1: \begin{abstract}

2: We present complex-valued Convolutional Neural Networks (CNNs) for RF

3: fingerprinting that go beyond translation invariance and appropriately account for

4: the inductive bias with respect to multipath propagation channels, a phenomenon that is

5: specific to the fields of wireless signal processing and communications.

6: We focus on the problem of fingerprinting

7: wireless IoT devices in-the-wild using Deep Learning (DL) techniques. Under

8: these real-world conditions, the multipath environments represented in the train

9: and test sets will be different. These differences are due to the physics

10: governing the propagation of wireless signals, as well as the limitations of

11: practical data collection campaigns.

12: 

13: Our approach follows a group-theoretic framework, leverages prior work on DL

14: on manifold-valued data, and extends this prior work to the wireless signal

15: processing domain. We introduce the Lie group of transformations that a signal

16: experiences under the multipath propagation model and define operations that

17: are equivariant and invariant to the frequency response of a Finite Impulse

18: Response (FIR) filter to build a ChaRRNet. We present results using synthetic and

19: real-world datasets, and we benchmark against a strong baseline model, that show

20: the efficacy of our approach. Our results provide evidence of the benefits of

21: incorporating appropriate wireless domain biases into DL models. We hope to

22: spur new work in the area of robust RF machine learning, as the 5G revolution

23: increases demand for enhanced security mechanisms.

24: 

25: \end{abstract}

26: