Secure Communication via Modulation Order Confusion

TL;DR

This work addresses modulation-classification threats in wireless security by proposing Modulation Order Confusion (MOC), which disguises the legitimate modulation as another order to mislead eavesdroppers while preserving Bob's data recovery. It develops two single-antenna schemes—Symbol Random Mapping (SRM) for low-to-high order confusion and Symbol Time Diversity (STD) for high-to-low order confusion—and two receiver-transparent multi-antenna schemes—Taylor-series/series expansion and Constellation Path Design (CPD)—with extensions to RIS-assisted systems. A convex optimization framework for mapping probabilities and a dynamic programming-based deconfusion receiver are proposed, along with joint beamformer and RIS design to enhance secrecy. Numerical results demonstrate that the proposed MOC schemes defeat both deep-learning and expert-knowledge modulation classifiers, with security improving at higher SNRs and a tunable balance between spectral efficiency and reliability.

Abstract

With the increasing threat posed by modulation classification to wireless security, this paper proposes a secure communication framework based on modulation order confusion (MOC), which intentionally disguises the original modulation as a higher- or lower-order one to mislead eavesdroppers. For single-antenna systems, two schemes are developed: symbol random mapping and symbol time diversity, enabling modulation order confusion with customized receivers. For multi-antenna systems, receiver-transparent MOC schemes are proposed, including series-expansion-based and constellation-path-based signal designs, and are further extended to RIS-assisted systems with joint beamformer and RIS reflection design. Numerical results show that the proposed schemes effectively defeat both deep-learning-based and expert-knowledge-based modulation classifiers without degrading communication performance.

Figures (9)

• Figure 1: Illustration of the single-antenna system.
• Figure 2: Illustration of the symbol design strategy based on CPD. In subfigure (a), the antenna $1$ transmits signals from the set $\mathcal{B}_1 = \{ 1+1\jmath, 1-1\jmath, -1+1\jmath, -1-1\jmath\}$, while antennas $2$ and $3$ transmit signals from identical sets $\mathcal{B}_2 = \mathcal{B}_3 = \{2,\,-2,\,2\jmath,\,-2\jmath\}$. In subfigure (b), the antenna $1$ transmits symbols from the set ${\mathcal{B}_1} = \{ {e^{\frac{3}{8}\pi }},{e^{\frac{7}{8}\pi }},{e^{ - \frac{1}{8}\pi }},{e^{ - \frac{5}{8}\pi }}\}$, while antenna $2$ uses the set ${\mathcal{B}_2} = \{ \sqrt {2 - \sqrt 2 } ,-\sqrt {2 - \sqrt 2 } ,\sqrt {2 - \sqrt 2 }\jmath ,-\sqrt {2 - \sqrt 2 }\jmath \}$.
• Figure 3: Illustration of the RIS-assisted multi-antenna system.
• Figure 4: The classification accuracy is evaluated w.r.t SNR for low-to-high-order confusion in the single-antenna system. In the simulation, QPSK is disguised as 16QAM. The three subplots correspond to different symbol mapping probabilities, i.e., (a) ${\boldsymbol{p}} = [0, 0, 0, 1]$, (b) ${\boldsymbol{p}} = [0.1, 0.2, 0.3, 0.4]$, and (c) ${\boldsymbol{p}} = [0.25, 0.25, 0.25, 0.25]$. Additionally, we assume that Eve employs four types of deep-learning based modulation classifiers, i.e., VGGVGGo2018over, SCGNetSCGNettunze2020sparsely, WSMFWSMFqi2020automatic, and ChainNetChainNethuynh2020chain.
• Figure 5: The classification accuracy is evaluated w.r.t SNR for high-to-low-order confusion in the single-antenna system. In the simulation, 16PSK is disguised as 9GAM. The two subplots correspond to the cases before and after MOC, i.e., (a) without confusion, and (b) after confusion. Here, VGGVGGo2018over, SCGNetSCGNettunze2020sparsely, WSMFWSMFqi2020automatic, and ChainNetChainNethuynh2020chain are also assumed to be employed at Eve.