Digit. Signal Process.

146 (2024) 104363

Contents lists available at ScienceDirect

Digital Signal Processing

journal homepage: www.elsevier.com/locate/dsp

Radio frequency fingerprint extraction and authentication towards open set

in noisy channels
Renhui Huang a, Xinyong Peng b, Zhi Chai a, Mingye Li a, Jiawei Ren c, Xuelin Yang a, *
a
State Key Laboratory of Advanced Optical Communication Systems and Networks, Shanghai Jiao Tong University, Shanghai, China
b
Guangxi Key Laboratory of Digital Infrastructure, Nanning, China
c
Henan Key Laboratory of Visible Light Communications, Zhengzhou, China

A R T I C L E I N F O A B S T R A C T

Keywords: Radio frequency fingerprint (RFF) identification has emerged as a promising technique for physical-layer identity
Radio frequency fingerprint identification authentication. However, most RFF identification schemes tend to make the restrictive closed-set assumptions
FSST and focus solely on RFF classification. In this paper, we propose a novel open-set RFF identification scheme, with
Supervised contrastive learning
both device classification and rogue device detection capability. Moreover, to reduce the open space risks and
Open set recognition
mitigate the impact of channel noise, an RFF extraction scheme based on Fourier-based synchrosqueezing
transform (FSST) and supervised contrastive learning (SCL) is proposed to acquire more concentrated and
distinctive features. The performance of the proposed scheme is experimentally demonstrated using the datasets
from 11 commercial wireless network interface controllers (WNICs) as well as open-source datasets. The results
verify the significant improvements in high-openness and low-SNR scenarios, with high identification accuracy
of 82 %, at a low SNR of 5 dB, while the openness can be as high as 21 %, and an accuracy of approximately 100
% when SNR>20 dB.

Introduction manufacturing and vary between devices, resulting in a unique RFF for
each device. Thus, RFF can be beneficially utilized as a method for
The proliferation of the Internet of Things (IoT) has led to the identity authentication.
connection of massive wireless devices in communication systems. Traditional RFF identification schemes manually extracted features
However, due to the inherent broadcasting of wireless networks, it is from the received signals, relying on the expert knowledge [5–7]. In
feasible for malicious actors to engage in imitation and illegitimate ac­ recent years, deep learning-based schemes, such as convolutional neural
cess, resulting in privacy leakage and property loss [1]. The existing network (CNN) [8–11], generative adversarial network (GAN) [12,13],
access authentication schemes, which mainly rely on protocols and etc., achieved remarkable recognition accuracy. They can automatically
encryption algorithms, have been proved to be defective [2]. Therefore, extract features and directly classify the devices based on the input
there is an urgent need to develop more advanced and straightforward signals, without the requirement of feature engineering. In addition,
security authentication schemes to keep pace with the continual various signal representations, such as raw I/Q samples [14–16], spec­
expansion of IoT devices. trums [17,18], and time-frequency spectrograms [19,20], were inves­
Radio frequency fingerprint (RFF) identification is a promising tigated as the inputs of the neural networks.
technique for physical-layer identity authentication. It extracts unique However, the existing RFF recognition schemes were mostly limited
features from the distortion of received signals, which is caused by to closed-set classification scenarios, where the prior probabilities of all
transmitter flaws. For instance, the carrier frequency offset (CFO) pre­ potential devices were known, and device categories were consistent
sents as a time-domain phase shift of the signal, and I/Q imbalances during both training and deployment phases. This generally leads to
result in deviations of the constellation diagram. Similarly, the non- invalid identification in practical applications, since a rogue device is
linearity effects inherent to the imperfect power amplifiers and filters likely to be classified as the most similar device in the training set.
were observed in either time-domain or frequency-domain representa­ Therefore, a complete RFF authentication system is required to detect
tions [3,4]. These signal distortions are created during hardware and reject the rogue devices as they appear. Open-set recognition (OSR)

* Corresponding author.
E-mail address: [email protected] (X. Yang).

https://doi.org/10.1016/j.dsp.2023.104363

Available online 23 December 2023

1051-2004/© 2023 Elsevier Inc. All rights reserved.
R. Huang et al. Digital Signal Processing 146 (2024) 104363

schemes were introduced into RFF authentication to address this issue All devices are included in the testing, the legal devices are correctly
[21,22]. In [21], the rouge devices were detected by comparing the identified, while the rogue devices are rejected.
Euclidean distances with the known devices in training, using the preset In the training stage, an encoder is pre-trained as the RFF extractor
thresholds. Machine learning algorithms like K-Nearest Neighbors based on the SCL scheme. The RFF extractor maps the FSST spectrogram
(KNN) were further employed based on the Euclidean distances of RFF to a feature vector, which can be regarded as the hardware fingerprint.
features to improve distance measurement [22]. Nevertheless, these Subsequently, with the extracted features, an open-set classifier is
OSR schemes relied mainly on algorithmic or conventional machine trained based on the DOC scheme. Simultaneously, the decision
learning techniques and still suffered from feature dispersion due to boundaries of each legal device are obtained.
noise and increased openness. In the test stage, RFFs of the test devices are extracted using the
In this paper, we aim to reduce the impact of channel noise and trained extractor and are identified by the trained open-set classifier.
improve recognition accuracy in the scenarios with high openness. We The obtained decision boundaries of legal devices are utilized for rogue
propose an RFF extraction scheme that leverages Fourier-based syn­ device detection. Any features that fall outside all decision boundaries
chrosqueezing transform (FSST) and supervised contrastive learning are designated as rogue devices.
(SCL) to generate more concentrated and distinctive RFFs. Specifically, When registering a new device, the RFF extractor does not need the
the received signals are initially transformed into spectrograms through retraining, while the open-set classifier must undergo retraining. In
FSST. Then, an extractor, pre-trained by SCL, extracts RFFs from the order to meet the demands of the IoT devices, we design a lightweight
FSST spectrograms. Contrastive learning was initially proposed for the classifier, employing a fully connected network with only one hidden
unsupervised [23,24] or semi-supervised [25] tasks, which results in layer. However, to harness the advantages of SCL, the retraining of RFF
suboptimal model performance. Hence, based on the practical RFF extractor can lead to a performance improvement. Given that the feature
application scenarios, we incorporate supervised information into the extractor model is relatively large, training it in the cloud environment is
contrastive learning process to enhance the feature extraction capabil­ a viable consideration. Further details of the proposed scheme are pro­
ities of the model. vided in the subsequent section.
Moreover, based on the extracted RFFs, a novel OSR scheme called
deep open classification (DOC) is employed for both legal device clas­ Methodology
sification and rogue device detection. A one-vs-rest sigmoid layer is
utilized as the output layer to mitigate the risk of open space, and Signal processing
Gaussian fitting is used to further tighten the rejection decision
boundaries. The proposed scheme is demonstrated by extensive exper­ Assuming the ideal transmitted signal is x(t), the received signal can
imental verifications, with excellent accuracy in legal device classifica­ be expressed as
tion, as well as outstanding capability in rogue device detection.
r(t) = f (x(t)) ∗ h(t) + n(t) , (1)
Compared to traditional machine learning algorithms [21], the pro­
posed scheme exhibits enhanced capabilities in terms of nonlinear where f(⋅) represents the distortion caused by the hardware impair­
fitting and generalization, while eliminating the necessity for predefined ments, h(t) represents the channel impulse response, the symbol ∗ de­
threshold settings. notes the convolution operation and n(t) represents the channel noise.
This paper is structured as follows. Section II provides an overview of Here, the distortion model is assumed to be stationary. However, it has
the proposed RFF authentication system. In Section III, we present the been found that the distortion may vary with time or temperature [26],
specific details of the signal processing, RFF extraction, and RFF leading to variations in the RFF features.
authentication methods. Section IV is dedicated to conducting a The received RF signals are pre-processed first, including synchro­
comprehensive evaluation of the proposed scheme and analyzing the nization, preamble extraction, CFO compensation, equalization, and
experimental results in detail. Finally, Section V summarizes the key normalization. A brief overview of these algorithms is provided below,
findings and concludes the paper. and the specific descriptions can be found in [27,28].

System overview 1) Synchronization: The synchronization process serves to locate the

precise beginning of a packet, and an imprecise synchronization will
The proposed RFF authentication scheme is depicted in Fig.1. introduce noise into the samples. The synchronization algorithm is
Initially, a large number of labeled signals are collected from different implemented with a double sliding window, and the autocorrelation
sources of legal and rogue devices. The signals are then suitably pre­ of the preamble is utilized to estimate the starting point of the packet.
processed and transformed into FSST spectrograms. All the legal devices 2) Preamble Extraction: To avoid the model learning identity-related
are involved in the training stage, and the model records the features of information, such as MAC addresses, only the preamble is extrac­
legal devices, while the rogue devices do not participate in this process. ted for RFF identification, rather than utilizing the entire packet.
Consequently, the training procedure resembles a registration process. 3) CFO Compensation: As indicated by relevant research [26], tem­
perature changes significantly affect the crystal oscillators, and any
drift in oscillator frequency can significantly degrade the overall
performance of the system. To avoid this problem, coarse CFO
compensation and further fine CFO compensation are performed
utilizing the preamble.
4) Equalization: Since there is a communication channel h(t), and signal
transmission in time-varying channels may experience significant
distortions, which can affect the performance. Equalization helps to
mitigate the impact of channels. In the proposed method, the Mini­
mum Mean Square Error (MMSE)-based algorithm is employed for
channel equalization.
5) Normalization: Normalization prevents the model from learning
power-related characteristics, which are more related to distance

Fig. 1. Framework of the proposed RFF authentication method.

2
R. Huang et al. Digital Signal Processing 146 (2024) 104363

than the device itself. The packets are normalized by dividing their regarded as negative pairs. The extra labeling information maximizes
root mean square. the feature extraction capabilities of the model.
The diagram of the SCL model is presented in Fig. 3. Given a batch of
The pre-processing is briefly described by labeled samples {xi , li }i∈I , a group of augmented samples {̃ xi , li }i∈I are
generated firstly, where xi represents the input data of the model (FSST
s(t) = Pre(r(t)) . (2)
spectrogram of the processed signal), li is its label, and I represents the
After that, the pre-processed signal is subjected to the FSST. FSST is a index set. During the data augmentation phase, a random level of
time-frequency (TF) analysis algorithm, which resembles a type of Gaussian white noise is added to the input samples to force the model to
reassignment method used for the short-time Fourier transform (STFT) extract noise-independent features. Subsequently, the encoder network
spectrogram [29]. For the signal s(t), the STFT is defined as maps the ̃ xi to a representation yi , and the projection network further
∫ maps yi to the computation space of SCL loss, which is represented as zi .
S(f , t) = s(τ)g(t − τ)e− 2πif (τ− t) dτ, (3) The encoder is optimized by minimizing the SCL loss and forward
propagated calculation, where the SCL loss is formulated as
where fand t denote frequency and time parameters respectively, and ∑∑ 1
( / )
exp zi ⋅zj τ
g(t − τ) is a sliding window. L SCL = − log ∑ , (6)
|P(i)| =i exp(zi ⋅zk /τ)
FSST transforms the TF coefficients according to the map (f,t)→( w i∈I j∈P(i) k∈I,k∕
̂ (f,
t), t), where w ̂ is defined by
where z is the output of the projection network and i, j, k are the
( )
1 ∂t S(f , t) indices. i is the index of the anchor sample, P(i) represents the index set
w
̂ (f , t) = Re . (4)
2πi S(f , t) of all positive samples in the batch, and |P(i)| denotes its cardinality. τ
refers to a temperature parameter. The SCL loss minimizes the distance
Here, Re(⋅) represents the real part of a complex value, and i is the
between the positive samples in the embedding space, while maximizing
imaginary unit.
the distance between the negative samples.
Then the FSST is given as
In addition, a larger batch size often leads to better results because a
∫
1 larger batch sizes introduces more positive and negative sample pairs.
T(ω, t) = S(f , t) δ(ω − w
̂ (f , t))df , (5)
g(0) However, the batch size is constrained by GPU/TPU memory, so a trade-
off needs to be found between the performance and available memory.
where δ(⋅) denotes the Dirac function, and ω is the frequency parameter The architecture of the encoder is depicted in Fig. 4. Seven con­
of FSST. volutional layers, an average pooling layer, and a dense layer make up
FSST redistributes the energy of the time-frequency spectrogram by the encoder and rectified linear unit (ReLU) is applied as the activation
moving the energy to nearby high-energy points. This process serves to function. The projection network is composed of a single linear layer
highlight the primary components of the signal. Fig.2 shows the time- with a size of 64, and the open-set classifier is a linear network that
frequency spectrogram generated by STFT and FSST, respectively. The contains one dense hidden layer with a size of 64.
signal originates from the short training field (STF) segment in Wi-Fi
preamble with a SNR of 20 dB. The application of FSST results in a
more concentrated TF representation and a sharper spectrogram, which
can effectively alleviate the impact of noise.

RFF extraction based on SCL

In this phase, the SCL scheme is applied to pre-train an encoder as the

RFF extractor. The objective of the encoder is to transform the input into
an embedding vector, in which the features from the same device are
aggregated, while those from different devices are separated [30]. In
contrast to self-supervised contrast learning, where positive pairs are
created through data augmentation, SCL uses the samples that have the
Fig. 3. Framework of the applied SCL model.
same label as the positive pairs, while those have different labels are

Fig. 2. Time-frequency spectrogram. (a) STFT; (b) FSST.

3
R. Huang et al. Digital Signal Processing 146 (2024) 104363

Algorithm 1
The training procedure of DOC.
Input: Training set D.
Output: Predicted probabilities p(yj = li ), i, j ≤ m; threshold ti for each class.
1. Set hyperparameters: epochs; batch size; α = 3.
2. Train the model by minimizing L DOC .
3. Get the predicted probabilities p(yj = li ).
4. for each class k = 1 : m, do
5. for each point p(yj = li ), a corresponding mirror point 1 + (1 − p(yj = li )).
6. Estimate the standard deviation σk by Gaussian fitting.
7. Get the probability threshold tk = max(0.5, 1 − ασk ).
8. end

Fig. 4. The architecture of the encoder.

Algorithm 2
RFF authentication based on DOC The test procedure of DOC.
Input: Test set, containing the samples that do not belong to any classes in training.
A complete RFF authentication system requires both functions of Output: Predicted label l ∈ {l1 , l2 , ..lm , rejection}.
device classification and rogue detection. The majority of existing RFF 1. Get the predicted probabilities p(yj = li ) using the trained model.
identification schemes operate under the closed-set assumption, which 2. for each sample k = 1 : n, do
requires that the test devices must be present in training. Ideally, the RFF 3. the predicted probabilities p(yj = li ), i ≤ m, are compared with the threshold σi .
4. if all p(yj = li ) < σi , i ≤ m :
classifier should not only classify the test devices into the correct pre­
5. the label of sample k = rejection.
sented classes, but also detect those that do not belong to any existing 6. else:
class. Accordingly, a novel deep open-set classification approach called 7. the label of sample k = max(p(yj = li ), i ≤ m).
DOC is employed to ensure the rogue device detection. 8. end
In conventional multi-class classifiers, the final output layer typically
uses softmax, which is unable to detect unknown classes. To address this
test (DUTs). An Ettus USRP N210 was utilized for signal reception, as
issue, for the given training set D = {(r1 ,y1 ),(r2 ,y2 ),...,(rN ,yN )}, where rn
shown in Fig. 5. The USRP N210 was configured with the center fre­
is the extracted RFF and yi ∈ {l1 , l2 , ..lm } is the label information, DOC
quency of 2.462 GHz, the bandwidth of 20 MHz, and the sample rate of
approach constructs a (m + 1)-classes classifier with the classes {l1 , l2 , .
20MSps. The WNICs emitted Wi-Fi signals based on 802.11n protocol
.lm , rejection} to indicate the unknown classes. In addition, a 1-vs-rest
with the carrier frequency of 2.462 GHz, the bandwidth of 20 MHz. Both
layer containing m sigmoid functions is leveraged as the final output
the transmitter and receiver were stationary and were positioned at a
layer to export the predicted probabilities [31]. The objective of the
distance of approximately 1 m from each other with a line of sight (LOS)
model training is to minimize the total logarithmic loss of m sigmoid
channel. Over 1000 packets were collected for each DUT and the pre­
functions, which is expressed as
ambles, which included the STF and the long training field (LTF) were
m ∑
∑ N
( ) ( ) extracted. With the aforementioned sample rate, each sample has a
L DOC = − I yj = li logp yj = li length of 320. All the samples were randomly divided into a training set
and a test set, with a ratio of 7:3.
i=1 j=1

( ) ( ( )) For the data augmentation stage in SCL, Gaussian white noise with
= li log 1 −
− I yj ∕ p yj = li , (7)
SNR between 10 and 20 dB was added to the signals, and the temper­
where I is the indicator function and p(yj = li ) represents the probability ature parameter τ in SCL loss was set to 0.1. The models were trained for
output from i-th sigmoid function on the j-th sample. 200 epochs with a batch size of 256, using the Adam optimizer in Keras,
Instead of using a uniform probability threshold, we apply a tight­ and the initial learning rate was set to 0.001.
ened decision boundary for each class by the Gaussian fitting to further In the experiments, all conditions remained consistent except for the
reduce the open space risk. In the training, we assume that the predicted variations in signal representations and methods used. This included
probabilities points of each class follow a half of Gaussian distribution factors such as data volume and the parameter settings. In addition to
centered around a probability of 100 %. For each existing point p(yj = the unavoidable environmental noise (>35 dB), several levels of artifi­
li ), a corresponding mirror point is artificially created as 1 + (1 − p(yj = cial noise were introduced to the raw signal to simulate the identifica­
tion performance in noisy channels.
li )). The standard deviation σi of i-th class is then estimated using both
the existing points and the created points, and the probability threshold
ti = max(0.5, 1 − ασ i ). We set α = 3, which is commonly used in sta­
tistics and has been verified to be effective in our experiments. Note that
different classes li obtain a different threshold ti due to Gaussian fitting.
In the test, all the predicted probabilities from m sigmoid functions
are compared with the threshold. If all the predicted probabilities are
lower than the corresponding thresholds, the sample will be rejected;
otherwise, its prediction class will be the one with the highest proba­
bility. The detailed training and test procedure of the DOC is described
in Algorithm 1 and Algorithm 2.

Experiments and analysis

Experimental setup

A total number of 11 IEEE 802.11n-based wireless network interface

controllers (WNICs) with the same model were used as the devices under
Fig. 5. Experimental devices. (a) WNICs; (b) USRP N210.

4
R. Huang et al. Digital Signal Processing 146 (2024) 104363

Closed-set identification √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅

2 × NTR
openness = 1 − , (8)
In closed-set identification, all the test devices appear in the training NTR + NTE
stage, which means the device category in the training and testing sets
where NTR represents the number of device classes involved in training
are the same. The recognition accuracy achieved by the proposed
and NTE represents that contained in the test set, NTR ≤ NTE . The higher
method is presented in Fig. 6. We compared the performance of the FSST
openness means more unknown devices appear in the test.
spectrogram to that of the raw I/Q signals and the STFT spectrogram. To
Normalized accuracy (NA) is calculated by weighing up the accuracy
evaluate the feature extraction capability of SCL, we compared the
of known devices (AKS) and unknown devices (AUS). NA is expressed as
proposed method to the scheme that does not utilize SCL, where the
follows
encoder connects to the classifier directly, and the entire model is
∑N
optimized by a cross-entropy loss (which makes up a conventional i=1 (TPi + TNi )
ResNet model). Additionally, we train a traditional closed-set classifier AKS = ∑N , (9)
(TPi + TNi + FPi + FNi )
using softmax and compare its performance with the proposed open-set
i=1

classifier. FU
The results indicate that all schemes attain high accuracy when SNR AUS = , (10)
TU + FU
reaches 20 dB, but a significant accuracy degradation appears in the case
of SNR reduction. Compared with raw I/Q signals and STFT spectro­ NA = λAKS + (1 − λ)AUS. (11)
gram, FSST spectrogram maintains higher identification accuracy when
Here, TP, TN FP and FN represent the true positive, false positive,
SNR decreases, indicating that FSST alleviates the impact of noise
true negative, and false negative known samples, respectively, while TU
effectively. Furthermore, it can be demonstrated that the model trained
and FU denote true unknown and false unknown samples, respectively.
by SCL can extract more distinguishable RFFs and achieves better
N is the number of known devices and λ is a regularization constant,
recognition accuracy than the schemes without SCL, especially at low
where 0 < λ < 1. Normalized accuracy provides a valid performance
SNRs. Finally, the proposed open-set classifier performs similarly to the
assessment when the unknown devices play the dominant role.
traditional closed-set classifier in the closed-set recognition task, but it
To assess the validity and reliability of the proposed OSR scheme, we
can also handle OSR tasks.
conducted extensive experiments in varying degrees of openness and
To evaluate the effectiveness of the proposed RFF extraction scheme,
SNRs. The results are presented in Figs. 8 and 9. We compared the
we used the t-Distributed Stochastic Neighbor Embedding (t-SNE) al­
proposed RFF extraction scheme with the scheme without both FSST and
gorithm to visualize the feature distribution, when the SNR is 10 dB, and
SCL and the scheme without only SCL. The scheme without FSST means
the number of devices is five. Fig. 7 presents the results of the visuali­
the raw I/Q samples are served as the signal representation, while the
zation, where the features extracted from the different devices are rep­
scheme without SCL means that the RFF extractor is trained in a tradi­
resented with different colors. The visualization clearly shows that the
tional supervised learning manner. Meanwhile, we compared the per­
FSST provides more concentrated features than raw I/Q signals, while
formance of the proposed DOC approach with the classical OSR
the model trained in SCL manner extracts features with a compact intra-
algorithm OpenMax [33]. The results confirmed that the proposed
class distance and a separable inter-class distance, explaining the higher
scheme achieves significant performance improvement in the scenarios
accuracy of the proposed method at low SNRs and high openness.
of high openness and low SNRs, since more concentrated and distinctive
RFFs are extracted. Moreover, in the case of high openness, the proposed
Open-set identification
DOC approach outperforms the other OSR algorithm in terms of iden­
tification accuracy, demonstrating that it reduces the open space risk
Since identification in an open set quite differs from that in a closed
effectively.
set, openness and normalized accuracy are adopted to better describe
In order to verify the generalization capabilities of the proposed
the OSR problem and assess the performance accurately [32]. Openness
method, we conducted experiments on additional open-source datasets,
is calculated as
namely, FIT/CorteXlab dataset [34] and ADS-B dataset [35]. The former
dataset was collected from USRP N2932 devices transmitting random
OFDM packets, while the latter was obtained from moving aircraft in the
real world. We also compare the proposed DOC scheme with the Open­
Max algorithm, and OSR algorithm proposed in [21] and [22]. The re­
sults are presented in Table 1. As anticipated, the proposed DOC scheme
provides better recognition accuracy, especially in the case of high
openness. It can be proved that the proposed scheme retains effective­
ness on various datasets.

Conclusion

In this paper, we propose and demonstrate a novel RFF extraction

and authentication scheme, where Fourier-based synchrosqueezing
transform and supervised contrastive learning are applied to extract the
distinctive hardware features, and a deep open classifier is employed for
legal device classification and rouge device detection. The excellent
performances of the proposed RFF scheme are experimentally verified
with high identification accuracies at low SNRs and high openness,
using the dataset collected from 11 practical WNICs as well as some
open-source datasets. The proposed RFF scheme is an excellent candi­
date in the practical applications of security enhancement via
identification.

Fig. 6. Closed-set recognition accuracy.

5
R. Huang et al. Digital Signal Processing 146 (2024) 104363

Fig. 7. Feature visualization. (a) raw I/Q signals; (b) FSST spectrogram; (c) features extracted by ResNet; (d) features extracted by SCL scheme.

Table 1
Recognition accuracy on open-source datasets.
openness 0 0.1 0.2 0.3 0.4 0.5

FIT/CorteXlab (proposed) 1 0.98 0.95 0.90 0.81 0.74

FIT/CorteXlab (OpenMax) 1 0.94 0.83 0.78 0.60 0.46
FIT/CorteXlab (proposed in [21]) 1 0.88 0.67 0.46 0.30 0.14
FIT/CorteXlab (proposed in [22]) 1 0.93 0.80 0.65 0.43 0.25

ADS-B 1 0.99 0.98 0.96 0.93 0.90

(proposed)
ADS-B (OpenMax) 1 0.98 0.92 0.88 0.85 0.78
ADS-B (proposed in [21]) 1 0.97 0.84 0.70 0.59 0.32
ADS-B (proposed in [22]) 1 0.98 0.90 0.85 0.77 0.60

Credit author statement

Renhui Huang: Conceptualization, Methodology, Software, Writing,

Xinyong Peng: Investigation. Zhi Chai: Visualization. Mingye Li: Su­
pervision, Software. Jiawei Ren: Validation: Xuelin Yang: Writing,
Reviewing and Editing

Fig. 8. Open-set recognition accuracy varies with openness, where SNR=15 dB.
Declaration of Competing Interest

The authors declare the following financial interests/personal re­

lationships which may be considered as potential competing interests:
Xuelin Yang reports financial support was provided by National Natural
Science Foundation of China (No. 62371281). Xuelin Yang reports
financial support was provided by China Aeronautical Science Fund (No.
2022Z066057002). Xuelin Yang reports financial support was provided
by Open Projects of Guangxi Key Laboratory of Digital Infrastructure
(No. GXDIOP2023011). Xuelin Yang reports financial support was pro­
vided by Guangxi Information Center, Nanning, China, and Henan Key
Laboratory of Visible Light Communications (No. HKLVLC2023-B09),
Zhengzhou, China. If there are other authors, they declare that they
have no known competing financial interests or personal relationships
that could have appeared to influence the work reported in this paper.
Fig. 9. Open-set recognition accuracy varies with SNR, when openness=0.16
(six devices for training and 11 devices for testing). Data availability

Data will be made available on request.

6
R. Huang et al. Digital Signal Processing 146 (2024) 104363

Acknowledgments [21] G. Shen, J. Zhang, A. Marshall, J.R. Cavallaro, Towards scalable and channel-
robust radio frequency fingerprint identification for LoRa, IEEE Trans. Inf.
Forensics Secur. 17 (2022) 774–787.
This work was partly supported by National Natural Science Foun­ [22] R. Xie, W. Xu, Y. Chen, J. Yu, A. Hu, A generalizable model-and-data driven
dation of China (No. 62371281), China Aeronautical Science Fund (No. approach for open-Set RFF authentication, IEEE Trans. Inf. Forensics Secur. 16
2022Z066057002), Open Projects of Guangxi Key Laboratory of Digital (2021) 4435–4450.
[23] Z. Wu, Y. Xiong, S.X. Yu, D. Lin, Unsupervised Feature Learning via Non-
Infrastructure (No. GXDIOP2023011), Guangxi Information Center, parametric Instance Discrimination, in: 2018 IEEE/CVF Conference on Computer
Nanning, China, and Henan Key Laboratory of Visible Light Communi­ Vision and Pattern Recognition, Salt Lake City, UT, USA, 2018, pp. 3733–3742.
cations (No. HKLVLC2023-B09), Zhengzhou, China. [24] X. Liu, et al., Self-supervised learning: generative or contrastive, IEEE Trans.
Knowl. Data Eng. 35 (1) (2023) 857–876, 1 Jan.
[25] Z. Wu, F. Wang, B. He, Specific emitter identification via contrastive learning, IEEE
References Commun. Lett. 27 (4) (2023) 1160–1164. April.
[26] W. Hou, X. Wang, J. Chouinard, A. Refaey, Physical layer authentication for mobile
[1] V. Mathy, F. Piessens, Key reinstallation attacks: forcing nonce reuse in WPA2, in: systems with time-varying carrier frequency offsets, IEEE Trans. Commun. 62 (5)
Proceedings of the 2017 ACM SIGSAC Conference on Computer and (2014) 1658–1667. May.
Communications Security, 2017. [27] J. Li, Guoqing Liu, G.B. Giannakis, Carrier frequency offset estimation for OFDM-
[2] Y. Zou, J. Zhu, X. Wang, L. Hanzo, A survey on wireless security: technical based WLANs, IEEE Signal Process. Lett. 8 (3) (2001) 80–82. March.
challenges, recent advances, and future trends, Proc. IEEE 104 (9) (2016) [28] P.H. Moose, A technique for orthogonal frequency division multiplexing frequency
1727–1765. Sept. offset correction, IEEE Trans. Commun. 42 (10) (1994) 2908–2914. Oct.
[3] Y. Shi, M. Jensen, Improved radiometric identification of wireless devices using [29] T. Oberlin, S. Meignen, V. Perrier, The Fourier-based syn-chrosqueezing transform,
MIMO transmission, IEEE Trans. Inf. Forensics Secur. 6 (4) (2011) 1346–1354. Dec. in: 2014 IEEE International Conference on Acoustics, Speech and Signal Processing
[4] J. Zhang, R. Woods, M. Sandell, M. Valkama, A. Marshall, J. Cavallaro, Radio (ICASSP), 2014, pp. 315–319.
frequency fingerprint identification for narrowband systems, modelling and [30] K. He, H. Fan, Y. Wu, S. Xie, R. Girshick, Momentum contrast for unsupervised
classification, IEEE Trans. Inf. Forensics Secur. 16 (2021) 3974–3987. visual representation Learning, in: 2020 IEEE/CVF Conference on Computer Vision
[5] J. Hua, H. Sun, Z. Shen, Z. Qian, S. Zhong, Accurate and efficient wireless device and Pattern Recognition (CVPR), 2020, pp. 9726–9735.
fingerprinting using channel state information, in: IEEE INFOCOM 2018 - IEEE [31] S. Lei, H. Xu, and B. Liu. "Doc: deep open classification of text documents." arXiv
Conference on Computer Communications, 2018, pp. 1700–1708. preprint arXiv, 2017.
[6] P. Liu, P. Yang, W. Song, Y. Yan, X. Li, Real-time identification of rogue WiFi [32] C. Geng, S. Huang, S. Chen, Recent advances in open set recognition: a survey, IEEE
connections using environment-independent physical features, in: IEEE INFOCOM Trans. Pattern Anal. Mach. Intell. 43 (10) (2021) 3614–3631. Oct.
2019 - IEEE Conference on Computer Communications, 2019, pp. 190–198. [33] A. Bendale, T.E. Boult, Towards open set deep networks, in: 2016 IEEE Conference
[7] A.C. Polak, D.L. Goeckel, Wireless device identification based on RF oscillator on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 1563–1572.
imperfections, IEEE Trans. Inf. Forensics Secur. 10 (12) (2015) 2492–2501. Dec. [34] C. Morin, L.S. Cardoso, J. Hoydis, J.M. Gorce, T. Vial, Transmitter classification
[8] G. Shen, J. Zhang, A. Marshall, M. Valkama, J.R. Cavallaro, Toward length- with supervised deep learning. Cognitive Radio-Oriented Wireless Networks, 2019,
versatile and noise-robust radio frequency fingerprint identification. IEEE Trans. pp. 73–86. Aug.
Inf. Forensics Secur. 18, 2023, pp. 2355–2367. [35] T. Ya, L. Yun, Z. Haoran, J. Zhang, W. Yu, G. Guan, M. Shiwen, Large-scale real-
[9] J. Yu, A. Hu, G. Li, L. Peng, A robust RF fingerprinting approach using world radio signal recognition with deep learning, Chin. J. Aeronaut. 35 (9) (2022)
multisampling convolutional neural network, IEEE Internet Things J. 6 (4) (2019) 35–48. Sep.
6786–6799. Aug.
[10] L. Peng, J. Zhang, M. Liu, A. Hu, Deep learning based RF fingerprint identification Renhui Huang is currently pursuing his M.S. with the State Key Laboratory of Advanced
using differential constellation trace figure, IEEE Trans. Veh. Technol. 69 (1) Optical Communication Systems and Networks, Shanghai Jiao Tong University. He
(2020) 1091–1095. Jan. received his B.S. degree in Information Engineering in 2021 from Donghua University. His-
[11] R. Das, A. Gadre, S. Zhang, S. Kumar, J. Moura, A deep learning approach to IoT research interests include wireless communications, wireless physical layer security, and
authentication, in: 2018 IEEE International Conference on Communications (ICC), deep learning
2018, pp. 1–6.
[12] J. Gong, X. Xu, Y. Lei, Unsupervised specific emitter identification method using
radio-frequency fingerprint embedded InfoGAN, IEEE Trans. Inf. Forensics Secur. Xinyong Peng is currently employed in the Joint Innovation Laboratory of Digital Guangxi
15 (2020) 2898–2913. Smart Infrastructures, Guangxi, China. His-research interests include network security and
[13] J. Gong, X. Xu, Y. Qin, W. Dong, A generative adversarial network based big data applications
framework for specific emitter characterization and identification, in: 2019 11th
International Conference on Wireless Communications and Signal Processing Zhi Chai is currently pursuing his Ph.D. degree with the State Key Laboratory of Advanced
(WCSP), 2019, pp. 1–6. Optical Communication Systems and Networks, Shanghai Jiao Tong University. He
[14] J. Yu, A. Hu, F. Zhou, Y. Xing, Y. Yu, G. Li, L. Peng, Radio frequency fingerprint received his B.S. degrees in Communication Engineering from Xidian University in 2021.
identification based on denoising autoencoders, in: 2019 International Conference His-research interests include physical-layer security, digital signal processing, secure key
on Wireless and Mobile Computing, Networking and Communications (WiMob), generation and distribution and machine learning.
2019, pp. 1–6.
[15] K. Sankhe, M. Belgiovine, F. Zhou, S. Riyaz, S. Ioannidis, K. Chowdhury, ORACLE:
optimized radio classification through convolutional neural networks, in: IEEE Mingye Li is currently pursuing his Ph.D. degree with the State Key Laboratory of
INFOCOM 2019 - IEEE Conference on Computer Communications, 2019, Advanced Optical Communication Systems and Networks, Shanghai Jiao Tong University.
pp. 370–378. He received his master’s degree in Optical Engineering from Nanjing University of In­
[16] B. He, F. Wang, Cooperative specific emitter identification via multiple distorted formation, Science&Technology in 2022. His-research interests include physical-layer
receivers, IEEE Trans. Inf. Forensics Secur. 15 (2020) 3791–3806. security, digital signal processing, optical network safety and machine learning.
[17] W. Wang, Z. Sun, S. Piao, B. Zhu, K. Ren, Wireless physical-layer identification:
modeling and validation, IEEE Trans. Inf. Forensics Secur. 11 (9) (2016) Jiawei Ren is currently employed in Henan Key Laboratory of Visible Light Communica­
2091–2106. Sept. tions, Zhengzhou, China. His-research interests include deep learning and communication
[18] L. Kang, L. Zhang, X. Huang, W. Hu, X. Yang, Hardware fingerprint authentication security.
in optical networks assisted by anomaly detection, IEEE Photonics Technol. Lett.
34 (19) (2022) 1030–1033. Oct.
[19] G. Shen, J. Zhang, A. Marshall, L. Peng, X. Wang, Radio frequency fingerprint Xuelin Yang is currently a Professor with the State Key Laboratory of Advanced Optical
identification for LoRa using spectrogram and CNN, in: IEEE INFOCOM 2021 - Communication Systems and Networks, Shanghai Jiao Tong University. His-interests
IEEE Conference on Computer Communications, 2021, pp. 1–10. mainly focused on ultrafast all-optical signal processing in optical fiber communication,
[20] J. Zhang, F. Wang, O. Dobre, Z. Zhong, Specific emitter identification via applications of semiconductor optical amplifiers, security of optical networks optical
Hilbert–Huang transform in single-hop and relaying scenarios, IEEE Trans. Inf. orthogonal frequency division multiplexing transmission, and passive optical networks.
Forensics Secur. 11 (6) (2016) 1192–1205. June. He serves on the Editorial Board of Optics & Quantum Electronics.