IEEE Solid-States Circuits Magazine - Fall 2022 - 33

raw I/Q data. The recorded packets
were truncated to the first 1,000 I/Q
points (40 bits) to isolate the fixed
BLE preambles and access addresses.
The recorded data were further processed
to simulate the impact of
quantization and receiver sampling
rate on the RFF system. The lightweight
CNN model was optimized
with post-training quantization to
speed up the inference on IoT edge
devices, and the power usage for
RFF classification was reported on a
Raspberry Pi to estimate the energy
and resource overhead that the RFF
system brings.
Dataset Normalization
Two major
receiver specifications
considered in this example are the
ADC sampling rate and the bit resolution.
To demonstrate the required
ADC sampling rate in the receiver
for sufficient classification accuracy,
we decimate the original raw
data to sampling rates of 1, 2, 5,
10, and 15 MSPS. A higher sampling
rate indicates a higher ADC speed
requirement and a higher ML classifier
input data length, which could
preserve more RFF features from
the signal with a cost of consuming
more power. The bit resolution sets
the precision of the digital domain
representations of the analog signal
and is also critical for determining
the RFF features available to the
classifier. In this case, the data are
quantized to 6 bits, 8 bits, 10 bits,
12 bits, 14 bits, and 16 bits to test
their impact on the classifier accuracy.
Because the ADC sampling
TABLE 1. THE CNN ARCHITECTURE.
LAYER
OUTPUT
DIMENSIONS
Input
16 Ch 1 × 5
Conv, stride = 2
16 Ch 1 × 5
Conv, stride = 2
Fully connected
Fully connected
2 × input size (In)
16 × In/2
16 × In/2
128
220 (for 220 PA
configurations)
Sil6 @ =
The bit resolution sets the precision of the
digital domain representations of the analog
signal and is also critical for determining the
RFF features available to the classifier.
rate and bit resolution are directly
related to the energy consumption
and cost of radios, achieving high
classification accuracy at a low
sampling rate and low bit resolution
is desirable for energy-efficient
RFF usage.
The data normalization is done by
linearly scaling and shifting the raw
packets:
roundc
max^^ hh
Q
absS
Si
6 @
$Qm
,
where S is all of the recorded signals
in the dataset, Si6 @ is a packet in the
recorded signal, Sil6 @ is the corresponding
normalized packet, and
,
Q 21N
=- which scales and shifts
the data to a range of 6-11 , @ with a
desired quantization level of N bits
to become the input to the CNN.
CNNs for RFF Classification
The structure of the CNN model is
shown in Table 1. The first 1,000
packets were selected to transmit
from each PA configuration from the
data recorded at the baseline SNR of
35 dB to form the CNN dataset with
a total size of 220,000. The packets
from each configuration were allocated
with a ratio of 4:3:3 to form
training, validation, and test sets,
respectively. Real and imaginary
samples are taken in by the CNN
as separate channels. The length of
the input vector N is 40*SPS (samples
per BLE symbol), as a result of
each packet being truncated to the
40-symbol-long preamble/access address
sequence. The raw dataset is
processed and normalized using the
aforementioned bit resolutions and
sampling rates. An Adam optimizer
with a learning rate of 0.001 and
binary cross-entropy loss function
was chosen from our experimental
trials for training the CNN model,
with a batch size of 128. The training
process was done with an Nvidia
V100 GPU. Each training epoch took
around 2.7-4 s, depending on the input
vector length. After training for
300 epochs, the model parameters
with the highest validation accuracy
were selected to be tested with the
test set, and the results are shown
in Figure 8.
The classification accuracy approaches
an asymptote for bit resolutions
100
99.5
99
98.5
98
97.5
97
96.5
96
95.5
95
68 10
12
Resolution Bit
FIGURE 8: The CNN model's 220-configuration RFF classification accuracy with different
receiver sampling rates and receiver bit resolutions.
IEEE SOLID-STATE CIRCUITS MAGAZINE
FALL 2022
33
14
SPS = 1
SPS = 2
SPS = 5
SPS = 10
SPS = 15
SPS = 25
16
Classification Accuracy (%)
IEEE Solid-States Circuits Magazine - Fall 2022

Table of Contents for the Digital Edition of IEEE Solid-States Circuits Magazine - Fall 2022
