PaperlayerView exact revision on arXiv

Source-linked AI summary

Physical Layer Security in Massive MIMO

Y. Ozan Basciftci, C. Emre Koksal, Alexei Ashikhmin

arXiv:1505.00396v2cs.ITmath.IT

TL;DR

The paper studies how simultaneous jamming and eavesdropping affect secure communication in single-cell massive MIMO. It compares attacks that spare pilot training with attacks that jam it, then proposes hiding pilot assignments; secure DoF matches the no-attack maximum without training jamming, while the proposed defense recovers that benchmark under training jamming.

  • Problem

    The paper asks how jamming and eavesdropping affect massive-MIMO downlink security, particularly when an adversary can attack pilot training.

    Source: §I · p. 1

  • Method

    The paper analyzes training-phase and data-only jamming, develops δ-conjugate beamforming, and proposes hiding randomized pilot assignments from the adversary.

    Source: §I · p. 1, A. Counter strategy against… · p. 9, §VI · p. 12

  • Results

    Secure DoF matches the no-attack maximum without training-phase jamming, falls to zero under training-phase jamming, and can reach the no-attack maximum when pilot assignments are hidden and the pilot set is extended.

    Source: §I · p. 1, §VI · p. 12

  • Takeaways & Limitations

    Pilot-assignment encryption is presented as a candidate defense, with computational cryptography offering security comparable to information-theoretic methods for practical purposes.

    Source: §I · p. 1, §VI · p. 12

Abstract

from arXiv · show

We consider a single-cell downlink massive MIMO communication in the presence of an adversary capable of jamming and eavesdropping simultaneously. We show that massive MIMO communication is naturally resilient to no training-phase jamming attack in which the adversary jams only the data communication and eavesdrops both the data communication and the training. Specifically, we show that the secure degrees of freedom (DoF) attained in the presence of such an attack is identical to the maximum DoF attained under no attack. Further, we evaluate the number of antennas that base station (BS) requires in order to establish information theoretic security without even a need for Wyner encoding. Next, we show that things are completely different once the adversary starts jamming the training phase. Specifically, we consider an attack, called training-phase jamming in which the adversary jams and eavesdrops both the training and the data communication. We show that under such an attack, the maximum secure DoF is equal to zero. Furthermore, the maximum achievable rates of users vanish even in the asymptotic regime in the number of BS antennas. To counter this attack, we develop a defense strategy in which we use a secret key to encrypt the pilot sequence assignments to hide them from the adversary, rather than encrypt the data. We show that, if the cardinality of the set of pilot signals are scaled appropriately, hiding the pilot signal assignments from the adversary enables the users to achieve secure DoF, identical to the maximum achievable DoF under no attack. Finally, we discuss how computational cryptography is a legitimate candidate to hide the pilot signal assignments. Indeed, while information theoretic security is not achieved with cryptography, the computational power necessary for the adversary to achieve a non-zero mutual information leakage rate goes to infinity.

I. INTRODUCTION

The paper argues that massive-MIMO security depends chiefly on whether the adversary jams pilot training: without training-phase jamming, secure DoF matches the no-attack maximum, while pilot contamination can undermine security. It proposes hiding pilot assignments as a defense and sets these results in a single-cell, TDD downlink model.

I. INTRODUCTION

  • The paper treats massive-MIMO security separately because ordinary MIMO security models can overlook vulnerabilities tied to how the system obtains CSI.It argues that the adversary’s access to CSI is less decisive than whether it can interfere with pilot training.Source: §I · pp. 1–2
  • Encrypting pilot assignments requires the adversary to decrypt them before training begins, pushing the computational power needed for useful cryptanalysis toward infinity.The paper discusses standard Diffie–Hellman key sharing and says computational cryptography can provide security comparable to information-theoretic methods for practical purposes.Source: §I · p. 2
  • Without training-phase jamming, the maximum secure DoF equals the maximum DoF under no adversary.The adversary may eavesdrop on training and data while jamming only data communication.Source: §I · pp. 1–2
  • δ-conjugate beamforming provides full security without Wyner encoding, and the paper evaluates the BS antenna count needed to satisfy security constraints.The strategy uses the zero correlation between estimated user-channel gains and adversary-channel gains when the adversary does not jam training.Source: §I · p. 2
  • Training-phase jamming that creates non-zero correlation between the adversary’s channel and a user’s estimated channel reduces maximum secure DoF to zero.With sufficiently large correlation, the user’s maximum achievable rate also vanishes as the BS antenna count grows.Source: §I · p. 2
  • Hiding pilot assignments and scaling up the pilot-signal set lets users approach the no-attack maximum DoF under training-phase jamming.The defense encrypts pilot assignments rather than transmitted data.Source: §I · pp. 1–2

II. SYSTEM MODEL AND PROBLEM STATEMENT

  • The system is a single-cell downlink with an M-antenna BS, K single-antenna users, and an adversary.The BS sends a distinct message to each user.Source: §I · p. 1, §II · pp. 2–3

A. Channel Model

  • In the TDD channel model, users send pilots during training and the BS estimates their channels before transmitting data in the block’s remaining channel uses.Channels are block-fading: gains remain constant within a block and are independent across blocks.Source: A. Channel Model · p. 3

B. Attack Model

The section distinguishes attacks by whether the full-duplex adversary jams during training, then defines the code and security measures used to analyze them. It focuses on secure rates and secure degrees of freedom under those attack models.

B. Attack Model

  • The two attack models differ in training: the adversary jams only data in no training-phase jamming, but jams and eavesdrops both phases in training-phase jamming.In the first model it eavesdrops training and data while remaining silent during training; in the second, its data-phase strategy is unchanged and it uses structured training signals.Source: B. Attack Model · p. 4

C. Code Definition

  • A code maps each user's message to a data sequence and then uses linear beamforming to form the transmitted channel input.Encoding functions may be deterministic or stochastic; the paper considers conjugate and δ-conjugate beamforming.Source: C. Code Definition · p. 4

D. Figures of Merit

  • The paper defines information-theoretic security through decodability and secrecy constraints, and measures secure DoF by the least user's secure degrees of freedom.Its DoF tracks rate growth with the number of BS antennas, rather than transmission power; the analysis characterizes maximum secure DoF and defenses against attacks that otherwise reduce it to zero.Source: D. Figures of Merit · p. 5

III. ADVERSARY NOT JAMMING THE TRAINING PHASE

Without training-phase jamming, the paper shows that massive MIMO retains the maximum DoF and can meet information-theoretic security constraints without stochastic encoding. It also gives antenna requirements and shows how δ-conjugate beamforming can bring the no-stochastic-encoding DoF arbitrarily close to the maximum.

A. Resilience of massive MIMO

  • The maximum secure DoF under no training-phase jamming equals the maximum DoF without an adversary: Td/T.The attack does not degrade DoF because the adversary remains silent during training, leaving estimated BS-to-user channel gains independent of its channel.Source: A. Resilience of massive MIMO · p. 5
  • With fixed Me, achievable secure rates grow unboundedly as M increases, while at fixed M they decrease as Me increases.The figure's example uses ρk = 1, ρf = 10, Td/T = 0.99, ρjam = 1, and a = 0.9.Source: A. Resilience of massive MIMO · p. 6, Figure 2 · p. 6

B. Establishing security without Wyner encoding

  • Deterministic encoding with δ-conjugate beamforming satisfies security constraints without stochastic encoding when M is sufficiently large.For ϵ = 0.05, T/Td = 5/4, Me = 1, ρk = 1, and Rk = 0.2, the required antenna threshold is minimized at 60 when δ = 0.78.Source: B. Establishing security without Wyner… · pp. 6, 7, 11, Figure 4 · p. 7
  • For any δ > 0, M ≥ S(ϵ) lets any deterministic code satisfy the secrecy constraint; using M ≥ max(V(R), S(ϵ)) also meets decodability without stochastic encoding.S(ϵ) is the antenna count needed to make conditional entropy ϵ-close to Rk for every user.Source: B. Establishing security without Wyner… · pp. 6, 11
  • Choosing δ close to zero makes the deterministic-encoding DoF arbitrarily close to the maximum secure DoF.Figure 3 relates S(ϵ) to ϵ; for its stated parameters, 100 BS antennas suffice to make equivocation exceed Rk − 0.05 for any Rk.Source: B. Establishing security without Wyner… · pp. 6–7, Figure 3 · p. 7

IV. ADVERSARY JAMMING THE TRAINING PHASE

Training-phase jamming can make secure communication fundamentally different from data-only jamming: a nonzero correlation between the adversary channel and an estimated user channel can eliminate secure DoF. Under stronger correlation conditions, a user's maximum rate also vanishes as the number of BS antennas grows.

IV. ADVERSARY JAMMING THE TRAINING PHASE

  • A nonzero correlation between an estimated user channel and the adversary channel limits the maximum secure DoF to zero.The theorem's proof states that the corresponding rate bound divided by log M tends to zero as M grows.Source: §IV · pp. 7–8
  • Without training-phase jamming, the estimated user channel and adversary channel are independent, so the correlation needed for the zero-DoF attack is absent.The paper states that the adversary must jam training to create nonzero correlation between those channel gains.Source: §IV · p. 8
  • A user's maximum achievable rate vanishes as M grows when the BS-to-user and adversary-channel correlation exceeds the correlation involving the estimated user channel.This rate result is stated under the additional assumptions of Corollary 2.Source: §IV · p. 8
  • Under training-phase jamming, achievable rates and maximum secure DoF are determined by the resource race between the adversary and users.For MMSE estimation with mutually orthogonal pilots, matching a target user's pilot satisfies the theorem's assumptions and yields zero secure DoF.Source: §IV · pp. 8–9

V. SECURE COMMUNICATION UNDER TRAINING-PHASE JAMMING

The section introduces a defense against training-phase jamming that enlarges the pilot set and hides user assignments from the adversary. It also describes computational cryptography for protecting those assignments and security without Wyner encoding.

V. SECURE COMMUNICATION UNDER TRAINING-PHASE JAMMING

  • The defense expands the pilot-signal set and keeps pilot assignments hidden from the adversary.The assignments are encrypted rather than the data communication.Source: §V · p. 9
  • With δ-conjugate beamforming, the BS can satisfy the security constraints under training-phase jamming without Wyner encoding.The section overview presents computational cryptography as a way to communicate pilot assignments.Source: §V · p. 9
  • The proposed defense uses computational cryptography to hide pilot assignments, rather than encrypting the data.The paper discusses the security achieved when computational cryptography protects the assignments.Source: §V · p. 9

A. Counter strategy against training-phase jamming

The counterstrategy randomly assigns users pilots from an enlarged set while keeping assignments hidden from the adversary. With an appropriately scaled pilot set, it can approach the no-attack secure DoF, at the cost of increased training overhead.

A. Counter strategy against training-phase jamming

  • The BS randomly assigns each user a distinct pilot from a set of L mutually orthogonal signals, with L at least K.The BS keeps these assignments hidden from the adversary while communicating them reliably to users.Source: A. Counter strategy against… · p. 9
  • When the pilot-set size scales appropriately with the number of BS antennas, the achieved secure rate per log M can approach the maximum no-attack DoF.The theorem's lower bound does not depend on how many pilot signals the adversary contaminates.Source: A. Counter strategy against… · p. 10
  • For γ = 1, the secure rate per log M can approach the no-attack maximum DoF by extending the training phase and setting the pilot-set size to T_r.The stated condition is T_r ≥ max(K, G(ϵ)); the resulting longer training phase sacrifices secure throughput.Source: A. Counter strategy against… · p. 10
  • Increasing T_r toward T makes ϵ vanish and brings the secure rate per log M closer to T_d/T.The section identifies increased training overhead as the cost of this approach.Source: A. Counter strategy against… · p. 10
  • With hidden assignments and a pilot set that scales with M, the adversary's jamming power must scale with L to make an impact.The strategy shifts the resource race from the adversary versus the target user to the adversary versus the BS.Source: A. Counter strategy against… · p. 10

B. Establishing security without Wyner encoding

This section shows that deterministic encoding can satisfy secrecy and decodability under training-phase jamming without Wyner encoding, given sufficient antennas and training. It then describes hiding pilot assignments with computational cryptography; the conclusion states that expanding the pilot set can recover the no-attack secure DoF.

B. Establishing security without Wyner encoding

  • The deterministic-encoding strategy uses δ-conjugate beamforming to meet the secrecy and decodability constraints during training-phase jamming.The section presents this as the counterpart to the no-training-jamming result, using a different mitigation strategy.Source: B. Establishing security without Wyner… · pp. 10–11
  • Choosing γ trades antenna requirements against training overhead: γ near 1 minimizes S1(ϵ) but increases the required training length.The training constraint is Tr ≥ max(M^γ, K).Source: B. Establishing security without Wyner… · p. 11
  • Under the stated antenna and training-length conditions, code sequences satisfy secrecy and decodability constraints for any rate tuple without stochastic encoding.Corollary 3 requires M ≥ max(V1(R), S1(ϵ)) and Tr ≥ M^γ.Source: B. Establishing security without Wyner… · p. 11

C. How do we hide the pilot signal assignments?

  • The BS shares a secret key with each user to encrypt that user's pilot assignment; the paper rules out channel-gain-based key generation because training is needed to observe those gains.It proposes standard private- or public-key methods, including Diffie-Hellman and RSA.Source: C. How do we hide the pilot signal… · p. 11
  • Pilot-assignment encryption is not shown to provide information-theoretic security, though the authors argue the attacker's useful cryptanalysis window can be arbitrarily small.The pilot must be identified before training starts; the paper contrasts this with encrypting messages, which remain vulnerable to later cryptanalysis.Source: C. How do we hide the pilot signal… · pp. 11–12

VI. CONCLUSION

  • Hiding pilot assignments and expanding the pilot-signal set can achieve secure DoF equal to the maximum under no attack.The conclusion identifies computational cryptographic key sharing as a candidate way to hide those assignments.Source: §VI · p. 12

APPENDIX A PROOF OF THEOREM 1

The proof bounds secure rates by simplifying the system under favorable assumptions, then constructs codes and evaluates their secrecy, decodability, and degrees of freedom. Its achievability argument uses channel estimation, beamforming, and coding analysis.

APPENDIX A PROOF OF THEOREM 1

  • The upper-bound argument assumes one user, no adversary, perfect channel knowledge, and user knowledge of the received pilots, reducing the model to a MISO channel.The resulting channel capacity upper-bounds secrecy rates in the original model.Source: Appendix A · p. 12
  • The achievability construction uses orthogonal pilots and MMSE channel estimation, then maps codewords to channel inputs with conjugate beamforming.The proof also specifies codebook generation and typical-set decoding.Source: Appendix A · p. 13
  • The proof analyzes decoding error and equivocation before evaluating the users' degrees of freedom.It establishes a sequence of codes satisfying the secrecy and decodability constraints.Source: Appendix A · pp. 13–15

A. Proof of Theorem 2

The appendix proves secrecy and decodability results for deterministic encoding, with separate arguments for Theorem 2 and Corollary 1. The corollary imposes antenna and training-length conditions on the achievable rate tuple.

A. Proof of Theorem 2

B. Proof of Corollary 1

  • For Corollary 1, the proof requires M ≥ max(V(R), S(ϵ)); under this condition it constructs codes satisfying secrecy and decodability.The proof maps each message directly to a codeword, without randomization in that mapping.Source: B. Proof of Corollary 1 · p. 15

APPENDIX C

This appendix derives an upper bound on achievable rates under secrecy and decodability constraints, then uses limiting arguments to bound secure degrees of freedom. The proof invokes covariance bounds and properties of the channel variables.

APPENDIX C

  • The proof begins from an achievable rate under the secrecy and decodability constraints and derives an upper bound on that rate.The derivation uses conditioning and memorylessness properties of the channel.Source: A. Proof of Theorem 3 · pp. 15–16
  • The rate bound is simplified using covariance-matrix inequalities, Jensen's inequality, and monotonicity properties.The proof defines covariance matrices for the transmitted signals and applies cited bounds to them.Source: A. Proof of Theorem 3 · pp. 16–17
  • The secure-DoF bound follows by taking limits after terms in the rate bound vanish as the block count grows.The proof states that γB, ϵB, and δB tend to zero as B grows.Source: A. Proof of Theorem 3 · pp. 16–18

B. Proof of Corollary 2

The Corollary 2 proof uses boundedness and dominated convergence, while Appendix D derives channel-estimation and achievable-rate bounds for Theorem 4. Those bounds distinguish two regimes of γ and use a pilot set sized to the training phase.

B. Proof of Corollary 2

  • The Corollary 2 argument assumes conjugate beamforming without loss of generality and extends to δ-conjugate beamforming by the same argument.Source: B. Proof of Corollary 2 · p. 18
  • The Corollary 2 proof applies dominated convergence after bounding g(M) and showing that the second logarithm term has a finite lower-bounded expectation.Source: B. Proof of Corollary 2 · p. 18

APPENDIX D PROOF OF THEOREM 4

  • Setting the pilot-set size L to T_r allows the base station to generate L mutually orthogonal pilot signals when the training length is at least L.Source: Appendix D · pp. 18–19
  • The base station estimates each user's channel by projecting the training signal onto that user's pilot and normalizing the projection.The event Π_i = 1 indicates that the adversary contaminates that user's pilot in block i.Source: Appendix D · p. 19
  • The achievable rate obtained with stochastic encoding and conjugate beamforming does not depend on J.Source: Appendix D · p. 19
  • The proof derives asymptotic lower bounds on R_k log M in the γ≤1 and γ>1 regimes once M is sufficiently large.Source: Appendix D · p. 19

APPENDIX E

Theorem 5 and Corollary 3 use a pilot set as long as the training phase and channel-gain estimation as in Theorem 4. Their proofs establish asymptotic bounds and construct codes meeting decodability and secrecy constraints under specified parameter conditions.

A. Proof of Theorem 5

B. Proof of Corollary 3

Related papers

Loading 1505.00396v2…