IEEE - Aerospace and Electronic Systems - June 2022 - 27

Bose
Figure 7.
Authentic and spoofed mixed signal datasets generation.
OFFLINE TRAINING
The offline training is conducted using a Software-Defined
Radio (SDR) [21], as shown in Figure 8. The training datasets
are fed to a frontend that prepares the correct two-bit
digital signal for an SDR. Next, the combined authentic
and spoofed signals are supplied to the acquisition and
tracking loops of the SDR. Once the track lock is established,
the receiver signal and measurement processing
blocks create the GPS features that are extracted for inputs
to the NN. The NN output is then compared to the desired
output to form an error signal. This error signal is used in
the training algorithm discussed previously to adjust the
weights, including biases. To validate, multiple datasets
with different delays and amplitudes are used.
ONLINE TESTING
The NN is trained offline using a combination of direct
authentic and delayed false signals. Once the NN is trained,
it is ready for online testing, as shown in Figure 9. A GPS
simulator acquires the authentic GPS signals, which are then
fed into the Mixed Signal Generator shown in Figure 7 to
create the combined authentic and spoofed signals. These
combined signals can be transmitted by a GPS spoofer transmitter.
The targeted GPS receiver gets the direct authentic
signals from the LOS GPS satellites as well as the delayed
spoofing signals from the spoofer's transmitter. The spoofing
signals are of slightly higher power than the authentic signals.
Consequently, the targeted receiver will track most if
not all of the spoofed signals. The spoofing detector extracts
features to be inputted to the NN, which continuously investigates
the presence of spoofing signals, detects the spoofing
signals, and then alerts the receiver ofthe presence ofspoofing
signals. As long as spoofing exists, the process is
repeated. In all tests, the generated spoofing signals and their
corresponding authentic signals can be analyzed offline.
SPOOFING DETECTION SIMULATION RESULTS
SIMULATION SETUP
The SDR simulation was performed in a computer with a
dual-core 3.5 GHz and 8 GB RAM. The SDR was set up
with transmission center frequency at 1575.42MHz with inphase
and quadrature sampling frequency of 2.048 MHz.
The GPS receiver code was C/A code set up for all-in-view,
i.e., to look for all visible GPS satellites. The total number of
training examples varied from 500 to 2500. The NN architecture
consisted of one input layer, one or two hidden
layers, and one output layer. The input layer neurons equal
the number of input features, whereas the output layer was a
single neuron. Hidden layers neurons varied from four to
twenty. Three activation functions examined included the
Sigmoid,the tanh, and the ReLUdiscussed previously. Three
learning algorithms implemented included the SGD,
RMSProp,and the Adam discussed earlier. Simulations
were performed using various combinations of the three
activation functions and the three learning algorithms. Best
results were obtained from using the ReLUactivation function
and the Adam learning algorithm. The model was
Figure 8.
Offline training ofspoofing detector.
JUNE 2022
IEEE A&E SYSTEMS MAGAZINE
27
IEEE - Aerospace and Electronic Systems - June 2022

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - June 2022
