PaperlayerView exact revision on arXiv

Source-linked AI summary

Secure Transmission with Multiple Antennas: The MISOME Wiretap Channel

Ashish Khisti, Gregory Wornell

arXiv:0708.4219v1cs.IT

TL;DR

The paper asks how multiple antennas can support secure communication in the wiretap channel, particularly for MISOME systems where existing capacity characterizations are difficult to evaluate. It derives practical secrecy-capacity results using generalized eigenvalues and beamforming, then analyzes channel-ignorant coding, large-antenna limits, and fading extensions. The results identify capacity-achieving beamforming and antenna-ratio regimes in which eavesdroppers are ineffective or can eliminate secrecy.

  • Problem

    Existing secrecy-capacity characterizations for nondegraded multiple-antenna broadcast channels are expressed through optimized auxiliary random variables that are difficult to evaluate explicitly.

  • Method

    The paper analyzes the MISOME wiretap channel using generalized-eigenvalue methods, capacity bounds, beamforming, masked beamforming, and extensions to ergodic fading.

  • Results

    For fixed known channels, secrecy capacity is characterized by the largest generalized eigenvalue and achieved by rank-one beamforming; masked beamforming is close to optimal at high SNR.

  • Takeaways & Limitations

    In the large-antenna limit, eavesdroppers with fewer antennas are comparatively ineffective, whereas at least twice as many antennas as the sender suffice to defeat transmission security.

Abstract

from arXiv · show

The role of multiple antennas for secure communication is investigated within the framework of Wyner's wiretap channel. We characterize the secrecy capacity in terms of generalized eigenvalues when the sender and eavesdropper have multiple antennas, the intended receiver has a single antenna, and the channel matrices are fixed and known to all the terminals, and show that a beamforming strategy is capacity-achieving. In addition, we show that in the high signal-to-noise (SNR) ratio regime the penalty for not knowing eavesdropper's channel is small--a simple ``secure space-time code'' that can be thought of as masked beamforming and radiates power isotropically attains near-optimal performance. In the limit of large number of antennas, we obtain a realization-independent characterization of the secrecy capacity as a function of the number $β$: the number of eavesdropper antennas per sender antenna. We show that the eavesdropper is comparatively ineffective when $β<1$, but that for $β\ge2$ the eavesdropper can drive the secrecy capacity to zero, thereby blocking secure communication to the intended receiver. Extensions to ergodic fading channels are also provided.

I. INTRODUCTION

The paper develops physical-layer techniques for securing transmissions with multiple antennas and provides practical secrecy-capacity characterizations for the MISOME channel. It also studies robustness to unknown eavesdropper channels, large-antenna limits, and ergodic fading.

  • I. INTRODUCTION: Multiple antennas are applied to the less explored problem of enhancing wireless communication security through physical-layer transmission techniques.The paper develops and optimizes these techniques and analyzes their resulting performance characteristics.
  • I. INTRODUCTION: Practical characterizations are needed because existing nondegraded broadcast-channel secrecy-capacity expressions use optimized auxiliary random variables that are difficult to evaluate explicitly.The paper addresses this difficulty for the MISOME configuration.
  • I. INTRODUCTION: The paper studies MISOME channels, where the sender and eavesdropper have multiple antennas while the intended receiver has one.The eavesdropper antennas may form one array or represent geographically dispersed, perfectly colluding single-antenna eavesdroppers.
  • I. INTRODUCTION: For fixed complex channel gains known to all terminals, the paper develops the MISOME secrecy capacity and tightly bounds the capacity expression’s auxiliary-variable optimization.This result addresses an open problem concerning the optimum auxiliary random variable.
  • I. INTRODUCTION: A masked beamforming scheme that does not require the eavesdropper’s channel is close to optimal in the high-SNR regime.The scheme is described as a secure space-time code that transmits isotropically while protecting the message with artificial noise in orthogonal directions.
  • I. INTRODUCTION: In the large-antenna Rayleigh-fading limit, the eavesdropper is comparatively ineffective with fewer antennas, while at least twice as many antennas as the sender suffice to defeat transmission security.The analysis characterizes secrecy capacity and masked-beamforming rates as functions of the eavesdropper-to-sender antenna ratio.
  • I. INTRODUCTION: The results extend to fast ergodic Rayleigh-fading channels, including upper and lower secrecy-capacity bounds for finite and asymptotically large antenna arrays.In this model, the receiver’s channel state is known to all parties, whereas the eavesdropper’s channel state is known only to the eavesdropper.

A. Upper Bound on Achievable Rates

The paper derives an analytically tractable upper bound for MISOME secrecy capacity, proves it achievable by generalized-eigenvector beamforming, and characterizes SNR-dependent behavior. It also evaluates masked beamforming, showing near-optimal high-SNR performance without eavesdropper-channel knowledge.

  • Upper bound: The genie-aided channel yields an upper bound through conditional mutual information I(x; y_r|y_e), optimized over compatible joint channel distributions.The bound is obtained by giving the receiver the eavesdropper’s observation and then tightening over joint distributions with the required marginals.
  • Capacity characterization: The upper bound is achievable, so MISOME secrecy capacity equals the largest generalized-eigenvalue expression and is attained by rank-one beamforming.The beamformer follows the generalized eigenvector associated with λmax, with message encoding using a scalar Gaussian wiretap code.
  • SNR regimes: At high SNR, secrecy capacity is SNR-limited when h_r has no component in the eavesdropper’s null space, but grows by 1 b/s/Hz per 3 dB when such a component exists.In the latter case, the transmitter can steer a null toward the eavesdropper without nulling the intended receiver.
  • SNR regimes: At low SNR, the optimal beamforming direction approaches the principal eigenvector of h_r^†h_r relative to H_e^†H_e, rather than generally aligning with h_r.Thus, ignoring the eavesdropper is generally suboptimal even in the low-SNR regime.
  • Eavesdropper-ignorant coding: Masked beamforming uses isotropic transmission with message signaling along h_r and synthesized spatio-temporal white noise, requiring only intended-channel knowledge.It is generally suboptimal but asymptotically near-optimal at high SNR, with beamforming loss at most log n_t b/s/Hz, equivalently 10 log10 n_t dB in SNR.
  • Eavesdropper-ignorant coding: Masked beamforming’s covariance is eavesdropper-channel independent, but its scalar wiretap-code rate depends on that channel, creating an insecurity zone around the sender.Eavesdroppers outside this zone are secure from the transmission, whereas those inside may not be.

D. Example

The example illustrates how secrecy rates depend on eavesdropper antennas, SNR, and channel knowledge, while the large-system analysis identifies three antenna-ratio regimes and extends the results to fading channels.

  • D. Example: For nt = 2, one eavesdropper antenna allows arbitrarily large secure rates with sufficient power, whereas two antennas cap the secure rate.The example’s solid capacity curves show the contrast directly.
  • D. Example: At high SNR, masked beamforming approaches capacity without transmitter knowledge of the eavesdropper channel.Its performance loss approaches 3 dB for one eavesdropper antenna and 0 dB when both are used.
  • E. Scaling Laws in the Large System Limit: In the large-system limit, secrecy rates are studied as functions of β, the eavesdropper-to-transmitter antenna ratio, under fixed received SNR scaling.The channel ensemble uses independent Rayleigh-fading gains known to all terminals.
  • E. Scaling Laws in the Large System Limit: For β < 1, the eavesdropper is effectively thwarted because the transmitter can steer a null toward it and achieve any desired receiver rate with enough power.This regime gives the eavesdropper proportionally fewer antennas than the sender.
  • E. Scaling Laws in the Large System Limit: For 1 ≤ β < 2, the eavesdropper caps the secure rate; at β = 1.5 it is no more than 1 b/s/Hz, and at received SNR at most 10 dB it is below 1/2 b/s/Hz.These bounds hold regardless of unlimited transmitter power in the stated regime.
  • E. Scaling Laws in the Large System Limit: For β ≥ 2, the eavesdropper can drive secrecy capacity to zero even with unlimited transmitter power.Theorem 5 gives the asymptotic capacity upper bound as 0 for β ≥ 2, −log(β −1) for 1 < β < 2, and infinity for β ≤ 1.
  • F. Capacity Bounds in Fading: For fast fading, the paper provides upper and lower secrecy-capacity bounds, with the lower bound achieved by adaptive masked beamforming.The bounds generally do not coincide, although capacity is calculable in the joint high-SNR and large-system limit.

V. UPPER BOUND DERIVATION

The upper-bound derivation replaces the original channel with a genie-aided degraded channel, then shows that Gaussian inputs and a suitable covariance optimization characterize the relevant bound.

  • V. UPPER BOUND DERIVATION: The genie-aided channel gives the receiver both yr and ye, so its secrecy capacity upper-bounds the original MISOME channel.Because the resulting broadcast channel is degraded, its secrecy capacity is expressed through max I(x; yr|ye).
  • V. UPPER BOUND DERIVATION: Gaussian input maximizes the conditional mutual information for each admissible noise covariance.The proof reduces the optimization to maximizing h(yr|ye), with Gaussianity attaining the entropy maximum under the relevant second moments.
  • V. UPPER BOUND DERIVATION: The MISOME capacity is derived through high- and low-SNR asymptotes, with rank-one input covariance maximizing the relevant quadratic form.For λmax > 1 the resulting expression is log λmax, while for λmax ≤ 1 it is zero.
  • V. UPPER BOUND DERIVATION: When He is not full column rank, a transformed full-rank channel is shown to have the same secrecy capacity as the original channel.This reduces the singular case to the preceding analysis and supports the zero-capacity conclusion when applicable.

B. Proof of Corollary 1

The proof of Corollary 1 bounds the high-SNR behavior by analyzing generalized-eigenvector directions and separating components in and orthogonal to the eavesdropper’s null space.

  • B. Proof of Corollary 1: The proof restricts attention to λmax > 1, the regime in which secrecy capacity is nonzero.The generalized-eigenvalue characterization supplies the relevant starting point.
  • B. Proof of Corollary 1: A vector ψ is decomposed into orthogonal projections onto the null space of He and its complement.The proof bounds the receiver gain using the smallest nonzero singular value of He and Cauchy–Schwarz.
  • B. Proof of Corollary 1: The lower-bound argument controls the generalized Rayleigh quotient by exploiting the eavesdropper-null-space component.The resulting inequalities establish the needed asymptotic lower bound.
  • B. Proof of Corollary 1: The upper-bound argument shows that the maximizing direction’s contribution from the eavesdropper-null-space component is controlled as power grows.A vanishing error term ε(P) is used with ε(P) → 0 as P → ∞.

C. Proof of Corollary 2

The proof of Corollary 2 obtains the low-SNR behavior by expanding the generalized-eigenvalue expression around P = 0.

  • C. Proof of Corollary 2: The proof studies the limit P → 0 and uses O(P) for terms whose ratio to P vanishes.This establishes the asymptotic scale used in the subsequent expansion.
  • C. Proof of Corollary 2: A Taylor expansion of (I + PH†eHe)^−1 reduces the generalized-eigenvalue expression to its first-order behavior in P.The derivation assumes P is sufficiently small that the eigenvalues of PH†eHe are below unity.
  • C. Proof of Corollary 2: Continuity of eigenvalues and λ(I + A) = 1 + λ(A) transform the matrix expansion into an expansion of the dominant eigenvalue.A Taylor expansion of the logarithm then yields the rate asymptote.
  • C. Proof of Corollary 2: Taking the limit P → 0 produces the claimed low-SNR result.The final step follows directly from the preceding expansion.

VII. MASKED BEAMFORMING SCHEME ANALYSIS

The masked beamforming scheme uses a deliberately chosen input distribution rather than directly maximizing the secrecy rate. Its rate analysis relies on convenient normalizations and evaluates the resulting mutual-information difference.

  • Masked beamforming uses a suboptimal choice of p_u and p_x|u instead of the distributions that maximize secrecy rate.The achievable secrecy rate is maximized over input distributions, whereas masked beamforming selects a different pair.
  • The scheme is specified through particular input distributions under the power constraint E[|x|^2] ≤ P.
  • Convenient normalizations are chosen to put the masked beamforming construction into a tractable form for rate evaluation.

A. Proof of Proposition 1

The proposition’s proof evaluates the masked beamforming secrecy rate by computing the receiver and eavesdropper mutual informations. Combining these expressions yields the stated rate, with the largest generalized eigenvalue providing the final simplification.

  • The proof evaluates the receiver mutual information and then the eavesdropper mutual information for the masked beamforming distribution.
  • The two mutual-information calculations combine to produce the secrecy-rate expression in (15).
  • The final equality uses the special representation of the largest generalized eigenvalue.
  • The resulting expression implies that the relevant gap approaches zero as P →∞, yielding (16).

VIII. SCALING LAWS DEVELOPMENT

The scaling-law analysis combines random-matrix limits with the generalized-eigenvalue distribution of Gaussian channel matrices. It derives asymptotic expressions for secrecy capacity and masked beamforming as the antenna dimensions grow with β fixed.

  • Random-matrix facts provide limiting tools for quadratic forms, η-transforms, and spectra of large Gaussian matrices.
  • For i.i.d. Gaussian receiver and eavesdropper channels with n_e > n_t, the generalized-eigenvalue distribution is characterized through an F-distribution.
  • With β = n_e/n_t fixed, strong-law limits for the chi-squared variables yield the stated asymptotic eigenvalue relation.
  • The secrecy-capacity scaling law follows by rewriting the generalized eigenvalue using its quadratic-form representation and applying the random-matrix limits.
  • An analogous derivation gives the masked beamforming scaling law, including the nonsingular-eavesdropper case and the limit n_e, n_t →∞ with n_e/n_t = β fixed.

A. Proof of (27a)

The proof of (27a) treats the fading channel as parallel channels indexed by the intended receiver’s gains and uses an adaptive masked beamforming distribution. The lower bound is evaluated under the power constraint, while the upper-bound construction uses correlated worst-case noises with unchanged marginals.

  • The fading channel is represented as parallel channels indexed by the intended receiver’s channel gains, with the eavesdropper observing (y_e, H_e).
  • Adaptive masked beamforming selects p_u|h_r and p_x|u,h_r subject to the long-term average power constraint.The construction adapts the masked beamforming distributions to the intended receiver’s channel state.
  • Evaluating the achievable-rate expression with these distributions yields (27a) together with (29a).
  • The converse introduces correlated receiver and eavesdropper noises while preserving the required marginal Gaussian distributions, then upper-bounds the original channel’s secrecy rate.The selected worst-case noise depends on the channel gains, power allocation, and generalized-eigenvector construction.

2) Upper bound on the auxiliary channel:

The section derives bounds for the auxiliary channel using secrecy conditions, Markov relations, worst-case Gaussian noise, concavity, Jensen’s inequality, and random-matrix convergence arguments.

  • Upper bound on the auxiliary channel: The derivation upper-bounds I(x(t); y_r(t) | y_e(t), H^n_r) separately for each time index t.The bound uses the Markov structure and a worst-case noise distribution for the Gaussian channel.
  • Upper bound on the auxiliary channel: Concavity of the capacity function and Jensen’s inequality support the averaging steps used in the bound.The cited passage identifies C(P) as concave in P and attributes key steps to Jensen’s inequality.
  • Upper bound on the auxiliary channel: The large-array argument combines almost-sure convergence with convergence in expectation to establish the stated lower and upper bounds.The proof selects ρ(h_r)=P for the lower bound and invokes Theorems 4 and 5 for the respective bounds.
  • Concluding remarks: The paper identifies tighter fast-fading bounds and extension to the general MIMOME channel as opportunities for further work.The high-SNR general-MIMOME result is noted as characterized elsewhere using generalized singular value analysis.

APPENDIX I PROOF OF LEMMA 1

The appendix proves Lemma 1 by applying standard information-theoretic inequalities, introducing time sharing, and reducing the result through generalized-eigenvalue algebra.

  • APPENDIX I PROOF OF LEMMA 1: The proof begins with a sequence of codes satisfying the stated asymptotic reliability and secrecy conditions.The argument assumes the existence of (2^nR, n) codes for every ε>0 and sufficiently large n.
  • APPENDIX I PROOF OF LEMMA 1: Fano’s inequality and mutual-information chain rules convert the coding assumptions into bounds involving the relevant information quantities.The proof combines the Fano term with subsequent inequalities and uses conditioning, Markov relations, and memorylessness.
  • APPENDIX I PROOF OF LEMMA 1: A uniformly distributed time-sharing variable q single-letterizes the n-letter variables while preserving the average input-power constraint.Conditioned on q=t, the constructed tuple has the distribution of the corresponding time-indexed variables, with E[∥x∥2]≤P.
  • APPENDIX I PROOF OF LEMMA 1: The remaining algebra uses convexity in θ and the definition of generalized eigenvalues to transform the intermediate expression into the claimed result.The proof substitutes the preceding identities and performs minor algebra to obtain the right-hand side.
Loading 0708.4219v1…