Enhancing WiFi CSI Fingerprinting: A Deep Auxiliary Learning Approach

Yong Huang, Wenjing Wang, Dalong Zhang, Junjie Wang, Chen Chen, Yan Cao, Wei WangIEEE Internet of Things Journal 2025 · DOI

TL;DR

This paper introduces CSI2Q, a novel system that significantly enhances WiFi Channel State Information (CSI) fingerprinting for device identification. By transforming CSI into a time-domain signal and leveraging deep auxiliary learning from readily available IQ data, CSI2Q achieves performance comparable to traditional IQ-based methods, even in challenging "open-world" scenarios where unknown devices may appear.

What Problem Does This Paper Address?

Radio Frequency (RF) fingerprinting offers a promising alternative to cryptography for device authentication in wireless networks, exploiting unique hardware imperfections. However, existing RF fingerprinting techniques often rely on specialized, costly equipment to capture in-phase and quadrature (IQ) samples. While recent advancements allow the use of Channel State Information (CSI) from commercial WiFi devices, these methods struggle with two key challenges: the "coarse granularity" of CSI measurements (meaning they lack fine-grained detail) and "open-world recognition" (the ability to identify devices not seen during training). CSI2Q aims to bridge this gap by enabling accurate device identification using readily available CSI, overcoming these limitations.

Methods

CSI2Q employs a multi-stage approach:

  • CSI Measurement Transformation:
    • Phase Unwrapping and Correction: Raw CSI phases are processed to ensure continuity and correct for jitters caused by hardware imperfections and environmental factors.
  • Cyclic Shift Division: A novel scheme is used to mitigate channel interference by dividing CSI measurements, aiming to isolate the transmitter's unique RF fingerprints.
  • Time-Domain Sample Generation: The processed CSI data is transformed into a high-dimensional time-domain feature vector that shares the same feature space as preamble IQ samples, addressing the coarse granularity issue.
  • Deep Auxiliary Learning:
    • Auxiliary IQ Dataset: A dataset of IQ samples from commercial WiFi devices is used to train a source feature extractor ($E_s$) and an IQ discriminator ($D$).
  • Knowledge Transfer: The learned feature extraction capabilities of $E_s$ are transferred to a target feature extractor ($E_r$) for CSI data via an auxiliary network ($A_T$). This process aims to imbue the CSI model with the ability to capture fine-grained details that might be missed by CSI alone.
  • CSI Classifier ($C$): A classifier is trained on the enhanced CSI features.
  • Open-World Recognition (OpenMax):
    • An OpenMax function is integrated into the CSI classifier's output layer. This function calibrates the uncertainty of known devices and estimates the likelihood of unknown devices, enabling robust identification even when encountering devices not present in the training set.

Datasets

The paper utilizes a combination of existing and synthetically generated datasets:

  • Auxiliary IQ Dataset: Derived from the public WiSig dataset, containing IQ samples from 174 commercial WiFi devices.
  • Synthetic CSI Dataset: Generated from the auxiliary IQ dataset by performing standard channel estimation, comprising CSI measurements from 85 devices.
  • In-Lab CSI Dataset: Collected from 10 off-the-shelf WiFi routers in a controlled environment using the PicoScenes tool.
  • In-The-Wild CSI Dataset: Collected from commercial WiFi access points in an uncontrolled, real-world environment, spanning four months and 25 transmitters.

Key Insights and Ideas

  • CSI and IQ Data are Correlated: The paper demonstrates that CSI and IQ data, despite their differences, share underlying correlations that can be exploited for improved fingerprinting.
  • Transforming CSI is Crucial: Directly using raw CSI is challenging due to its coarse granularity. Transforming it into a time-domain signal that aligns with IQ data's feature space is a key enabler for better performance.
  • Auxiliary Learning Bridges the Gap: Transferring knowledge from a well-trained IQ fingerprinting model to a CSI model significantly boosts the CSI model's ability to extract subtle, device-specific features.
  • Open-World Recognition is Achievable: The integration of OpenMax effectively addresses the challenge of identifying unknown devices, making the system more practical for real-world deployment.

Key Results

CSI2Q demonstrates significant performance improvements across various datasets:

  • Synthetic CSI Dataset:
    • Closed-world accuracy: 79% to 96%
    • Open-world accuracy: 72% to 95%
  • In-Lab CSI Dataset:
    • Closed-world accuracy: 66% to 86%
    • Open-world accuracy: 54% to 82%
  • In-The-Wild CSI Dataset:
    • Closed-world accuracy: 74% to 90%
    • Open-world accuracy: 57% to 82%

The paper also shows that the TCN (Temporal Convolutional Network) architecture, when used as the feature extractor, offers a good balance of performance and computational efficiency.

Limitations and Future Work

  • Limitations:
    • The performance of CSI fingerprinting is still influenced by device brands and their unique RF imperfections.
  • While the system shows improvement, the accuracy for unknown devices in the in-the-wild scenario can still be moderate (e.g., 44.73% to 69.37%).
  • The paper acknowledges that the specific deep learning model architectures are not fixed, suggesting CSI2Q is a framework that can be generalized.
  • Future Work:
    • The paper's framework can be extended to other neural network architectures.
  • Further research could explore more advanced techniques for handling complex channel conditions and improving the robustness of open-world recognition.

Summary updated

BibTeX
@article{huang2025enhancing,
  author = {Yong Huang and Wenjing Wang and Dalong Zhang and Junjie Wang and Chen Chen and Yan Cao and Wei Wang},
  title = {{Enhancing WiFi CSI Fingerprinting: A Deep Auxiliary Learning Approach}},
  journal = {IEEE Internet of Things Journal},
  year = {2025},
  doi = {10.1109/jiot.2025.3625062},
}
Figure