Lightweight power terminal authentication method and system fusing ris and meta-learning

Abstract

The application provides a lightweight power terminal authentication method and system fusing RIS and meta-learning. The method comprises the following steps: according to the identity of the power terminal, a reconfigurable intelligent surface (RIS) is controlled to adopt a reflection configuration corresponding to the power terminal, and a channel response of the power terminal after reflection of the RIS is received, the channel response is taken as an enhanced spatial channel fingerprint of the power terminal; a pre-trained lightweight feature extraction network is used to perform space-time feature extraction on the enhanced spatial channel fingerprint, and a corresponding low-dimensional space-time feature vector is obtained; the low-dimensional space-time feature vector is input into a lightweight authenticator pre-trained based on a meta-learning framework, and an authentication result of the power terminal is output. The authentication scheme provided by the application effectively balances the relationship among authentication security, small sample adaptation efficiency and computational overhead, and provides a technical path with advanced theory and engineering feasibility for lightweight and secure access of a single-antenna terminal in a new power communication network.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to a lightweight power terminal authentication method and system that integrates RIS and meta-learning. Background Technology

[0002] As new power systems rapidly evolve towards digitalization and intelligence, power communication networks are bearing an unprecedented demand for massive terminal access, and network boundaries are continuously expanding. Hundreds of millions of smart meters, distributed sensors, and monitoring devices constitute the nerve endings of the ubiquitous power Internet of Things. However, the vast majority of these devices are constrained by strict cost and power consumption limitations, and generally adopt a single-antenna design with highly limited resources. This hardware reality makes it difficult to directly deploy traditional access authentication protocols based on asymmetric cryptography on resource-constrained power terminals due to computational complexity and high energy consumption.

[0003] Physical layer authentication (PLE) technology has emerged as a promising lightweight alternative. Its core idea is to leverage the inherent spatial uniqueness and time-varying nature of wireless channels to transform physical characteristics such as channel state information into natural device fingerprints, thereby establishing identity authentication at the communication link. It boasts inherent advantages of low overhead and low latency. However, current mainstream PLE schemes typically assume that the terminal is equipped with a multi-antenna array to acquire rich spatial domain features for high-precision differentiation, making them unsuitable for single-antenna terminal scenarios. In a single-antenna configuration, the dimensionality of extractable PLE features collapses significantly, resulting in a severe lack of distinguishability between legitimate terminals and potential attackers, limiting authentication security strength and making it difficult to effectively resist identity spoofing threats such as Sybil attacks. Furthermore, the power communication environment is characterized by typical heterogeneity and high dynamism, with diverse terminal types, complex deployment scenarios, and drastic channel time-varying characteristics. This presents deep learning-based authentication models with the dual challenges of high costs for acquiring labeled data and weak model generalization capabilities.

[0004] In recent years, the rise of Reconfigurable Intelligent Surface (RIS) technology has provided a new physical layer foundation for overcoming the aforementioned bottlenecks. RIS consists of a large number of low-cost, software-tunable passive reflective units. It can actively modulate the amplitude and phase of incident electromagnetic waves by adjusting the phase shift of each unit in real time, generating customizable and difficult-to-clone dynamic fingerprints for single-antenna terminals. This promises to fundamentally solve the inherent problem of insufficient feature resolution in single-antenna scenarios. However, the physical features enhanced by RIS alone are insufficient to constitute a complete solution. The frequent access and dynamically changing environment of massive terminals in the power Internet of Things (IoT) require authentication systems to have the ability to learn rapidly and adapt robustly from a very small number of samples. Meta-learning, as an advanced paradigm of learning by doing, aims to enable models to acquire meta-knowledge that can quickly adapt to new tasks through cross-task training, providing a powerful algorithmic framework for addressing the challenges of small-sample learning and rapid adaptation. Simultaneously, to meet the stringent requirements of power services for real-time authentication and limited terminal resources, the models used for feature extraction and decision-making must be lightweight to ensure efficient execution at the edge. Further analysis of the existing technology is as follows: Reconfigurable smart surfaces (RSS), as two-dimensional electromagnetic metasurfaces composed of numerous low-cost, passive programmable reflective units, can actively reshape the wireless propagation environment by dynamically controlling the phase and amplitude of incident electromagnetic waves, providing a novel fingerprint generation mechanism for physical layer authentication. Early research on RIS-assisted physical layer authentication primarily focused on how to utilize its programmability to address the problem of insufficient channel entropy. Hu et al. proposed a physical layer key generation scheme based on discrete phase-shifted RIS, constructing a dynamic time-varying channel by rapidly and randomly switching the phase of RIS units, achieving key updates in a static environment, and solving the problem of low key generation rate caused by insufficient channel time-varying properties. The paper "Security Interruption Analysis of RIS-Assisted Communication under Discrete Phase Control" further explored the quantitative relationship between RIS phase optimization and the security performance of legitimate users, proving that by rationally designing the RIS reflection coefficient, it is possible to enhance the response of legitimate channels while expanding their spatial uncorrelation with eavesdropping channels. However, most of the above studies assume that the RIS is in an ideal channel estimation state, failing to fully consider the impact of discrete phase shift and quantization errors on authentication accuracy in actual deployments. Faced with increasingly complex malicious attacks, researchers have begun to explore the unique advantages of RIS in resisting active deception attacks. Liu et al. designed a RIS-cascaded channel spatial authentication mechanism to counter pilot spoofing attacks, utilizing the directionality of reflected beams to suppress attackers' forgery of spatial channel features. The papers "Physical Layer Authentication Based on Deep Learning and Reconfigurable Smart Surfaces" and "Label-Based Physical Layer Authentication in RIS-Assisted Communication Systems" extract channel-related physical layer attributes as authentication features and utilize RIS-embedded spatial imprints as a supplementary authentication dimension, combined with specific algorithms to improve recognition accuracy and effectively enhance resistance to spoofing attacks. However, the performance of the above schemes highly depends on the real-time and accurate optimization of the RIS phase, facing the challenge of excessively high computational complexity in large-scale array scenarios.

[0005] In recent years, the integration of artificial intelligence and RIS (Reliable Surface Imaging) has become a research hotspot. Papers such as "Deep Reinforcement Learning for Practical Phase Offset Optimization of RIS-Assisted Networks in Short Packet Communication" and "Intelligent Reflective Surfaces Based on Deep Reinforcement Learning for Secure Wireless Communication" introduce deep reinforcement learning frameworks. By dynamically and adaptively adjusting the RIS phase through channel perception, they maximize the rate difference between legitimate and eavesdropping channels, realizing a shift in security strategies from passive adaptation to active shaping. Papers such as "Injecting Reliable RF Fingerprints into the Internet of Things Using Metasurfaces" and "Reconfigurable Intelligent Surface-Assisted Fingerprint-Based Millimeter-Wave Positioning System" propose RIS fingerprint extraction algorithms based on convolutional neural networks (CNNs), directly learning nonlinear features from the original received signal, avoiding cumbersome explicit channel estimation. Paper "Robust Generative Defense Against Adversarial Attacks in Intelligent Modulation Recognition" enhances the model's robustness against unknown attacks by simulating attacker behavior using generative adversarial networks. Overall, current research on RIS-assisted physical layer authentication still faces challenges such as insufficient modeling of actual deployment losses, difficulty in balancing the real-time performance and robustness of intelligent optimization algorithms, and the lack of a general defense framework for multi-attack scenarios. Furthermore, research on lightweight implementations and cross-layer collaborative authentication mechanisms for large-scale RIS arrays has not yet been systematically developed.

[0006] Meta-learning, also known as learning to learn, is a machine learning paradigm that trains a model on a large number of related tasks, enabling it to acquire prior knowledge, parameter initialization strategies, optimization processes, or learning rules that allow it to quickly adapt to new tasks. This allows for effective generalization to new tasks with only a small number of samples and minimal gradient update steps. Unlike traditional machine learning, which primarily addresses specific problems, meta-learning accelerates learning efficiency for new tasks by learning general strategies and principles. Typical meta-learning methods fall into three categories: metric-based, model-based, and optimization-based meta-learning. Among them, Model-Independent Meta-Learning (MAML) achieves rapid adaptation by optimizing parameter initial points that are highly sensitive to task gradients, achieving high performance on new tasks with only a few gradient descent steps. In the field of identity authentication, the potential of meta-learning has been initially validated. The paper "Temporal Metametric Learning for Multi-User Physical Layer Authentication in Mobile Industrial IoT" reconstructs multi-user physical layer authentication into a temporal few-sample classification task using a temporal metametric learning framework. By combining a temporal decay mechanism with historical CIR samples to learn meta-parameters, it improves the accuracy and robustness of authentication in mobile IoT scenarios. The paper "MCRFF: A Meta-Contrastive Learning-Based RFID Fingerprinting Method" proposes a meta-contrastive learning-based RFID fingerprinting method (MCRFF). This method combines the generalization optimization of meta-learning with the discriminative representation learning of contrastive learning, employing a two-stage training process of device perception and device generalization to adapt to a lightweight model. However, existing works mostly apply meta-learning to a single stage of the authentication process; no research has yet deeply integrated it with the controllable physical environment of RIS and efficient lightweight feature extraction networks to construct a complete, end-to-end adaptive authentication system.

[0007] Convolutional Neural Networks (CNNs), with their powerful local feature extraction and hierarchical representation learning capabilities, have become the standard tool for processing gridded data such as images and time-series signals. Unlike traditional machine learning models that require manual feature design, CNNs can achieve end-to-end feature learning and task adaptation without manual intervention in feature engineering, greatly improving application efficiency in complex scenarios. At the architectural level, CNNs continue to evolve towards lightweight and efficient designs. Efficient CNNs aim to reduce computational resources and memory consumption, enabling deployment on resource-constrained terminal devices. Furthermore, incorporating attention mechanisms into CNN models can improve the effectiveness of feature extraction. The paper "Few-Sample Mobile Phone Screen Defect Classification Based on Attention-Relation Networks" mentions that in the mobile phone screen defect classification task, Attention-Relation Networks improved metric learning by introducing attention mechanisms, achieving defect classification with a small number of samples and demonstrating better recognition results than traditional CNNs. Currently, researchers are combining meta-learning with CNNs to improve the model's generalization ability under different machines or operating conditions. The paper "Adaptive Model-Independent Meta-Learning Network for Cross-Machine Fault Diagnosis with Limited Samples" introduces Adaptive Model-Independent Meta-Learning (AMAML) into a lightweight spatial bilateral channel attention mechanism, improving the accuracy and robustness of cross-machine fault diagnosis with limited samples. Typical applications of CNNs in physical layer security authentication fall into three categories: channel feature-based authentication, hardware-inherent feature-based authentication, and multi-scenario customized authentication. Among them, CNN authentication schemes based on Channel State Information (CSI) differentiate devices by extracting unique fingerprint features of the wireless channel, achieving extremely high attack detection accuracy in static environments. The paper "CNN-Based Physical Layer Authentication Method for Underwater Acoustic Sensor Networks" utilizes the location fingerprint formed by the multipath structure of underwater acoustic channels to propose a CNN-based collaborative authentication framework. By analyzing CIR to extract features, it adapts to underwater scenarios with limited bandwidth, and its malicious packet detection accuracy outperforms traditional support vector machine (SVM) methods. However, the application of deep learning methods often requires a large amount of training data, and the variability of wireless scenarios, such as antenna positions and different occlusion conditions, can cause serious shifts in data distribution, resulting in a sharp decline in performance in new environments. At the same time, the deployment of complex network structures on resource-constrained terminal devices remains a significant challenge.

[0008] While existing research has made progress in its respective fields, it has failed to systematically solve the core contradiction faced by single-antenna power IoT terminals: the balance between limited hardware capabilities, dynamic and complex environments, and high-intensity security requirements.

[0009] In summary, the continuous expansion of new power communication networks presents a severe challenge to the secure access of massive numbers of single-antenna, resource-constrained power IoT terminals. Traditional cryptographic-based authentication protocols are difficult to apply due to high computational and communication overhead, while physical layer authentication schemes relying on multi-antenna arrays cannot extract distinguishable spatial features for single-antenna terminals to resist Sybil attacks. Furthermore, the heterogeneous and complex deployment environment of power terminals makes traditional deep learning authentication models face bottlenecks such as high training data annotation costs and poor environmental adaptability. Summary of the Invention

[0010] The purpose of this invention is to provide a lightweight power terminal authentication method and system that integrates RIS and meta-learning, aiming to solve the above-mentioned problems in the prior art.

[0011] This invention provides a lightweight power terminal authentication method that integrates RIS and meta-learning, comprising: Based on the identity identifier of the power terminal, the reconfigurable smart surface RIS adopts a reflection configuration corresponding to the power terminal, and receives the channel response of the power terminal after reflection by the RIS, and uses the channel response as the enhanced spatial channel fingerprint of the power terminal. The enhanced spatial channel fingerprint is subjected to spatiotemporal feature extraction using a pre-trained lightweight feature extraction network to obtain the corresponding low-dimensional spatiotemporal feature vector. The low-dimensional spatiotemporal feature vector is input into a lightweight authenticator pre-trained based on a meta-learning framework, and the authentication result of the power terminal is output.

[0012] This invention provides a lightweight power terminal authentication system that integrates RIS and meta-learning, comprising: The RIS control and fingerprint acquisition module is used to control the reconfigurable smart surface RIS to adopt a reflection configuration corresponding to the power terminal according to the identity of the power terminal, and to receive the channel response of the power terminal after reflection by the RIS, and to use the channel response as the enhanced spatial channel fingerprint of the power terminal. The spatiotemporal feature extraction module is used to extract spatiotemporal features from the enhanced spatial channel fingerprint using a pre-trained lightweight feature extraction network to obtain the corresponding low-dimensional spatiotemporal feature vector. The meta-learning authentication module is used to input the low-dimensional spatiotemporal feature vector into a lightweight authenticator pre-trained based on the meta-learning framework, and output the authentication result of the power terminal.

[0013] This invention also provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the above-described lightweight power terminal authentication method integrating RIS and meta-learning.

[0014] This invention also provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor, implements the steps of the aforementioned lightweight power terminal authentication method integrating RIS and meta-learning.

[0015] The embodiments of this invention can include the following beneficial effects: This invention proposes a lightweight authentication method integrating reconfigurable smart surfaces, model-independent meta-learning, and a CNN-GRU hybrid network. First, by optimizing the reflection coefficient matrix of the reconfigurable smart surface, a highly discriminative spatial channel fingerprint is actively constructed for each single-antenna terminal at a fixed location, fundamentally solving the problem of insufficient feature resolution of single-antenna devices at the physical level. Second, a model-independent meta-learning framework is introduced, enabling the authentication system to learn rapidly adaptive meta-knowledge from limited historical data. Only a small number of samples are needed to achieve accurate and rapid adaptation of the authentication model, significantly improving the system's scalability and environmental robustness. Finally, a lightweight CNN-GRU deep feature extraction network is designed, efficiently fusing the spatial structure and microscopic time-varying information of the channel response with extremely low parameter count and computational complexity. Millisecond-level inference can be achieved at the network edge, meeting the real-time requirements of power services. Simulation experiments show that the proposed method exhibits good authentication performance and generalization ability in complex scenarios such as time-varying channels and Sybil attacks, providing reliable technical support for lightweight access authentication in new power communication networks. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a lightweight power terminal authentication method integrating RIS and meta-learning according to an embodiment of the present invention. Figure 2 This is a schematic diagram of a novel power communication network system model according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating the authentication process of an embodiment of the present invention; Figure 4This is a schematic diagram of the CNN-GRU architecture according to an embodiment of the present invention; Figure 5 This is a schematic diagram comparing the spatial feature distribution under different RIS regulation strategies in an embodiment of the present invention; Figure 6 This is a schematic diagram comparing the authentication accuracy as a function of signal-to-noise ratio (SNR) in an embodiment of the present invention. Figure 7 This is a schematic diagram comparing the authentication accuracy under different sample sizes in an embodiment of the present invention; Figure 8 This is a schematic diagram comparing the adaptation convergence speed of different methods in embodiments of the present invention; Figure 9 This is a schematic diagram comparing the registration delay and energy consumption of different methods in embodiments of the present invention; Figure 10 This is a multi-dimensional performance evaluation and comparative analysis diagram of the lightweight authentication framework according to an embodiment of the present invention; Figure 11 This is a schematic diagram of a lightweight power terminal authentication system that integrates RIS and meta-learning according to an embodiment of the present invention. Detailed Implementation

[0018] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.

[0019] Method Implementation Examples According to embodiments of the present invention, a lightweight power terminal authentication method integrating RIS and meta-learning is provided. Figure 1 This is a flowchart of a lightweight power terminal authentication method integrating RIS and meta-learning according to an embodiment of the present invention, as shown below. Figure 1 As shown, the lightweight power terminal authentication method integrating RIS and meta-learning according to an embodiment of the present invention specifically includes: Step S101: Based on the identity identifier of the power terminal, control the reconfigurable smart surface RIS to adopt a reflection configuration corresponding to the power terminal, and receive the channel response of the power terminal after reflection by the RIS, using the channel response as the enhanced spatial channel fingerprint of the power terminal; specifically including: A unique RIS reflection coefficient matrix is ​​pre-generated for each legitimate power terminal, and a mapping relationship is established between the power terminal's identity and the RIS reflection coefficient matrix; specifically including: With the goal of maximizing the distinguishability of channel fingerprints among different power terminals, the minimum Euclidean distance criterion is adopted to optimize and determine the reflection coefficient matrix for each power terminal that maximizes the minimum Euclidean distance between its normalized channel response vector and the normalized channel response vectors of other power terminals. The mapping relationship is queried based on the current power terminal's identity to determine the corresponding target RIS reflection coefficient matrix, and each reflection unit of the RIS is controlled to perform phase adjustment according to the target RIS reflection coefficient matrix; After the RIS completes phase adjustment, it receives the detection signal transmitted by the power terminal; wherein, the detection signal includes a reflected link signal formed by the RIS after reflection in the adjusted phase state and a direct link signal that directly reaches the base station from the power terminal. The reflected link signal and the direct link signal are coherently superimposed at the base station receiver to form a composite channel response. The composite channel response is used as the enhanced spatial channel fingerprint of the power terminal.

[0020] Step S102 involves using a pre-trained lightweight feature extraction network to extract spatiotemporal features from the enhanced spatial channel fingerprint, obtaining the corresponding low-dimensional spatiotemporal feature vector; specifically including: Enhanced spatial channel fingerprints are collected at multiple consecutive time points to form a time-series channel matrix; The temporal channel matrix is ​​input into a pre-trained lightweight feature extraction network; wherein the lightweight feature extraction network includes a spatial feature extraction unit and a temporal feature extraction unit; The spatial feature extraction unit independently extracts spatial structure features from the composite channel response at each time step and outputs a spatial feature sequence. The temporal feature extraction unit is used to perform temporal dependency modeling on the spatial feature sequence, capture the dynamic evolution law of the composite channel response, and output a low-dimensional spatiotemporal feature vector. The spatial feature extraction unit adopts a one-dimensional depthwise separable convolutional structure to independently extract spatial structural features from the composite channel response at each time step. The temporal feature extraction unit is a gated recurrent unit (GRU), which is used to perform temporal dependency modeling on spatial feature sequences and take the hidden state of the final time step as a temporal context summary. The lightweight feature extraction network also includes a fully connected layer for mapping the temporal context summary into a low-dimensional spatiotemporal feature vector.

[0021] Step S103 involves inputting the low-dimensional spatiotemporal feature vector into a lightweight authenticator pre-trained based on a meta-learning framework, and outputting the authentication result of the power terminal; specifically including: Obtain the initial parameters of the lightweight validator trained using the Model-Independent Meta-Learning MAML framework during the offline phase; specifically including: In the offline phase, multiple power terminals are randomly selected from the historical dataset to construct meta-tasks, and the channel response measurement samples of each terminal in each meta-task are divided into support set and query set; In the inner loop, the lightweight validator is initialized with the current meta-parameters, the cross-entropy loss is calculated using the support set of each meta-task, and gradient descent is performed for a preset number of steps to obtain the adaptation parameters specific to that meta-task. In the outer loop, the loss of each meta-task on the query set is calculated based on the adaptation parameters, and the current meta-parameters are updated with the goal of minimizing the sum of the losses of all meta-tasks on the query set. The inner and outer loops are executed iteratively until convergence, thus obtaining the initial parameters of the lightweight authenticator; For a newly connected power terminal, a preset number of channel samples of the power terminal are collected as a support set. The initial parameters are updated finitely times using the support set to obtain the exclusive authentication parameters of the power terminal. The authentication threshold of the power terminal is set based on the minimum confidence value of the samples in the support set. Based on the exclusive authentication parameters, the currently extracted low-dimensional spatiotemporal feature vector is input into the lightweight authenticator to calculate the confidence score that the power terminal belongs to a legitimate device. The confidence score is compared with the authentication threshold. If the confidence score is greater than or equal to the authentication threshold, a valid authentication result is output; otherwise, an invalid authentication result is output.

[0022] The following describes in detail the above-mentioned technical solution of the present invention with reference to the specific details of the lightweight power terminal authentication method that integrates RIS and meta-learning in the embodiments of the present invention.

[0023] This invention aims to solve the core challenge of secure, lightweight, and adaptive access authentication for single-antenna, resource-constrained power terminals in complex environments. It proposes a collaborative authentication framework that deeply integrates RIS, meta-learning, and a lightweight CNN-GRU network. This framework constructs an authentication model where RIS generates physical layer features, meta-learning drives rapid model adaptation, and a lightweight CNN achieves efficient authentication. In other words, the key to this invention lies in: 1. A RIS spatial fingerprint active enhancement method is proposed: by establishing an optimization model with feature distinguishability as the direct objective, a unique and environmentally robust enhanced channel fingerprint is dynamically constructed for each single-antenna terminal, thereby creating distinguishable spatial features for the single-antenna terminal.

[0024] 2. A fast small-sample adaptive mechanism for the authentication model based on model-independent meta-learning was designed: In response to the challenges of model generalization caused by the heterogeneity of power Internet of Things terminals and the time-varying environment, the MAML meta-learning framework was introduced into the authentication system, enabling the model to quickly and accurately adapt to new terminals with only a very small number of samples, which greatly reduces the system's dependence on labeled data and the operation and maintenance costs.

[0025] 3. A lightweight spatiotemporal feature extraction and authentication model for RIS-enhanced features was constructed: A lightweight network integrating CNN and gated recurrent units was designed to extract the spatial structure and micro-time-varying features of the channel response with extremely low computational complexity; and combined with a meta-learning framework, a highly efficient authentication system capable of millisecond-level inference in resource-constrained environments was finally realized.

[0026] In other words, the embodiments of this invention will break through the limitations of existing isolated research and construct a new lightweight authentication paradigm with inherent synergy by deeply integrating RIS, meta-learning, and a lightweight CNN-GRU network, as detailed below: I. System Model This invention provides a lightweight access authentication system model for novel power communication networks, such as... Figure 2 As shown, this model includes single-antenna power terminals, reconfigurable smart surfaces, base stations (access points), and potential Sybil attackers. In this model, a large number of resource-constrained single-antenna power terminals are typically deployed in typical power environments such as substations and distribution rooms, and periodically initiate access authentication requests to the base station. The base station, as the access and control node at the network edge, deploys a lightweight convolutional neural network authentication model trained using a meta-learning framework, and is responsible for the global coordination and dynamic configuration of the RIS reflection coefficient matrix.

[0027] To overcome the inherent limitation of insufficient spatial feature resolution in single-antenna terminals, a reconfigurable intelligent surface (RIS) consisting of N independently adjustable reflective elements is introduced into the model. Deployed in the wireless propagation environment between the terminal and the base station, RIS can receive configuration commands from the base station in real time via a control link, maximizing the distinguishability of the channel fingerprint. For each legitimate terminal initiating an authentication request, the base station assigns an appropriate reflection coefficient matrix to the RIS based on the terminal's identity, thus establishing a one-to-one mapping between the terminal's location and the RIS configuration at the physical layer. The pilot signal transmitted by the terminal, after intelligent reflection by the RIS, is coherently superimposed with the direct path signal at the base station receiver, forming a high-dimensional, unique, and time-varying stable composite channel response—the RIS-enhanced spatial channel fingerprint. This fingerprint not only possesses physical non-cloning properties and strong location correlation but also, through the RIS's active beamforming, constructs distinguishable spatial features for single-antenna terminals.

[0028] When a Sybil attack occurs in the system, a malicious node attempts to impersonate one or more legitimate terminals to initiate authentication. Since the attacker's physical location is necessarily different from that of a legitimate terminal, their signal, after being reflected by the same RIS configuration, creates a composite channel response at the base station that differs statistically from the preset fingerprint of a legitimate terminal. When the base station receives the authentication request signal, it first extracts the channel state information enhanced by the RIS, and then inputs it into a lightweight CNN-GRU authentication model for real-time inference. This model is trained offline using a meta-learning paradigm, enabling it to learn quickly from small samples and adapt to new tasks. For registered terminals, the model can directly perform high-confidence identity determination; for newly registered terminals, the model can quickly fine-tune the authenticator based on a few samples, achieving accurate and low-overhead registration and subsequent authentication. Finally, the base station compares the model's output confidence score with a preset threshold to determine the legitimacy of the access request, thus effectively resisting Sybil attacks with low latency and low computational overhead, achieving lightweight and highly secure access authentication.

[0029] II. Lightweight Certification This invention proposes a lightweight physical layer authentication scheme for single-antenna, resource-constrained power terminals. This scheme deeply integrates channel modulation using a reconfigurable smart surface (RIS), feature extraction using a convolutional neural network (CNN), and the rapid adaptive capabilities of meta-learning, aiming to achieve secure, lightweight, and environmentally adaptable terminal authentication. The scheme consists of two phases: an offline training phase and an online rapid authentication phase. The system flow is as follows: Figure 3 As shown.

[0030] 1. Offline training phase In this stage, a general CSI feature extraction network with high discriminativeness is constructed by utilizing the historical channel state information (CSI) of known terminals, and a meta-learning authenticator that can quickly adapt to new terminals based on a small number of samples is trained.

[0031] (1) RIS-enhanced channel fingerprint generation To enhance the uniqueness and distinguishability of physical layer channel fingerprints, a dedicated RIS reflection coefficient matrix is ​​pre-calculated for each legitimate terminal during the offline phase to optimize the spatial characteristics of its channel response.

[0032] Assume the system has a total of Given a known terminal, the base station is equipped with M receiving antennas, and the RIS consists of N programmable reflector units. For the terminal... Its composite channel response (i.e., spatial channel fingerprint) with RIS assistance. It can be represented as: (1); in, This represents the direct channel vector from the terminal to the base station. This represents the channel matrix from RIS to the base station. This represents the terminal-specific RIS reflection coefficient matrix. This represents the channel vector from the terminal to the RIS.

[0033] To maximize the distinguishability of channel fingerprints from different terminals in the feature space, the minimum Euclidean distance criterion is adopted to optimize the RIS configuration for each terminal. .

[0034] (2); (3); in, The objective function for measuring the distinguishability of channel fingerprints is... and Representing terminals respectively and Normalized CSI channel vectors under the corresponding RIS configuration.

[0035] By all Each terminal solves the above optimization problem sequentially, resulting in a set of globally optimized RIS configurations. Subsequently, channel state information of each terminal under its dedicated RIS configuration was collected under various typical channel environments to form a historical dataset. It is used to train the feature extraction network and provides a data foundation for the construction of meta-learning tasks.

[0036] (4); in, This indicates the number of samples per terminal.

[0037] (2) Lightweight CSI feature extraction network based on CNN-GRU To effectively extract discriminative features that combine spatial structure and temporal evolution information from CSI channel responses, this invention proposes a lightweight convolutional neural network-gated recurrent unit (CNN-GRU) hybrid feature extraction network. The network first extracts the spatial structure features of the channel response at each time step using a CNN, and then uses a GRU to model the temporal dependency of the feature sequence. This allows it to capture both the spatial distribution pattern and micro-time-varying features of the channel response with extremely low computational complexity, generating a more discriminative and robust terminal identity fingerprint. Its architecture is as follows: Figure 4 As shown.

[0038] Based on the constructed historical dataset The terminal will continuously within a short time window Secondary channel measurements are constructed as spatiotemporal input samples. For the terminal The i-th sample, its time-series channel matrix Represented as: (5); To adapt to real-valued neural networks, each complex channel vector is... Convert to a two-dimensional real-valued matrix This ultimately forms the input tensor. : (6); (7); One-dimensional depthwise separable convolutions are used to process the input slices at each time step independently. It efficiently extracts the channel spatial distribution pattern under RIS modulation and outputs a compact spatial feature vector. By processing T time points sequentially, the CSI spatial feature sequence is obtained. .

[0039] The sequence is input into a single-layer gated recurrent unit (GRU) in chronological order. Its gating mechanism captures the dynamic evolution of CSI features, and the hidden state at the final time step is obtained. As a summary of the temporal context: (8); in, It is in a hidden state.

[0040] Finally, the final state of the GRU is transmitted through a fully connected layer. Mapped to the terminal's low-dimensional spatiotemporal fingerprint : (9); in, To output the weight matrix, This is the output bias vector.

[0041] (3) Training of meta-learners for fast adaptation To enable the system to quickly build high-precision authenticators for new terminals based on a very small number of samples, this stage employs a Model-Agnostic Meta-Learning (MAML) framework. This is done within a fixed feature extraction network. Based on this, train a lightweight authenticator. This allows it to obtain optimal initialization parameters. This enables it to quickly adapt to new terminals.

[0042] Based on historical datasets Constructing a meta-training task distribution Randomly select N terminals and use... Extract each terminal The features of each sample are used as the support set. and extract other The features of non-overlapping samples are used as the query set. Each meta-task is defined as This simulates a low-sample registration and authentication scenario for new terminals. MAML learns initial parameters through a two-layer optimization using both inner and outer loops. .

[0043] In the inner loop, for each meta-task With the current meta-parameter Initialize the authenticator to obtain the authentication network. Utilizing the support set for this task Calculate the cross-entropy loss and perform... Step gradient descent to obtain task adaptation parameters : (10); in, For the inner loop learning rate, To support the cross-entropy loss of the set.

[0044] In the outer loop, the meta-parameters are updated by minimizing the sum of the query set losses for all tasks. : (11); in, For the outer loop learning rate, The cross-entropy loss is applied to the query set. After optimization, the initial parameters of the validator with fast adaptive priors are finally obtained. .

[0045] 2. Online authentication stage When a terminal requests network access, the system utilizes the feature extraction network pre-trained in the offline stage and meta-learning prior knowledge to securely and quickly complete identity registration and subsequent real-time authentication with minimal consumption of samples and computing resources.

[0046] (1) Quick Registration When a new terminal is detected When a request for access is made, the system initiates a registration process to create a unique authentication configuration for that user. The registration process is as follows: Step 1: Collect channel fingerprint. According to formulas (1)-(2), the base station controls the RIS to switch to the new terminal based on the pre-calculated reflection coefficient matrix. It guides the terminal to send probe signals within a short period of time, and the collected channel responses constitute an online support set. and query set .

[0047] Step 2: Extract CSI channel features. Using a fixed feature extraction network according to formulas (5)-(9). Process all samples to obtain the corresponding low-dimensional spatiotemporal feature vectors. .

[0048] Step 3: Meta-learning adaptation. The initial meta-learning parameters obtained in the offline phase are... As a starting point, with The features are taken as input, and the formulas (10)-(11) are executed. Gradient descent is used to quickly adapt and obtain the dedicated lightweight authenticator parameters for this terminal. .

[0049] Step 4: Confidence and Threshold Calculation. For Dedicated Authentication Networks Its output is a softmax-normalized posterior probability vector, and the posterior probability of the terminal's corresponding class is defined as the confidence score. : (12); Based on support set Calculate the confidence scores for all samples, and introduce a safety factor based on the lowest confidence score. Determine the initial authentication threshold: (13); in, For safety, a threshold of 0.95 is preferred, and this threshold can be dynamically adjusted based on the false alarm rate during subsequent authentication. At this point, the terminal... Upon completion of registration, the system will store unique parameters for the user. .

[0050] (2) Lightweight real-time authentication After registration is completed, the terminal It is included in the legitimate set. In each subsequent communication, the system performs real-time identity authentication on the terminal. The specific authentication process is as follows: Step 1: RIS Configuration and Channel Acquisition. When the terminal transmits a signal, the base station controls the RIS to switch to the terminal's dedicated reflection coefficient matrix. The current composite channel response is obtained through formula (1). .

[0051] Step 2: Extract CSI channel features. Use an offline-trained feature extraction network. Perform a single forward propagation and extract its features. : (14); Step 3: Calculate the confidence score. Enter the terminal's proprietary authenticator. The posterior probability of the corresponding category of the terminal is calculated according to formula (12) and used as the confidence score for real-time authentication. .

[0052] Step 4: Authentication Decision. The confidence score is compared with the threshold set during the registration phase. Compare and make certification decisions; such as If so, the power terminal's access authentication is successful; If the power terminal is deemed an illegal device, access will be refused.

[0053] III. Simulation Experiments and Result Analysis To comprehensively verify the integrated performance of the novel lightweight access authentication technology for power communication networks based on reconfigurable smart surfaces and meta-learning, this invention will construct an experimental environment that closely resembles real-world scenarios using the MATLAB simulation platform. Control experiments will be designed around four core objectives: RIS spatial fingerprint enhancement effect, rapid registration and adaptation with small samples, defense against various Sybil attacks, and system robustness under complex environments. This will verify the effectiveness and advancement of the proposed solution in resource-constrained and dynamically changing power environments, providing reliable experimental support for the actual deployment of the technology and its security applications at the edge.

[0054] 1. Experimental Environment and Scene Setup This experiment, based on the MATLAB simulation platform and deep learning toolbox, simulates a real-world access scenario for a novel power communication network in a complex electromagnetic environment. The simulation scenario is set as a smart substation. The validator (Bob) is simulated as an industrial-grade data concentrator (DCU) deployed at the edge, equipped with M=8 receiving antennas. The legitimate user (Alice) is simulated as a single-antenna power IoT terminal, randomly distributed within an area approximately 30-50m from Bob. To address line-of-sight (LoS) link obstruction caused by metal shielding in the power environment, a RIS (N=64 elements) is deployed at a key reflection point on a wall at a height of 5m. The Sybil attacker (Eve) is configured as a malicious node located in Alice's neighborhood (distance <0.5m), attempting to bypass physical layer authentication through power simulation or close-range camouflage. The experiment employed an Orthogonal Frequency Division Multiplexing (OFDM) system with a carrier frequency of 2.4 GHz and a bandwidth of 20 MHz. Channel modeling comprehensively considered path loss and multipath effects: the Alice-Bob direct link used Rayleigh fading to simulate non-line-of-sight transmission, while reflection links involving RIS, such as Alice-RIS and RIS-Bob, used Rician fading with a Rice factor of K=6 dB. The path loss exponent was set to 2.2–3.5 to reflect the severe multipath scattering characteristics within substations. During the meta-learning training phase, each meta-task contained K=1–10 support set samples, and the Adam optimizer was used with a learning rate set to the inner loop. =0.01, outer loop =0.001, thus ensuring that the model has the ability to quickly adapt to different noise levels (SNR of 0~25dB) and heterogeneous terminal access.

[0055] 2. Dataset Construction To verify the effectiveness of the lightweight authentication scheme, this experiment constructs an experimental dataset by simulating a typical scenario of secure access to power IoT terminals under a Sybil attack, adapting to the full experimental verification requirements of RIS feature enhancement, meta-learning few-shot adaptation, and defense against multiple types of Sybil attacks. The dataset uses Channel State Information (CSI) as the core feature, collecting signals at T=10 consecutive time steps to form a spatiotemporal tensor. Each sample has a dimension of 8×2×10 (corresponding to the number of antennas, complex real and imaginary parts, and time step), and simultaneously records Received Signal Strength Indicator (RSSI) as a benchmark for comparison with traditional schemes. Data acquisition covers the entire signal-to-noise ratio range of 0–25 dB, including 30 legitimate terminal locations and Sybil attackers physically close to them (<0.5 m), and correlates the reflection coefficient matrix of RIS under three strategies: Proposed enhancement, Max-SNR, and random phase. For the meta-learning training requirements, the dataset is organized into several independent meta-tasks, each strictly divided into a support set and a query set to simulate a plug-and-play registration scenario for new terminals ranging from 1 to 50 shots. The entire dataset is divided into three subsets: RIS spatial fingerprint enhancement, small sample fast adaptation, and Sybil attack lightweight authentication, covering three attack variants: power simulation, close-range camouflage, and dynamic environment. All samples are labeled with terminal identity, attack type, and SNR label to ensure the rigor of model training and performance evaluation.

[0056] 3. Evaluation Indicators To comprehensively and quantitatively evaluate the performance and security of the authentication system, this embodiment of the invention uses authentication accuracy, false alarm rate (FAR), true positive rate (TPR), ROC curve, and registration latency as core statistical indicators.

[0057] (1) Accuracy, which mainly measures the model's ability to correctly classify legitimate users and Sybil attackers globally, is calculated as follows: (15); in, and These represent the number of samples that correctly accepted legitimate users and correctly rejected attackers, respectively. and These represent the number of samples that were mistakenly identified as legitimate users (false positives) and the number of legitimate users that were mistakenly identified as attackers (false negatives), respectively.

[0058] (2) False Alarm Rate (FAR), defined as the probability that attacker Eve is mistakenly identified by the system as the legitimate identity Alice, directly reflects the system's robustness against Sybil attacks. The calculation formula is as follows: (16); (3) True Positive Rate (TPR) reflects the probability that the system correctly authenticates legitimate users. The calculation formula is as follows: (17); (4) False Rejection Rate (FRR) represents the probability that a legitimate user is incorrectly rejected by the system. It reflects the impact of the authentication mechanism on the availability of normal power services. The calculation formula is as follows: (18); (5) Convergence speed, as a dynamic adaptation performance evaluation index, the convergence speed is quantified by the number of gradient iteration steps, which represents the number of gradient update steps required for the certification model to converge from the initial state to the industrial-grade high accuracy level. The fewer the iteration steps, the stronger the model's rapid adaptation capability, and the better it can meet the high real-time access requirements of power services.

[0059] (6) Registration delay, which is quantified by time, represents the total time taken for a new terminal to complete identity registration and model adaptation, and directly reflects the access efficiency of the system.

[0060] (7) Total adaptation energy consumption, which is the total resource consumption during the terminal's identity authentication adaptation process. The lower the value, the more suitable the solution is for resource-constrained power terminal deployment scenarios.

[0061] 4. Experimental Results and Analysis (1) RIS-assisted spatial channel fingerprinting enhancement experiment To investigate the impact of different RIS configuration strategies on authentication performance under Sybil attack scenarios, this experiment constructed a simulation environment with a signal-to-noise ratio (SNR) range of 0~25dB. The focus was on comparing the feature discrimination capability and attack resistance performance of four typical authentication schemes: Scheme A (Proposed) based on the active fingerprint construction scheme that maximizes the minimum Euclidean distance, Scheme B (Max-SNR) traditional RIS phase alignment scheme, Scheme C (Random-Phase) random RIS phase scheme, and Scheme D (No-RIS) traditional physical layer authentication benchmark scheme. The feature recognition and security protection capabilities of each scheme under Sybil attack were systematically evaluated.

[0062] First, the effectiveness of the proposed RIS spatial fingerprint active enhancement scheme in improving the physical layer feature resolution was verified. Figure 5This paper demonstrates the distribution of legitimate terminals and attackers in the feature space under a Sybil attack scenario. It is evident that under the traditional Max-SNR and random phase criteria, the lack of optimization for identity recognition leads to significant overlap in feature distribution between legitimate users and physically adjacent attackers. This low distinguishability makes traditional methods ineffective against elaborate identity spoofing. In contrast, the active enhancement algorithm proposed in this invention significantly widens the feature gap between legitimate and illegitimate devices in the feature space by optimizing the RIS reflection coefficient matrix, forming clearly separable clustered regions. This physically demonstrates that RIS possesses the ability to actively reshape the wireless propagation environment and construct highly distinguishable spatial fingerprints, effectively compensating for the inherent deficiency of insufficient feature resolution in single-antenna terminals.

[0063] Figure 6 The results show a comparison of authentication accuracy as a function of signal-to-noise ratio (SNR). Experiments demonstrate that the proposed scheme exhibits optimal authentication performance across the entire SNR range. Even in extreme environments with low SNR, the proposed scheme maintains an accuracy rate of over 70%, demonstrating strong noise resistance. Furthermore, as the SNR increases, the accuracy of the proposed scheme rapidly rises and eventually converges to over 99%, with both convergence speed and final accuracy significantly outperforming traditional comparative schemes. Experimental results demonstrate the strong robustness of the proposed algorithm against noise interference. By actively constructing physical layer features through RIS, it can provide high-quality input information for subsequent lightweight authentication models, thereby ensuring the access security of power communication networks in complex electromagnetic environments.

[0064] (2) Experiment on fast adaptation of small samples based on meta-learning To verify the advantages of the proposed meta-learning framework over traditional deep learning in terms of learning speed and sample requirements, this experiment is designed to support multi-dimensional comparative verification based on sample size, gradient iteration steps, computational and energy consumption. Two typical authentication models are selected as benchmarks for comparative experiments: a standard CNN-GRU model using a traditional pre-training-fine-tuning strategy and a Vanilla CNN model trained from scratch without pre-training weights. By quantifying the authentication accuracy, registration latency, and energy consumption under different experimental conditions, the performance advantages of the proposed scheme over traditional models in small sample adaptation, dynamic environment response, and resource-constrained scenarios are verified, providing effective methodological support for lightweight secure access of single-antenna terminals in the power IoT.

[0065] from Figure 7As can be seen, the proposed scheme maintains absolute leadership across the entire sample range and exhibits extremely strong generalization ability under very small sample conditions. In the extreme constraint scenario of K=1, the authentication accuracy of the proposed scheme reaches over 88%, far exceeding Baseline B and Baseline C. With the increase in the support set size K, the performance of the proposed scheme rapidly increases, exceeding 97% accuracy when K=5 and entering a stable saturation period. This fully verifies that meta-learning, by capturing the prior meta-knowledge of the enhanced channel fingerprint through RIS, overcomes the dependence of traditional deep learning on massive labeled data. Although Baseline B ultimately reaches approximately 93% accuracy, its growth is slow in the K<10 range, reflecting the difficulty of achieving rapid feature adaptation in the absence of gradient-sensitive initialization in standard fine-tuning strategies. Baseline C performs the worst, barely approaching the level of Baseline B when K=50, highlighting the cold-start bottleneck of traditional models due to the scarcity of samples in the early stages of power terminal access.

[0066] To verify the real-time response capability of the proposed scheme under dynamic power environments, this experiment fixed the support set size K=5 and quantified the relationship between authentication accuracy and the number of gradient iteration steps. The results are as follows: Figure 8 As shown, the proposed solution exhibits significant rapid adaptation characteristics. Its curve starts at a high level, achieving an accuracy of over 76% with zero steps. Furthermore, after only 1-2 gradient updates, the accuracy rapidly spikes to over 97%, achieving identity registration within seconds or even milliseconds. In contrast, the Baseline B curve using the standard fine-tuning strategy has a lower slope, requiring approximately 10 iterations to reach around 95%. This difference verifies that the MAML framework, by optimizing the initial parameter distribution, positions the model at an optimal initial position highly sensitive to authentication tasks, significantly shortening the online learning cycle for new terminal access and meeting the high real-time requirements of the power industry.

[0067] For resource-constrained power IoT terminals, this experiment further evaluated the computational efficiency of the solution, including registration latency and total adaptation energy consumption. The results are as follows: Figure 9As shown. Regarding registration latency, due to the minimal number of iterations resulting from meta-learning, the proposed solution exhibits extremely low latency, significantly lower than Baseline B which employs a traditional fine-tuning strategy. Baseline C, requiring training from scratch, has the highest latency, showing an order-of-magnitude difference compared to the proposed solution. In terms of energy consumption, the proposed solution also maintains the lowest resource consumption level, with its total adaptation overhead significantly lower than both Baseline B and Baseline C. Experimental results demonstrate that although the CNN-GRU architecture has slightly higher computational complexity during inference than Vanilla CNN, meta-learning significantly reduces the number of adaptation iterations, thus achieving extremely high energy efficiency at the system level. The proposed solution ensures high-strength security authentication while also meeting the lightweight and low-power operation requirements of power edge devices.

[0068] (3) Performance comparison experiment of lightweight authentication framework To verify the comprehensive performance advantages of the lightweight authentication scheme integrating RIS, meta-learning, and CNN-GRU proposed in this invention under complex attack scenarios, dynamic environments, and resource-constrained conditions, this experiment designs a multi-dimensional comparative experiment. Traditional RSSI authentication (Scheme A), CSI-related physical layer authentication (Scheme B), and standard CNN single-antenna authentication (Scheme C) are selected as benchmarks. Quantitative verification is carried out from four core dimensions: attack resistance capability, environmental adaptation efficiency, real-time performance, and lightweight nature. By comparing the authentication accuracy under different attack types, the number of supported samples required for new terminal adaptation, authentication latency, and the number of model parameters, the applicability and superiority of the proposed scheme in the secure access scenario of single-antenna terminals in the power IoT are systematically evaluated, providing data support for technology implementation.

[0069] from Figure 10 As can be seen, the proposed solution exhibits significant performance superiority across all scenarios. In the ROC curve of the close-range attack scenario, solution D maintains optimal recognition performance within a wide FAR range, and when FAR < In the low false alarm rate range, scheme C, with its superior model fitting advantage of single-antenna CSI fingerprint, currently has an ROC curve above that of scheme D, but the TPR difference between the two is small. When FAR ≥ At this point, the performance advantage of Scheme D began to emerge and continued to expand. With the improvement of FAR, the TPR growth rate of Scheme D was significantly faster than that of Scheme C, and the curve quickly overtook and maintained its lead, further widening the gap with Schemes B and A. This verifies that RIS actively constructs a highly discriminative spatial channel fingerprint for single-antenna terminals by optimizing the reflection coefficient matrix, improving the anti-spoofing and stability of features at the physical layer. Furthermore, the meta-learning framework endows the model with the ability to quickly adapt, enabling it to maintain the stability of feature discriminative power over a wide FAR range, effectively compensating for the lack of natural feature dimensions in single-antenna terminals. In contrast, Scheme C relies solely on single-antenna natural CSI fingerprint training and lacks a physical layer feature enhancement mechanism. As FAR expands, its insufficient model generalization ability becomes apparent, and the TPR growth tends to level off. Schemes A and B, due to the lack of intelligent feature extraction and adaptive learning mechanisms, struggle to distinguish between legitimate terminals and attacker features under near-range attacks, and their overall performance remains at a low level.

[0070] In terms of access efficiency and environmental adaptability, the Model-Independent Meta-Learning (MAML) framework introduced in this invention demonstrates a significant prior advantage when handling new terminal access. Conventional deep learning models, lacking meta-knowledge initialization, exhibit a slow increase in accuracy, typically requiring more than two training samples to barely reach the 90% industrial-grade certification standard. This invention, by learning a general representation of channel feature distribution offline, requires only about four support samples for fine-tuning, achieving a leap in accuracy from 90% to over 95%, perfectly adapting to the high-efficiency plug-and-play access requirements of massive heterogeneous terminals in power grids.

[0071] In the experiment balancing real-time performance and lightweight design, the conventional large model (Scheme C) achieved high accuracy, but its inference latency was as high as 13.2ms, exceeding the 10ms real-time threshold for power control services. The proposed solution, however, compresses the number of parameters to around 14k and controls the inference latency to 4.1ms, achieving an ideal balance between model size and inference latency. This advantage stems from the lightweight architecture design of the CNN-GRU hybrid network, which combines one-dimensional depthwise separable convolutions with a single-layer GRU to ensure feature extraction capabilities while reducing computational complexity to less than 50% of traditional models. Combined with the minimal iteration steps of meta-learning, it enables efficient deployment on resource-constrained terminals.

[0072] For three typical Sybil attack variants in the power Internet of Things (IoT), the proposed solutions maintain an extremely high detection accuracy of over 95.8%, demonstrating a balanced and robust defense capability. Specifically, in power simulation attacks, traditional signal strength-based solution A, due to its single criterion, achieves a detection rate of only 18.5% when the attacker precisely controls the transmission power, rendering it almost completely ineffective. Solution D, however, extracts the spatiotemporal fine-grained features of the channel response, rendering such strength simulation methods ineffective. In the most threatening close-range spoofing attack, the extremely close physical proximity of the legitimate party and the attacker leads to a sharp increase in natural channel correlation. In this case, solutions B and C, lacking an effective spatial differentiation dimension, see their detection rates drop to 40.5% and 62.1%, respectively. Solution D, however, fully leverages the active fingerprint construction capability of RIS (Reactive Fingerprint Analysis), forcibly introducing spatial diversity at the physical layer through phase matrix optimization, maintaining a detection rate of 95.8%. Furthermore, when facing dynamic environment attack scenarios, the performance of the traditional solution B, which uses a fixed statistical template, drops sharply to 28.6% because it cannot track environmental drift. In contrast, the solution D of this invention, with its rapid adaptive capability of the meta-learning framework, can instantly capture the time-varying characteristics of the environment and maintain a high level of authentication of 96.4%.

[0073] In summary, the experimental results fully verify the collaborative design concept of the proposed scheme, which actively constructs high-discrimination fingerprints through RIS, achieves rapid small-sample adaptation through meta-learning, and efficiently extracts spatiotemporal features through CNN-GRU. It has achieved breakthroughs over traditional schemes in terms of attack resistance, environmental adaptation efficiency, real-time performance, and lightweight design, and provides a reliable technical solution for secure access of single-antenna, resource-constrained power terminals.

[0074] System Implementation Examples According to embodiments of the present invention, a lightweight power terminal authentication system integrating RIS and meta-learning is provided. Figure 11 This is a schematic diagram of a lightweight power terminal authentication system integrating RIS and meta-learning according to an embodiment of the present invention, as shown below. Figure 11 As shown, the lightweight power terminal authentication system integrating RIS and meta-learning according to an embodiment of the present invention specifically includes: The RIS control and fingerprint acquisition module 1100 is used to control the reconfigurable smart surface RIS to adopt a reflection configuration corresponding to the power terminal according to the identity identifier of the power terminal, and to receive the channel response of the power terminal after reflection by the RIS, and to use the channel response as the enhanced spatial channel fingerprint of the power terminal. The spatiotemporal feature extraction module 1102 is used to extract spatiotemporal features from the enhanced spatial channel fingerprint using a pre-trained lightweight feature extraction network to obtain the corresponding low-dimensional spatiotemporal feature vector. The meta-learning authentication module 1104 is used to input the low-dimensional spatiotemporal feature vector into a lightweight authenticator pre-trained based on the meta-learning framework, and output the authentication result of the power terminal.

[0075] The embodiments of the present invention are system embodiments corresponding to the above method embodiments. The specific operation of each module can be understood by referring to the description of the method embodiments, and will not be repeated here.

[0076] In summary, this invention addresses the security challenges of strong spatial correlation, dynamic environmental changes, and highly covert Sybil attacks faced by wireless access authentication in novel power communication networks and power Internet of Things scenarios. It proposes a lightweight physical layer authentication framework that deeply integrates RIS, meta-learning, and a lightweight CNN-GRU hybrid network. This framework fundamentally solves the inherent defects of insufficient spatial features and low discriminability of single-antenna devices from a physical perspective by constructing a RIS reflection coefficient matrix optimization model with feature distinguishability as the core objective. This proactively constructs a spatial channel fingerprint with strong uniqueness and anti-spoofing capabilities for single-antenna terminals. Furthermore, by introducing the MAML mechanism and combining it with a lightweight CNN-GRU spatiotemporal feature extraction network, the authentication model achieves rapid and accurate adaptation in small-sample scenarios, while significantly improving the system's robustness and scalability in complex time-varying communication environments. Simulation results demonstrate that the proposed scheme maintains over 95% attack detection accuracy in typical Sybil attack scenarios, including power simulation, close-range camouflage, and dynamic environmental interference. Furthermore, the model has low parameter count and inference latency is controlled within milliseconds, perfectly meeting the resource-constrained characteristics of power edge devices and the real-time requirements of power services. The proposed authentication scheme effectively balances authentication security, small-sample adaptation efficiency, and computational overhead, providing a theoretically advanced and engineering-feasible technical path and implementation scheme for lightweight secure access of single-antenna terminals in new power communication networks. Future development can further optimize the RIS dynamic phase modulation strategy and expand its attack resistance dimensions, integrating edge computing, blockchain, and other technologies to construct a distributed collaborative authentication system, and improving loss modeling and environmental adaptability verification in practical deployments.

[0077] Device Example 1 This invention provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it performs the steps described in the method embodiment.

[0078] Device Example 2 This invention provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor, performs the steps described in the method embodiment.

[0079] The computer-readable storage media described in this embodiment include, but are not limited to, ROM, RAM, disk, or optical disk.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A lightweight power terminal authentication method integrating RIS and meta-learning, characterized in that, include: Based on the identity identifier of the power terminal, the reconfigurable smart surface RIS adopts a reflection configuration corresponding to the power terminal, and receives the channel response of the power terminal after reflection by the RIS, and uses the channel response as the enhanced spatial channel fingerprint of the power terminal. The enhanced spatial channel fingerprint is subjected to spatiotemporal feature extraction using a pre-trained lightweight feature extraction network to obtain the corresponding low-dimensional spatiotemporal feature vector. The low-dimensional spatiotemporal feature vector is input into a lightweight authenticator pre-trained based on a meta-learning framework, and the authentication result of the power terminal is output.

2. The method according to claim 1, characterized in that, Based on the identity identifier of the power terminal, the reconfigurable smart surface RIS adopts a reflection configuration corresponding to the power terminal, and receives the channel response of the power terminal after reflection by the RIS. Using the channel response as the enhanced spatial channel fingerprint of the power terminal specifically includes: A unique RIS reflection coefficient matrix is ​​pre-generated for each legitimate power terminal, and a mapping relationship is established between the power terminal's identity and the RIS reflection coefficient matrix. The mapping relationship is queried based on the current power terminal's identity to determine the corresponding target RIS reflection coefficient matrix, and each reflection unit of the RIS is controlled to perform phase adjustment according to the target RIS reflection coefficient matrix; After the RIS completes phase adjustment, it receives the detection signal transmitted by the power terminal; wherein, the detection signal includes a reflected link signal formed by the RIS after reflection in the adjusted phase state and a direct link signal that directly reaches the base station from the power terminal. The reflected link signal and the direct link signal are coherently superimposed at the base station receiver to form a composite channel response. The composite channel response is used as the enhanced spatial channel fingerprint of the power terminal.

3. The method according to claim 2, characterized in that, The pre-optimization and generation of a unique RIS reflection coefficient matrix for each legitimate power terminal specifically includes: With the goal of maximizing the distinguishability of channel fingerprints among different power terminals, the minimum Euclidean distance criterion is adopted to optimize and determine the reflection coefficient matrix for each power terminal that maximizes the minimum Euclidean distance between its normalized channel response vector and the normalized channel response vectors of other power terminals.

4. The method according to claim 1, characterized in that, The enhanced spatial channel fingerprint is subjected to spatiotemporal feature extraction using a pre-trained lightweight feature extraction network to obtain the corresponding low-dimensional spatiotemporal feature vector, specifically including: Enhanced spatial channel fingerprints are collected at multiple consecutive time points to form a time-series channel matrix; The temporal channel matrix is ​​input into a pre-trained lightweight feature extraction network; wherein the lightweight feature extraction network includes a spatial feature extraction unit and a temporal feature extraction unit; The spatial feature extraction unit independently extracts spatial structure features from the composite channel response at each time step and outputs a spatial feature sequence. The temporal feature extraction unit is used to perform temporal dependency modeling on the spatial feature sequence, capture the dynamic evolution law of the composite channel response, and output a low-dimensional spatiotemporal feature vector.

5. The method according to claim 4, characterized in that, The spatial feature extraction unit adopts a one-dimensional depthwise separable convolutional structure to independently extract spatial structural features from the composite channel response at each time step. The temporal feature extraction unit is a gated recurrent unit (GRU), which is used to perform temporal dependency modeling on spatial feature sequences and take the hidden state of the final time step as a temporal context summary. The lightweight feature extraction network also includes a fully connected layer for mapping the temporal context summary into a low-dimensional spatiotemporal feature vector.

6. The method according to claim 1, characterized in that, The low-dimensional spatiotemporal feature vector is input into a lightweight authenticator pre-trained based on a meta-learning framework, and the authentication result of the power terminal is output, specifically including: Obtain the initial parameters of the lightweight validator trained using the Model-Independent Meta-Learning MAML framework during the offline phase; For a newly connected power terminal, a preset number of channel samples of the power terminal are collected as a support set. The initial parameters are updated finitely times using the support set to obtain the exclusive authentication parameters of the power terminal. The authentication threshold of the power terminal is set based on the minimum confidence value of the samples in the support set. Based on the exclusive authentication parameters, the currently extracted low-dimensional spatiotemporal feature vector is input into the lightweight authenticator to calculate the confidence score that the power terminal belongs to a legitimate device. The confidence score is compared with the authentication threshold. If the confidence score is greater than or equal to the authentication threshold, a valid authentication result is output; otherwise, an invalid authentication result is output.

7. The method according to claim 6, characterized in that, The initial parameters for obtaining the lightweight validator trained using the Model-Independent Meta-Learning MAML framework during the offline phase specifically include: In the offline phase, multiple power terminals are randomly selected from the historical dataset to construct meta-tasks, and the channel response measurement samples of each terminal in each meta-task are divided into support set and query set; In the inner loop, the lightweight validator is initialized with the current meta-parameters, the cross-entropy loss is calculated using the support set of each meta-task, and gradient descent is performed for a preset number of steps to obtain the adaptation parameters specific to that meta-task. In the outer loop, the loss of each meta-task on the query set is calculated based on the adaptation parameters, and the current meta-parameters are updated with the goal of minimizing the sum of the losses of all meta-tasks on the query set. The inner and outer loops are executed iteratively until convergence, thus obtaining the initial parameters of the lightweight authenticator.

8. A lightweight power terminal authentication system integrating RIS and meta-learning, characterized in that, include: The RIS control and fingerprint acquisition module is used to control the reconfigurable smart surface RIS to adopt a reflection configuration corresponding to the power terminal according to the identity of the power terminal, and to receive the channel response of the power terminal after reflection by the RIS, and to use the channel response as the enhanced spatial channel fingerprint of the power terminal. The spatiotemporal feature extraction module is used to extract spatiotemporal features from the enhanced spatial channel fingerprint using a pre-trained lightweight feature extraction network to obtain the corresponding low-dimensional spatiotemporal feature vector. The meta-learning authentication module is used to input the low-dimensional spatiotemporal feature vector into a lightweight authenticator pre-trained based on the meta-learning framework, and output the authentication result of the power terminal.

9. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the lightweight power terminal authentication method integrating RIS and meta-learning as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an implementation program for information transmission, which, when executed by a processor, implements the steps of the lightweight power terminal authentication method integrating RIS and meta-learning as described in any one of claims 1-7.

Images