Physical layer user classification authentication method based on deep learning

Abstract

The invention provides a physical layer user classification authentication method based on deep learning, which combines a decision tree and a random forest algorithm to classify and identify channel characteristics under different environmental factors, and realizes preliminary authentication through classification confidence. The method comprises the following steps: collecting feature data of each user feature type of each fixed position user with labels as a data set, wherein the user feature types comprise channel distribution type features and channel features; the label is used for distinguishing whether the user is legal or the identity of each individual legal user; according to the method, large-scale and small-scale channel indexes, such as path loss, shadow effect and time delay extension, are utilized to carry out rapid primary selection on a multi-user scene of a long-term fixed position, and on this basis, an amplitude phase analysis method is adopted to complete final authentication. Simulation and actual measurement results show that the method can realize a high user identification rate in a simulation environment, and has the characteristics of high efficiency, robustness and light weight.

Description

Technical Field

[0001] This invention pertains to physical layer security and user authentication technologies, and particularly relates to a physical layer authentication technology applicable to multiple environments and multiple user scenarios. Background Technology

[0002] In application scenarios such as the Internet of Things (IoT) and Industrial Internet of Things (IIoT), unlike traditional single-user authentication scenarios, these scenarios typically involve a massive number of devices and nodes accessing different target channel environments. Examples include sensors and robots in industrial automation systems and monitoring nodes in smart cities. Due to the wide distribution of users, significant differences in channel distribution and environment, and diverse communication standards, current physical layer authentication models mostly assume that the system operates in a single target channel environment. They only consider the design and learning of authentication schemes or machine learning models under a single channel model, resulting in poor adaptability to different target channel environments. In multi-user authentication scenarios, the deployment cost is high, and it is difficult to perform rapid authentication for multiple users simultaneously, making it difficult to handle scenarios with multiple target channel user authentication.

[0003] For example, such as Figure 1 As shown, in a highly automated production plant, Alice is a monitoring device, and Bob is an automated robotic arm. Bob connects to Alice via a wireless network to complete intelligent automated collaborative production of wood processing. Real-time classification and authentication of multiple Bob identities is required. The automated robotic arms with different Bob identities are located in different positions with varying distribution types. How the classification and authentication method adapts to the multi-target channel environment based on deep learning is of great significance for promoting the development of physical layer security technology. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a solution that can quickly and accurately perform physical layer classification and authentication of multiple users at one time in a multi-target scenario, and enhance the resistance to spoofed users and relay attacks in the process.

[0005] The technical solution adopted by this invention to solve the above-mentioned technical problems is a physical layer user classification and authentication method based on deep learning, comprising the following steps:

[0006] The dataset is constructed by collecting feature data of each user feature type from users at fixed locations with labels. The user feature types include channel distribution type features and channel features. The labels are used to distinguish whether a user is legitimate or the identity of a legitimate individual user.

[0007] The CART decision tree classification mechanism is used, with the collected dataset as the root node. The Gini index is calculated for all user feature types. Then, the feature with the smallest Gini index is selected as the splitting feature of the current node. If the current node has child nodes, the dataset is divided into sub-datasets according to the splitting feature. The child nodes are recursively built for each sub-dataset until the leaf nodes are built. Finally, all leaf nodes are output as the preliminary user classification and authentication results.

[0008] By introducing random forest ensemble learning, multiple CART decision trees are constructed through bootstrap sampling and random feature subsets, and the final user classification and authentication results are output using majority voting, thereby improving the robustness and generalization ability of the system.

[0009] Furthermore, a confidence threshold is set, and for classification results below the confidence threshold in the final user classification authentication results, amplitude phase support vector machine (SVM) authentication is further performed.

[0010] User characteristic types include path loss, shadowing fading, delay spread, channel characteristics, and signal-to-noise ratio.

[0011] The beneficial effects of this invention are that, by combining decision trees and random forests, it achieves fast and accurate authentication in multi-environment, multi-user scenarios, and is suitable for multi-user authentication in long-term fixed node, multi-target channel environments such as the Internet of Things and industrial wireless networks. Compared with traditional methods, it has the following advantages:

[0012] High efficiency: Supports rapid initial selection by multiple users at long-term fixed locations;

[0013] High accuracy: Random forest ensemble reduces the risk of overfitting and improves the ability to detect camouflage and simulation attacks;

[0014] Low complexity: The authentication process is lightweight, making it suitable for resource-constrained IoT and industrial wireless networks. Attached Figure Description

[0015] Figure 1 User maps for different target channel environments;

[0016] Figure 2 This is a schematic diagram of a tree structure;

[0017] Figure 3 This is a diagram of the random forest structure.

[0018] Figure 4 This is a CART decision tree structure diagram;

[0019] Figure 5 Simulate user distribution diagram for multi-user communication;

[0020] Figure 6 To verify the success rate comparison;

[0021] Figure 7 Here are the confusion matrices for different methods: (a) Logistic Regression, (b) Nearest Neighbors, (c) Support Vector Machine, (d) CART Decision Tree, and (e) Random Forest. Detailed Implementation

[0024] A deep learning-based authentication method for multi-target channel environments, adaptable to various scenarios, not only effectively addresses the shortcomings of traditional methods but also significantly contributes to the development of physical layer security technologies. This invention proposes a lightweight channel classification scheme combining decision trees, deep learning, and multi-source fusion authentication. This scheme enables rapid simultaneous authentication of multiple target user channels, enhancing the generalization ability of physical layer authentication systems and playing a crucial role in promoting the application of physical layer security authentication in multi-target channel scenarios.

[0025] Considering the spatiotemporal context of user presence and channel distribution theory, when users remain in the same location for an extended period, decision trees can be used to classify user distribution for preliminary authentication. If each user is located in a different location, their distribution type will certainly be different, allowing decision trees to directly classify users and achieve direct authentication.

[0026] Considering that devices such as those in the Internet of Things (IoT) and industrial wireless networks are located in static environments, the construction of decision trees and the differentiation of users essentially boil down to feature selection, based on the fixed characteristics of their locations. This invention differentiates and authenticates users through channel distribution type features, achieving classification and identification of the same user and completing authentication for multiple users. Table 1 lists the channel distribution type features.

[0027] Table 1

[0028]

[0029] Channel distribution characteristics include large-scale features, small-scale features, and channel type. Large-scale features include path loss, shadowing fading, and environmental characteristics. Small-scale features include delay spread and the Doppler effect. Channel type is essentially a small-scale fading distribution, including Rayleigh, Rician, and Nakagami distributions.

[0030] Figure 2The system model is illustrated, where Alice communicates with Bob deployed in multiple target channel scenarios, and Eve represents a potential malicious spoofing user. Considering typical application scenarios such as unmanned factories and workshops, Alice acts as the communication center, while Bob1, 2, ..., n, etc., can be access users who are mechanical operation nodes, residing in the same location for extended periods. The target channel environment is quasi-static with minimal disturbance, thus providing a stable wireless channel environment. Bob and Alice have already established initial identity and access authorization through upper-layer keys. Now, Alice needs to accurately verify the continuous data packets sent by Bob1, 2, ..., n. Based on this data, a decision tree mechanism can be used to quickly perform initial user selection and classification by combining and classifying channel distribution types.

[0031] Based on the characteristics of channel distribution types, it is proposed to construct CART (Classification and Regression Trees) decision trees and random forest optimization schemes to achieve user classification and authentication through decision trees. The specific scheme is described as follows.

[0032] Figure 4 This describes the structure of the decision tree in the model. The CART decision tree is a classic binary tree structure where the root node represents the input dataset, containing collected data on the channel distribution types and features of each user. Internal nodes represent feature judgment conditions, and leaf nodes represent user categories. The precision of each category can be individual user data to authenticate each user and classify whether a user is legitimate. The presence or absence of feature A determines whether a random distribution contains a subclass distribution feature; if it does, a child node is formed.

[0033] CART decision trees select the optimal feature for splitting based on the Gini index. First, the Gini index for all features is calculated. For each feature category in dataset S, the Gini index G(S) is defined as:

[0034]

[0035] Where C is the number of feature categories, p i This represents the proportion of samples with the i-th feature in the dataset. This index measures the impurity of samples within a node; a smaller value indicates a purer node.

[0036] For each candidate feature A, the formula for calculating the split Gini gain ΔG(S,A) is as follows:

[0037]

[0038] Where V(A) represents all possible values ​​of candidate feature A, S vThis represents the subset of data when feature A takes the value v. |S v | represents a subset S of data v The number of data points in dataset S, |S| represents the number of data points in dataset S. The feature that minimizes the Gini index is selected as the splitting feature for the current node. The weighted Gini index G is then calculated after splitting feature A. A for:

[0039]

[0040] CART decision trees first calculate the Gini index for all channel distribution type features, and then select the feature G with the smallest Gini index. A This serves as the splitting characteristic of the current node.

[0041]

[0042] If the current node has child nodes, the dataset is divided into several sub-datasets according to the selected features, and child nodes are recursively constructed for each sub-dataset; otherwise, the current node is directly constructed as a leaf node.

[0043] The recursive termination condition is usually that all remaining features belong to the same distribution type or there are no more features that can be split. For example, splitting terminates when the node purity G(S) ≤ the preset purity ∈ or the number of samples ≤ the minimum sample value N_min. In this case, the leaf node is directly output as the final classification result. That is:

[0044] G(S)≤∈, or |S|≤N min (5)

[0045] When all branches meet the termination condition, the entire CART decision tree model is finally generated, which is used to automatically classify and identify different channel types.

[0046] The CART decision tree algorithm of this invention first calculates the Gini index for all channel distribution type features, and then recursively constructs leaf nodes if child nodes exist; otherwise, it continues to construct leaf nodes until the remaining features belong to the same distribution type and there are no child nodes, at which point the user classification result is output.

[0047] In practical wireless communication systems, different users are located in different physical environments. Their long-term fixed locations and the inherent characteristics of the channel can provide information for multi-user identification. Therefore, prior knowledge of users is used for decision tree training and user classification. Table 2 shows the user feature types used in the embodiments; more features can also be used to train the decision tree. In the preferred embodiment, the user feature types use both large-scale and small-scale features from the channel distribution type features, as well as channel features. Channel features include the channel itself and the signal-to-noise ratio. The channel itself uses the channel frequency response represented by a complex number.

[0048] Table 2

[0049]

[0050] The model design considers two approaches: decision tree-based and random forest-based. During training, labeled historical data is used to construct the decision tree. The model automatically learns the optimal splitting rules from the data to achieve accurate classification of new data.

[0051] Figure 2 The diagram illustrates the effect of decision tree classification, which performs more refined user classification and identification based on the number of features in the input dataset. After the data is input into the tree, the preferred feature is determined, and the smaller the Gini coefficient, the greater the reduction in Gini impurity after partitioning. In the example dataset, after inputting the data into the decision tree, as shown in Table 3, the path loss has the smallest Gini coefficient. In the dataset, the path loss of different users usually varies significantly, so the range of path loss is first determined. Then, in the subset after path loss partitioning, shadow fading, caused by environmental occlusion, reflects the differences in the micro-environment in which the user is located. At this point, the discriminative power of shadow fading exceeds that of other features. Further feature partitioning is then performed sequentially, constructing multi-layer decision nodes based on features such as shadow fading, delay spread, and signal-to-noise ratio. Each node selects the feature with the highest discriminative power for user identity partitioning based on the Gini coefficient. The Gini coefficients of user features are shown in Table 3.

[0052] Table 3

[0053]

[0054] Finally, tree growth is controlled through pre-pruning strategies, and model complexity is limited by hyperparameters: the maximum depth does not exceed 4 levels, and the minimum number of samples in a split node is no less than 50.

[0055] In practical systems, since each judgment requires traversing each layer of nodes sequentially, the computational complexity is high for large amounts of user data, leading to reduced recognition efficiency. To address this shortcoming, the Random Forest algorithm is introduced as an optimization method. A Random Forest consists of multiple decision trees, each of which randomly selects a subset of features and sample data for modeling during training.

[0056] In classification and recognition tasks in multi-target channel environments, a single decision tree may be affected by overfitting and has limitations when handling complex, high-dimensional, and multi-class data. To further improve the robustness and generalization ability of classification, the random forest ensemble learning method improves the accuracy and stability of classification by constructing multiple decision trees and uses a voting method for final decision-making during prediction, making it particularly suitable for user identity classification and authentication in multi-target channel environments.

[0057] Random forest construction begins with a training dataset S and a total feature set F. The total feature set F contains all user feature types. For each tree t, the random forest generates a subset S of samples from the dataset S through a statistical resampling bootstrap. t At the same time, a subset F is randomly selected from the total feature set. t (where |F) t |=m<|F|). This process can be represented as:

[0058]

[0059] For each tree t, the sample S obtained by random sampling is used. t and feature subset F t Construct decision trees. Each decision tree internally selects the best splitting feature based on indicators such as Gini index, information gain, or chi-square test, and recursively constructs nodes. The formula for selecting splitting features is consistent with that described above.

[0060] The weighted Gini index for candidate feature A can be expressed as:

[0061]

[0062] p i Let v represent the proportion of samples of the i-th class of candidate features v, where v represents the current value in candidate feature V(A). The final selection is made such that G... A Split based on the smallest feature:

[0063]

[0064] After the random forest is constructed, for a given input sample x, each tree h t (x) yields a predicted category y. In classification tasks, the final prediction result is determined through a majority voting mechanism, and its mathematical expression is:

[0065]

[0066] Where T is the total number of decision trees, and 1 / {·} is the indicator function, when h t The value is 1 when (x) = y, and 0 otherwise.

[0067] Furthermore, a key challenge in multi-target channel classification authentication is preventing malicious attackers from spoofing identities, such as in Man-in-the-Middle (MITM) attacks. An attacker might forge the channel characteristics of a legitimate user to pass traditional single-target authentication. Random forests can leverage the collective voting mechanism of multiple decision trees, combined with temporal channel characteristics, to improve the ability to detect spoofed channels. In MAC spoofing attacks, attackers might attempt to clone the MAC address of a legitimate user for access. Random forests, by combining multi-dimensional features such as channel state information, signal amplitude characteristics, and temporal characteristics, can effectively distinguish between legitimate and spoofed MAC addresses using channel features.

[0068] In wireless channel environments, a user's channel characteristics change over time and space, while users who remain in the same location for a long period exhibit relatively stable channel characteristic distributions. Random forests can be used to build long-term channel models for users, enabling simultaneous authentication in multi-target environments.

[0069] Figure 3 This diagram showcases the decision tree structure generated using the Random Forest method. The Random Forest directly verifies user identity through decision trees with different features. Each tree determines the user type based on the judgment result of different features. The results from all decision trees are then aggregated and output to verify the target's identity. Multiple decision trees can simultaneously recognize the input data, significantly improving recognition speed through parallel computing. Because the training process of each decision tree is random, the Random Forest effectively avoids overfitting that can occur with a single decision tree, improving the model's generalization ability. In the user identification stage, the Random Forest obtains the final classification result by aggregating the results from each decision tree.

[0070] For classification results below the confidence threshold in the end-user classification authentication results, amplitude-phase support vector machine (SVM) authentication is further performed.

[0071] To directly verify the efficiency of the subsequent model in classifying and recognizing users when multiple users are running simultaneously, a scenario of multiple users working concurrently was also simulated. User-related parameters were designed for the simulation, including path loss models, noise power, shadowing fading models, and delay spread models under the same environment.

[0072] The simulation process first sets system parameters, such as the number of users, carrier frequency, user and base station altitudes, transmit power, noise power, shadow fading standard deviation, and time correlation coefficient. Then, specific user parameters are configured, including the individual user's altitude, location, and latency. The base station location is set as the layout center point, with a position of (0,0,10).

[0073] The simulation steps for a multi-user channel dataset include:

[0074] Input channel prior information; after system parameter initialization, calculate the three-dimensional distance d from all users u to the base station; calculate the path loss PL according to different users; generate shadow fading S; generate delay spread; calculate the received power P and convert it to SNR; generate user channel H; store user data in all data, and output multi-user channel [H1,H2,…,Hn].

[0075] User parameter settings are shown in Table 4:

[0076] Table 4

[0077]

[0078] Subsequently, the path loss in the scenario was simulated using the COST231-Hata path loss model, as well as the time-dependent shadowing fading and exponentially distributed delay spread. Finally, the signal-to-noise ratio was calculated and the dataset was packaged.

[0079] Figure 5 This paper demonstrates the locations of users and base stations during simulated multi-user communication. Based on the different spaces and current states of different users, I have collected a dataset with multiple features for each user.

[0080] To compare with traditional methods, experimental results were conducted comparing classic classification and recognition schemes such as logistic regression, k-nearest neighbor algorithm (kNN), and support vector machine (SVM). Table 5 shows the classification success rate and its F1 score:

[0081] Table 5

[0082]

[0083]

[0084] Figure 6 Table 5 shows a comparison of the validation success rates of several schemes. The results are as follows: Logistic Regression achieved an accuracy of 82.94% and an F1 score of 82.61%. The Nearest Neighbor algorithm achieved an accuracy of 81.64% and an F1 score of 81.57%. Support Vector Machine achieved an accuracy of 98.18% and an F1 score of 98.18%. Decision Tree and Random Forest both achieved 100% validation. This is mainly because, under ideal conditions, there are some parameters that are difficult to obfuscate, such as the path loss value being different for each user.

[0085] Figure 7The confusion matrices of different algorithms are presented, reflecting the overall prediction performance of the models and the confusion between categories. It can be seen that the decision tree classification performs well overall. Through Matlab simulations and CSI tool tests, the recognition performance of this method in multi-user simultaneous operation and large-scale fading simulation scenarios is verified. Comparisons with traditional algorithms such as logistic regression, kNN, and SVM show that decision trees and random forests achieve 100% recognition rate in ideal scenarios.

Claims

1. A physical layer user classification and authentication method based on deep learning, characterized in that, Including the following steps: The dataset is constructed by collecting feature data of each user feature type from users at fixed locations with labels. The user feature types include channel distribution type features and channel features. Tags are used to distinguish whether a user is legitimate or the identity of a legitimate individual user; The CART decision tree classification mechanism is used, with the collected dataset as the root node. The Gini index is calculated for all user feature types. Then, the feature with the smallest Gini index is selected as the splitting feature of the current node. If the current node has child nodes, the dataset is divided into sub-datasets according to the splitting feature. The child nodes are recursively built for each sub-dataset until the leaf nodes are built. Finally, all leaf nodes are output as the preliminary user classification and authentication results. We introduce random forest ensemble learning, construct multiple CART decision trees through bootstrap sampling and random feature subsets, and output the final user classification and authentication results using majority voting.

2. The method as described in claim 1, characterized in that, A confidence threshold is set, and for classification results that are lower than the confidence threshold in the final user classification authentication results, amplitude phase support vector machine (SVM) authentication is further performed.

3. The method as described in claim 2, characterized in that, The amplitude-phase support vector machine (SVM) authentication uses the user channel signal preprocessed with wavelet threshold denoising. It extracts the normalized amplitude histogram kurtosis and instantaneous phase difference entropy as user features and inputs them into the support vector machine (SVM) based on the RBF kernel function for pattern recognition and decision-making.

4. The method as described in claim 1, characterized in that, The channel distribution type features in user feature types include large-scale features and small-scale features.

5. The method as described in claim 1, characterized in that, Large-scale features include path loss and shadow fading; small-scale features include delay spread; and channel features include signal-to-noise ratio.

6. The method as described in claim 1, characterized in that, During the construction of the CART decision tree, the splitting is terminated when the Gini index of the current node is less than or equal to the preset purity, or when the number of samples is less than or equal to the preset minimum sample value, and the current node becomes a leaf node.

7. The method as described in claim 1, characterized in that, During the training of the CART decision tree, a pre-pruning strategy is used to control tree growth. The maximum depth of the decision tree does not exceed 4 layers, and the minimum number of samples in the split node is not less than 50.

Images