CN-121985334-A - LoRa terminal identification method integrating double-frame carrier frequency offset characteristics

CN121985334ACN 121985334 ACN121985334 ACN 121985334ACN-121985334-A

Abstract

The invention relates to the technical field of Internet of things safety and wireless equipment identity recognition, and discloses a LoRa terminal recognition method integrating double-frame carrier frequency offset characteristics, which comprises the steps of firstly utilizing software radio equipment USRP to collect and perform frame detection and analysis on signals transmitted by the LoRa terminal; and finally, inputting the characteristics into an integrated learning classifier consisting of a support vector machine, a random forest, a linear discriminant analysis and a K nearest neighbor classifier, realizing the collaborative decision of a plurality of classifiers through the joint weighting of static weights and dynamic confidence, and completing the identity recognition of registered equipment and the rejection of unknown equipment by combining a double-threshold rejection mechanism. The invention can improve the accuracy, stability and robustness of the identity recognition of the LoRa equipment, reduce the risk of false recognition of unknown equipment and is suitable for the LoRa network security access scene under the condition of resource limitation.

Inventors

ZHAO ZICHEN
Zan Zhaopeng
LIU BOWEN
CAO RUI
CHEN ZICHAO
WU WENJIA

Assignees

东南大学

Dates

Publication Date: 20260505
Application Date: 20260408

Claims (9)

1. The method for identifying the LoRa terminal by fusing the frequency offset characteristics of the double-frame carrier is characterized by comprising the following steps: Step (1) signal acquisition, which acquires wireless signals transmitted by a LoRa terminal by using software radio equipment to obtain a complex baseband signal sequence, and realizes the determination of the initial position of a LoRa frame and the analysis of a signal frame by detecting the related of a local reference Chirp signal, Step (2) continuous double-frame construction, signal preprocessing and feature extraction, constructing double-frame samples from two frames of signals continuously transmitted by the same equipment, synchronizing, estimating and compensating carrier frequency offset and normalizing amplitude, extracting double-frame carrier frequency offset and average value and difference thereof to form feature vectors, And (3) integrating learning identification, inputting the feature vectors into a plurality of basic classifiers to carry out joint decision, carrying out weighted fusion based on the classifier accuracy and prediction entropy, and realizing equipment identity identification and unknown equipment rejection by combining a confidence threshold.
2. The method for identifying a LoRa terminal by fusing dual-frame carrier frequency offset characteristics according to claim 1, wherein the step (1) specifically comprises: (11) The wireless signal emitted by the LoRa terminal is collected by using the USRP of the software radio equipment, and the sampling rate is set The complex baseband signal sequence is obtained by: Wherein the method comprises the steps of As an in-phase component, As the orthogonal component(s), Representing imaginary units (12) And carrying out LoRa frame detection on the acquired continuous signals, and determining a frame starting position according to a correlation result of the received signals and the local reference Chirp signals, so as to analyze a complete LoRa signal frame.
3. The method for identifying a LoRa terminal by fusing dual-frame carrier frequency offset characteristics according to claim 1, wherein the step (2) specifically comprises: (21) After obtaining a single LoRa signal frame, combining two frames of signals continuously transmitted by the same equipment according to time sequence to form a continuous double-frame sample, and by constructing the continuous double-frame sample, keeping the homologous hardware characteristics of the equipment under the condition of approximately consistent short-time channel by double frames, providing a basis for the subsequent extraction of radio frequency fingerprint characteristics, (22) Respectively carrying out synchronization, carrier frequency offset estimation and compensation and amplitude normalization processing on continuous double-frame signals to obtain a standardized signal sequence after time alignment, frequency offset compensation and amplitude normalization so as to reduce the influence of receiving gain fluctuation and environmental change on subsequent feature extraction, (23) After the signal preprocessing is completed, carrier frequency offset estimation is carried out on the continuously received first frame signal and the continuously received second frame signal respectively, and corresponding frequency offset estimation values are obtained And And constructing a continuous double-frame frequency offset feature vector based on the frequency offset estimation value: Wherein: representing a carrier frequency offset estimate of the first frame signal; Representing a carrier frequency offset estimate of the second frame signal; the average value characteristic of the continuous double-frame frequency offset is used for inhibiting random noise interference; the frequency offset difference characteristic between the continuous double frames is represented and is used for describing the frequency drift characteristic of equipment in the short-time continuous transmission process; The continuous double-frame frequency deviation feature vector is used for representing local oscillation frequency deviation characteristics of the LoRa equipment and frequency stability difference under a continuous transmitting state, and is used as equipment radio frequency fingerprint features to be input into a subsequent classification recognition module.
4. The method for identifying a LoRa terminal by fusing dual-frame carrier frequency offset characteristics according to claim 1, wherein the step (3) specifically comprises: (31) Training basic classifier, namely taking the fused fingerprint feature vector F output in the step (23) as input, selecting 4 heterogeneous basic classifiers of a Support Vector Machine (SVM), a Random Forest (RF), linear Discriminant Analysis (LDA) and K Nearest Neighbor (KNN), independently training on a training set, enabling each classifier to output posterior probability vectors of equipment categories, (32) The combined estimation of the static weight and the dynamic confidence coefficient comprises the steps of determining the static base weight according to the identification accuracy of each base classifier on a verification set, calculating the dynamic confidence coefficient according to the entropy value of the current prediction output of each classifier, indicating that the prediction is more definite when the entropy value is lower, taking the normalized product of the static base weight and the dynamic confidence coefficient as the final combined weight, (33) Weighted probability fusion and equipment identity output, namely weighting and summing posterior probabilities output by 4 classifiers according to joint weights to obtain integrated posterior probability: wherein F is the feature vector of the sample currently to be identified, C is the candidate device class identification, the range of values is {1, 2,.. The number of devices in the training set is C }, C is the total number of devices in the training set, K is the number of basic classifiers (in this embodiment k=4, SVM, random forest, LDA and KNN respectively); for the joint weight of the kth classifier, satisfy =1; A posterior rate estimation is carried out on the samples F belonging to the category c for the kth classifier; is the posterior probability after integration. The integrated posterior probability synthesizes the judgment of all the basic classifiers on each class, and differential weighting is carried out according to the reliability, so that the influence of single classifier errors on the final result is effectively reduced, and the equipment class with the maximum integrated posterior probability is taken as the identification output: (34) The unknown equipment rejection mechanism based on the confidence coefficient comprises the steps of calculating the difference between the maximum probability value and the second highest probability value of the integrated posterior probability vector to form an inter-class confidence coefficient, judging that the unknown equipment is illegally accessed if the maximum integrated posterior probability is lower than a preset absolute threshold or the inter-class confidence coefficient is lower than a preset relative threshold, and outputting equipment identity if the maximum integrated posterior probability is not lower than the preset relative threshold.
5. The method for identifying a LoRa terminal with a double-frame carrier frequency offset feature as claimed in claim 4, wherein the training of the basic classifier in the step (31) specifically includes: (311) A Support Vector Machine (SVM), a nonlinear decision boundary is established by adopting a radial basis function, a decision function value is converted into posterior probability through Platt scaling, the method is suitable for high-dimensional small sample classification, (312) Random Forest (RF), formed by integrating a plurality of decision trees by Bagging, outputs posterior probability according to the proportion of ticket numbers of various categories, has good robustness to noise and characteristic fluctuation, (313) Linear Discriminant Analysis (LDA), constructing linear projection maximizing the ratio of inter-class divergence to intra-class divergence, classifying based on Gaussian posterior probability, low computational overhead and strong interpretability, (314) And the K nearest neighbor classifier (KNN) adopts a Markov distance measure to estimate posterior probability according to the class distribution of nearest neighbor samples, and effectively retains local feature distribution information. The four classifiers have different learning norms, so that the diversity of an integrated framework is ensured, and the probability of common errors of all the classifiers is effectively reduced.
6. The method for identifying a LoRa terminal with integrated dual-frame carrier frequency offset feature according to claim 4, wherein the estimating by combining static weights and dynamic confidence coefficients in step (32) specifically comprises: (321) Static basis weight calculation assuming that the device identification accuracy of the kth classifier on the verification set is Its static basis weight is defined as: the static basis weight reflects the comprehensive discrimination capability of each classifier in the training phase, Representing the accuracy of the identification of the jth classifier alone, Is the static base weight of the kth classifier, composed of Determining the ratio of the accuracy of the classifier to the sum of the accuracy of all the classifiers, and meeting the normalization condition The larger the value, the stronger the overall discrimination capability of the classifier on the verification set. (322) Calculating prediction entropy, namely outputting a posterior probability vector to a current input sample F by a kth classifier, and calculating the prediction entropy: Wherein M is the total number of registered device classes, predictive entropy Smaller values indicate more concentrated confidence in a classifier for a class, larger values indicate more fuzzy determinations, lower weights should be assigned, (323) Dynamic confidence factor calculation: Wherein the method comprises the steps of For the super-parameters, the sensitivity degree of the confidence factors to the entropy values is controlled, (324) Joint weight normalization: Joint weights Meanwhile, the historical accuracy (static robustness) of the classifier and the prediction certainty (dynamic confidence) of the current sample are considered, so that the weight distribution gives consideration to global and local information. Wherein, the For the static basis weight of the kth classifier (determined by validation set accuracy), To its dynamic confidence factor Is the information entropy of the posterior probability vector, 0 Is a temperature super parameter), K is the total number of the basic classifier, and denominator For the normalized term, j is the classifier index variable, which traverses all K basic classifiers; The comprehensive contribution degree of the jth classifier to the current sample is represented, and the historical discrimination capability (static state) of the jth classifier on the verification set is fused, and the certainty degree (dynamic state) of the jth classifier on the current sample prediction is fused.
7. The method for identifying a LoRa terminal with a double-frame carrier frequency offset feature as claimed in claim 4, wherein the unknown equipment rejection mechanism based on the confidence level in the step (34) specifically comprises: (341) In order to realize effective discrimination of unknown equipment and illegal access, the integrated posterior probability is subjected to double-threshold rejection test before the identity of the equipment is output, and inter-class confidence difference is defined: inter-class confidence difference The confidence advantage of the optimal category relative to the suboptimal category is measured, the distinguishing degree of the recognition result is reflected, wherein F is the continuous double-frame carrier frequency offset characteristic vector of the current sample, For the integrated posterior probability obtained after weighted fusion by equation (7), In order to be of the optimal class, For the maximum of the integrated posterior probability, Posterior probability for suboptimal class; the larger the value of the sample is, the more reliable the identification result is, and the smaller the sample is more likely to belong to unknown equipment or illegal access. (342) Refusing the judgment rule, namely judging the current sample as unknown equipment or illegal access if any one of the following conditions is met (i) the absolute confidence condition is not met, namely Less than a preset absolute threshold (Ii) the relative confidence condition is not satisfied, i.e Less than a preset relative threshold If both conditions are satisfied, outputting the equipment identity mark The dual-threshold mechanism jointly judges from two dimensions of absolute confidence and inter-class distinction, and compared with a single-threshold scheme, the dual-threshold mechanism has stronger rejection capability to unknown equipment and lower false recognition rate.
8. An electronic device comprising a memory, a processor and a computer program stored on the memory and operable on the processor, wherein the processor when executing the program implements a method for identifying a LoRa terminal incorporating dual-frame carrier frequency offset characteristics as claimed in any one of claims 1 to 7.
9. A computer readable storage medium having stored thereon computer instructions which, when executed by a processor, implement a method for identifying a LoRa terminal incorporating dual-frame carrier frequency offset features as claimed in any one of claims 1 to 7.

Description

LoRa terminal identification method integrating double-frame carrier frequency offset characteristics Technical Field The invention belongs to the technical field of Internet of things security, and particularly relates to a LoRa equipment identity recognition method based on radio frequency fingerprints (Radio Frequency Fingerprinting, RFF), which is particularly suitable for a low-power-consumption wide area network (LPWAN) scene with limited resources and time-varying channels. Background With the rapid expansion of the field of the Internet of things, the LoRa technology establishes the key position of the LoRa technology in a low-power-consumption wide area network by virtue of the remarkable characteristics of long transmission distance, low energy consumption, open source codes and the like, and is widely popularized in the fields of environment monitoring, intelligent metering, asset management tracking and the like. However, the large-scale application of the LoRa network also causes a security crisis which cannot be ignored, such as means of hidden channel invasion, beacon spoofing attack and the like, and causes great hidden danger to the operation security and stability of the Internet of things system. Software identifications like MAC addresses are very vulnerable to tampering, cloning or forging and cannot effectively support access standards of highly trusted devices, while encryption authentication mechanisms can enhance security protection, they often involve additional computational effort and memory resource consumption, which is very disadvantageous for pursuing low cost, low energy consumption and resource-constrained LoRa terminal devices. The radio frequency fingerprint identification technology utilizes the inherent physical layer difference of the wireless equipment hardware to carry out identity authentication, has the advantages of difficult counterfeiting, no need of modifying an upper layer protocol, low realization cost and the like, and therefore becomes an important research direction of LoRa equipment security authentication. The existing radio frequency fingerprint identification methods are mainly divided into two types: And the deep learning-based method is to automatically extract fingerprint features by using a convolutional neural network or a cyclic neural network. The method has the following defects of (1) sensitivity to channel time variation, remarkably reduced recognition accuracy in time-span acquisition, (2) poor model interpretability, difficulty in meeting the requirements of patent declaration on technical principle definition, and (3) large training calculation overhead, and is not suitable for frequently updated Internet of things scenes. And (3) performing equipment identification by utilizing physical layer characteristics such as carrier frequency offset, phase noise and the like based on the characteristic engineering method. The method has the advantages of simple realization and better interpretability. However, the existing method based on feature engineering still has the defects that (1) single frame features are easily affected by noise and instantaneous channel disturbance, stability is insufficient, (2) carrier frequency offset is easily drifted along with temperature and local oscillation state change, distinguishing capability is limited when the method is used independently, and (3) the existing distinguishing mode mainly adopts a simple classifier, and effective suppression on unknown equipment and complex scene false recognition is lacking. In feature engineering-based research, existing research has generally been divided into two categories, steady state features refer to those hardware fingerprints that remain relatively stable over a longer time scale, and semi-steady state features refer to those hardware features that are relatively stable over a short period of time, but that may change over a longer period of time. Based on this classification, differential adjacent double-frame radio frequency fingerprints are proposed for the prior art, and the method uses differential phase noise between adjacent double frames as a main fingerprint feature. Particularly, it should be pointed out that although the existing double-frame method represented by differential adjacent double-frame radio frequency fingerprints utilizes the hardware correlation of adjacent frames, the technical scheme mainly relies on single differential logic to extract fingerprint features, and the differential strategy cannot be adaptively adjusted according to the real-time channel quality and signal state in a complex interference environment, so that the robustness of the method in a complex internet of things scene is limited. Aiming at the problems, the invention provides the LoRa terminal identification method integrating the double-frame carrier frequency offset characteristics and the integrated learning, which overcomes the defect of insufficient robustness