AN-Aided Beamforming for Secure Wireless Links

• AN-aided beamforming is a physical-layer security technique that combines confidential data transmission with artificial noise injection to degrade eavesdroppers' SINR.
• It employs joint optimization of beamforming vectors and noise covariance, using methods such as smooth objective approximation and the GPI algorithm to maximize secrecy rates.
• The design delivers robust and scalable secrecy performance under imperfect CSI, achieving up to 40% higher sum secrecy rates compared to traditional methods.

Artificial Noise (AN)-Aided Beamforming is a suite of joint precoding and interference-shaping strategies for physical-layer security in wireless communication systems. It combines the principles of information-theoretic secrecy with spatial signal processing by concurrently directing confidential data streams to legitimate users and injecting artificial Gaussian or structured noise to degrade potential eavesdroppers' signal-to-interference-plus-noise ratios (SINR). This technique leverages the superposition property of the wireless medium and exploits multiple antenna degrees of freedom to maximize secrecy rates, even under imperfect eavesdropper channel knowledge or in the presence of multiple adversaries.

1. Theoretical Foundations and Problem Formulation

At the core, AN-aided beamforming extends classic wiretap and broadcast channel models to the multi-user, multi-antenna (MU-MIMO) setting with multiple eavesdroppers. Given a base station (BS) with $N$ antennas, $K$ single-antenna legitimate users, and $M$ eavesdroppers, the BS transmits a superimposed signal

$$\mathbf{x} = \sum_{k=1}^K \mathbf{w}_k s_k + \mathbf{n}_{\mathrm{AN}},$$

where $\mathbf{w}_k$ is the beamformer for stream $k$, $s_k$ the confidential symbol, and $$\mathbf{n}_{\mathrm{AN}} \sim \mathcal{CN}(0, \boldsymbol{\Sigma}_{\mathrm{AN}})$$ is the artificial noise injected according to covariance $\boldsymbol{\Sigma}_{\mathrm{AN}} \succeq 0$. The total transmit power constraint $E[\|\mathbf{x}\|^2] \leq P$ must be met.

The secrecy rate for each legitimate user is formulated as

$$R_{s,k} = \left[ R_k - \max_{m=1,\ldots,M} R_{m,k}^e \right]^+,$$

with user $k$'s rate $R_k = \log_2(1+\mathrm{SINR}_k)$ and eavesdropper $m$'s rate for stream $k$, $R_{m,k}^e = \log_2(1+\mathrm{SINR}_{m,k}^e)$. The overall sum-secrecy rate is

$$R_s = \sum_{k=1}^K \left[ R_k - \max_{m=1,\ldots,M} R_{m,k}^e \right]^+.$$

This non-smooth, non-convex optimization problem demands advanced algorithmic treatment (Choi et al., 2022).

2. Joint Precoding and Artificial Noise Synthesis

The principal design challenge in AN-aided beamforming is the simultaneous, joint optimization of all beams $\{\mathbf{w}_k\}$ and the AN covariance $\boldsymbol{\Sigma}_{\mathrm{AN}}$ to maximize $R_s$. The approach in (Choi et al., 2022) innovates by:

• Smooth Objective Approximation: The non-smooth $\max_m\{R_{m,k}^e\}$ is relaxed via a log-sum-exp surrogate, controlled by a parameter $\alpha>0$, leading to a differentiable proxy for the sum-secrecy rate:

$$\tilde{R}_s = \sum_{k=1}^K \left\{ R_k - \alpha \log\left( \sum_{m=1}^M \exp\left( R_{m,k}^e/\alpha \right) \right) \right\}.$$

• Nonlinear Eigenvalue Reformulation: Variables are aggregated into $$v=[\mathbf{w}_1^\top, \ldots, \mathbf{w}_K^\top, \vec(\boldsymbol{\Sigma}_{\mathrm{AN}})^\top ]^\top$$ (unit norm). Each $R_k$ and $R_{m,k}^e$ is expressed as the logarithmic difference of quadratic forms $v^\mathrm{H} A_k v$, $v^\mathrm{H} B_k v$ (and $C_{m,k}$, $D_{m,k}$ for the eavesdroppers). The stationary condition is a nonlinear (generalized) eigenvalue problem:

$$B_{\mathrm{KKT}}^{-1}(v) A_{\mathrm{KKT}}(v) v = \lambda(v) v,$$

where $A_{\mathrm{KKT}}(v)$, $B_{\mathrm{KKT}}(v)$ are adaptive matrices formed from current iterates (Choi et al., 2022).

• Generalized Power Iteration (GPI): This NEP is solved via GPI: at each step, invert-block diagonal $B_{\mathrm{KKT}}$, multiply by $A_{\mathrm{KKT}}$ and $v^{(t-1)}$, and normalize. Upon convergence, $\{\mathbf{w}_k\}$ and $\boldsymbol{\Sigma}_{\mathrm{AN}}$ are extracted from $v^*$.

This approach contrasts with traditional alternating optimization, which fixes one variable set while optimizing the other, but may fall short of global or near-global secrecy rate maximization (Choi et al., 2022, Yu et al., 2017).

3. Algorithmic Frameworks and Complexity

Designs for AN-aided beamforming hinge on efficiently solving high-dimensional, non-convex, and in some cases, semi-definite programs:

• GPI Algorithm: Each GPI iteration involves block-diagonal matrix inversion and matrix-vector products. If the problem dimension is $d = K N + \dim_{\mathrm{AN}}$, each step has $O(d^2)$ cost, but by leveraging block structure, actual per-iteration overhead is $O(\max\{K,J\} N^3)$ (Choi et al., 2022). Typically, convergence is rapid (5–10 iterations in practice).
• Alternating Iterative Structure (AIS): For directional modulation networks, AIS alternates between optimizing the beam for the confidential message and the AN projection matrix. Each subproblem is approached via GPI (for the quadratic-fractional AN subproblem) and generalized eigenvalue analysis (for the beam) (Yu et al., 2017).
• Initialization: "Leakage-based" initializations, using eigen-solutions for maximizing "leakage" to eavesdroppers, significantly accelerate convergence compared to random initialization (Yu et al., 2017).

4. Impact of Channel State Information and Robustness

The secrecy performance and tractability of AN-aided beamforming depend critically on the quality and type of Channel State Information at the Transmitter (CSIT):

• Perfect CSIT of Eavesdroppers: When the BS has instantaneous knowledge of all $g_m$, each term in the secrecy rate can be computed exactly; the joint optimization directly incorporates the true eavesdropper channels.
• Partial or Statistical CSIT: If only statistical knowledge (mean/covariance) of the eavesdroppers is available, as in practical scenarios, one replaces $|g_m^H w|^2$ with $w^H R_m w$, where $R_m$ is the covariance. The same GPI-based framework applies, but now with deterministic surrogates (Choi et al., 2022). This design remains robust, retaining most secrecy gains even under partial CSIT.
• Robustness to Imperfect Legitimate User CSI: By recasting constraints as worst-case or probabilistic on bounded uncertainties, using convex relaxations or Bernstein-type inequalities, designs can maintain secrecy guarantees in the presence of channel estimation errors (Zhang et al., 2018).

5. Performance Evaluation, Trade-offs, and Scalability

Extensive simulations highlight the fundamental trade-offs and practical guidelines in AN-aided beamforming:

• Performance Gains: Joint designs yield 20–40% higher sum secrecy rate than classic RZF, zero-forcing plus null-space AN, or alternating approaches. The improvement scales with the number of antennas $N$ and the number of eavesdroppers $M$ (Choi et al., 2022).
• Complexity vs. Performance: For $N \gg K+M$, only 1–2 dimensions of optimized AN are needed to achieve most of the secrecy gain. Optimized beams are primary; AN is a supporting tool for robust secrecy (Choi et al., 2022).
• Scalability and Online Feasibility: Fast GPI convergence (typically <10 iterations) enables practical, real-time implementation in 5G/6G base stations. The structure of the block-diagonal matrices can be exploited for efficient computation.
• Initialization Benefits: Using leakage-based initialization for the beams and AN halves or better the number of iterations required for convergence (Yu et al., 2017).
• Secrecy-Performance under Partial CSI: With only covariance CSIT of eavesdroppers, significant secrecy gains are retained over non-robust or purely beamforming-only (no AN) designs (Choi et al., 2022).

6. Extensions, Applications, and Open Problems

AN-aided beamforming principles have been generalized across various architectures and communication paradigms:

• Directional Modulation Systems: The GPI-AIS design maximizes secrecy in line-of-sight, angularly-resolved link scenarios. It exhibits 10–20% secrecy rate gains over null-space and leakage-based techniques in medium/high SNR (Yu et al., 2017).
• Massive MIMO and mmWave/THz: For extremely large array systems and near-field scenarios, the spatial shaping of AN can be performed in both angle and range domains, enabling 3D focus and additional secrecy degrees of freedom.
• Secure SWIPT (Simultaneous Wireless Information and Power Transfer): AN is co-designed to degrade eavesdroppers while also meeting energy harvesting targets at legitimate users.
• Robust Beamforming: With partial, bounded, or statistical uncertainties in CSI, robustified AN-aided beamforming uses SDP/LMI representations and enjoys rank-one optimality so that vector beamformers can be uniquely extracted (Zhang et al., 2018).
• Design Guidelines: For large $N$, allocate most degrees of freedom to beams, with 1–2 to AN; tune the soft-max approximation parameter $\alpha$ in $[0.1,1]$ for accuracy and stability; favor joint over alternating optimization; leverage leakage-based initialization for tuning.

A key open direction is the extension of these frameworks to distributed, multi-cell, and network MIMO secrecy topologies with active or passive eavesdropper cooperation, multi-hop, and relay networks.

7. Summary Table: Core Elements in AN-Aided Beamforming Optimization

Component	Mathematical Object	Role in Secrecy Design
Beams $\{\mathbf{w}_k\}$	$N\times 1$ unit-norm vectors	Transmit confidential data to $K$ users
AN Covariance $\boldsymbol{\Sigma}_{\mathrm{AN}}$	$N\times N$ PSD matrix	Injects spatially-shaped noise to jam eavesdroppers
Smooth Approx. Parameter $\alpha$	Scalar $\approx 0.1$–$1$	Controls trade-off in log-sum-exp max approx.
Channel Info	$\{\mathbf{h}_k\}$, $\{\mathbf{g}_m\}$, $\{R_m\}$	Determines legitimate/eavesdropper SINR calculation
GPI Iteration	Block-diagonal matrix inversion	Accelerated joint optimization over beams + AN
Performance Metric	$R_s$ (sum secrecy rate)	Benchmarks improvement over classic beamforming

The joint GPI-based framework for AN-aided beamforming achieves near-optimal secrecy performance, demonstrates robustness to CSI limitations, and is computationally viable for large-scale deployments in advanced wireless networks (Choi et al., 2022, Yu et al., 2017).