Abstract
Sharing of the spectrum is defined as using a specific area and time by various wirelesstechniques. Protecting the spectrum from intruders is a must for maximum utilization of
the spectrum. As the software-defined radio (SDR) technologies are being easily available
at low cost, shared spectrum faces security attacks. As a result, malign users can pass data
easily on the shared spectrum which causes issues for genuine users.
The existing way to deal with intruders is done manually, but the ideal way is to
correctly detect multiple transmitters. It is important to localize multiple transmitters at a
time to save the spectrum resources because a technique to locate a single transmitter could
be avoided by an offender by using many devices. This can be solved using the multiple
transmitter localization problem which deals with the sensors and transmitters. It is of great
benefit in applications where the detection of intruders is the goal. It deals with solving the
problem using a deep learning-driven approach to solve the problem using MTL.
In this project, we study the effects of the adversarial attack and adversarial training
on the multiple transmitter localization and evaluate the performance of the CNN model.
Three adversarial attacks are performed i.e., Fast Gradient Signed Method (FGSM) Attack,
Projected Gradient Descent (PGD) Attack, Sparse L1 Descent Attack and we have
implemented adversarial training on the CNN model by collecting adversarial samples
from the white-box attacks. The results clearly show that these white-box attacks affect the
model’s performance in a significant way. The project also proves that adversarial training
increases the robustness of the CNN model after the adversarial attacks are implemented.
Pytorch and Cleverhans library are used to implement this project.