Session 8: Robust CNN-based Automatic Modulation Classification
Presentation Type
Oral
Student
Yes
Track
Other
Abstract
Automatic Modulation Classification (AMC) is a crucial task in wireless communication systems and is used to determine the modulation scheme used in a received signal in order to retrieve the original signal. Conventional AMC techniques are based on statistical models and using Convolutional Neural Networks (CNN) to implement AMC has grown in popularity. However, these deep learning models are susceptible to adversarial attacks and perturbations in the signals. These attacks introduce small changes to the input data and can cause a model to misclassify a signal. Adversarial attacks can have significant consequences and are a large concern in the field of machine learning, as they can undermine the reliability of a model. We propose the use of adversarial training to create a robust CNN classifier to classify the modulation of BPSK, QPSK, 16-QAM and 64-QAM OFDM signals affected by an adversarial entity. The robust CNN consists of several convolutional layers with an SGD optimizer. The CNN undergoes adversarial training to increase its robustness against attacks. Future work would involve the use of GAN to further increase the robustness of an OFDM AMC system. We seek to analyze the performance of the CNN against adversarial attacks after adversarial training. Safeguarding DNN based AMC systems is a crucial problem and improving the robustness of a system against such attacks is what we aim to solve.
Start Date
2-6-2024 2:30 PM
End Date
2-6-2024 3:30 PM
Session 8: Robust CNN-based Automatic Modulation Classification
Dakota Room 250 A/C
Automatic Modulation Classification (AMC) is a crucial task in wireless communication systems and is used to determine the modulation scheme used in a received signal in order to retrieve the original signal. Conventional AMC techniques are based on statistical models and using Convolutional Neural Networks (CNN) to implement AMC has grown in popularity. However, these deep learning models are susceptible to adversarial attacks and perturbations in the signals. These attacks introduce small changes to the input data and can cause a model to misclassify a signal. Adversarial attacks can have significant consequences and are a large concern in the field of machine learning, as they can undermine the reliability of a model. We propose the use of adversarial training to create a robust CNN classifier to classify the modulation of BPSK, QPSK, 16-QAM and 64-QAM OFDM signals affected by an adversarial entity. The robust CNN consists of several convolutional layers with an SGD optimizer. The CNN undergoes adversarial training to increase its robustness against attacks. Future work would involve the use of GAN to further increase the robustness of an OFDM AMC system. We seek to analyze the performance of the CNN against adversarial attacks after adversarial training. Safeguarding DNN based AMC systems is a crucial problem and improving the robustness of a system against such attacks is what we aim to solve.