cs.AIFeb 16, 2026

Secure and Energy-Efficient Wireless Agentic AI Networks

Authors: Yuanyan Song, Kezhi Wang, Xinmian Xu

Organizations: Department of Computer Science, Brunel University of London, UB8 3PH, UK · Nanjing University of Posts and Telecommunications

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

In this paper, we introduce a secure wireless agentic AI network comprising one supervisor AI agent and multiple other AI agents to provision quality of service (QoS) for users' reasoning tasks while ensuring confidentiality of private knowledge and reasoning outcomes. Specifically, the supervisor AI agent can dynamically assign other AI agents to participate in cooperative reasoning, while the unselected AI agents act as friendly jammers to degrade the eavesdropper's interception performance. To extend the service duration of AI agents, an energy minimization problem is formulated that jointly optimizes AI agent selection, base station (BS) beamforming, and AI agent transmission power, subject to latency and reasoning accuracy constraints. To address the formulated problem, we propose two resource allocation schemes, ASC and LAW, which first decompose it into three sub-problems. Specifically, ASC optimizes each sub-problem iteratively using the proposed alternating direction method of multipliers (ADMM)-based algorithm, semi-definite relaxation (SDR), and successive convex approximation (SCA), while LAW tackles each sub-problem using the proposed large language model (LLM) optimizer within an agentic workflow. The experimental results show that the proposed solutions can reduce network energy consumption by up to 59.1% compared to other benchmark schemes. Furthermore, the proposed schemes are validated using a practical agentic AI system based on Qwen, demonstrating satisfactory reasoning accuracy across various public benchmarks.

Figures & tables

Fig. 1: A typical workflow of agentic AI to execute a specific reasoning task.
Fig. 2: The framework of a secure wireless agentic AI network.
 \begin{split}\mathcal{L}^{\textit{LR}}(\bm{\alpha},\mu,\lambda)&=\sum_{n\in\mathcal{N}}(\hat{E}_{n}^{\textit{exe}}+E_{n}^{\textit{tr}})+\rho\|\bm{\alpha}-\hat{\bm{\alpha}}^{r_{\textit{ADMM}}-1}\|_{2}^{2}+\mu(\sum_{n\in\mathcal{N}}\alpha_{n}-N_{\textit{min}})+\lambda(\sum_{n\in\mathcal{N}}\alpha_{n}\textit{Acc}_{n}-\textit{Acc}_{\textit{min}}\sum_{n\in\mathcal{N}}\alpha_{n})\\ &=\sum_{n\in\mathcal{N}}\rho\alpha_{n}^{2}+\alpha_{n}(p_{n}^{tr}t_{n}^{ul}+p_{n}^{\textit{exe}}t_{n}^{\textit{inference}}-\frac{s+D}{B\cdot\hat{c}^{dl}}-\underset{\begin{subarray}{c}\forall n\in\mathcal{N}\\ \hat{\alpha}_{n}^{r_{\textit{ADMM}}-1}=1\end{subarray}}{\max}\{\frac{o_{n}}{B\cdot c_{n}^{ul}}\}+\mu+\lambda(\textit{Acc}_{n}-\textit{Acc}_{\textit{min}})-2\rho\hat{\alpha}_{n}^{r_{\textit{ADMM}}-1})\\ &\quad+\rho(\hat{\alpha}_{n}^{r_{\textit{ADMM}}-1})^{2}.\end{split} (27)
Table 3
 {−r^ndl}ub≜\displaystyle\{-\hat{r}_{n}^{dl}\}^{ub}\triangleq−log2(1+∣∣hBS,nHw∣∣2∑n′∈N(1−α^n′rADMM−1)hn′,n2pn′tr+σ2)\displaystyle-{\rm log}_{2}\left(1+\dfrac{||\bm{h}_{\textit{BS},n}^{H}\bm{w}||^{2}}{\sum_{n^{\prime}\in\mathcal{N}}(1-\hat{\alpha}_{n^{\prime}}^{r_{\textit{ADMM}}-1}){h}_{n^{\prime},n}^{2}{p_{n^{\prime}}^{tr}}+\sigma^{2}}\right) (39)
  −∑n′∈N∣∣hBS,nHw∣∣2hn′,n2pn′tr(α^n′−α^n′rADMM−1)(∑n′∈N(1−α^n′rADMM−1)hn′,n2pn′tr+σ2)(∑n′∈N(1−α^n′rADMM−1)hn′,n2pn′tr+σ2+∣∣hBS,nHw∣∣2)\displaystyle-\sum_{n^{\prime}\in\mathcal{N}}\frac{||\bm{h}_{\textit{BS},n}^{H}\bm{w}||^{2}{h}_{n^{\prime},n}^{2}{p_{n^{\prime}}^{tr}}(\hat{\alpha}_{n^{\prime}}-\hat{\alpha}_{n^{\prime}}^{r_{\textit{ADMM}}-1})}{\left(\sum_{n^{\prime}\in\mathcal{N}}(1-\hat{\alpha}_{n^{\prime}}^{r_{\textit{ADMM}}-1}){h}_{n^{\prime},n}^{2}{p_{n^{\prime}}^{tr}}+\sigma^{2}\right)\left(\sum_{n^{\prime}\in\mathcal{N}}(1-\hat{\alpha}_{n^{\prime}}^{r_{\textit{ADMM}}-1}){h}_{n^{\prime},n}^{2}{p_{n^{\prime}}^{tr}}+\sigma^{2}+||\bm{h}_{\textit{BS},n}^{H}\bm{w}||^{2}\right)} 
Table 4
 rdl,lb\displaystyle r^{dl,lb}=min⁡∀n∈Nαn=1{log2(∣∣hBS, nHw∣∣2+∑n′∈N(1−αn′)hn’, n2pn′tr,r−1+σ2)+∑n′∈N(1−αn′)hn’, n2(pn′tr−pn′tr,r−1)ln2(∣∣hBS, nHw∣∣2+∑n′∈N(1−αn′)hn’, n2pn′tr,r−1+σ2)\displaystyle=\underset{\begin{subarray}{c}\forall n\in\mathcal{N}\\ \alpha_{n}=1\end{subarray}}{\min}\{{\rm log}_{2}(||\bm{h}_{\textit{BS, n}}^{H}\bm{w}||^{2}+\sum_{n^{\prime}\in\mathcal{N}}(1-\alpha_{n^{\prime}}){h}_{\textit{n', n}}^{2}{p_{n^{\prime}}^{tr,r-1}}+\sigma^{2})+\frac{\sum_{n^{\prime}\in\mathcal{N}}(1-\alpha_{n^{\prime}}){h}_{\textit{n', n}}^{2}(p_{n^{\prime}}^{tr}-p_{n^{\prime}}^{tr,r-1})}{{\rm ln}2(||\bm{h}_{\textit{BS, n}}^{H}\bm{w}||^{2}+\sum_{n^{\prime}\in\mathcal{N}}(1-\alpha_{n^{\prime}}){h}_{\textit{n', n}}^{2}{p_{n^{\prime}}^{tr,r-1}}+\sigma^{2})} (54)
  −log2(∑n′∈N(1−αn′)hn’, n2pn′tr+σ2)},\displaystyle-{\rm log}_{2}(\sum_{n^{\prime}\in\mathcal{N}}(1-\alpha_{n^{\prime}}){h}_{\textit{n', n}}^{2}{p_{n^{\prime}}^{tr}}+\sigma^{2})\}, 
Table 5
Fig. 3: The framework of LAW.
NotationValueNotationValue
pBStrp_{\text{BS}}^{\text{tr}}10 WBB5 MHz
pntr,maxp_{n}^{\text{tr,max}}2 Wσ2\sigma^{2}-107 dBm
pnexep_{n}^{\text{exe}}[10, 20] Wknk_{n}[2560, 4096]
DD10 Mbitsτ\tau3000 s
TABLE I: Simulation parameters [ 28 , 33 ] .
Fig. 4: The execution energy consumption versus different numbers of AI agents.
SchemesNmin=10N_{\text{min}}=10Nmin=15N_{\text{min}}=15Nmin=20N_{\text{min}}=20
RandAS234174110748999
GA150192998338402
LAW136932843636113
ASC118682667334996
TABLE II: The execution energy consumption (J) under different minimum numbers of reasoning AI agents ( NminN_{\text{min}} ).
SchemesAccmin=60%Acc_{\text{min}}=60\%Accmin=70%Acc_{\text{min}}=70\%Accmin=80%Acc_{\text{min}}=80\%
RandAS212222341727520
GA131201521916189
LAW120221424214959
ASC104231186812966
TABLE III: The execution energy consumption (J) under different average reasoning accuracy thresholds ( AccminAcc_{\text{min}} ).
Fig. 5: The transmission energy consumption versus different numbers of AI agents.
SchemesL=10L=10L=20L=20L=30L=30
FixedTP310.76299.20225.38
FixedBF260.27246.01171.85
DE155.95139.2154.967
LAW146.01105.2153.785
ASC109.8777.21949.872
TABLE IV: The transmission energy consumption (J) under different numbers of BS antennas ( LL ).
SchemesARC-EARC-CBoolQ
ASC85%75%83%
LAW76%65%82%
GA66%56%80%
RandAS56%41%75%
TABLE V: The reasoning accuracy under different public benchmarks.
More in this topic

Wireless Resource Allocation

Wireless resource allocation assigns spectrum, power, beams, and computing capacity to meet throughput, latency, fairness, and energy goals.

See all 174 papers

Explore similar work

CardsList

  1. Agentic AI-Empowered Wireless Agent Networks With Semantic-Aware Collaboration via ILAC

    Apr 1, 2026Zhouxiang Zhao, Jiaxiang Wang, Zhaohui Yang +4Wireless CommunicationsMulti-Agent Communication

    College of Information Science and Electronic Engineering, Zhejiang University, and also with Zhejiang Provincial Key Laboratory of Info. Proc., Commun. & Netw. (IPCAN), Hangzhou 310027, China · Department of Engineering, King’s College London, London, WC2R 2LS, UK · Department of Electrical and Computer Engineering and Institute for Data Science and Computing, University of Miami, Coral Gables, FL 33146, USA +1

  2. Federated Agentic AI for Wireless Networks: Fundamentals, Approaches, and Applications

    Mar 2, 2026Lingyi Cai, Yu Zhang, Ruichen Zhang +5Wireless CommunicationsAgentic RL

    Research Center of 6G Mobile Communications, School of Cyber Science and Engineering, Huazhong University of Science and Technology, Wuhan, 430074, China · College of Computing and Data Science, Nanyang Technological University, Singapore · School of Engineering, Edith Cowan University, Perth, WA 6027, and the School of Computer Science and Engineering, University of New South Wales (UNSW), Sydney, NSW 2033, Australia +2

  3. Networking-Aware Energy Efficiency in Agentic AI Inference: A Survey

    Apr 9, 2026Xiaojing Chen, Haiqi Yu, Wei Ni +5Energy-Efficient MLLLM Inference Efficiency

    Shanghai University, Shanghai, China · Nanyang Technological University, Singapore · Edith Cowan University, Perth, Australia +2