Spatial de-correlation of generated keys from wireless channels using adversarial deep learning

Abstract

Physical-layer Key Generation (PKG) is the major candidate for use in secure wireless communications and Internet of Things (IoT) systems. Using Deep Learning (DL) and the Band Feature Mapping (BFM) method leads to reciprocal features, which is an essential requirement for the key generation in Orthogonal Frequency-Division Multiplexing Frequency Division Duplexing systems. Additionally, randomness and spatial de-correlation are two other essential requirements of secure PKG schemes. When the distance of an eavesdropper from a legal user is short, the eavesdropper can experience a correlated fading and generate the secret key.Other works assume that the adversary is far away from legitimate users, whereas the proposed scheme allows the adversary to approach the legitimate users without sacrificing the security Conventional DL-based BFM includes an offline training stage using a pre-collected dataset. To solve the spatial correlation problem, this paper simultaneously uses the concepts of physical layer security and adversarial training. Moreover, a DL-based adversary in the PKG model is considered which has not been studied yet. Simulation results confirm the effectiveness of the proposed Adversarial DL (ADL) key generation scheme in terms of Key Error Rate and Key Generation Rate. Our results show that using the proposed training strategy the illegal user can only generate a random key with an error rate of about 0.5. In the meantime, this method maintains the performance of the generated key by the legal users under a certain level. The mentioned features make ADL key generation scheme an appealing candidate for applications, such as secure cloud-based communications, low-size networks, and resource-constrained IoT.