Search

CN-121980357-A - IoT (internet traffic control) device RFF (radio frequency identification) physical enhancement recognition method for improving domain generalization capability

CN121980357ACN 121980357 ACN121980357 ACN 121980357ACN-121980357-A

Abstract

The invention discloses an RFF physical enhancement recognition method for an IoT device with improved domain generalization capability, which comprises the following steps of 1, signal data acquisition and data set construction, 2, source domain feature pre-training, 3, self-adaption during online test, 4, model performance test and evaluation, and 5, recognizing RFFs to be recognized through an optimal RFF recognition model of the IoT device. The invention has the advantages that the invention is verified on the mobile phone equipment of the IEEE 802.11 OFDM system, so that the average RFF identification performance is improved, and other RFF methods can be embedded to improve the upper limit of the performance, and the invention has the stabilization generalization capability and engineering practicability.

Inventors

  • WEN ZHENYU
  • LI YIYANG
  • LI TAOTAO
  • SU JIE
  • HONG ZHEN

Assignees

  • 浙江工业大学

Dates

Publication Date
20260505
Application Date
20260402

Claims (7)

  1. 1. An IoT device RFF physical enhancement identification method that improves domain generalization capability, comprising: step 1, adopting software radio equipment as a signal capturing tool to capture two paths of orthogonal components of I/Q of a signal, recording original I/Q data, and integrating the original I/Q data into a radio frequency signal data set; Step 2, combining channel invariant feature extraction with prototype learning, establishing a channel invariant radio frequency fingerprint and a source domain prototype, adopting ResNet series of model networks as feature extractors, obtaining a feature extractor with fixed parameters through prototype comparison learning pre-training, calculating prototype points of source domain categories, storing the prototype points in a prototype memory, and constructing an RFF (radio frequency) identification model of the IoT equipment consisting of the feature extractor and the prototype classifier; Step 3, aiming at the unlabeled test data stream, keeping the parameters of a feature extractor fixed, realizing domain self-adaption by only dynamically updating a prototype memory library, namely executing CFO and CPO physical enhancement on a test sample to generate an enhanced sample, wherein CFO refers to carrier frequency offset, CPO refers to carrier phase offset, extracting features of an original sample and the enhanced sample by using the fixed feature extractor, fusing, screening high-confidence samples based on information entropy, calculating a target domain prototype, updating the prototype memory library by using an index sliding average formula, and performing similarity matching classification by using updated prototype points to obtain an optimal model of an adaptive target domain; Step 4, selecting a proper radio frequency signal data set, carrying out a test aiming at the generalization capability of a model domain, and comprehensively evaluating the performance of the RFF recognition model of the IoT device; And 5, identifying the RFF to be identified through the optimal RFF identification model of the IoT device.
  2. 2. The IoT device RFF physical enhancement identification method that improves domain generalization capability according to claim 1, wherein the step 1 comprises: Step 101, realizing real-time capturing and caching of Wi-Fi signals by adopting a multithreading architecture through a software radio peripheral USRP, setting an acquisition frequency band to be 2.4GHz-2.5GHz, and respectively acquiring data of a plurality of different types of mobile phone equipment in five scenes of channels, temperatures, positions, non-line-of-sight and moving; Step 102, recording two paths of amplitude values of I/Q of the acquired radio frequency signals, capturing an effective Wi-Fi data packet from the I/Q data by a double sliding window power detection method, and integrating the effective Wi-Fi data packet into a data set to be stored as a pkt file.
  3. 3. The IoT device RFF physical enhancement identification method that improves domain generalization capability according to claim 1, wherein the step 2 comprises: Step 201, channel invariant radio frequency fingerprint extraction is performed based on an IEEE 802.11 OFDM system preamble signal, wherein the preamble signal comprises a short training symbol STS, a guard interval GI and a long training symbol LTS, channel influence is eliminated through fast Fourier transform and logarithmic operation, and a steady state radio frequency fingerprint symbol pair and a transient state radio frequency fingerprint symbol pair are extracted; and calculating to obtain a steady-state RFF symbol pair by using STS and LTS: , RFF Stable is a steady state RFF symbol pair, Is the 2 nd to 5 th STS, Is the 1 st LTS, and, Is the 5 th to the 9 th STS, Is the 2 nd LTS, F log (·) is a complex operation of the fast Fourier transform and the logarithmic operation; calculating by using STS, GI and LTS to obtain a transient RFF symbol pair: , RFF Transient is a transient RFF symbol pair, Is the 0 th to 4 th STS, Is the 9 th to 10 th STS, Is the guard interval GI of the time period, All LTS, concat (-) refers to the splicing operation of a plurality of signal segments; Step 202, integrating RFF Stable and RFF Transient into The classification task was performed with ResNet series of models as feature extractor F θ (·) and the source domain D s class prototype was calculated by l 2 normalized mean, with the formula: , w j is the source domain class prototype, Meaning that l 2 norms are normalized, I (·) represents an indication function, y i =j is to determine whether the label of the I-th sample is equal to the class j; Finally, the model prototype is adopted to compare the learning loss l contr , and the source domain prototype P S is stored in a memory bank after training.
  4. 4. The IoT device RFF physical enhancement identification method that improves domain generalization capability according to claim 1, wherein the step 3 comprises: Step 301, designing a physical enhancement strategy based on the core causes CFO and CPO of RFF offset, carrying out random frequency offset and phase offset enhancement on a test sample, fusing the characteristics of an original sample and an enhanced sample, calculating a target domain prototype point and guiding model self-adaption by the target domain prototype point; step 302, realizing physical information perception prototype integration, firstly, through original test data Enhancement of samples To jointly perceive the prototype point distribution and calculate the fusion characteristics Fusion probability : , F θ (. Cndot.) is the fusion characteristic of the sample, Probability of being a sample; then, according to the physical perception prototype, a pseudo label is allocated to each sample, and the high confidence samples are extracted by using an entropy compression method, wherein the calculation formula is as follows: , Refers to the entropy of the information of the ith test sample, y s refers to the total number of device classes, σ (·) is a softmax function, Is the probability value of the j-th category after fusion; finally, the first K high confidence samples are selected to calculate out the physical perception prototype points of the test set, and the calculation formula is as follows: , refers to the physical awareness prototype point of class j, The method is characterized in that l 2 norms are normalized, I (-) represents an indication function, and F θ (·) ens is a fused feature vector; Step 303, after the physical perception prototypes of the test set are available, moving prototypes P M in a prototype library to gradually adapt to the class distribution of test data, dynamically updating prototypes of a repository by an exponential moving average method, and adapting to the test domain distribution; and finally, calculating the classification probability by using the prototype library updated for the t time.
  5. 5. The IoT device RFF physical enhancement method for improving domain generalization capability according to claim 4, wherein in step 301, CFO is randomly enhanced by first obtaining a frequency offset range through coarse/fine estimation , The rough estimate of the frequency offset is referred to as, Finger fine estimating frequency offset, max representing maximum value, using a variable subject to uniform distribution Controlling the enhanced frequency offset alpha delta f to obtain a radio frequency signal after CFO enhancement, wherein the formula is as follows: , Refers to the signal after carrier frequency offset enhancement, f k (·) is the impairment function, x b (t) is the baseband signal, t represents time, e is a natural constant, τ refers to the sampling phase, Δw k is the original carrier frequency offset, Is the carrier phase offset, h (t) refers to the radio channel response function, U refers to uniform distribution; CPO random enhancement, namely controlling and enhancing phase bias beta delta phi by using variable beta obeying uniform distribution to obtain a radio frequency signal after CPO enhancement, wherein the formula is as follows: , in order for the CPO enhanced signal to be useful, 。
  6. 6. The method for improving the domain generalization capability of the IoT device RFF physical enhancement recognition method according to claim 4, wherein in step 303, after the physical perception prototypes of the test set are available, prototypes P M in the prototype library are moved, so that prototypes in the prototype library gradually adapt to the class distribution of the test data, and the prototypes in the repository are dynamically updated by an exponential sliding average method, so as to adapt to the test domain distribution, where the formula is: , 、 Respectively refer to the first Secondary, the first The next updated prototype library, gamma is the momentum coefficient, Is a new prototype of the current batch, and P S is an initial prototype library obtained by source domain pre-training.
  7. 7. The IoT device RFF physical enhancement identification method that improves domain generalization capability according to claim 1, wherein the step 4 comprises: Step 401, developing comprehensive performance evaluation on the constructed time-varying RFF dataset, uniformly adopting ResNet series models as a characteristic extractor baseline, comparing the recognition accuracy of a main stream RFF recognition algorithm, a plurality of signal characteristic representation methods and an advanced TTA method, verifying domain generalization capability of an RFF recognition model of an IoT device under the conditions of time-crossing, scene-crossing and device-crossing; In step 402, the calculation formula of the accuracy is as follows: , Wherein Acc refers to the ratio of the number of positive predictive devices to the total number of test devices, TP, TN, FP, FN representing true positive, true negative, false positive and false negative, respectively.

Description

IoT (internet traffic control) device RFF (radio frequency identification) physical enhancement recognition method for improving domain generalization capability Technical Field The invention relates to the field, in particular to an RFF physical enhancement identification method for an IoT device for improving domain generalization capability. Background With the rapid development of internet of things (IoT) technology, application scenes such as smart home and smart city are increasingly popular, and intelligent upgrading is promoted by connection of mass wireless devices, but the openness of wireless channels makes IoT devices face security threats such as unauthorized access, and device identity verification becomes a core requirement. Radio Frequency Fingerprinting (RFF) identification techniques utilize inherent hardware defects of a device radio frequency front end (e.g., IQ imbalance, clock jitter, etc.) as unique identity, providing a reliable physical layer authentication scheme for IoT devices. In recent years, deep Learning (DL) has been widely used for RFF recognition due to its strong feature extraction and classification capabilities, achieving higher accuracy. However, most DL-basedRFF methods in the prior art assume that RFF characteristics in a training and testing stage meet ideal conditions of stability, independence and uniform distribution (i.i.d.), and that the environment (position, shielding, etc.) and running state (application load, electric quantity, etc.) of IoT devices in a real scene change at any time, so that significant domain deviation occurs in the RFF characteristics, the generalization performance of the model is rapidly reduced, and the large-scale application of the model is severely limited. The current domain generalization capability promotion of IoT device RFF identification faces two major core challenges, and both existing schemes have obvious limitations. The dynamic time-varying characteristic of the wireless channel leads to random deviation of amplitude, frequency and phase characteristics of a received signal to destroy RFF stability, and the Device State Effect (DSE) is that fluctuation of the working state of the device (such as temperature change caused by mobile phone application load) can lead to deviation of RFF core characteristics (such as CFO and CPO) distribution, and the RFF characteristics of the same device can also change along with the state. Aiming at the problems, the prior art has the defects that the traditional method only focuses and relieves the channel effect, DSE is not considered, domain Generalization (DG) is difficult to cope with the unseen equipment state, domain Adaptation (DA) needs to carry label test data and is not suitable for an on-line scene, the self-adaptation (TTA) method in the main stream test is a computer vision design, RFF data characteristics are not adapted, and the performance is unstable or even reduced after application. The joint action of the channel effect and the Device State Effect (DSE) ensures that the RFF distribution deviation problem in a time-varying scene cannot be effectively solved, and the RFF technology falls to the ground. The DL-basedRFF model identification accuracy fluctuates, the generalization performance is deteriorated, the safety risks such as identity misjudgment and authentication failure are also caused, and the real-time reliable authentication requirement of the IoT device is difficult to meet. The prior art has obvious limitation, so that a novel domain generalization method which is adaptive to the physical characteristics of RFF and can simultaneously cope with two major effects is urgently needed, the technical bottleneck is broken through, and the large-scale application of the method in the field of IoT security authentication is promoted. Disclosure of Invention The invention aims to provide an IoT device RFF physical enhancement recognition method for improving domain generalization capability, so as to solve the problems in the background art. In order to achieve the above purpose, the present invention provides the following technical solutions: An IoT device RFF physical enhancement identification method that improves domain generalization capability, comprising: step 1, adopting software radio equipment as a signal capturing tool to capture two paths of orthogonal components of I/Q of a signal, recording original I/Q data, and integrating the original I/Q data into a radio frequency signal data set; Step 2, combining channel invariant feature extraction with prototype learning, establishing a channel invariant radio frequency fingerprint and a source domain prototype, adopting ResNet series of model networks as feature extractors, obtaining a feature extractor with fixed parameters through prototype comparison learning pre-training, calculating prototype points of source domain categories, storing the prototype points in a prototype memory, and constructing an RFF (r