Polarization-Regularized Adversarial Pruning for Efficient Radio Frequency Fingerprint Identification on IoT Devices

Area of Science:

• Cybersecurity
• Machine Learning
• Internet of Things (IoT)

Background:

• Radio Frequency Fingerprint Identification (RFFI) offers robust security for IoT devices.
• Deep neural networks enhance RFFI but face deployment challenges on resource-limited edge devices due to high complexity.
• Current pruning methods for RFFI models often result in significant accuracy degradation.

Purpose of the Study:

• To develop an effective pruning method for deep neural networks in RFFI applications.
• To address the performance recovery issue in pruned RFFI models.
• To enable efficient deployment of RFFI on constrained edge devices.

Main Methods:

• Proposed a novel pruning method integrating adversarial learning and polarization regularization.
• Utilized polarization regularization with learnable soft masks to differentiate channels for pruning.
• Implemented an adversarial learning strategy to align feature distributions and recover performance.

Main Results:

• Successfully pruned ResNet18 and VGG16 models on RFFI datasets.
• Achieved substantial reductions in model complexity.
• Demonstrated only minor losses in identification performance post-pruning.

Conclusions:

• The proposed method effectively reduces model complexity for RFFI.
• Adversarial learning and polarization regularization enhance performance recovery after pruning.
• The technique facilitates the deployment of accurate RFFI on resource-constrained IoT devices.