PaperlayerView exact revision on arXiv

Source-linked AI summary

Intelligent Reflecting Surface Assisted Anti-Jamming Communications: A Fast Reinforcement Learning Approach

Helin Yang, Zehui Xiong, Jun Zhao, Dusit Niyato, Qingqing Wu, H. Vincent Poor, Massimo Tornatore

arXiv:2004.12539v1eess.SP

TL;DR

Smart jamming threatens legitimate wireless transmissions, motivating anti-jamming methods that operate despite dynamic and unknown jammer behavior. This paper jointly optimizes BS power allocation and IRS reflecting beamforming with fuzzy WoLFPHC learning, achieving reported gains in system rate and service protection level over existing approaches.

  • Problem

    Smart jamming threatens legitimate wireless transmissions, while the jamming model and behavior may be dynamic and unknown.

  • Method

    The paper jointly optimizes BS transmit power allocation and IRS reflecting beamforming using a fuzzy WoLFPHC learning approach.

  • Results

    21.29% and 13.36% improvements are reported for system rate and service protection level, respectively, over existing approaches.

  • Takeaways & Limitations

    The IRS-assisted learning approach improves anti-jamming performance in terms of system rate and service protection level compared with existing approaches.

Abstract

from arXiv · show

Malicious jamming launched by smart jammers can attack legitimate transmissions, which has been regarded as one of the critical security challenges in wireless communications. With this focus, this paper considers the use of an intelligent reflecting surface (IRS) to enhance anti-jamming communication performance and mitigate jamming interference by adjusting the surface reflecting elements at the IRS. Aiming to enhance the communication performance against a smart jammer, an optimization problem for jointly optimizing power allocation at the base station (BS), and reflecting beamforming at the IRS is formulated while considering quality of service (QoS) requirements of legitimate users. As the jamming model and jamming behavior are dynamic and unknown, a fuzzy win or learn fast-policy hill-climbing (WoLFPHC) learning approach is proposed to jointly optimize the anti-jamming power allocation and reflecting beamforming strategy, where WoLFPHC is capable of quickly achieving the optimal policy without the knowledge of the jamming model, and fuzzy state aggregation can represent the uncertain environment states as aggregate states. Simulation results demonstrate that the proposed anti-jamming learning-based approach can efficiently improve both the IRS-assisted system rate and transmission protection level compared with existing solutions.

I. INTRODUCTION

Wireless communications are vulnerable to smart jamming, while existing anti-jamming approaches face limitations in spectrum use, energy consumption, hardware cost, and unknown attack behavior. The paper introduces IRS-assisted anti-jamming with joint BS power allocation, IRS reflecting beamforming, and fuzzy reinforcement learning, reporting improved system rate and service protection.

  • Smart jammers can transmit interference over legitimate channels, threatening wireless communication security.
  • Frequency hopping may fail against simultaneous multi-channel attacks and requires extra spectrum resources.
  • Existing anti-jamming methods can incur excessive hardware cost, system complexity, and energy consumption.
  • IRS uses adaptively controlled passive reflecting elements to adjust reflected-wave strength and direction.
  • The paper jointly optimizes BS transmit power allocation and IRS reflecting beamforming while satisfying legitimate users' QoS requirements.
  • A fuzzy WoLFPHC approach learns anti-jamming policies in dynamic, unknown environments by aggregating uncertain states.
  • 21.29% and 13.36% improvements are reported for system rate and service protection level, respectively, over existing approaches.

A. System Model

The system uses an IRS-assisted multi-user link in which a smart jammer disrupts legitimate transmissions. IRS reflection, BS power allocation, and beamforming are modeled under quasi-static fading to improve desired reception and mitigate jamming.

  • The system contains one BS, K single-antenna legitimate UEs, an IRS with M reflecting elements, and a multi-antenna smart jammer.
  • Cell-edge links can suffer severe attenuation or blockage, motivating IRS-assisted communication over the frequency band.
  • The jammer sends faked or replayed signals through N_J antennas to degrade legitimate communication performance.
  • The IRS reflection matrix is diagonal, with each element combining reflection amplitude ω_m and phase shift θ_m; the model fixes ω_m = 1.
  • The BS allocates transmit powers P_k subject to a maximum-power constraint and uses unit-norm beamforming vectors w_k for the UEs.
  • Each UE receives the BS signal, IRS-reflected signal, jammer interference, inter-user interference, and additive Gaussian noise, yielding a received SINR.

B. Problem Formulation

The paper formulates joint BS power-allocation and IRS reflecting-beamforming optimization to maximize system rate under power and minimum-SINR constraints. Because the variables are coupled in a non-convex objective and jamming dynamics are uncertain, the formulation is transformed into a reinforcement-learning problem.

  • B. Problem Formulation: The optimization jointly selects BS transmit powers {P_k} and IRS reflecting matrix Φ to maximize the achievable rate of all UEs against smart jamming.
  • B. Problem Formulation: The formulation imposes a BS total-power limit and minimum received SINR thresholds for every legitimate UE.
  • B. Problem Formulation: The problem is difficult because its objective is non-concave in Φ and power variables are intricately coupled with Φ.
  • B. Problem Formulation: Alternating optimization and semidefinite-relaxation methods can address single-time-slot instances but may produce suboptimal, greedy-search-like behavior by ignoring historical information and long-term benefit.
  • B. Problem Formulation: The transmit beamforming vector w_k is fixed by maximizing output SINR, while power allocation and IRS reflecting beamforming remain the joint optimization variables.
  • B. Problem Formulation: Unknown jamming models, channel variation, and dynamic uncertain behavior further increase the difficulty of solving the optimization problem.
  • A. Optimization Problem Transformation Based on RL: The RL formulation treats the IRS-assisted system as the environment and the BS controller as the learning agent, with states, actions, transitions, policies, and rewards.
  • A. Optimization Problem Transformation Based on RL: States encode discretized observed information, including previous jamming power and estimated channel coefficients, while actions select power allocation and IRS phase shifts.

B. Fuzzy WoLF-PHC-Based Joint Power Allocation and Reflecting Beamforming

The proposed approach combines WoLF-PHC reinforcement learning with fuzzy state aggregation to jointly learn anti-jamming power allocation and IRS reflecting beamforming under unknown, dynamic jamming. It uses fuzzy representations and exploration–exploitation updates to adapt decisions in uncertain environments.

  • The approach is designed for unknown jamming behavior, where conventional optimization may converge to suboptimal solutions and the optimal anti-jamming policy is difficult to obtain without the jamming model.
  • ε-greedy selection balances exploitation of the highest-Q action with random exploration to avoid non-optimal policies.
  • WoLF-PHC updates the action policy with separate learning rates for winning and losing states, using ξ_loss > ξ_win.
  • Fuzzy state aggregation represents continuous system states with a fixed number of aggregate states, reducing the state space handled by WoLF-PHC.
  • The fuzzy state-action value function combines fuzzy-state values according to each state’s degree of relationship to the fuzzy states.
  • The trained agent observes estimated jamming power, SINR values, and channel coefficients, then selects joint power and reflecting-beamforming actions and updates from rewards and new states.
  • After convergence, the learned model searches for BS power allocation and IRS reflecting-matrix strategies according to the selected action.

IV. SIMULATION RESULTS AND ANALYSIS

The simulations evaluate IRS-assisted anti-jamming approaches under dynamic smart-jamming settings, comparing fuzzy WoLF-PHC learning with iterative optimization, fast Q-Learning, and optimal power allocation without IRS. Experiments vary transmit power, SINR targets, IRS size, and jamming behavior across averaged realizations.

  • Learning procedure: Algorithm 1 observes system states, selects ε-greedy actions, receives rewards and next states, updates policies and fuzzy Q-values, and outputs joint power allocation and reflecting beamforming.The resulting fuzzy WoLF-PHC model is loaded to produce the joint control strategy.
  • Simulation setup: The experiments vary BS maximum transmit power from 15 to 40 dBm, SINRmin from 10 to 25 dB, and IRS elements from 20 to 100.Simulation results are averaged over 500 independent realizations.
  • Simulation setup: The simulated smart jammer has unknown current jamming power, while the BS estimates previous power levels from historical information.The jamming power ranges from 15 dBm to 40 dBm according to jammer behavior.
  • Simulation setup: The evaluation compares fuzzy WoLF-PHC joint power allocation and reflecting beamforming with system-rate maximization, fast Q-Learning, and optimal power allocation without IRS.The proposed approach and alternatives are assessed in an IRS-assisted system with a jammer.

A. Convergence Comparisons of Different Approaches

The proposed fuzzy WoLF-PHC learning approach converges faster than the comparison methods while improving achievable rate and SINR protection. IRS assistance supports stronger performance, whereas the fastest no-IRS method has the worst overall performance.

  • Convergence performance: The proposed fuzzy WoLF-PHC approach accelerates convergence and enhances system rate and SINR protection compared with fast Q-Learning and Baseline 1.The paper attributes this to WoLF-PHC and fuzzy state aggregation improving learning efficiency in dynamic environments.
  • Convergence trade-offs: Optimal power allocation without IRS converges fastest but has the worst performance because it lacks IRS-based performance improvement and jamming resistance.The paper notes that slow convergence may fail to protect SINR in real-time systems.
  • Convergence speed: Within 11 iterations, the proposed approach converges, compared with about 24 iterations for fast Q-Learning and 15 for Baseline 1.The proposed method also saves 36.36% convergence time compared with Baseline 1.
  • Convergence performance: The proposed approach improves system achievable rate by 21.29% and SINR protection level by 13.36% at a stable level.These improvements are reported relative to Baseline 1.
  • Convergence performance: At convergence, SINR protection levels are about 1, 0.89, 0.93, and 0.86 for the compared approaches, respectively.The passage reports these values in the order of the approaches discussed in the convergence comparison.

B. Performance Comparisons versus Maximum Transmit Power

Increasing maximum transmit power improves both system rate and SINR protection. The proposed learning approach and Baseline 1 achieve strong system rates, while the proposed method provides higher SINR protection than the baselines and fast Q-Learning.

  • Overall trends: System rate and SINR protection level improve as maximum transmit power Pmax increases.The comparisons use K = 4, M = 60, and SINRmin = 10 dB.
  • System rate: The proposed learning approach and Baseline 1 have good system-rate values across Pmax and greatly outperform the other approaches.This comparison concerns system-rate performance over different maximum transmit powers.
  • SINR protection: Baseline 1 has obviously lower SINR protection than the proposed and fast Q-Learning approaches because it performs only single-time-slot optimization.The paper links this limitation to ineffective QoS guarantee.
  • IRS benefit: IRS-based performance improvement over no IRS increases with Pmax, indicating an advantage of deploying IRS against smart jamming.The optimal no-IRS power-allocation approach can have comparable SINR protection to IRS-assisted fast Q-Learning, but assumes perfect CSI and a known jamming model.
  • Method comparison: The proposed approach has higher system rate and SINR protection level than fast Q-Learning.The paper attributes this to WoLF-PHC and fuzzy state aggregation searching the joint strategy in dynamic uncertain environments.

C. Performance Comparisons versus Number of Reflecting Elements

Increasing the number of IRS elements improves anti-jamming performance, with the proposed learning approach widening its rate advantage and reaching full SINR protection earlier than the alternatives. The reported gain over no IRS grows substantially with IRS size.

  • Overall trends: All IRS-based approaches improve with the number of reflecting elements and greatly outperform optimal power allocation without IRS.The paper attributes this to the IRS providing higher degrees of freedom for performance optimization.
  • Achievable-rate gain: At M = 20, the proposed approach gains about 2.36 bits/s/Hz over optimal power allocation without IRS, increasing to 12.21 bits/s/Hz at M = 100.These values quantify the growing achievable-rate benefit of IRS assistance.
  • Mechanism: Increasing M improves the IRS reflecting beamforming design, helping strengthen desired signals and mitigate smart-jammer interference.The paper links the performance gain to higher IRS power and reflecting beamforming gain.
  • Achievable rate: The proposed learning approach outperforms fast Q-Learning and Baseline 1 in achievable rate, with the gap increasing significantly as M grows.This comparison is reported in Fig. 6(a).
  • SINR protection: As reflecting elements increase, the proposed approach is the first among the compared methods to achieve 100% SINR protection level.The comparison is reported in Fig. 6(b).
  • Conclusion: Employing IRS improves anti-jamming communication performance against smart jamming.The conclusion follows the reported comparisons across IRS size settings.

D. Impact of SINR Target

As the UE SINR target increases, system rate and SINR protection decline slightly at first, then drop significantly beyond a threshold. IRS-assisted approaches outperform the no-IRS baseline, with the proposed fuzzy WoLF-PHC approach achieving the best performance.

  • SINR-target effects: At low SINR targets, system rate and SINR protection level decline slightly as the target increases, then drop significantly beyond a threshold.The decline is attributed to interference becoming the performance bottleneck at high UE SINR targets.
  • IRS benefit: IRS-assisted approaches achieve higher system rate and SINR protection than the optimal power-allocation approach without IRS.The SINR protection gap increases significantly as the SINR target rises.
  • IRS benefit: IRS reflecting beamforming enhances desired signal power and mitigates interference from the smart jammer.This provides the mechanism behind the stronger performance of IRS-assisted approaches.
  • Learning approaches: The proposed fuzzy WoLF-PHC and fast Q-learning approaches significantly improve SINR protection against smart jamming compared with Baseline 1.Both learning approaches use a SINR-aware reward function.
  • Joint optimization: The proposed learning approach achieves the best performance, indicating that jointly optimizing power allocation and reflecting beamforming is crucial.The result supports combining practical reflecting beamforming with more sophisticated optimization.

V. CONCLUSIONS

The paper proposes IRS-assisted anti-jamming communication by jointly optimizing BS power allocation and IRS reflecting beamforming. A fuzzy WoLF-PHC learning approach addresses the non-convex optimization and unknown jamming model, while simulations show improved system rate and service protection level.

  • Contribution: The IRS enhances legitimate-user signals while mitigating smart-jammer interference in wireless communication systems.The proposed system uses IRS assistance to improve anti-jamming performance.
  • Optimization: The optimization jointly designs BS transmit power allocation and IRS reflecting beamforming under a non-convex formulation.The joint design targets the anti-jamming strategy rather than either control variable alone.
  • Learning method: Fuzzy WoLF-PHC learning is proposed to obtain an anti-jamming strategy when the jamming model is unknown.WoLF-PHC and fuzzy state aggregation are described as enabling rapid policy learning in uncertain environments.
  • Results: Simulation results verify improved system rate and service protection level compared with existing approaches.The conclusion attributes these improvements to the IRS-assisted system and proposed learning approach.
Loading 2004.12539v1…