PaperlayerView exact revision on arXiv

Source-linked AI summary

APEG: Adaptive Physical Layer Authentication with Channel Extrapolation and Generative AI

Xiqi Cheng, Rui Meng, Xiaodong Xu, Haixiao Gao, Ping Zhang, Dusit Niyato

arXiv:2603.21923v1eess.SP

TL;DR

Dynamic wireless channels challenge reliable physical-layer authentication because existing methods may not adapt to changing CSI distributions. APEG addresses this with collaborator-conditioned channel extrapolation and generative models, with CADM achieving the strongest reported fingerprint-generation metrics while CCMDM offers faster convergence.

  • Problem

    Existing PLA schemes struggle to maintain robust authentication under rapidly time-varying channel conditions and shifting CSI distributions.

  • Method

    APEG uses nearby collaborators’ CSI fingerprints as conditions for channel extrapolation and GAI-based generation of a target device’s CSI fingerprints, including CCMDM and CADM.

  • Results

    CADM achieves SSIM 0.960 and cosine similarity 0.996 at 20 dB SNR, while the diffusion-based schemes show stronger training stability than CAVAE.

  • Takeaways & Limitations

    The framework shows potential for robust collaborative authentication in time-varying 6G IIoT scenarios.

Abstract

from arXiv · show

With the rapid advancement of 6G, identity authentication has become increasingly critical for ensuring wireless security. The lightweight and keyless Physical Layer Authentication (PLA) is regarded as an instrumental security measure in addition to traditional cryptography-based authentication methods. However, existing PLA schemes often struggle to adapt to dynamic radio environments. To overcome this limitation, we propose the Adaptive PLA with Channel Extrapolation and Generative AI (APEG), designed to enhance authentication robustness in dynamic scenarios. Leveraging Generative AI (GAI), the framework adaptively generates Channel State Information (CSI) fingerprints, thereby improving the precision of identity verification. To refine CSI fingerprint generation, we propose the Collaborator-Cleaned Masked Denoising Diffusion Probabilistic Model (CCMDM), which incorporates collaborator-provided fingerprints as conditional inputs for channel extrapolation. Additionally, we develop the Cross-Attention Denoising Diffusion Probabilistic Model (CADM), employing a cross-attention mechanism to align multi-scale channel fingerprint features, further enhancing generation accuracy. Simulation results demonstrate the superiority of the APEG framework over existing time-sequence-based PLA schemes in authentication performance. Notably, CCMDM exhibits a significant advantage in convergence speed, while CADM, compared with model-free, time-series, and VAE-based methods, achieves superior accuracy in CSI fingerprint generation. The code is available at https://github.com/xiqicheng192-del/APEG

I. Introduction

6G’s dynamic, heterogeneous wireless environments make identity authentication increasingly important, while cryptographic methods face complexity and resource limitations. PLA reduces these burdens but must overcome degraded reliability under rapidly changing channels, motivating APEG’s collaborative channel-extrapolation and GAI approach.

  • Cryptography-based authentication requires key management and intensive computation, creating latency and resource burdens for future high-mobility, low-power 6G devices.
  • PLA complements cryptographic authentication by using physical features such as CSI fingerprints, avoiding encryption overhead and frequent key updates.
  • Related work: Existing CSI-based PLA evolved from statistical hypothesis testing to ML classifiers and deep learning for device discrimination and channel adaptation.
  • Research challenge: Rapid channel and environmental changes shift CSI distributions and temporal correlations, weakening fixed thresholds and classifier decision boundaries.
  • Proposed direction: APEG uses nearby collaborators’ CSI fingerprints as conditions for channel extrapolation, enabling GAI to generate legitimate users’ real-time fingerprints in dynamic environments.
  • System model: The system models Alice as the device requiring identification, Jack as a verified nearby collaborator, Bob as the receiver, and Eve as a spoofing attacker.

C. Proposed Adaptive PLA Scheme With Channel Extrapolation and GAI (APEG)

APEG trains a GAI model to learn the joint CSI representation of a dynamic target device and a known neighboring collaborator. During authentication, the collaborator’s CSI conditions prediction of the target’s fingerprint for identity decisions.

  • Bob uses Jack’s known CSI with the trained model to predict Alice’s real-time CSI fingerprint and support authentication decisions.
  • APEG trains a GAI model on Alice’s and Jack’s estimated CSI fingerprints to capture their spatially correlated joint representation.

1) Training Phase:

The training workflow collects paired CSI fingerprints, converts them into real-valued image tensors, and trains a GAI model for later fingerprint prediction and authentication. Frequency-domain image representations preserve phase information while enabling convolutional modeling across antennas and subcarriers.

  • Training and authentication workflow: Training collects Alice’s and Jack’s CSI fingerprint samples across different time slots, while authentication receives signals and obtains Jack’s CSI.
  • Training and authentication workflow: The trained GAI model predicts Alice’s CSI fingerprint from Jack’s CSI for subsequent authentication.
  • Authentication decision: Authentication compares predicted and received fingerprints using SSIM, PSNR, cosine similarity, and NMSE, then ranks samples by distance.
  • Continuous operation: The workflow uses Jack as an environmental anchor for offline spatial-correlation learning, online extrapolation, and selective fine-tuning against environmental drift.
  • CSI fingerprints as images: Alice’s and Jack’s CSI fingerprints become paired M × N images with real and imaginary channels, forming real-valued tensors for 2D convolution.
  • Domain selection: The frequency domain preserves full-resolution amplitude and phase information, whereas Angle-Delay transformations may introduce leakage and quantization errors.

B. Proposed CCMDM-based Fingerprint Generation Scheme

CCMDM generates Alice’s CSI fingerprints by applying masked diffusion to paired Alice–Jack fingerprints, preserving Jack’s collaborator fingerprint as a clean condition. Its reverse denoising process learns to reconstruct Alice’s fingerprint from progressively corrupted channel data.

  • Model construction: CCMDM transforms Alice’s and Jack’s CSI fingerprints into a paired two-channel image for conditional fingerprint generation.The fingerprints are concatenated along the last dimension, while masking preserves the collaborator region as clean during diffusion.
  • Reverse diffusion: The reverse process starts from a Gaussian state and progressively denoises it to reconstruct Alice’s original CSI fingerprint.Each reverse transition is modeled as a learned conditional Gaussian distribution, with the network predicting noise from the noisy input and diffusion timestep.
  • Forward diffusion: The forward diffusion process progressively adds Gaussian noise to Alice’s region while leaving Jack’s region clean.The mask is defined as m = [1_A, 0_J], and the forward process supports closed-form sampling at arbitrary diffusion steps.
  • Training objective: CCMDM trains by minimizing the discrepancy between the true reverse transition and the learned reverse process using KL divergence.The variational objective approximates the true reverse trajectory through a Gaussian-chain formulation and an ELBO-based loss.

C. Proposed CADM-based Fingerprint Generation Scheme

CADM improves collaborator-conditioned CSI generation by explicitly modeling dependencies between Jack’s and Alice’s features through cross-attention. The mechanism extracts multi-scale Jack features and uses them to guide iterative denoising of Alice’s fingerprint.

  • Conditional generation: CADM uses Jack’s CSI fingerprint as characteristic-aligned conditioning information rather than merely concatenating it with Alice’s fingerprint.The conditioning signal is intended to guide generation toward realistic and structure-consistent CSI outputs for Alice.
  • Generation procedure: CADM extracts multi-scale features from Jack’s CSI fingerprint and uses them while denoising Alice’s noisy fingerprint.The generation phase repeatedly encodes Jack’s fingerprint and applies the denoising update before returning Alice’s generated CSI fingerprint.
  • Design rationale: CADM’s predicted denoising mean is more biased toward the distribution indicated by the conditioning signal than the corresponding mean in the preceding approach.This design explicitly addresses CCMDM’s limited modeling of the statistical dependency between Alice’s and Jack’s fingerprints.
  • Cross-attention mechanism: Cross-attention captures dependencies between collaborator and user features during denoising.Each cross-attention layer uses four attention heads and can query structural features from Jack’s region at each spatial position.

D. Complexity Analysis of the Proposed Schemes

The complexity analysis represents the model input as two concatenated CSI fingerprints and defines the main architectural quantities used to characterize computation.

  • Complexity setup: The input tensor concatenates two CSI fingerprints along the final dimension, producing size (C, H, 2W).The analysis denotes P as the channel multiplication sequence and L as the number of layers.

1) Complexity Analysis of CCMDM:

The complexity analysis decomposes CCMDM computation across its time-embedding, convolutional, encoder, decoder, bottleneck, and attention components. It retains dominant convolution and attention terms for an aggregate approximation.

  • Convolutional components: The first convolution has complexity O(9BC_bCS), where C is the input-channel count and S = H × 2W is the spatial resolution.The kernel size is 3 × 3, and B and C_b denote batch size and base channel size.
  • Encoder and decoder: CCMDM complexity includes encoder and decoder costs across L layers, with layer-specific channel counts C_l = C_b · p_l and spatial resolutions S_l.The encoder also accounts for the number of ResNet blocks and whether self-attention is used at each layer.
  • Bottleneck: The bottleneck contributes a separate computation term involving the deepest-layer channel and spatial dimensions.The analysis expresses this component using B, C_L, S_L, γ, and C_b.
  • Aggregate complexity: The total CCMDM complexity is approximated by retaining dominant convolutional and attention terms across the network.C_max and S_max denote the maximum channel count and spatial resolution, while T is the number of diffusion steps.
  • CADM comparison: CADM processes Alice’s and Jack’s fingerprints through dual encoder-decoder branches and adds cross-attention at every ResNet block.Self-attention remains present, so each block includes both self-attention and cross-attention costs on H × W feature maps.

2) Complexity Analysis of CADM:

CADM’s computational complexity includes encoder, bottleneck, decoder, and attention terms, with dual branches and cross-attention increasing the dominant cost.

  • CADM’s encoder complexity includes self-attention and cross-attention terms for each block.
  • The bottleneck uses two ResNetBlocks per branch and one cross-attention operation.
  • The decoder complexity is expressed separately before aggregating all network components.
  • The dominant total complexity is O(2Cmax^2 Smax + 4Smax^2), where Cmax and Smax denote maximum channel number and spatial resolution.The factor 2 reflects the dual-branch design, while the factor 4 reflects adding cross-attention to self-attention.

3) Latency and Power Consumption Analysis of CADM:

The authors evaluate CADM inference on a workstation and estimate both latency and energy consumption for authentication.

  • CADM inference was evaluated on an NVIDIA RTX A6000 GPU and Intel Xeon w7-2495X CPU.
  • Inference Latency: 9.03 ms per denoising step is measured on average, and 20-step DDIM reduces total end-to-end inference time to approximately 180.6 ms.The estimate contrasts with standard diffusion’s 1,000 steps.
  • Inference Latency: 180.6 ms falls within the sub-second authentication latency requirements stated for many 6G applications.
  • Energy Consumption: 525 W is the estimated peak workstation power, yielding approximately 94.8 J for one 180.6 ms authentication instance.The energy estimate is described as a conservative upper bound based on general-purpose workstation hardware.

A. Simulation Settings

The simulation uses dynamic DeepMIMO channels and evaluates CADM with a collaborator-conditioned U-Net, multiple baselines, and CSI-generation and authentication metrics.

  • Dataset and Scenario: The dataset contains 12,000 CSI samples from continuous Alice and Jack trajectories, split into training and testing sets at a 9:1 ratio.
  • Dataset and Scenario: Alice and Jack move in the same direction at constant speed with fixed relative distance, while five Eves are randomly placed around Alice and k is set to 0.5.
  • CADM Architecture: CADM uses a U-Net with encoder, bottleneck, and decoder components containing ResNet modules and attention mechanisms.
  • CADM Architecture: The model takes noisy channel ht and collaborator channel hJ as inputs, predicts noise, and removes it to obtain the denoised channel.
  • CADM Architecture: CADM uses a dual-branch encoder, cross-attention, and multi-resolution fusion to model the spatial mapping from Jack to Alice.These mechanisms target alignment and reconstruction of multipath channel details.
  • Comparison Schemes: The comparison schemes include GRU-based PLA, LSTM-based PLA, direct collaborator fingerprints, and CAVAE-based fingerprint generation.
  • Evaluation Metrics: CSI generation is assessed with SSIM, PSNR, Cosine Similarity, and NMSE, while authentication uses F1 score and authentication error rate RE.SSIM captures perceptual structure, PSNR pixel-level amplitude differences, Cosine Similarity vector direction, and NMSE normalized amplitude error.
  • Evaluation Metrics: F1 combines Precision and Recall, which are computed from true positives, false positives, and false negatives.

D. Simulation Results of Fingerprint Prediction

CADM reconstructs frequency-domain fingerprints accurately, while CCMDM converges faster; CADM maintains quality across noise levels and benefits from four attention heads.

  • Fingerprint Reconstruction: CADM frequency-domain fingerprints closely match ground truth in structural distribution and amplitude range for outdoor and indoor scenarios.
  • Fingerprint Reconstruction: Angle-delay outputs show slight background noise but recover key signal positions because the domain is sparse.
  • Loss Convergence: CCMDM stabilizes around 200 epochs, whereas CADM converges after approximately 1200 epochs and LSTM and GRU converge most rapidly.
  • Loss Convergence: CADM’s longer training reflects the additional optimization burden of cross-attention, while its added training cost yields superior fingerprint generation quality.
  • Noise Robustness: PSNR distributions vary minimally across SNR levels, indicating consistently high structural fidelity under different noise conditions.
  • Noise Robustness: At 20 dB, NMSE values concentrate in lower negative ranges, while the 5 dB and 20 dB distributions differ by only about 0.1 at NMSE = −15 dB.The reported negative dB values correspond to smaller normalized reconstruction errors.
  • Attention-Head Ablation: Increasing cross-attention heads from 2 to 4 substantially improves SSIM, whereas increasing from 4 to 8 provides marginal gain and adds computational overhead.
  • Attention-Head Ablation: Four attention heads are selected as the optimal configuration for balancing generation performance and computational complexity.

5) Fingerprint Generation Performance Compared with Comparative schemes:

CADM provides the strongest CSI fingerprint generation and authentication performance across the evaluated metrics and SNR conditions. Its cross-attention-based feature alignment preserves channel structure and fidelity more effectively than time-series, VAE-based, and other extrapolation schemes.

  • Fingerprint generation: At 20 dB SNR, CADM achieves SSIM 0.960 and cosine similarity 0.996, outperforming GRU, LSTM, CAVAE, CCMDM, and CA.These results indicate strong preservation of spatial structure and directional consistency in generated CSI fingerprints.
  • Fingerprint generation: CADM exceeds 31 dB PSNR and achieves NMSE below -20 dB, demonstrating strong signal fidelity and error suppression.Its PSNR and NMSE outperform temporal models and CAVAE, while CCMDM lacks comparable fine-grained feature integration.
  • Authentication performance: CADM maintains an F1 score of 1.0 across SNR levels when authentication uses cosine similarity and PSNR.GRU, LSTM, CAVAE, and CCMDM remain lower or less consistent, while CA performs worst in the comparisons.
  • Authentication performance: CADM maintains an F1 score of 1.0 across all SNR conditions when authentication uses SSIM.CA performs acceptably at 5 dB because shared environmental noise increases local structural similarity, but its performance weakens as SNR increases.
  • Authentication performance: CADM holds Re = 0 from 5 dB to 20 dB under Euclidean-distance authentication, whereas comparative schemes show errors or fluctuations.GRU declines to roughly 0.30, LSTM rises to about 0.43 at 20 dB, and CA remains near 1.0 across the SNR range.
  • Framework design: The framework combines collaborator-supported channel extrapolation with CCMDM and CADM to enable dynamic device identification and align auxiliary features.The stated design uses a collaborator near Alice and CADM’s feature alignment for user-domain channel extrapolation.
Loading 2603.21923v1…