Adversarial Machine Learning for NextG Covert Communications Using Multiple Antennas.

Area of Science:

• Wireless Communications
• Cybersecurity
• Artificial Intelligence

Background:

• Eavesdroppers utilize deep learning (DL) classifiers to detect wireless transmissions.
• Protecting wireless communication privacy against intelligent eavesdropping is a significant challenge.

Purpose of the Study:

• To investigate the use of adversarial perturbations generated by a cooperative jammer (CJ) to enhance wireless communication privacy.
• To assess the effectiveness of a multi-antenna CJ in fooling DL-based eavesdroppers.

Main Methods:

• A cooperative jammer (CJ) with multiple antennas transmits adversarial perturbations.
• Power allocation and channel diversity are optimized across the CJ's antennas.
• An upper bound is imposed on perturbation strength to minimize impact on legitimate receiver's bit error rate (BER).

Main Results:

• Adversarial perturbations successfully cause eavesdroppers to misclassify signals as noise with high probability.
• The legitimate receiver experiences only a slight increase in bit error rate (BER).
• Utilizing multiple antennas significantly enhances the effectiveness of the adversarial perturbation.

Conclusions:

• Adversarial perturbations generated by a multi-antenna CJ offer a viable method for enhancing wireless communication privacy.
• The proposed technique effectively disrupts eavesdropping attempts by DL classifiers while maintaining acceptable communication quality.
• Multi-antenna systems provide a performance advantage for adversarial jamming strategies.