PaperlayerView exact revision on arXiv

Source-linked AI summary

The Secrecy Capacity Region of the Gaussian MIMO Multi-receiver Wiretap Channel

Ersen Ekrem, Sennur Ulukus

arXiv:0903.3096v1cs.IT

TL;DR

The paper asks how to characterize secrecy capacity for Gaussian MIMO multi-receiver wiretap channels with confidential messages and an external eavesdropper, where standard scalar broadcast converses are insufficient. It develops MMSE/Fisher-information converse tools, extends them through degraded and aligned MIMO channels using channel enhancement, and obtains the general capacity region, achievable with Gaussian dirty-paper coding.

  • Problem

    The secrecy capacity region of the general Gaussian MIMO multi-receiver wiretap channel is difficult to characterize, and existing Gaussian broadcast-channel converse techniques do not extend directly to the secrecy setting.

  • Method

    The paper uses MMSE–mutual-information and Fisher-information–differential-entropy relationships, then applies channel enhancement and limiting arguments across degraded, aligned, and general MIMO channels.

  • Results

    The secrecy capacity region of the general Gaussian MIMO multi-receiver wiretap channel is characterized and is achievable with a variant of dirty-paper coding using Gaussian signals.

  • Takeaways & Limitations

    Gaussian dirty-paper coding provides the capacity-achieving scheme for the characterized general MIMO multi-receiver wiretap channel.

  • Takeaways & Limitations

    The analysis assumes perfect secrecy and formulates capacity under covariance constraints, with other constraints handled through corresponding covariance-matrix sets.

Abstract

from arXiv · show

In this paper, we consider the Gaussian multiple-input multiple-output (MIMO) multi-receiver wiretap channel in which a transmitter wants to have confidential communication with an arbitrary number of users in the presence of an external eavesdropper. We derive the secrecy capacity region of this channel for the most general case. We first show that even for the single-input single-output (SISO) case, existing converse techniques for the Gaussian scalar broadcast channel cannot be extended to this secrecy context, to emphasize the need for a new proof technique. Our new proof technique makes use of the relationships between the minimum-mean-square-error and the mutual information, and equivalently, the relationships between the Fisher information and the differential entropy. Using the intuition gained from the converse proof of the SISO channel, we first prove the secrecy capacity region of the degraded MIMO channel, in which all receivers have the same number of antennas, and the noise covariance matrices can be arranged according to a positive semi-definite order. We then generalize this result to the aligned case, in which all receivers have the same number of antennas, however there is no order among the noise covariance matrices. We accomplish this task by using the channel enhancement technique. Finally, we find the secrecy capacity region of the general MIMO channel by using some limiting arguments on the secrecy capacity region of the aligned MIMO channel. We show that the capacity achieving coding scheme is a variant of dirty-paper coding with Gaussian signals.

1 Introduction

The paper studies secrecy capacity for Gaussian multi-receiver wiretap channels, emphasizing that the general MIMO problem requires new converse techniques beyond scalar broadcast-channel arguments. It proceeds through scalar, degraded, aligned, and general MIMO cases, adapting channel enhancement and estimation-theoretic tools.

  • Problem: The multi-receiver wiretap channel extends secure communication to one transmitter, multiple legitimate users, and an external eavesdropper.Its general secrecy capacity region is challenging because even the corresponding arbitrary-user broadcast channel is not fully characterized.
  • Scalar case: The scalar case shows that existing Gaussian broadcast-channel converse proofs cannot be straightforwardly extended to the secrecy setting.A stand-alone entropy-power inequality does not resolve the ambiguity involving auxiliary random variables.
  • Proof technique: The new converse approach uses relationships among MMSE, mutual information, Fisher information, and differential entropy.The paper presents MMSE-based and Fisher-information-based proof counterparts for the scalar channel.
  • MIMO roadmap: The MIMO analysis treats degraded channels first, then aligned channels, and finally the general case.Channel enhancement is used to generalize from the degraded to the aligned setting, while the general channel follows through limiting arguments.
  • Related work: The paper also situates its approach relative to prior Gaussian MIMO wiretap results based on Sato-type outer bounds and noise-correlation optimization.A later alternative proof used channel enhancement for the single-user MIMO wiretap channel.

2 Multi-Receiver Wiretap Channels

This section defines the general multi-receiver wiretap channel and its secrecy-capacity framework, then specializes to degraded channels. For degraded channels, the capacity region is characterized by auxiliary-variable rate bounds and optimal joint distributions.

  • Channel model: The general channel has one transmitter, K legitimate receivers, an external eavesdropper, and a memoryless transition probability.Each user receives a confidential message that must remain secret from the eavesdropper.
  • Coding framework: A code contains K message sets, one encoder, and K decoders, with one decoder at each legitimate receiver.Achievability is defined through vanishing error probability as blocklength grows.
  • Capacity definition: The secrecy capacity region is the closure of achievable rate tuples under perfect secrecy.The formulation considers secrecy for any subset of the users’ messages.
  • Degraded channel: For degraded channels, the secrecy capacity region is given by a union of rate tuples satisfying Rk ≤ I(Uk; Yk|Uk+1, Z).The bound applies for k = 1, . . . , K, with U1 = X and UK+1 = φ.
  • Equivalent characterization: The degraded-channel formulation can equivalently be written as Rk ≤ I(Uk; Yk|Uk+1) − I(Uk; Z|Uk+1).For multiple users, the boundary is traced by optimizing joint distributions involving X and the auxiliary variables.

3 Gaussian MIMO Multi-receiver Wiretap Channel

The Gaussian MIMO model is developed from degraded and aligned subclasses to the fully general channel under covariance constraints. Positive-semidefinite noise ordering enables degraded analysis, while unordered aligned channels require dirty-paper coding and channel enhancement.

  • Degraded model: The degraded Gaussian MIMO channel uses additive Gaussian observations Yk = X + Nk under a covariance constraint on X.Its noise covariance matrices satisfy a positive-semidefinite ordering, allowing the channel to be treated as degraded.
  • Degraded model: Noise-vector correlations do not affect the capacity-equivocation region, so they can be adjusted to enforce a suitable Markov chain.This preserves the corresponding secrecy capacity region while enabling use of the degraded-channel characterization.
  • Aligned model: The aligned Gaussian MIMO channel has equal receiver antenna counts but no positive-semidefinite ordering among noise covariance matrices.Its noise covariance matrices are strictly positive-definite, and the lack of ordering prevents a degraded-channel interpretation.
  • Aligned model: For the aligned channel, dirty-paper coding with stochastic encoding is shown to be optimal instead of degraded-channel superposition coding.The aligned setting therefore requires a coding strategy adapted to unordered noise covariances.
  • General model: The general Gaussian MIMO channel allows arbitrary channel-gain matrices, receiver dimensions, and strictly positive-definite noise covariance matrices.Its input is constrained by a positive-semidefinite covariance matrix.
  • Covariance constraints: Capacity regions under total-power or per-antenna constraints can be obtained from the covariance-constrained region when the constraints form compact sets of input covariance matrices.The per-antenna set is defined by Sii ≤ Pi for each transmit antenna.
  • Covariance constraints: A singular positive-semidefinite covariance constraint can be replaced by an equivalent degraded or aligned channel with fewer transmit and receive antennas.For strictly positive-definite constraints, the paper uses the equivalent formulation without loss of generality in the degraded and aligned cases.

4 Gaussian SISO Multi-receiver Wiretap Channel

The Gaussian SISO multi-receiver wiretap channel has a secrecy capacity region characterized using Gaussian signaling, but standard scalar broadcast-channel converse proofs do not extend directly. The paper instead develops MMSE- and Fisher-information-based converses to establish the required bounds.

  • Channel model: The two-user Gaussian SISO wiretap channel is defined under a power constraint and ordered Gaussian noise variances, allowing a degraded-channel representation.The noise correlations can be adjusted without changing the secrecy capacity region so that the relevant Markov chain holds.
  • Capacity region: The secrecy capacity region is the union of rate pairs satisfying the theorem’s bounds over α ∈ [0, 1], with Gaussian signaling achieving the region.The complementary parameter is ᾱ = 1 − α.
  • Converse difficulty: Existing entropy-power-inequality converses for the Gaussian scalar broadcast channel cannot be straightforwardly extended because the secrecy term I(X; Z|U2) complicates the optimization.In the secrecy setting, the entropy-power inequality can produce the opposite of the inequality needed to establish Gaussian signaling optimality.
  • MMSE and Fisher-information converses: The converse uses the relationship between MMSE and conditional mutual information, together with a proposition controlling the zeros and monotonicity of an MMSE-derived function.The proposition also has a Fisher-information counterpart based on the relationship between Fisher information and differential entropy.
  • Converse result: Applying these tools yields the desired secrecy-rate bounds for both users and completes the converse proof of the SISO capacity region.The first-user bound is obtained through alternative case-based derivations, while the second-user bound follows from the single-letter expression and Markov-chain bounds.
  • Outlook for MIMO: The SISO analysis motivates extending the new converse technique to degraded and aligned MIMO wiretap channels before treating the general MIMO case.The MIMO development uses channel enhancement after the degraded case.

5 Degraded Gaussian MIMO Multi-receiver Wiretap Channel

This section establishes the secrecy capacity region for degraded Gaussian MIMO multi-receiver wiretap channels and develops the converse tools needed for arbitrary numbers of users.

  • Main result: The degraded-channel secrecy capacity region is characterized as a union of rate tuples over positive semi-definite covariance matrices.The achievability uses jointly Gaussian auxiliaries, while the converse is developed first for two users and then extended to arbitrary K.
  • Two-user specialization: For two users, the capacity region is the union of rate pairs over covariance matrices K1 satisfying 0 ⪯ K1 ⪯ S.The corresponding Gaussian construction uses X = U2 + V with independent Gaussian components having covariance matrices S − K1 and K1.
  • Converse tools: The converse relies on MMSE, Fisher information, differential entropy, and worst additive noise optimization results.The Fisher-information tools include matrix inequalities, behavior under addition of independent vectors, and a Fisher-information–differential-entropy relationship.
  • Channel assumptions: The degraded model assumes ordered Gaussian noise covariance matrices, 0 ≺ Σ1 ⪯ Σ2 ⪯ ΣZ, under a covariance constraint on X.Conditional optimization results identify a covariance matrix K* for bounding the relevant entropy differences.
  • Converse completion: The two-user converse is completed by deriving the desired bounds on R2 and R1, then the arbitrary-K converse uses the developed tools and one additional lemma.The union of these bounds over admissible covariance matrices gives an outer bound for the secrecy capacity region.

6 Aligned Gaussian MIMO Multi-receiver Wiretap Channel

This section establishes the secrecy capacity region for aligned Gaussian MIMO multi-receiver wiretap channels using channel enhancement, with achievability based on stochastic dirty-paper coding.

  • Channel enhancement: Channel enhancement extends the degraded-channel result to the aligned setting while accounting for the external eavesdropper and multiple legitimate users.The converse constructs a degraded channel whose capacity region contains the original region and whose boundary coincides at the selected weighted point.
  • Coding interpretation: Dirty-paper coding uses an encoding order π, with the indexed rate expression corresponding to the user positioned in that order.Additional dummy messages consume the eavesdropper’s decoding capability and are used solely to confuse it.
  • Main result: The aligned-channel secrecy capacity region is given by the convex closure of a union over covariance allocations and encoding permutations.The theorem applies to aligned Gaussian MIMO multi-receiver wiretap channels without requiring an ordering of the noise covariance matrices.
  • Achievability: Achievability combines Marton’s broadcast-channel scheme with stochastic encoding for secrecy and jointly Gaussian auxiliary random vectors.The resulting construction produces Gaussian X and verifies the secrecy constraints through a Markov chain and entropy bounds.
  • Secrecy verification: The constructed code achieves perfect secrecy for the stated rates, with the secrecy verification relying on vanishing decoding error and Fano’s lemma.The analysis bounds the secrecy expression’s terms and obtains error terms that vanish as n tends to infinity.
  • Converse optimization: The converse maximizes weighted sums of rates and analyzes the resulting non-convex covariance optimization through KKT conditions.The KKT conditions support construction of enhanced noise covariance matrices satisfying the required ordering and boundary matching.

7 General Gaussian MIMO Multi-receiver Wiretap Channel

The paper proves the secrecy capacity region for the general Gaussian MIMO multi-receiver wiretap channel by reducing it to aligned channels and taking a limiting argument. The resulting region equals the dirty-paper-coding region and is achievable with stochastic Gaussian dirty-paper coding.

  • The proof constructs an aligned channel indexed by α and shows that its secrecy capacity region converges to one containing the original region as α → 0.The construction relies on an equivalent channel obtained through invertible transformations and preserves the dirty-paper-coding region.
  • The theorem’s region is achieved by dirty-paper coding with stochastic encoding under an encoding order specified by a permutation.The rate notation follows the encoding order rather than directly indexing the kth user.
  • The general channel’s secrecy capacity region is given by the convex closure of a union over one-to-one user permutations.
  • The constructed channel is an aligned Gaussian MIMO multi-receiver wiretap channel, so its secrecy capacity region follows from the aligned-channel result.
  • The limiting argument uses degradedness relations and continuity of log determinants to transfer decodability and perfect secrecy from the equivalent channel.The Markov chain X → Ȳ_k → Ŷ_k ensures that rates decodable in the equivalent channel remain decodable in the constructed channel.

8 Conclusions

The paper characterizes the Gaussian MIMO multi-receiver wiretap channel’s secrecy capacity region and establishes a proof methodology based on estimation-theoretic identities, channel enhancement, and limiting arguments. A Gaussian dirty-paper-coding variant achieves the region.

  • The Gaussian MIMO multi-receiver wiretap channel’s secrecy capacity region is characterized.
  • Existing scalar broadcast-channel converse extensions leave ambiguity about auxiliary random variables in the secrecy setting.
  • MMSE–mutual-information and Fisher-information–differential-entropy relationships resolve this ambiguity in the converse proof.
  • The degraded MIMO result is generalized to arbitrary channels using channel enhancement and limiting arguments.
  • A variant of dirty-paper coding with Gaussian signals achieves the characterized region.

A Proof of Lemma 11

The proof evaluates an inner integral by integration by parts and then applies the stated assumption to complete the lemma’s derivation.

  • The inner integral is evaluated using integration by parts.
  • Substituting the resulting expression into the preceding equation completes the proof under the assumption in (183).

B Proof of Lemma 12

The proof establishes the lemma’s parts through properties of conditional probability densities and conditional expectations, then selects g(U) = E[X|U] for the final claim.

  • The first part follows because f(x|u) is a valid probability density function.
  • The second equality uses that the inner expectation is zero by the lemma’s first part.
  • The final part follows by choosing g(U) = E[X|U] in the lemma’s second part.

C Proof of Lemma 14

The proof establishes the lemma by manipulating conditional densities, differentiating the resulting identities, and using vanishing boundary terms and symmetry to complete the argument.

  • C Proof of Lemma 14: Conditional densities are introduced with subscripts denoting the random vector whose density they represent.For example, f_X(x|u) denotes the conditional density of X.
  • C Proof of Lemma 14: Differentiating the density identities and using conditional independence yields the required intermediate relations.
  • C Proof of Lemma 14: Vanishing conditional densities at infinity removes the boundary terms needed for the subsequent calculation.The proof uses that probability density functions f_X(x|u) and f_Y(w − x|u) vanish at infinity.
  • C Proof of Lemma 14: Symmetry supplies the corresponding relation, completing the proof.

D Proof of Lemma 15

The proof derives the lemma through conditional Fisher information identities, invoking conditional independence and earlier lemmas before completing the algebraic substitution.

  • D Proof of Lemma 15: The argument starts from the definition of the conditional Fisher information matrix.
  • D Proof of Lemma 15: Conditional independence of X and Y given U supplies one key equality in the Fisher-information calculation.
  • D Proof of Lemma 15: Earlier lemmas and the definition of conditional Fisher information provide the remaining intermediate identities.
  • D Proof of Lemma 15: Substituting the derived relations into the main identity completes the proof.

E Proof of Lemma 17

The proof uses Markov-chain and factorization properties to analyze Fisher-information cross-terms, showing that the cross-terms vanish under the stated differentiation–integration assumption.

  • E Proof of Lemma 17: The derivation uses the Markov chain V → U → X and the joint-density factorization f(x,u,v) = f(x,v)f(u|x,v).
  • E Proof of Lemma 17: The resulting identities are substituted into the main expression before analyzing its cross-terms.
  • E Proof of Lemma 17: The inner integral is evaluated using an assumption that justifies interchanging differentiation and integration.
  • E Proof of Lemma 17: The evaluated cross-term is zero, and substituting this result completes the proof.

F Proof of Lemma 18

The proof derives covariance-ordering and definiteness properties from KKT conditions, positive-semidefinite inequalities, ordered multipliers, and an inductive reverse-order argument.

  • F Proof of Lemma 18: KKT conditions yield relations involving K_j, M_j, and neighboring covariance terms.
  • F Proof of Lemma 18: The resulting relations imply 0 ⪯ ˜Σ_j ⪯ Σ_j for j = 1, . . . , m.
  • F Proof of Lemma 18: Positive semidefiniteness and the ordering µ_m ≥ µ_{m−1} imply ˜Σ_{m−1} ⪯ ˜Σ_m ⪯ Σ_Z.
  • F Proof of Lemma 18: Checking the lemma's equations in reverse order establishes the corresponding bounds for earlier indices.
  • F Proof of Lemma 18: The argument concludes by proving ˜Σ_1 ≻ 0 and verifying the remaining lemma parts through KKT consequences and analogous reasoning.
Loading 0903.3096v1…