Rapid correction method for radiation source individual identification model

Through the dual-teacher-student forgetful learning framework, using loss-maximizing noise injection and knowledge distillation technology, the performance degradation of deep learning models under data pollution is solved, and the rapid correction of individual recognition systems of radiation sources and data security improvement is achieved.

CN119003991APending Publication Date: 2024-11-22SOUTHEAST UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411202515.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-29
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

Deep learning models have problems of performance degradation in the face of data pollution, especially in radiation source individual identification systems that are vulnerable to attacks, resulting in security risks. Existing methods such as retraining models are expensive and impractical.

Method used

Using a dual teacher-student forgetting learning framework, the model is forced to forget poisoned data through a loss-maximizing noise injection mechanism, and transfer knowledge from capable and incompetent teacher models through knowledge distillation techniques, repairing and optimizing model performance.

Benefits of technology

Achieve high accuracy on the retained data set and extremely low accuracy on the forgotten data set, which significantly improves model update speed, adaptability and stability, provides an effective solution to data security problems, and avoids high computations for retraining cost.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119003991A_ABST
    Figure CN119003991A_ABST
Patent Text Reader

Abstract

The invention relates to a rapid correction method for an individual identification model of a radiation source. The method comprises the following steps: step 1, proposing a noise injection mechanism with maximum loss; step 2, carrying out multi-stage model repair and performance optimization; step 3, carrying out a knowledge distillation strategy of a double-teacher model; according to the scheme, by generating the noise interference original model with the maximum loss, the method can effectively forget the poisoned category, knowledge is transmitted from the energetic teacher model and the non-energetic teacher model through the knowledge distillation technology, and the performance of the models on the remaining data set is repaired and optimized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of machine forgetting learning and radiation source individual identification, and in particular to a model updating method based on a dual-teacher-student forgetting framework, which is used for quickly correcting a radiation source individual identification model when data pollution occurs. Background Art

[0002] Radiation source individual identification technology is a technology that uses radio signal fingerprints to identify different emission sources. Fingerprint features are usually generated by hardware defects in the transmitter components, and these features can be used to distinguish different devices. With the development of deep learning technology, the radiation source individual identification method based on deep learning has been able to automatically extract fingerprint features from the input data and identify different emitters, showing higher performance and robustness compared to traditional methods.

[0003] However, deep learning models are vulnerable to data security issues. For example, in the Automatic Dependent Surveillance-Broadcast (ADS-B) system, aircraft identification information can be tampered with, leading to potential security risks. Similarly, in the Automatic Identification System (AIS), the information carried by the signal can also be corrupted by hostile behavior. Deep learning methods often fail to fully consider system security and are therefore vulnerable to carefully designed attack techniques such as adversarial attacks, backdoor attacks, and data obfuscation.

[0004] These security threats pose a potential challenge to the radiation source individual identification system. An attacker may use triggers to perform backdoor attacks, causing the model to be trained on poisoned samples, which will seriously affect the model performance. Although it is effective to retrain the model on a clean dataset that has been cleared of poisoned data, this approach is often impractical in practice due to the high computational cost. Summary of the invention

[0005] In order to solve the above problems, this paper proposes a dual teacher-student forgetting learning framework, which aims to solve the problem of model performance degradation caused by data poisoning in the task of radiation source individual identification. By generating noise interference with the original model with maximized loss, this method can effectively forget the poisoned category, and transfer knowledge from capable and incompetent teacher models through knowledge distillation technology to repair and optimize the performance of the model on the remaining data set. Experimental results show that this method not only achieves high accuracy on the retention set, but also achieves extremely low accuracy on the forgetting set, and significantly improves the model update speed. It shows good adaptability and stability on different data sets and network structures, providing an effective solution to the data security problem in the task of radiation source individual identification.

[0006] The technical solution of the present invention is as follows: A method for quickly correcting a radiation source individual identification model comprises the following steps:

[0007] Step 1: Propose a noise injection mechanism that maximizes the loss. A noise injection method for the poisoned category is proposed. By optimizing the problem, the noise that maximizes the classification loss is found, forcing the model to forget the poisoned data. Generate a noise matrix for the forget set. The original model uses the noise matrix to ignore C f The information corresponding to the category is forgotten. Taking into account the characteristics of the individual radiation sources, the learned noise is adaptive to the input signal with various modulation modes and other parameters. Specifically, the normal distribution is used Initialize the noise matrix,

[0008] The details are as follows:

[0009] Step 1-1 Data preparation: Collect and prepare a complete training dataset D c , containing N samples and M categories, identifying and separating the poisoned sample set D f and its corresponding category C f , the total number of forgotten categories N f , retain the sample set D that has not been poisoned r ,

[0010] Step 1-2 Noise matrix initialization: For the poisoned category C f , initialize the noise matrix w noise , each element is drawn from a standard normal distribution Extract from

[0011] Steps 1-3 Loss maximization noise generation:

[0012] Use the original model to analyze the poisoned sample set D f For training, the noise matrix is ​​generated by the following optimization problem:

[0013]

[0014] Among them, p i (y) is the true label probability of the i-th forgotten category, q i (u i ) is the output probability of the original model for the i-th forgotten category, w noise is the noise parameter, and λ is the parameter used to balance the optimization problem. The goal is to find the original model in C f L2-norm bounded noise that maximizes the upper classification loss.

[0015] At present, deep neural network models are designed based on minimizing loss functions. However, the present invention adopts maximizing loss functions in this step and establishes a usable forgetting noise matrix module based on loss maximization noise generation. Obviously, maximizing a loss function that is only for the poisoned category C f The noise matrix helps to forget the relevant specified knowledge in the next step, and the generated noise can be regarded as an anti-sample that destroys the previously learned knowledge, which is used to eliminate the poisoned sample set D in the next step. f information.

[0016] Step 2: Multi-stage model repair and performance optimization, as follows:

[0017] After noise injection, the performance of the model on the holdout set is gradually restored through a multi-stage model repair process while keeping the poisoned data forgotten.

[0018] Add the obtained noise to a randomly selected subset of the holdout set In the above equation, we form a perturbation set D n .

[0019] In the perturbation set D n Train the model for one epoch to damage the model on the poisoned set D f Recognition ability:

[0020]

[0021] In the subset of the holdout set Perform multiple repair training on to restore model performance:

[0022]

[0023] in, represents the original model, is the model after noise damage; M u (x;θ u ) is the model after noise damage and multiple repairs. It is observed that even a single epoch of damage is sufficient to destroy the model with respect to the retention set D r The weight of .

[0024] In this step, based on the noise with maximized loss, the present invention proposes a multi-stage model repair and performance optimization, the purpose of which is to induce the model to be poisoned by category C f The relevant model parameters are gradually restored to the performance of the model on the retention set Dr. It is worth noting that under normal circumstances, this process usually only needs to be performed in a small number of epochs and the relevant operations are performed on a small number of samples. Therefore, compared with retraining on the retention set Dr with a larger number of samples, this step greatly shortens the time to forget relevant samples.

[0025] Step 3: The knowledge distillation strategy of the dual teacher model is as follows:

[0026] After performing multiple repairs, the model is in the retention set D r The performance on the above is still weak. This paper proposes a knowledge distillation method based on teacher-student interaction to solve this problem.

[0027] Two teacher models are introduced: an incompetent teacher model that has been processed with noise injection to forget the poisoned data; and a competent teacher model that contains the knowledge of the complete dataset. Through this strategy, the student model is able to achieve a balance between forgetting and learning.

[0028] Knowledge distillation method:

[0029] Defining the Incompetent Teacher Model I (x;θ u ) and the capable teacher model T c (x;φ t ), and the student model fs(x;φt).

[0030] The knowledge transfer between the incompetent teacher and the student model is calculated by the Kullback-Leibler (KL) divergence:

[0031]

[0032] Where: τ is the distillation temperature; and They represent the output probability distribution of the incompetent teacher and student models for category i. I (x) and g s (x) are the classifiers for incompetent teachers and students, respectively.

[0033] The KL divergence calculation enables knowledge transfer between the teacher and student models:

[0034]

[0035] In the formula: represents the output probability distribution of a capable teacher for category i. c (x) is a classifier with a capable teacher.

[0036] Step 4: Optimize the objective function, as follows:

[0037] The knowledge transfer between the competent teacher and the student model is calculated by KL divergence, and the student model is optimized to minimize the KL divergence with the competent teacher and maximize the KL divergence with the incompetent teacher. The student model is trained to obtain the retention set D by optimizing the following objective function raccurate knowledge of the poisoned set D f Wrong knowledge:

[0038]

[0039] Among them, λ is a hyperparameter that balances the contribution of the two teacher models to the knowledge distillation process.

[0040] The optimization goal is to obtain the KL divergence between the student and the competent teacher on the reserved set D r Similarly, we minimize the KL divergence between the student and the incompetent teacher to obtain the exact knowledge about the poisoned set D f erroneous knowledge. Student model f s (x;φ t ) has information about the poisoned set D f Expertise but incompetent teachers I However, it has experienced the damage process of maximizing noise, which will seriously damage the information about the poisoned set D f Due to the influence of the balance of the two types of teacher hyperparameters, the student model is in the reserved set D r At the same time, the information about the poisoned set D was discarded f Therefore, the present invention innovatively applies the fast correction model to the task of individual radiation source identification, providing an efficient solution for forgetting poisoned data in the field of big data, avoiding the huge time overhead caused by model retraining, and quickly forgetting the poisoned information while retaining the original representation ability on the remaining data, that is, achieving balanced performance of the model on the poisoned set and the retained set.

[0041] Compared with the prior art, the advantages of the present invention are as follows: By implementing the above technical solution, the present invention can not only effectively delete the poisoned data from the radiation source identification model, reducing the negative impact of the poisoned data on the model performance, but also quickly restore the performance of the model on the retention set through the knowledge distillation technology, avoiding the high computational cost required to retrain the model. The test results of the Pulse-signal-IQ data set in Table 1 and the Oracle data set in Table 2 show that the present invention has achieved excellent performance in both ResNet18 and CNN networks. Compared with other baseline models, the present invention can achieve balanced performance on the forgetting set and the retention set regardless of single-category forgetting or multi-category forgetting. At the same time, Figure 1 The comparison of the acceleration ratios of each model shows that the present invention has the highest operating efficiency under different data sets. Therefore, the method of the present invention has verified its effectiveness on different data sets and network structures, proving the adaptability, stability and excellence of the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is the running time efficiency test result of the present invention;

[0043] Figure 2 It is a schematic diagram of a flow chart of the present invention. DETAILED DESCRIPTION

[0044] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0045] Embodiment 1:

[0046] This method can be used in the fields of radiation source individual identification, forgetting learning, etc.

[0047] Example 1 describes a case in which the method is applied to radiation source individual identification, a method for quickly correcting a radiation source individual identification model, comprising the following steps:

[0048] Step 1: Propose a noise injection mechanism that maximizes the loss. A noise injection method for the poisoned category is proposed. By optimizing the problem, the noise that maximizes the classification loss is found, forcing the model to forget the poisoned data. Generate a noise matrix for the forget set. The original model uses the noise matrix to ignore C f The information corresponding to the category is forgotten. Taking into account the characteristics of the individual radiation sources, the learned noise is adaptive to the input signal with various modulation modes and other parameters. Specifically, the normal distribution is used Initialize the noise matrix.

[0049] The details are as follows:

[0050] Step 1-1 Data preparation: Collect and prepare a complete training dataset D c , containing N samples and M categories, identifying and separating the poisoned sample set D f and its corresponding category C f , the total number of forgotten categories N f , retain the sample set D that has not been poisoned r ,

[0051] Step 1-2 Noise matrix initialization: For the poisoned category C f , initialize the noise matrix N, each element is from the standard normal distribution Extract from

[0052] Steps 1-3 Loss maximization noise generation:

[0053] Use the original model to analyze the poisoned sample set D f For training, the noise matrix is ​​generated by the following optimization problem:

[0054]

[0055] Among them, p i (y) is the true label probability of the i-th forgotten category, q i (u i ) is the output probability of the original model for the i-th forgotten category, w noise is the noise parameter, and λ is the parameter used to balance the optimization problem. The goal is to find the original model in C f L2-norm bounded noise that maximizes the upper classification loss.

[0056] Step 2: Multi-repair process:

[0057] Select a subset D from the retained set rsub , add the noise matrix N to this subset to form the perturbation set D n In the perturbation set D n Train the model for one epoch, and damage the model to the poisoned set D f recognition ability.

[0058] In the subset D of the retained set rsub Perform multiple training iterations on the same model to fix the model weights:

[0059]

[0060] Step 3: Knowledge Distillation:

[0061] Defining the Incompetent Teacher Model I (x;θ u ) and the capable teacher model T c (x;φ t ), and the student model fs(x;φt).

[0062] Calculate the KL divergence between the incompetent teacher and the student model:

[0063]

[0064] Calculate the KL divergence between the teacher and student models:

[0065]

[0066] Step 4: Optimize the objective function:

[0067] Train the student model to obtain the retention set D by optimizing the following objective function r accurate knowledge of the poisoned set D f Wrong knowledge:

[0068]

[0069] Experimental verification

[0070] Experiments are conducted on the self-built dataset Pulse-Signal-IQ and the public ORACLE dataset to verify the effectiveness of the method.

[0071] The dataset description is as follows:

[0072] Pulse-Signal-IQ: The dataset is constructed based on GNU Radio software and HackRF One hardware. In order to construct a dataset containing multiple different transmitter targets, the experiment uses seven identical HackRF One software defined radio modules as transmitters and a USRP B210 as a receiver for transmission and reception experiments. Due to the continuous development of radar systems, the operating parameters and modulation methods of radars in real electronic countermeasure environments are also emerging in an endless stream. This dataset considers nine major radar pulse modulation methods: unmodulated pulse CW, linear frequency modulation pulse LFM, nonlinear frequency modulation pulse NLFM, two-phase coded modulation pulse BPSK, four-phase coded modulation pulse QPSK, two-frequency modulation pulse BFSK, four-frequency modulation pulse QFSK, two-frequency two-phase coded modulation pulse BFSK_BPSK, two-frequency four-phase coded modulation pulse BFSK_QPSK. And at the same time, it meets the sample requirements of different modulation parameters of the same system. In the generated dataset, each transmitter sends the same number of pulse signals of different types of modulation methods, and each signal contains 1000 groups of pulses. The pulse width of the radar signal is 100us, the carrier frequency band covers 60MHz to 2GHz, and the maximum bandwidth of the radar signal is 1GHz. The baseband signal after the RF front-end processing at the receiving end is sampled at a sampling frequency of 16MHz to obtain sample data of 1600 sampling points for each pulse. The signal-to-noise ratio of the signal is 15dB, and the data set contains a total of 49,000 radar pulse signal samples.

[0073] ORACLE: This public dataset uses IQ samples collected from an experimental setup of USRP SDRs, with 1 USRP B210 as a receiver and 16 USRP X310 as transmitters. Frames compliant with the IEEE802.11a standard are generated through the MATLAB WLAN System Toolbox. The generated data frames contain random payloads but have the same address field, and are then streamed to the selected SDR for wireless transmission. The receiver samples the incoming 2.45GHz center frequency simulated WiFi signal at a sampling rate of 5MS / s. The collected complex IQ samples are divided into subsequences. This experiment further processes the dataset into IQ complex signals with a sample length of 2048, with a total sample size of 52,000.

[0074] Experiment with different deep learning models such as ResNet and CNN and evaluate the performance of the models.

[0075] After the baseline model training is completed, poisoning attacks are simulated and predefined poisoning samples are introduced into the training dataset.

[0076] According to the aforementioned implementation measures, the loss maximization noise injection, multiple restoration processes and knowledge distillation of the present invention are implemented on the model.

[0077] The model processed by the algorithm of the present invention is tested to evaluate its recognition accuracy on the poisoned category and the normal category.

[0078] The performance of the model processed by the algorithm of the present invention is compared with the baseline model, the model with only noise injection, and the retrained model.

[0079] The computation time during the implementation of the present invention was recorded and compared with the time required to retrain the model.

[0080] Performance Evaluation

[0081] This section will focus on the following key indicators:

[0082] Accuracy on the Retain Set (ADr): Evaluates the model’s performance on the normal category, which should be close to the baseline model or the retrained model.

[0083] Accuracy on the Forget Set (ADf): Evaluates the model’s performance on the poisoned classes. It should be significantly lower than the baseline model, indicating effective forgetting.

[0084] Run-time Efficiency: Evaluate the time required to implement the framework of the present invention and compare it with the time required to retrain the model.

[0085] Membership Inference Attack (MIA): MIA is used to evaluate whether the model retains the memory of the poisoned samples. A successful MIA should make it difficult to identify the poisoned samples as "members".

[0086] Through these detailed experimental verification and performance evaluation steps, the algorithm framework of the present invention can comprehensively demonstrate its advantages and effectiveness in data poisoning defense in different dimensions.

[0087] Through the above-mentioned implementation measures, the present invention can effectively delete the poisoned data from the radiation source individual identification model while maintaining the memory and recognition performance of normal data, avoiding the high computational cost required for retraining the model.

[0088] The test results are shown in Table 1, Table 2, Table 3, Figure 1 As shown:

[0089] Table 1 Pulse-signal-IQ dataset test results

[0090]

[0091] Table 2 Oracle data set test results

[0092]

[0093] Table 3 MIA member reasoning test results

[0094]

[0095]

[0096] Through these detailed experimental verification and performance evaluation steps, the algorithm framework of the present invention can comprehensively demonstrate its advantages and effectiveness in data poisoning defense in different dimensions.

[0097] The present invention may also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A method for quickly correcting a radiation source individual identification model, characterized in that: The method comprises the following steps: Step 1: Propose a noise injection mechanism that maximizes the loss. Step 2: Multi-stage model repair and performance optimization, Step 3: Knowledge distillation strategy of dual teacher model, Step 4: Optimize the objective function.

2. The method for rapidly correcting the radiation source individual identification model according to claim 1, characterized in that: Step 1 proposes a noise injection mechanism that maximizes the loss, as follows: Step 1-1 Data preparation: Collect and prepare a complete training dataset D c , containing N samples and M categories, identifying and separating the poisoned sample set D f and its corresponding poisoned category C f , the total number of forgotten categories N f , retain the sample set D that has not been poisoned r , Step 1-2 Noise matrix initialization: For the poisoned category C f , initialize the noise matrix w noise , each element is drawn from a standard normal distribution Extract from Steps 1-3 Loss maximization noise generation: Using the original model to analyze the poisoned sample set D f For training, the noise matrix is ​​generated by the following optimization problem: Among them, p i (y) is the true label probability of the i-th forgotten category, q i (u i ) is the output probability of the original model for the i-th forgotten category, w noise is the noise parameter, λ is the parameter used for balancing the optimization problem, the purpose is to find the original model in C f L2-norm bounded noise that maximizes the upper classification loss.

3. The method for rapidly correcting the radiation source individual identification model according to claim 2, characterized in that: The step 2: multi-stage model repair and performance optimization is as follows: After the noise injection, the model performance on the retention set is gradually restored through a multi-stage model repair process, while keeping the poisoned data in a forgotten state. Add the obtained noise to a randomly selected subset of the holdout set In the above equation, we form a perturbation set D n , in the perturbation set D n Train the model for one epoch to damage the model on the poisoned set D f Recognition ability: In the subset of the holdout set Perform multiple repair training on to restore model performance: in, represents the original model, is the model after noise damage; M u (x;θ u ) is the model after noise damage and multiple restorations.

4. The method for rapidly correcting the radiation source individual identification model according to claim 3, characterized in that: Step 3: The knowledge distillation strategy of the dual teacher model is as follows: Two teacher models are introduced: an incompetent teacher model that has been processed with noise injection to forget the poisoned data; and a competent teacher model that contains the knowledge of the complete dataset. Through this strategy, the student model can achieve a balance between forgetting and learning. Knowledge distillation method: Defining the Incompetent Teacher Model I (x;θ u ) and the capable teacher model T c (x;φ t ), and the student model fs(x;φt), the knowledge transfer between the incompetent teacher and the student model is calculated by the Kullback-Leibler (KL) divergence: Where: τ is the distillation temperature; and denote the output probability distribution of the incompetent teacher and student models for category i, respectively, and g I (x) and g s (x) are classifiers for incompetent teachers and students, respectively. The KL divergence calculation enables knowledge transfer between the teacher and student models: In the formula: represents the output probability distribution of a capable teacher for category i, g c (x) is a classifier with a capable teacher.

5. The method for rapidly correcting the radiation source individual identification model according to claim 2, characterized in that: Step 4: Optimize the objective function, as follows: The knowledge transfer between the competent teacher and the student model is calculated by KL divergence, and the student model is optimized to minimize the KL divergence with the competent teacher and maximize the KL divergence with the incompetent teacher. The student model is trained to obtain the knowledge about the reserved set D by optimizing the following objective function: r accurate knowledge of the poisoned set D f Wrong knowledge: Among them, λ is a hyperparameter that balances the contribution of the two teacher models to the knowledge distillation process.