the widespread adoption of mobile devices have led to the integration of numerous wireless devices, and have also brought about the emergence of new technologies such as IoT and 5G networks. However, deploying multiple wireless technologies on these devices presents significant security and privacy challenges, as the inherent broadcast nature of IoT devices creates vulnerabilities that can be exploited by malicious users. Therefore, managing and distributing encryption keys for these IoT devices becomes exceptionally challenging. Consequently, there is an urgent need for a novel, tamper-proof, and secure method for device authentication.

[0003] Radiosource identification (RFFI) refers to the process of identifying different radiosources by extracting subtle features containing personal information about the transmitting device from the received signal. These subtle features are called radio frequency fingerprints (RFF) and are attributed to the tolerance effects present in radio frequency devices. These effects lead to nonlinear features in the device response, and variations in these nonlinear features can distinguish different radiosource devices. Similar to human fingerprints, RFF is the basis for distinguishing radio frequency devices and can be used as a security method for device authentication [1].

[0004] Recently, deep learning has become the mainstream solution for solving the RFFI problem. However, the performance of data-driven methods based on deep learning in RFFI is limited because a large number of labeled signal samples cannot be obtained in non-cooperative scenarios. Few-shot learning (FSL), as a small-instance deep learning paradigm, has been proven to be very effective in image recognition tasks with limited data. It aims to solve the problems of insufficient generalization and poor adaptability to rare categories of typical classification methods. It can alleviate the severe performance degradation caused by the lack of labeled samples. Existing FSL-based RFFI methods can effectively solve the RFFI problem of closed sets with few training samples. Zhang et al. [2] proposed a data augmentation (DA)-assisted few-shot learning method, which improved the recognition accuracy and the robustness of the model. Wang et al. [3] proposed a new method based on hybrid data augmentation and deep metric learning (HDA-DML), which got rid of the dependence on auxiliary datasets.

[0005] RFFI continuously captures radio frequency signals of new target categories, which are unknown during training and can naturally be misclassified as known categories. To ensure reliable recognition, traditional RFFI models require retraining for both new and old categories. While adding training data for retraining in non-open scenarios has little impact on RFFI model performance, repeatedly training previous training samples leads to significant time consumption and wasted storage resources, which is unreasonable for real-time RFFI tasks. Furthermore, directly discarding training samples of old targets and fine-tuning the trained model with new targets can cause catastrophic forgetting, resulting in a decline in the ability to recognize old targets. Therefore, it is necessary to develop a new RFFI paradigm and design models for open scenarios that can quickly adapt to updated data and learn effective knowledge of new categories while maintaining the original classification capabilities.

[0006] Incremental learning is an advanced open-scene target recognition technique. First, the model is trained using the base class, and then new classes are added by providing the corresponding training data. Then, the model is incrementally trained using only the new class, thereby achieving unified recognition of the base class and the new class. Recent studies have shown that the RFFI method based on incremental learning can save computational resources, avoid catastrophic forgetting, and avoid repeated training and storage of old classes. Liu et al. [3] summarized a general RFFI framework and proposed a method to effectively reduce the amount of old data required for incremental learning-based RFFI, thereby saving training time and storage space. Liu et al. [4] proposed a distributed sensor system that uses incremental learning to solve the problem of radio frequency fingerprint recognition, which can significantly reduce training time and adapt to streaming data. Numerous experiments show that the method has high computational efficiency and satisfactory recognition accuracy, especially in cases of low signal-to-noise ratio. Zhou et al. [5] proposed an RFFI method based on knowledge extraction, which uses continuous data streams to continuously update the recognition model. Liu et al. [6] proposed a new channel separation scheme that does not use historical data, which is a strategy to automatically separate device fingerprints at different learning stages and avoid potential conflicts.

[0007] Driven by the aforementioned challenges, this invention proposes a novel RFFI method based on class incremental learning, which can adapt to new categories with a small number of training samples in an efficient and timely manner. The main challenge of RFFI in few-sample class incremental learning [7], [8] is that a small number of labeled samples can lead to severe overfitting to new targets and catastrophic forgetting of old samples.

[0008] [1]Zhang J,Woods R,Sandell M,et al.Radio Frequency FingerprintIdentification for Narrowband Systems,Modelling and Classification[J].IEEETransactions on Information Forensics and Security,2021,16:3974–3987.

[0009] [2]Zhang X,Wang Y,Zhang Y,et al.Data Augmentation Aided Few-ShotLearning for Specific Emitter Identification[C].In 2022IEEE 96th VehicularTechnology Conference(VTC2022-Fall),

[0010] 2022:1–5.

[0011] [3]Wang C,Fu X,Wang Y,et al.Few-Shot Specific Emitter Identificationvia Hybrid Data Augmentation and Deep Metric Learning[J].arXiv preprintarXiv:2212.00252,2022.

[0012] [4]Liu D,Liu C,Yuan M.Incremental Learning for Radio FrequencyFingerprint Identification[C].In 2021 18th International Computer Conferenceon Wavelet Active Media Technology and Information Processing(ICCWAMTIP),2021:326–329.

[0013] [5] Liu M, Wang J, Zhao N, et al. Radio Frequency Fingerprint Collaborative Intelligent Identification Using Incremental Learning[J]. IEEE Transactions on Network Science and Engineering, 2022, 9(5): 3222–3233.

[0014] [6] Zhou J, Peng Y, Gui G, et al. A Novel Radio Frequency Fingerprint Identification Method Using Incremental Learning[C]. In 2022 IEEE 96th Vehicular Technology Conference (VTC2022-Fall), 2022: 1–5.

[0015] [7] Tao X, Hong X, Chang X, et al. Few-Shot Class-Incremental Learning[C]. In 2020 IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020: 12180–12189.

[0016] [8] Zhang C, Song N, Lin G, et al. Few-Shot Incremental Learning with Continually Evolved Classifiers[C]. In 2021 IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021: 12450–12459. Summary of the Invention

[0017] To overcome the shortcomings of existing technologies, this invention aims to propose a method for identifying individual radiation sources in complex scenarios. The technical solution adopted in this invention is a small-sample incremental radiation source identification method based on knowledge distillation and graph models. First, it trains using known radiation source data, obtaining a backbone network by continuously optimizing the cross-entropy loss. Then, pseudo-class incremental training is performed: using known radiation source data, support sets and query sets are respectively divided for virtual incremental classes and virtual base classes. The support set is used to learn the classifier weights for different categories, while the query set is used to calculate the optimization loss. Two classifiers are built for the virtual base class and virtual incremental class using the support set. These two classifiers are then connected to the same adaptive graph model classifier. The parameters of the adaptive graph model classifier are updated through training. Using the updated adaptive graph model classifier, predictions are made for the query sets of the virtual base class and virtual incremental class, and then the loss is calculated to optimize the adaptive graph model. The system first establishes a classifier, then connects the optimized adaptive graphical model classifier to the backbone network. Next, it enters the incremental training phase. On one hand, the backbone network is frozen as the teacher network, and simultaneously replicated and trained using a small number of newly added radiation source individuals to obtain the student network. The teacher network is then used to perform knowledge distillation on the student network. The distilled student network becomes the trained backbone network, which is also replicated and frozen as the teacher network for the next incremental training iteration. This incremental training occurs whenever a small number of new radiation source individuals appear. On the other hand, the already trained adaptive graphical model classifier adaptively connects to the context information of the newly added radiation source individuals. Finally, the trained backbone network and its connected adaptive graphical model classifier are used to achieve radiation source individual identification for small sample increments.

[0018] The specific steps of knowledge distillation are as follows:

[0019] a) Initialization phase: In the initial phase, a basic backbone network is trained to solve the initial task, which involves all categories. This model will serve as the teacher network in subsequent class incremental learning.

[0020] b) Introduction of new categories: The frozen teacher network is trained using the newly introduced categories to form a student network, and the existing teacher network is used to predict the new category data to generate soft target labels. These soft target labels contain the probability distribution of the teacher network's predictions for the new category data.

[0021] c) Student network training: Student networks are trained using soft target labels to approximate the predictions of the teacher network;

[0022] d) Iterative process: When a new category is introduced, the above steps are executed repeatedly to gradually update the student network. In each iteration, knowledge is transferred from the teacher network to the student network through soft target labels.

[0023] Indicates in task T n The parameter matrix of the CNN classifier learned in the middle, where W n Each row vector in It is the weight corresponding to a specific category c, C n It is task T n The number of categories in the embedding vector, E is the dimension of the embedding vector. It can be viewed as the prototype vector of category c, by looking at all other categories. To adjust To obtain the value, we first need to collect the weight vectors of all other categories from previous tasks:

[0024]

[0025] Where N is the total number of tasks to date, then W N In this model, weight vectors are represented as nodes in the graph model. An adaptive graph model classifier is used to model the relationships between these prototype vectors and propagate contextual information. An attention mechanism is used to facilitate information propagation between nodes, enabling each node to receive contextual information from other nodes in the graph.

[0026] In the task of identifying individual radiation sources, fully connected graphs By node and edge j≠k. Where It is the j-th node, e jk It is a connection and Let I be the total number of nodes. For any two nodes... and Relationship coefficient e jk The calculation formula is

[0027]

[0028] Where λ and μ represent linear transformation functions used to project the original prototype representation onto a new metric space, and the functions <·,·> calculate the cosine similarity between two vectors. For clarity... and The task index in the subscript is omitted. Then, the obtained coefficients are normalized using the softmax function to obtain the final attention weights corresponding to the center node j.

[0029]

[0030] In the normalized attention coefficient a jk Based on a jkThe information of all nodes in the aggregate graph is then merged with the original node representation to obtain...

[0031]

[0032] Where U is the weight matrix of the linear transformation, the above operation is repeated to update the embeddings of all nodes in the graph, and finally the updated classifier is obtained.

[0033]

[0034] The features and beneficial effects of this invention are:

[0035] This invention proposes a method for small-sample incremental RFFI (Radiation Source Identification) based on knowledge distillation and graph models. This method not only effectively solves the problem that RFFI systems in real-world scenarios must constantly process new categories, but also addresses the difficulty of simultaneously collecting enough data to train an ideal classification model. A series of experiments were conducted on the ADS-B dataset to verify the recognition performance of the proposed method under small-sample incremental conditions. Figure 4 The figure shows the confusion matrix effect on the ADS-B dataset. The first 60 classes are the base classes, and the last 40 classes are the new classes. Each session adds a small number of new class data points (5 per session). This demonstrates that the method of this invention can learn a small number of new class data points without forgetting the old classes. In other words, it simultaneously solves the problems of overfitting with small samples and catastrophic forgetting in incremental learning. Attached image description:

[0036] Figure 1 Overall framework diagram of a small-sample incremental radiation source individual identification method based on knowledge distillation and graph model.

[0037] Figure 2 Flowchart of small sample incremental learning recognition.

[0038] Figure 3 Schematic diagram of the graphical model module.

[0039] Figure 4 The confusion matrix effect on the ADS-B dataset. Detailed Implementation

[0040] This invention proposes a small-sample incremental radiation source individual identification method based on knowledge distillation and graph model, which uses knowledge distillation and graph model to complete the identification of radiation source individuals in complex scenarios.

[0041] In the field of radiation source identification research, RFFI (Radiation Source Identification) tasks are typically modeled as classification problems within a closed set, meaning that training and test samples share the same label space. Therefore, these methods are only suitable for ideal scenarios where training and test sample distributions are similar. However, real-world RFFI tasks need to be performed in open scenarios, and acquiring RFF data is a gradual process. RFFI systems must constantly process new categories, especially when applied to time-sensitive tasks such as battlefield reconnaissance. Furthermore, it is often difficult to collect enough data simultaneously to train an ideal classification model. Therefore, there is an urgent need to develop a method for radiation source identification in complex scenarios with small sample increments. This invention utilizes the ideas of knowledge distillation and graphical model classifiers to simultaneously address the overfitting problem with small samples and the catastrophic forgetting problem of incremental changes in RFFI tasks.

[0042] To address the catastrophic forgetting problem in the full training process of existing intelligent recognition models, this paper proposes an incremental recognition method based on knowledge distillation, updating the recognition system only with new class data. First, a well-trained backbone network is obtained through initial training. During incremental training, this well-trained backbone network is frozen as the teacher network and used to guide the updating of the student backbone network. To address the difficulty in obtaining individual radiation source sample data and the limited availability of effective samples in real-world incremental scenarios, a classifier decoupling learning strategy is adopted. Each increment only updates the classifier, avoiding knowledge forgetting during the incremental process of individual radiation source classes. Furthermore, to ensure that the classifier learned in a single session is applicable to all classes, a continuously evolving classifier is proposed. This classifier uses a graph model to propagate contextual information between classifiers for adaptation. Finally, to achieve the learning of this classifier, a pseudo-incremental learning paradigm is designed. Data is sampled from the base dataset to periodically construct a pseudo-incremental learning task to optimize the graph parameters. Figure 1 The diagram shows the overall framework of a small-sample incremental radiation source individual identification method based on knowledge distillation and graph models. First, a base class model is trained using base class data. This initial base class model is obtained by continuously optimizing the cross-entropy loss. Then, features are extracted from the backbone network of the initial base class model for pseudo-class incremental training. An adaptive graph model classifier is learned through pseudo-class incremental training. During the incremental training phase, on the one hand, the backbone network obtained in the previous stage is frozen as the teacher network, and it is copied as the student backbone network. On the other hand, the already trained adaptive graph model classifier adaptively connects to the context information based on the newly emerging small number of radiation source individuals. Finally, small-sample incremental radiation source individual identification is achieved. The specific technical solution adopted in this invention is as follows:

[0043] 1. Small sample incremental learning settings

[0044] Few-sample incremental learning aims to design a machine learning algorithm that can continuously learn from a small number of new categories, memorizing a small number of new samples after new category data appears, while retaining knowledge of old categories. In real-world scenarios, the number of new radar source signals is usually very small, unlike ordinary incremental learning which has a large number of radar signals and labels for incremental training. This necessitates research into how to perform incremental learning with a small sample size.

[0045] The few-shot incremental learning process consists of a base session and multiple incremental learning sessions. During each session, only the data from the current session is available, and the model needs to retain knowledge from older tasks while learning new tasks. For example... Figure 2 The diagram illustrates the format of the partitioned dataset and the setup for the few-shot incremental experiment. For a typical few-shot incremental task, the model's learning process can be divided into multiple recognition tasks, which appear in a specific order.

[0046] This invention assumes an m-step small sample increment task. Let... and This represents the training and testing data for the session {0,1,...,δ}. We will use... The test samples in the test evaluate the model learned in the base session to obtain the classification accuracy A0. When the model enters the first session with small sample increments, there are only N classes and K labeled samples available for learning, each class coming from... Training. When this session ends, the test samples are... The model was evaluated using a test sample that included all the classes seen. The resulting classification accuracy is denoted as A1. The above process will terminate after learning δ incremental sessions.

[0047] 2. Knowledge Distillation

[0048] Knowledge distillation of backbone networks plays a crucial role in addressing the catastrophic forgetting problem in few-shot incremental class learning. Incremental class learning refers to progressively introducing new categories into an existing model so that the model can continuously adapt to new tasks. However, traditional incremental class learning methods are prone to catastrophic forgetting, i.e., forgetting previously learned categories when learning new ones. Knowledge distillation of backbone networks can help mitigate the catastrophic forgetting problem in few-shot incremental class learning. The specific steps are as follows:

[0049] e) Initialization Phase: In the initial phase, a basic backbone network (ResNet 20) is trained to solve an initial task involving all classes. This model will serve as the teacher network in subsequent class incremental learning.

[0050] f) New Category Introduction: When a new category is introduced, a new student network is initialized, and the existing teacher network is used to predict the new category data, generating soft target labels. These soft target labels contain the probability distribution of the teacher network's predictions for the new category data.

[0051] g) Student Network Training: The student network is trained using soft target labels to approximate the teacher network's predictions. This helps the student network capture the teacher network's knowledge, including information on learned categories and new categories.

[0052] h) Iterative Processing: When a new category is introduced, the above steps can be executed iteratively multiple times to gradually update the student network. Each iteration transfers knowledge from the teacher network to the student network via soft target labels.

[0053] In this approach, the student network can retain knowledge of older categories while progressively learning new categories. This knowledge transfer helps reduce catastrophic forgetting because knowledge of older categories is transferred to the student network through the teacher network's soft target labels. Thus, even when learning new categories, knowledge of older categories is not completely lost, resulting in better class incremental learning performance.

[0054] 3. Graph Model Module

[0055] In few-shot incremental tasks, RFF samples possess group information, meaning that a classifier learned in a single incremental recognition task can only provide reliable recognition results for the current task. As the number of learning tasks increases, information gained from previous tasks is gradually lost, making it difficult to ensure accurate recognition of targets from all previous categories. In evaluations involving all previous categories, directly connected classifiers may fail to provide sufficient discriminative power and could lead to incorrect decisions. Therefore, to ensure that the classifier's learning incorporates global contextual information from all previous tasks, designing a continuous learning mechanism to prevent the loss of learned contextual information is crucial. Figure 3 As shown, this invention proposes an Adaptive Graph Attention Network (GAT) module that can update classifier weights based on global context information.

[0056] GAT possesses desirable properties that make it a suitable method for encoding contextual information. First, graph node updates are performed through an attention mechanism, ensuring that the order of context encoding remains class-invariant during incremental learning. Second, GAT models can scale to any number of classes, allowing classifier updates to share the same learned GAT across any session. Furthermore, because the nodes in a GAT model are fully connected, it shares a similar structure with the Transformer, which also utilizes self-attention to propagate information.

[0057] let Indicates in task T n The parameter matrix of the CNN classifier learned in the middle. Let W represent a matrix. n Each row vector in It is the weight corresponding to a specific category c, C n It is task T n The number of categories in the embedding vector, E is the dimension of the embedding vector. This can be viewed as the prototype vector of category c. The values ​​of different dimensions essentially indicate the distinguishability of different channels. To improve the distinguishability of the classifier, this invention can examine the values ​​of all other categories... To adjust The value of . Therefore, this invention first needs to collect the weight vectors of all other categories from the previous task:

[0058]

[0059] Where N is the total number of tasks to date. Then, the present invention will... N The weight vectors in the graph model are represented as nodes, and GAT (Graph Attention) is used to model the relationships between these prototype vectors and propagate contextual information. An attention mechanism is used to facilitate information propagation between nodes, enabling each node to receive contextual information from other nodes in the graph.

[0060] In RFFI tasks, fully connected graphs By node and edge in It is the j-th node, e jk It is a connection and Let I be the total number of nodes. For any two nodes... and Relationship coefficient e jk The calculation formula is

[0061]

[0062] Where λ and μ represent linear transformation functions used to project the original prototype representation onto a new metric space. The functions <·,·> calculate the cosine similarity between two vectors. For clarity, and The task index in the subscript is omitted. Then, the obtained coefficients are normalized using the softmax function to obtain the final attention weights corresponding to the center node j.

[0063]

[0064] In the normalized attention coefficient ajk Based on a jk The information of all nodes in the aggregate graph is then merged with the original node representation to obtain...

[0065]

[0066] Where U is the weight matrix of the linear transformation. Repeat the above operation to update the embeddings of all nodes in the graph, finally obtaining the updated classifier.

[0067]

[0068] To handle new tasks with few-sample class increments, an adaptive module is proposed. This module updates the classifiers learned from previous and current tasks and concatenates them to predict all classes. To facilitate knowledge propagation, a multi-head attention mechanism is employed. Furthermore, the network input is embedded in a graph to promote the learning of contextual knowledge.

[0069] 4. Pseudo-class incremental learning

[0070] Introducing context encoding into the adaptive module of a classifier is crucial, especially when applying Gaussian Adaptive Transformation (GAT) in the context of incremental learning. However, in few-shot incremental learning scenarios, training can only utilize data from a single session, and incremental session data is always limited. To address this challenge, this invention, inspired by meta-learning, conceives a virtual incremental learning training method. In this method, virtual incremental tasks are constructed to simulate incremental testing scenarios by randomly selecting samples from the basic training set. Based on the data in the basic dataset, this invention constructs a streamlined incremental learning task, where some sampled classes serve as the base class and others as the incremental class. Specifically, this invention divides the virtual incremental class and virtual base class into support sets and query sets, respectively. The support set is used to learn the classifier weights for different categories, while the query set is used to calculate the optimization loss. First, this invention uses the support set to build two classifiers for the virtual base class and virtual incremental class, respectively. Then, these two classifiers are concatenated and introduced into the adaptive GAT module for parameter updates. Using the updated classifiers, predictions are made on the query sets of the virtual base class and virtual incremental class, and the loss is calculated to optimize the adaptive GAT module. To synthesize new classes, this invention randomly rotates the sampled virtual incremental class data, causing significant changes in the inter-class angles. This rotation causes the synthesized image to lose its original class features, but it retains semantic similarity. The advantage of this strategy is that it enhances the robustness of the model. After the adaptive GAT module is trained, this invention fixes its parameters and applies them to new incremental sessions.

[0071] A specific example of the present invention is as follows.

[0072] First, train on the base class dataset. Pre-trained model M, randomly initialized GAT modules

[0073] The preferred embodiment of the present invention is:

[0074] 1. Pseudo-incremental stage

[0075] Step 1: From Support and query sets for pseudo-base classes are generated by random sampling during training.

[0076] Step 2: From Support and query sets for pseudo-incremental classes are generated by random sampling during training.

[0077] Step 3: Randomly rotate the samples of each class by an angle γ to obtain...

[0078] Step 4: Use on M Learn the FC (Fully Connected Layer) to obtain W0.

[0079] Step 5: Learn the FC layer on M using a pseudo-incremental support set. Get W i .

[0080] Step 6: Using M, {W′0,W′1} and Calculate the target loss and obtain Where N represents the number of base classes, y i p(x) represents the true label of the sample. i ) represents the probability output obtained after Softmax.

[0081] Step 7: Optimize using SGD (stochastic gradient descent) Obtain the trained GAT module

[0082] 2. Knowledge Distillation and Incremental Stage of Graph Model Classifier

[0083] Step 1: Load a small amount of new individual data

[0084] Step 2: Use the old backbone network as the teacher network and the new backbone network as the student network to perform knowledge distillation.

[0085] Step 3: As new classes are added, the fully connected layer changes, and the GAT module... And so it changes.

[0086] Step 4: Obtain the new backbone network and GAT module

[0087] Step 5: Learn a new network model to achieve individual identification of radiation sources with small sample increments.

[0088] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying individual radiation sources using small-sample incremental methods based on knowledge distillation and graph models, characterized in that... First, a backbone network is trained using known radiation source data, continuously optimizing the cross-entropy loss. Then, pseudo-class incremental training is performed: using the known radiation source data, support and query sets are created for the virtual incremental and virtual base classes, respectively. The support set is used to learn the classifier weights for different classes, while the query set is used to calculate the optimization loss. Two classifiers are built for the virtual base and virtual incremental classes using the support set, respectively. These two classifiers are then connected to the same adaptive graphical model classifier. The parameters of the adaptive graphical model classifier are updated through training. Using the updated adaptive graphical model classifier, predictions are made on the query sets of the virtual base and virtual incremental classes, and the loss is calculated to optimize the adaptive graphical model classifier. The optimized adaptive graphical model classifier is then integrated with the backbone network. The network is connected; then, it enters the incremental training phase. On one hand, the backbone network is frozen as the teacher network, and it is replicated and trained with a small number of newly added radiation source individuals to obtain the student network. The teacher network is used to perform knowledge distillation on the student network, and the student network that has been distilled is the trained backbone network. This student network is also replicated and frozen as the teacher network for the next incremental training, that is, incremental training is performed every time a small number of newly added radiation source individuals appear. On the other hand, the already trained adaptive graphical model classifier will adaptively connect the context information according to the newly added small number of radiation source individuals. Finally, the trained backbone network and its connected adaptive graphical model classifier are used to realize the identification of radiation source individuals in small sample incremental classes.

2. The method for identifying individual radiation sources based on knowledge distillation and graph models using small-sample incremental methods as described in claim 1, characterized in that, The specific steps of knowledge distillation are as follows: a) Initialization phase: In the initial phase, a basic backbone network is trained to solve the initial task, which involves all categories. This model will serve as the teacher network in subsequent class incremental learning. b) Introduction of new categories: The frozen teacher network is trained using the newly introduced categories to form a student network, and the existing teacher network is used to predict the new category data to generate soft target labels. These soft target labels contain the probability distribution of the teacher network's predictions for the new category data. c) Student network training: Student networks are trained using soft target labels to approximate the predictions of the teacher network; d) Iterative process: When a new category is introduced, the above steps are executed repeatedly to gradually update the student network. In each iteration, knowledge is transferred from the teacher network to the student network through soft target labels.

3. The method for identifying individual radiation sources based on knowledge distillation and graph models using small-sample incremental methods as described in claim 1, characterized in that... The adaptive graph model of the backbone network and its connections, with the following specific connection patterns: Indicates in task T n The parameter matrix of the CNN classifier learned in the middle, where W n Each row vector in It is the weight corresponding to a specific category c, C n It is task T n The number of categories in the vector, where E is the dimension of the embedding vector. It can be viewed as the prototype vector of category c, by looking at all other categories. To adjust To obtain the value, we first need to collect the weight vectors of all other categories from previous tasks: Where N is the total number of tasks to date, then W N In this model, weight vectors are represented as nodes in the graph model. An adaptive graph model classifier is used to model the relationships between these prototype vectors and propagate contextual information. An attention mechanism is used to facilitate information propagation between nodes, enabling each node to receive contextual information from other nodes in the graph. In the task of identifying individual radiation sources, fully connected graphs By node and edge j≠k. Where It is the j-th node, e jk It is a connection and Let I be the total number of nodes, and for any two nodes... and Relationship coefficient e jk The calculation formula is Where λ and μ represent linear transformation functions used to project the original prototype representation onto a new metric space, and the functions <·,·> calculate the cosine similarity between two vectors. For clarity... and The task index in the subscript is omitted. Then, the obtained coefficients are normalized using the softmax function to obtain the final attention weights corresponding to the center node j. In the normalized attention coefficient a jk Based on a jk The information of all nodes in the aggregate graph is then merged with the original node representation to obtain... Where U is the weight matrix of the linear transformation, the above operation is repeated to update the embeddings of all nodes in the graph, and finally the updated classifier is obtained.

Images