Abstract
Recently, Artificial Intelligence (AI)-driven Physical-Layer Authentication (PLA), which focuses on achieving endogenous security and intelligent identity authentication, has attracted considerable interest. When compared with Discriminative AI (DAI), Generative AI (GAI) offers several advantages, such as fingerprint augmentation, reconstruction, and denoising. Inspired by these innovations, this paper provides a systematic exploration of GAI's integration with PLA. We commence with a concise review of identity authentication techniques and GAI models. Then, we contrast the limitations of DAI with the potential of GAI in addressing PLA challenges. Specifically, we introduce a structured taxonomy for GAI-enhanced PLA methodologies, encompassing three key stages: fingerprint collection, model training, and performance optimization within the PLA pipeline. Furthermore, we propose a novel PLA framework based on GAI and channel extrapolation for dynamic environments. To demonstrate GAI's efficacy in enhancing PLA robustness, we implement a case study using the Generative Diffusion Model (GDM). Finally, we outline potential future research directions for GAI-based PLA.
Introduction
As we transition toward 6G and the "Internet of Everything," traditional upper-layer security measures face a scaling crisis. Physical-Layer Authentication (PLA) offers a promising alternative by utilizing unique hardware "fingerprints"—such as Radio Frequency (RF) characteristics or Channel State Information (CSI)—to verify device identities. However, traditional AI-based PLA predominantly relies on Discriminative AI (DAI), which faces significant hurdles: data scarcity in real-world environments, extreme sensitivity to noise, and an inability to adapt to dynamic environmental changes.
Generative AI (GAI) represents a paradigm shift for PLA. Unlike DAI, which focuses on learning decision boundaries, GAI learns the underlying probability distributions of legitimate signals. This enables robust identity verification even in noisy, dynamic environments with limited labeled data.
Taxonomy of GAI-Enhanced PLA
This research proposes a structured taxonomy for integrating GAI into the PLA pipeline, categorized into three primary stages: fingerprint collection, model training, and performance optimization.
- Fingerprint Collection: GAI models like Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) can augment sparse datasets. Additionally, Transformers can reconstruct missing fingerprint data caused by signal interference or deep fading.
- Model Training: GAI facilitates unsupervised attack detection and "Open-Set Recognition," allowing systems to identify unknown attackers without requiring prior examples of their specific signal characteristics.
- Performance Optimization: Techniques such as "machine unlearning" help models forget the features of eliminated users, while parameter fine-tuning allows rapid adaptation to new scenarios without the overhead of full retraining.
Proposed Framework: GDM and Channel Extrapolation
The centerpiece of this work is a novel PLA framework combining Generative Diffusion Models (GDM) with channel extrapolation. In dynamic environments, a user's physical fingerprint changes rapidly. The proposed framework learns the correlation between a legitimate user's fingerprint and a collaborator's fingerprint. By using the collaborator's signal as a "condition," the GDM can predict the legitimate user's expected signal in a new environment, effectively denoising the authentication process and maintaining high fidelity.
Experimental Evaluation
The framework was evaluated using the DeepMIMO datasets, covering both "O1" Outdoor (large-scale, open) and "I3" Indoor (complex, high-interference) scenarios.
Key Results
The GDM-based scheme demonstrated exceptional performance compared to traditional baselines such as VAEs and time-series models (LSTM, GRU).
- Accuracy and Robustness: The proposed scheme maintained a perfect F1 score of 1.0 and an authentication error rate of 0 across a wide range of Signal-to-Noise Ratios (SNR), from 5 dB to 20 dB.
- Fidelity: Visual analysis confirmed that the GDM-generated fingerprints were structurally and energetically almost identical to true physical fingerprints.
- Computational Feasibility: While more intensive than simple DAI models, the system achieved an inference time of 180.6 ms per instance using a Denoising Diffusion Implicit Model (DDIM) accelerator, making it viable for high-security edge nodes.
Discussion and Future Directions
GAI offers inherent defensive advantages; for instance, GANs trained in a minimax game are naturally more resilient to adversarial perturbations. However, the computational intensity of GDM remains a challenge for resource-constrained IoT devices. Future research will focus on lightweight design through quantization, the integration of physical propagation laws into training, and the use of digital twins to enhance model generalization. By transitioning from drawing boundaries to learning distributions, GAI-enhanced PLA provides a robust foundation for secure 6G wireless communication.
Summary updated