KAN-ResNet-Enhanced Radio Frequency Fingerprint Identification with Zero-Forcing Equalization.

Area of Science:

• Cybersecurity
• Signal Processing
• Machine Learning

Background:

• Radio Frequency Fingerprint Identification (RFFI) is crucial for Internet of Things (IoT) security.
• Time-varying channels significantly degrade RFFI accuracy due to device aging and environmental shifts.

Purpose of the Study:

• To develop a robust RFFI method that overcomes time-varying channel interference.
• To enhance the accuracy and generalization capabilities of RFFI for IoT authentication.

Main Methods:

• Implemented Zero-Forcing (ZF) equalization on Wi-Fi preamble signals (IEEE 802.11) to mitigate channel effects.
• Introduced KAN-ResNet, a novel residual network incorporating a KAN module with B-spline basis functions and SiLU activation for complex nonlinear mapping.
• Enhanced model generalization using dynamic B-spline grid updates and L1 regularization.

Main Results:

• Achieved 99.4% accuracy on datasets collected 20 days apart, demonstrating robustness against channel variations.
• Reduced the RFFI error rate from 6.3% to 0.6%, significantly outperforming existing methods.
• The KAN-ResNet effectively improved classification of RFF features.

Conclusions:

• The proposed ZF equalization and KAN-ResNet method significantly enhances RFFI accuracy and reliability in dynamic environments.
• This approach offers a promising solution for secure and robust IoT device authentication.
• The integration of advanced signal processing and deep learning techniques addresses key challenges in RFFI.