PLA must account for the time-varying characteristics of wireless channels, including fading and multipath effects. Additionally, it should consider the mobility of both the transmitter and receiver. Device movement can cause fluctuations in channel characteristics, influenced by factors such as Doppler shift and coherence time. Fig. 1 depicts the standard Alice-Bob-Eve model in a wireless communication scenario. In this model, Alice, the legitimate transmitter, communicates with Bob, the legitimate receiver, in a multipath scattering environment. Meanwhile, a malicious device, Eve, attempts to impersonate Alice and gain illegitimate access to Bob. Such an attack diminishes the credibility of the legitimate transmitter, posing a threat to the security of the communication system.

Assuming that orthogonal frequency division multiplexing (OFDM) technology is utilized in this study and Ns subcarriers are assigned for communication, let the vectors X→ and Y→ denote the transmitted and received signals, respectively. We have the following equation: (1)Y→=Ha⋅X→+Wwhere Ha represents the channel frequency response (CFR) matrix with dimensions Na×Ns, where Na is the number of antennas and Ns is the number of subcarriers. This matrix can be estimated using pilot signals. The term W represents additive white Gaussian noise. For the i-th antenna and the j-th subcarrier, the CFR Haij is defined as a complex number: (2)Haij=|Haij|ej∠Haij=real(Haij)+j⋅img(Haij)

The channel matrix Haij typically represents the CSI. Changes in the wireless communication environment can affect some signal paths, while others remain invariant. The receiver (Bob) collects the CSI between itself and the transmitter (Alice), and uses this CSI as a reference for identifying legitimate communication. Given that Eve, an eavesdropper within communication range, may attempt to inject illegitimate messages, Bob must determine the legitimacy of a message source based on the real-time CSI characteristics.

To facilitate authentication, Bob labels the CSI data with unique identifiers to distinguish between transmitters. CSI from Eve is labeled as 0, while Alice’s data is labeled as 1. Bob uses historical CSI data and corresponding labels to train the lightweight SE-ConvNet network for authentication. When Bob receives a signal from an unidentified transmitter, he inputs the estimated CSI into the pretrained SE-ConvNet model. By activating the sigmoid function at the end of the fully connected layer, a prediction score mapped to between 0 and 1 is obtained. Finally, the score is compared to the threshold: (3)D={1,if score>ϵ00,if score≤ϵ0where D=1 denotes that there is a legitimate device, and D=0 denotes that there is no legitimate devices and threshold ϵ0 is obtained through experiments.

We set the learning rate to 1×10−3 to strike a balance between fast convergence and training stability. A higher learning rate may cause instability during training, whereas a lower value could result in unnecessarily prolonged training. To mitigate overfitting, we employ a weight decay coefficient of 5×10−4 for regularization, penalizing large weights and promoting simpler models that generalize better to unseen data. The model is trained for 50 epochs, a number selected based on monitoring both training and validation losses to ensure sufficient learning without overfitting. A batch size of 32 is used during training, balancing efficiency and memory constraints; smaller batches introduce noise in gradient updates, while larger ones may exceed memory capacity.

The integration of these hyperparameters is crucial to the training process and significantly impacts the model’s performance in the authentication task. This section first describes the source of the channel dataset, followed by a comparison of the performance of SE-ConvNet and a traditional CNN. During testing, both the SE-ConvNet and CNN models are evaluated by feeding them the test set. The comparison not only involves the probability of the received channel data belonging to the legitimate device (Alice), but also includes a series of other performance metrics.

Recall and False Positive Rate (FPR) are critical metrics for assessing the effectiveness of PLA schemes. Recall, also referred to as True Positive Rate (TPR), indicates the percentage of actual positive instances correctly identified by the model. Conversely, FPR reflects the model’s propensity to misclassify negative samples. A low FPR indicates better discrimination between positive and negative samples. This is particularly important in applications such as security monitoring and fraud detection, where minimizing false alarms is critical. Precision, on the other hand, measures the proportion of samples labeled as positive by the model that are truly positive. The formulas for these metrics are given below: (13)Recall(TPR)=TPTP+FN,FPR=FPFP+TN,Precision=TPTP+FP

In these equations, FP denotes instances where Bob incorrectly classifies Eve’s channel matrix as authenticated. TN indicates the number of illegitimate labels that Eve’s signal should have received. FN represents the number of legitimate channels that Bob misclassifies as illegitimate, while TP signifies the correct identification of legitimate tags from legitimate devices. The accuracy of the model is computed using the following formula: (14)Accuracy=TP+TNTP+TN+FP+FN

To comprehensively assess the overall performance of SE-ConvNet, this study also introduces the F1-score, which combines accuracy and recall. The F1-score offers a comprehensive assessment of the model’s performance, with values ranging from 0 to 1, where a higher score signifies better performance. The formula for the F1-score is as follows: (15)F1-score=2⋅Precision⋅RecallPrecision+Recall

The loss function used for training the model is the binary cross-entropy loss, expressed as: (16)Lcross=−[𝒞log⁡𝒞^+(1−𝒞)log⁡(1−𝒞^)]

In this equation, the classification labels for legitimate and illegitimate devices are assigned values of 1 and 0, respectively. The variable 𝒞 represents the true label of the data, while 𝒞^ indicates the predicted probability that the data comes from a legitimate device.

In this study, we employed the advanced wireless channel generator QuaDRiGa [32] to simulate channel environments and obtain CSI data. The QuaDRiGa platform incorporates a drift model that enables the smooth evolution of small-scale parameters, such as multipath power, delay, departure angles, and arrival angles, over short time intervals as mobile terminals move along a specified trajectory.

The receiver’s movement trajectory in the automobile assembly plant is illustrated in Fig. 6. Our experiment considers a 100 m × 100 m region, where a stationary receiver records channel data as the transmitter moves along different paths, introducing channel heterogeneity. CSI data gathered from various locations within this region are treated as originating from distinct emitters, thereby simulating the CSI of multiple transmitters.

QuaDRiGa generated two datasets for this study: Dataset A, using parameters aligned with [17], and Dataset B, employing parameters tailored for experimental validation as described in Table 1. The QuaDRiGa platform generated a complex CSI matrix of size 10,000×1024, with each CSI instance represented as a 1×1024 complex vector. Each vector was subsequently converted into a 32×32 channel matrix. The dataset was partitioned into three subsets for training, validation, and testing, with a distribution of 80%, 10%, and 10%, respectively. For comparative evaluation, we selected a CNN architecture from [17] as a baseline model.

Since SE-ConvNet requires a specific data format for training and testing, the CSI data must be normalized. Normalization refers to mapping the input values to the range of [0,1] to enhance the learning efficiency of the authenticator. The normalization formula is as follows: (17)H^anorml′(t)=H^al(t)−H^amin(t)H^amax(t)−H^amin(t)

In the 1-D case, Ha(t) is transformed by obtaining its maximum H^amax(t) and minimum H^amin(t) values. The 1-D is denoted as H^al(t), and the normalized value is H^anorml′(t). Moreover, before using the authenticator to identify devices in the wireless communication environment, the authenticator must be trained. Specifically, we input the preprocessed dataset into the authenticator for training to extract deep features, thereby acquiring an authentication model that can represent and differentiate spatiotemporal environmental features at different locations.

When tested on Dataset A, both CNN and the proposed SE-ConvNet achieved high accuracy, as shown in Table 2 below:

Fig. 7 illustrates the authentication accuracy comparison between SE-ConvNet and a traditional CNN across varying SNR conditions on Dataset B. When the SNR is 20 dB, both models achieve over 95% accuracy, demonstrating their effectiveness in high-quality signal environments. However, as the SNR drops to 0 dB, SE-ConvNet maintains high accuracy, whereas the CNN’s performance declines significantly. The robustness of SE-ConvNet at 0 dB SNR can be attributed to the SE module, which enhances the model’s ability to accurately process low-SNR channel data.

To further analyze the training process of SE-ConvNet, Fig. 8 illustrates the evolution of its accuracy and loss curves as the number of epochs increases under different SNR levels. The training and validation losses consistently decrease, while accuracy enhances, demonstrating that the model effectively fits the data with strong generalization ability. These results validate the training efficacy of SE-ConvNet under specific channel conditions and highlight its adaptability to the training data. Moreover, the stable trends suggest that the model avoids overfitting during training, providing a robust foundation for practical applications.

Fig. 9a presents the ROC curve of SE-ConvNet under different signal-to-noise ratios (SNR). At an SNR of 0 dB, the area under the ROC curve (AUC) reaches 0.980, indicating SE-ConvNet’s strong performance in distinguishing legitimate from illegitimate devices even in low-SNR environments. Fig. 9b illustrates the F1-score variations across different thresholds when the SNR is 0 dB. The F1-score remains above 0.9 for thresholds within the [0.5, 0.9] range, peaking near a threshold of 0.85, which signifies optimal classification accuracy under these challenging conditions.

Fig. 10 provides a comparative analysis of FLOPS and memory usage between SE-ConvNet and CNN with an input data size of 2×32×32. FLOPS serves as a key indicator of a model’s computational complexity, with lower values indicating greater efficiency in inference. Memory usage measures the amount of memory required for model parameters and intermediate activations during inference. Time per batch indicates the duration needed to process a specified number of input samples. A lower time per batch reflects higher throughput and faster response times, critical for real-time applications. Analyzing this metric helps identify optimization opportunities for improved computational efficiency. The comparison clearly shows that SE-ConvNet has significantly lower computational complexity than traditional CNN while maintaining accuracy. This finding not only demonstrates SE-ConvNet’s advantages in resource-constrained environments but also indicates its feasibility for large-scale deployment.