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

## TL;DR

This paper introduces a new method to reduce the size of deep learning models used for identifying IoT devices based on their radio signals, without losing much accuracy.

## Contribution

The novel approach combines adversarial learning and polarization regularization to improve pruning efficiency and maintain performance in RFFI tasks.

## Key findings

- The proposed method effectively prunes ResNet18 and VGG16 models used for RFFI.
- It achieves significant reductions in model complexity with minimal loss in identification accuracy.
- The adversarial learning strategy helps recover performance after pruning.

## Abstract

Radio frequency fingerprint identification (RFFI) based on physical-layer characteristics provides a reliable solution for secure authentication of Internet of Things (IoT) devices. Deep neural networks have demonstrated strong capability in improving RFFI performance; however, their high computational complexity and large parameter size pose significant challenges for deployment on resource-constrained edge devices. In RFFI tasks, existing pruning methods often lack effective performance recovery strategies, which leads to noticeable degradation in identification accuracy after pruning. To address this issue, this paper proposes a pruning method based on adversarial learning and polarization regularization. Polarization regularization is applied to learnable soft masks to effectively distinguish channels to be pruned from those to be retained. In addition, an adversarial learning-based performance recovery strategy is introduced to align the output feature distributions between the baseline network and the pruning network, thereby improving identification accuracy after pruning. Experimental results on multiple RFFI datasets demonstrate that the proposed method can effectively prune ResNet18 and VGG16, achieving substantial reductions in model complexity with only minor losses in identification performance.