Residual Channel Boosts Contrastive Learning for Radio Frequency Fingerprint Identification

TL;DR

The paper tackles robust radio frequency fingerprint identification (RFFI) under limited labeled data and unseen environments. It introduces a residual-channel data augmentation strategy that uses least-squares and MMSE channel estimations followed by equalization to generate diverse residual channels for contrastive learning with a lightweight SimSiam framework. The authors show that a mixed LS+MMSE approach yields higher clustering quality (NMI) and faster training, achieving performance close to supervised learning with only 1% labeled data. The method is computationally efficient and well-suited for real-time wireless security applications, reducing data labeling needs and training time while maintaining strong cross-environment generalization.

Abstract

In order to address the issue of limited data samples for the deployment of pre-trained models in unseen environments, this paper proposes a residual channel-based data augmentation strategy for Radio Frequency Fingerprint Identification (RFFI), coupled with a lightweight SimSiam contrastive learning framework. By applying least square (LS) and minimum mean square error (MMSE) channel estimations followed by equalization, signals with different residual channel effects are generated. These residual channels enable the model to learn more effective representations. Then the pre-trained model is fine-tuned with 1% samples in a novel environment for RFFI. Experimental results demonstrate that our method significantly enhances both feature extraction ability and generalization while requiring fewer samples and less time, making it suitable for practical wireless security applications.

Figures (4)

• Figure 1: We employ a lightweight CNN, modified from our previous work pan2024equalization, as the backbone model for SimSiam. The backbone has a parameter size of only 2.58 MB, making it well-suited for IoT deployment. The encoder consists of four blocks, an adaptive pooling layer, a flattened layer, and an L2 normalization layer, which collectively enhance feature extraction and representation learning. The settings for the projection MLP and prediction MLP follow the SimSiam study. Input data in a small mini-batch undergoes various data augmentation. The agreement of equalized signals $x_{LS}$ and $x_{MMSE}$ are expected to be maximized by recovering the same $x_{BB}$ in the present of $\Delta h$.
• Figure 2: Equalized sequences are fed into the pre-trained backbone model, which extracts features for RFFI features. The classification head is trained using the labeled dataset to fine-tune the model for downstream tasks.
• Figure 3: We visualize the extracted features using t-SNE and show the corresponding train and validation loss for the LS-only, Mixed, and MMSE-only methods. The Mixed method achieves the highest average NMI and the shortest average training time compared to MMSE-only, while LS-only performs the worst.
• Figure 4: The SNR ranges from 0 to 15 dB, with a supervised model used as the baseline for comparison.