Secure wireless communication evading eavesdropper via adversarial machine learning

Abstract

Acquiring secure communication link between legitimate receiver and transmitter has been an important issue for decades. Especially for wireless communication systems, physical layer security (PLS), which utilizes channel state information to achieve information-theoretic secrecy, is applied along with cryptography which takes place in the upper layer by encrypting the data with secret key. Recently, adversarial signal design, one of the machine learning techniques that fools the pretrained deep learning model, has been applied for deep learning-aided wireless communication systems to ensure the security, referred to as secure, yet adversarial signal. The objective of this secure, yet adversarial signal is to design an adversarial signal that lets the legitimate receiver classify accurately while the eavesdropper misclassifies. While such adversarial signal design can be thought of as joint-PLS-cryptography approach assuming coherent DL-based receiver, huge computational complexity for generating the adversarial signal prevents this coherence, i.e., channel becomes outdated while crafting the adversarial signal. We investigate to reduce this computational complexity via two schemes: (i) meta perturbation

(ii) universal surrogate model. Key idea of (i) the meta perturbation is to find (meta-learn) a good initialization of the perturbation signal for the adversarial signal to ensure coherent, secure communication between the legitimate receiver and the transmitter and (ii) the universal surrogate model is to learn parameters of a generative surrogate model in lieu of target models for crafting transferable secure, yet adversarial signal. Numerical results verify that the proposed approaches reduce computational complexity of the secure, yet adversarial signal design with a reasonable performance degradation.