Considerations for Radio Frequency Fingerprinting across Multiple Frequency Channels

Area of Science:

• Cybersecurity
• Wireless Communication
• Machine Learning

Background:

• Radio Frequency Fingerprinting (RFF) is a proposed method for wireless device authentication.
• Existing RFF models are typically trained and tested on single frequency channels, limiting their effectiveness in multi-channel operations.

Purpose of the Study:

• To evaluate the multi-channel performance of existing single-channel RFF models.
• To develop and validate a novel training data selection technique for creating robust multi-channel RFF models.

Main Methods:

• Evaluated four single-channel RFF models (discriminant analysis, three neural networks) on multi-channel data.
• Characterized performance using the multi-class Matthews Correlation Coefficient (MCC).
• Proposed and tested a new training data selection technique for multi-channel model development.

Main Results:

• Single-channel models showed significant performance degradation (MCC > 0.9 to < 0.05) when tested on unseen channels.
• The proposed multi-channel training technique improved cross-channel average MCC from 0.657 to 0.957, achieving channel-agnostic performance.
• Multi-channel neural networks demonstrated robustness in noisy conditions, maintaining or improving performance compared to single-channel counterparts.

Conclusions:

• Single-channel RFF models are unreliable for multi-channel wireless security.
• The developed multi-channel training strategy enhances RFF performance and robustness across frequency channels.
• Multi-channel neural networks offer a promising solution for secure and adaptable wireless authentication.