Machine Learning Enhanced Near-Field Secret Key Generation for Extremely Large-Scale MIMO

Aug 2, 2026

Next-generation wireless links are expected to rely on mmWave and THz bands together with extremely large-scale MIMO (XL-MIMO), shifting key generation from far-field to radiative near-field propagation. The work studies physical-layer secret key generation under the most difficult line-of-sight (LoS) scenario and injects artificial randomness to enable theoretical analysis of secret key rate. A zero-forcing (ZF) precoding method is derived, and its low secret key rate behavior is identified for low transmit power and small eavesdropper distances. To address these challenges, a low-complexity machine learning-based beam focusing (MLBF) scheme is proposed, achieving higher secret key rate than benchmark methods in simulations.

LanguageEnglish

CategoryTechnology

FormatPDF

Pages6

Reading Time15 min

AccessFree to download

Introduction

System Model

System Overview

Machine Learning Enhanced Near-Field Secret Key Generation for Extremely Large-Scale MIMO Chen Chen∗ , Junqing Zhang† ∗ Division of Network and Systems Engineering, KTH Royal Institute of Technology, Sweden Email: [chch2@kth.se](chch2@kth.se) † Department of Electrical Engineering and Electronics, University of Liverpool, Liverpool, L69 3GJ, United Kingdom Email: [Junqing.Zhang@liverpool.ac.uk](Junqing.Zhang@liverpool.ac.uk) Abstract—The next generation of communication systems are expected to operate at high frequency bands such as millimetre wave (mmWave) and terahertz (THz) bands, and use extremely large-scale multiple-input-multiple-output (XL-MIMO). This brings a paradigm shift from far-field to near-field communications. In this paper, we investigate physical-layer key generation in near-field XL-MIMO communications and focus on the most challenging line-of-sight (LoS) propagation scenario. Tobe specific, we introduce artificial randomness to enhance secret key generation and enable theoretical analysis of secret key rate (SKR). We provide the zero-forcing (ZF) precoding solution that can null the received signal at the eavesdropper. We show that the ZF precoding leads to a low SKR in challenging scenarios of low transmit powers and small eavesdropping distances. To improve the SKR in these challenging scenarios, we propose a novel low-complexity machine learning-based beam focusing (MLBF) scheme. Simulation results show that the proposed MLBF scheme achieves a higher SKR than the benchmark methods. Index Terms—Physical-layer key generation, extremely largescale MIMO, machine learning I. INTRODUCTION Physical-layer key generation based on reciprocal and random wireless channels has been envisioned as a promising technique to secure the Internet of Things (IoT) [1] . The spatial decorrelation prevents the eavesdroppers from generating the same keys. Different from conventional public key cryptography-based key distribution systems that are computationally expensive, physical-layer key generation provides light-weight solutions for low-cost IoT applications with low computation overhead and energy consumption. Moreover, it has been proved to be information-theoretically secure [2],[3] . To cater to emerging data-hungry applications such as autonomous vehicles, remote surgery and extended reality, millimetre wave (mmWave) bands (30-300 GHz) have been widely used in 5G [4] . Future 6G networks are expected to exploit terahertz (THz) bands (0.1-10 THz) . The tiny wavelengths enables the deployment of extremely large-scale multiple-input-multiple-output (XL-MIMO) array consisting of hundreds or thousands of antennas. In this case, it is very likely that wireless communications happen in radiative near-field region. Different from the planar wave propagation in conventional far-field region, electromagnetic propagation needs to be modeled as spherical wavefronts in radiative near-field region [5] . However, the existing work investigating physical-layer key generation in massive MIMO [6] and intelligent reflecting surface-aided systems [7] focused on farfield communications. Physical-layer key generation faces serious challenges in mmWave and THz communications. First, due to the high penetration loss, mmWave and THz transmissions are dominated by line-of-sight (LoS) channels [8], [9] that may remain static for a long time. Second, mmWave and THz links have limited scattering, and thus the Gaussian channel model is not applicable. This hinders the theoretical analysis on the secret key rate using Gaussian channel models. Artificial randomness has been explored to improve the secret key rate (SKR) in quasi-static environments [10], [11] . In [10], single-antenna legitimate users independently generated local randomness and communicated through a direct link or an untrusted relay. In [11], the authors investigated multiantenna systems and evaluated the SKR performance with zero-forcing (ZF) and maximum rat

FAQs

Why does the paper focus on near-field communications for XL-MIMO secret key generation?

Future mmWave/THz systems with XL-MIMO are likely to operate in the radiative near-field region, where spherical wavefront modeling differs from far-field planar propagation. This change motivates studying physical-layer key generation under near-field conditions.

What is the role of artificial randomness in the proposed framework?

Artificial randomness is introduced to enhance secret key generation in LoS XL-MIMO and to make a theoretical analysis of the secret key rate (SKR) possible.

How does zero-forcing (ZF) precoding perform in challenging scenarios?

ZF precoding can null the received signal at the eavesdropper, but it results in low SKR when transmit power is low and the eavesdropper is very close.

What is the purpose and benefit of the MLBF scheme?

The MLBF scheme is designed to improve SKR in the challenging low-power, small-distance LoS cases using low-complexity machine learning-based beam focusing. Simulations show it achieves a higher SKR than benchmark methods and can yield a positive SKR even when the eavesdropper is at the user’s location.

Aug 2, 2026

Upload by Olivia Brown