CrossRF: A Domain-Invariant Deep Learning Approach for RF Fingerprinting

Abstract: Radio Frequency (RF) fingerprinting offers a promising approach for drone identification and security, although it suffers from significant performance degradation when operating on different transmission channels. This paper presents CrossRF, a domain-invariant deep learning approach that addresses the problem of cross-channel RF fingerprinting for Unmanned Aerial Vehicle (UAV) identification. Our approach aims to minimize the domain gap between different RF channels by using adversarial learning to train a more robust model that maintains consistent identification performance despite channel variations. We validate our approach using the UAVSig dataset, comprising real-world over-the-air RF signals from identical drone models operating across several frequency channels, ensuring that the findings correspond to real-world scenarios. The experimental results show CrossRF's efficiency, achieving up to 99.03% accuracy when adapting from Channel 3 to Channel 4, compared to only 26.39% using conventional methods. The model maintains robust performance in more difficult multi-channel scenarios (87.57% accuracy adapting from Channels 1,3 to 2,4) and achieves 89.45% accuracy with 0.9 precision for controller classification. These results confirm CrossRF's ability to significantly reduce performance degradation due to cross-channel variations while maintaining high identification accuracy with minimal training data requirements, making it particularly suitable for practical drone security applications.

• The paper introduces a novel domain-invariant RF fingerprinting framework using adversarial training and knowledge distillation to mitigate cross-channel degradation.
• The methodology employs cascaded 1D convolutional layers with a GRL and multi-stage training, achieving up to 99.03% accuracy in single-channel adaptation.
• The results validate that CrossRF significantly improves UAV detection, yielding up to a 72% accuracy gain over traditional methods in varied RF environments.

Authoritative Analysis of "CrossRF: A Domain-Invariant Deep Learning Approach for RF Fingerprinting" (2505.18200)

Context and Motivation

RF fingerprinting has become a prominent technique in device identification due to its capacity to exploit hardware-level signal imperfections, offering a protocol-independent solution for non-invasive UAV detection. However, its practical deployment for drone security is undermined by significant performance degradation when models are applied across different frequency channels, a challenge directly tied to the domain gap induced by channel-specific signal characteristics. The paper systematically addresses the robustness limitations in cross-channel fingerprinting by leveraging adversarial domain adaptation, targeting the identification of UAVs under variable transmission channels.

Related Work and Research Gap

Recent literature has established deep learning as highly effective for RF fingerprinting, especially using raw I/Q data for device discrimination [Yang2022]. Prior approaches have shown susceptibility to channel-induced distribution shifts, leading to near-random classification accuracy in cross-channel settings [rff]. While adversarial domain adaptation methods, such as ADDA [Tzeng2017], have been deployed successfully in vision applications, their integration into RF fingerprinting frameworks—especially for channel-invariant drone identification—is not yet widely explored. Furthermore, most existing studies consider cross-device adaptation or static channel conditions, failing to address the real-world operational needs for robust identification when RF environments are dynamic.

Methodological Framework

The proposed CrossRF architecture is grounded in ADDA, comprising distinct source and target encoders, an adversarial discriminator featuring a Gradient Reversal Layer (GRL) [Ganin2014], and a classifier. The encoders are constructed from cascaded 1D convolutional layers (with BN, LeakyReLU, dropout), optimized for RF signal feature extraction. The GRL acts to encourage domain invariance via adversarial training, allowing the discriminator to guide the encoders toward extracting features invariant with respect to channel-induced shifts.

Training proceeds in two stages. The initial stage optimizes the source encoder and classifier on labeled source channel data. The adaptation stage applies adversarial loss and knowledge distillation (with KL divergence), weighted by $\lambda_{adv}$ and $\lambda_{distill}$, to preserve source knowledge while aligning the target encoder to the channel-invariant feature space. Hyperparameter optimization via Optuna refines architectural and learning parameters.

The real-world UAVSig dataset [Zhao2024] underpins the experimental protocol, capturing signals from multiple identical UAV models and controllers operating on four distinct 2.4 GHz ISM channels, thus providing a high-fidelity testbed for cross-channel adaptation assessment.

Experimental Findings

CrossRF demonstrates pronounced gains in cross-channel generalization. Strong numerical results include:

• Single-channel adaptation: CrossRF achieves up to 99.03% accuracy when adapting from Channel 3 to Channel 4, a substantial improvement over conventional baseline performance (26.39%). Channel 1 to Channel 2 adaptation yields 81.94%.
• Multi-channel adaptation: Training on Channels 1 and 3, then adapting to Channels 2 and 4, yields 87.57% accuracy—a notable performance under increased domain shift complexity.
• Controller classification: Standard classifier achieves 89.45% accuracy with 0.9 precision and 0.89 recall for multiclass controller identification, obviating the need for domain adaptation.

These results validate the efficacy of adversarial adaptation: significant mitigations of performance drop that commonly occurs with channel switching, and robust identification maintained with limited target domain labeling.

Claims and Contradictions

The paper asserts that CrossRF can "significantly reduce performance degradation due to cross-channel variations while maintaining high identification accuracy with minimal training data requirements." This claim is substantiated by strong empirical evidence, showing up to a 72% absolute accuracy improvement in cross-channel settings, contradicting traditional RF fingerprinting approaches that require extensive retraining or fail entirely under channel variation.

Practical and Theoretical Implications

The domain-invariant methodology has direct implications for operational drone security systems: models trained on one channel can be deployed across multiple channels without catastrophic accuracy loss, streamlining maintenance and reducing retraining overhead. This approach addresses real-world requirements, where channel conditions are variable and interference is prevalent.

Theoretically, the integration of adversarial discriminative adaptation and knowledge distillation provides an effective framework for domain-invariant representation learning in non-vision signal domains. The design principle may inform broader applications in wireless security, IoT device authentication, and spectrum management where domain shifts are inherent.

PyReason analysis further reveals factor relationship networks underlying adaptation success, suggesting distinctive solution strategies for single- vs multi-channel challenges. This points toward logical reasoning tools as valuable complements to empirical deep learning studies.

Speculation on Future Developments

Future trajectories in AI-based RF fingerprinting will likely include:

• Expansion to broader channel multiplexing, incorporating more dynamic interference modeling.
• Joint adversarial adaptation for multiple environmental factors (not limited to channel, but including hardware receiver variance).
• Integration with federated or continual learning paradigms to accommodate evolving RF landscapes with minimal manual intervention.
• Employment of lightweight architectures or edge deployment optimization for real-time drone security in constrained environments.

Such developments may further the applicability and scalability of RF fingerprinting in heterogeneous wireless environments.

Conclusion

CrossRF introduces a domain-invariant adversarial adaptation framework for robust UAV identification across varied RF channels, overcoming fundamental limitations of traditional fingerprinting models. The approach is validated on authentic multi-channel datasets and demonstrates substantial accuracy improvements in challenging operational conditions. The research establishes a foundation for practical, deployable, and theoretically robust RF fingerprinting systems under real-world channel variability, with implications for UAV security and further applications in wireless authentication domains (2505.18200).