Explainable multi agent reinforcement learning framework for secure and adaptive communication in UAV swarm based fanets.

Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Sci Rep . 2026 Mar 3;16:11830. doi: 10.1038/s41598-026-39366-x Search in PMC Search in PubMed View in NLM Catalog Add to search Explainable multi agent reinforcement learning framework for secure and adaptive communication in UAV swarm based fanets Hend Khalid Alkahtani Hend Khalid Alkahtani 1 Department of Information Systems, College of Computer and Information Sciences, Princess Nourah bint Abdulrahman University, Riyadh, 11671 Saudi Arabia Find articles by Hend Khalid Alkahtani 1 , Ybytayeva Galiya Ybytayeva Galiya 2 Department International Educational Corporation, Kazakh Leading Academy of Architecture and Civil Engineering, Almaty, Kazakhstan Find articles by Ybytayeva Galiya 2 , Bekarystankyzy Akbayan Bekarystankyzy Akbayan 3 School of Digital Engineering, Narxoz University, Almaty, Kazakhstan Find articles by Bekarystankyzy Akbayan 3 , Ayman Qahmash Ayman Qahmash 4 Department of Informatics and Computer Systems, King Khalid University, Abha, Saudi Arabia Find articles by Ayman Qahmash 4, ✉ Author information Article notes Copyright and License information 1 Department of Information Systems, College of Computer and Information Sciences, Princess Nourah bint Abdulrahman University, Riyadh, 11671 Saudi Arabia 2 Department International Educational Corporation, Kazakh Leading Academy of Architecture and Civil Engineering, Almaty, Kazakhstan 3 School of Digital Engineering, Narxoz University, Almaty, Kazakhstan 4 Department of Informatics and Computer Systems, King Khalid University, Abha, Saudi Arabia ✉ Corresponding author. Received 2025 Jul 26; Accepted 2026 Feb 4; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13065831  PMID: 41771935 Abstract The increasing use of the Unmanned Aerial Vehicle (UAV) swarms in real-time and mission-critical operations requires such communication infrastructure not only to meet security and adaptation demands, but also to be transparent, interpretable. This article gives an Explainable Multi-Agent Reinforcement Learning (EMARL) framework of an intelligent and safe Flying Ad Hoc Networks (FANETs) communication model. The offered system combines a decentralized learning system by Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm, a trust-based security system, and an Explainable AI (XAI) package of SHAP, LIME, and attention visualization techniques. The EMARL system allows every UAV agent to autonomously base their decision on the routing policy that is both interpretable and the result of a combination of local observations, learned policies, as well as trust estimates of the adjacent agents. Network modeling and mobility dynamics of UAV is simulated through NS-3, UAV mobility dynamics through AirSim, and a Python-based MARL engine in order to train policies and make decisions that are coordinated. The evaluation of the performance reveals that EMARL has always seen improved packet delivery ratio (PDR), improved accuracy, reduced delay, improved energy efficiency and false positive rate over the traditional protocols like, Ad hoc On-Demand Distance Vector Routing (AODV), Trust based, Q-Routing and Standard, MARL even under jamming and Sybil attack conditions. Exploitability-based measurements also validate the framework as an entity that provides clarity and accountability of decisions, thus enhancing human interpretability and credence. At the ablation studies, the presence of the XAI and trust modules is deemed to be essential to ensure the robustness of the system. Comprehensively, the EMARL framework is the vital step on the path aiming at safe, interpretable, and scalable UAV swarm communications in dynamic and hostile scenes. Keywords: Explainable AI, Multi-Agent reinforcement learning, UAV swarm communication, FANET security, Trust-Based routing Subject terms: Engineering, Mathematics and computing Introduction As it happened in the past couple of years, FANETs have already become one of the most important extensions of something we already know and love closes to Mobile Ad Hoc Networks (MANETs), only specific to UAVs communication 1 . As organizations continue to expand the use of UAVs in non-military use cases, including civilian and commercial applications, emergencies response, like disaster management, accuracy farming, environmental surveillance, and infrastructure examination, FANETs offer a scalable and dynamic network 2 . Drones’ swarms are one of the most advanced deployment paradigms in FANETs since they can be operating in a distributed, coordinated, and autonomous way 3 . One of the factors these swarms are based on is real-time communication, collaborative decision-making, and high rates of adaptability, which is perfect in time-sensitive and large-scale procedures 4 . Nevertheless, the agility and flexibility, which define UAV swarms in the first place, provide an alternative major vulnerability in terms of network security, coordination reliability, and communication resilience. The mobility characteristics, its decentralized architecture, resource limitations, and an open-air medium of communication of FANETs make it vulnerable to various kinds of cybersecurity threats, such as jamming, spoofing, eavesdropping, Sybil attacks, and denial-of-service (DoS) attacks 5 . Traditional centralized security mechanisms are ineffective in such dynamic and distributed environments due to high latency, single points of failure, and the lack of real-time adaptability 6 . Moreover, most existing defense systems operate as black-box models, offering little to no interpretability of their decision-making processes, thereby hindering trust, debuggability, and regulatory acceptance in mission-critical operations 7 . Despite the rapid advancements in machine learning and artificial intelligence for communication and security in ad hoc networks, there remains a significant gap in deploying interpretable, decentralized, and real-time secure communication frameworks for UAV swarms operating in FANETs 8 . Namely, most of the traditional network security protocols are insufficient due to the dynamic topology, a heterogeneous range of communication paths, limited computational resources, and lack of fixed infrastructure 9 . Reinforcement learning (RL), and, more precisely, Multi-Agent Reinforcement Learning (MARL) renders to be a potential possibility of modeling UAV swarm behavior in the form of a continuous learning and autonomous adjusting 10 . Nevertheless, the current MARL-based systems are not interpretable, and one cannot easily explain the reasons behind the decisions involved with routing, authentication, or a defense action, particularly in adversarial or unclear situations 11 . In addition, the aspect of real-time operation should not only require a system capable of adaptively responding to environmental pressures and threats, but, most importantly, offer transparent rationale behind its decisions to ease the trust and accountability in autonomous systems 12 . The absence of such a system contributes to the jeopardy of ruined missions, ineffective reaction to the hazard, as well as small-scale flexibility during the deployment of huge swarms. The combination of Explainable Artificial Intelligence (XAI) and MARL in FANETs brings about a revolutionary chance to overcome the challenges mentioned above 13 . The capability of MARL of distributing the intelligence with a group of autonomous agents and having them still operate collectively is invaluable to UAV swarm systems 14 . This enables UAVs to observe locally and have a distributed optimal communication route, intrusion detection, and network change adaptation without using centralized control 15 . The system can be auditable and transparent by inserting XAI methods in the decision-making chain - e.g., SHAP values, policy visualization, attention heatmaps, etc 16 . The experience of operator and mission planners is enriched with understanding behavior of the agent, verification of defensive strategies and anomaly detection 17 . Furthermore, explainable MARL systems fit the future of autonomous aerial systems where regulatory compliance and safety of the mission as well as human-AI collaboration are increasingly taking the center stage 18 . The possibility to explain the decisions, even ones as isolating a hacked UAV, data rerouting, or launching evasion maneuvers, results in increased trust in operations and expansion of AI-based FANETs to critical fields 19 . Also, the suggested framework could integrate with the notion of context-awareness so that the agents can learn in the setting of geographical, time-based, and mission-based factors to result in superior security and flexibility under practical limits 20 . The combination of explainability and multi-agent learning has not just technical application but also applied to fill in the gap between human supervision and autonomy between drone swarms. Since such systems, through their trends towards application in smart cities, intelligent transportation, and aerial IoT ecosystems, prove significant, their decision making-processes need to remain robust and explicable to ensure credibility in society and continuation in technology 14 . The study puts forward a proposal of explaining Multi-Agent Reinforcement Learning Framework to be deployed in secure and adaptive communication with UAV swarm-based FANETs. The important contributions of the work are: A decentralized MARL-based communication infrastructure that adapts in real time due to security attacks and environment adaptation. Adoption of XAI methods to be able to interpret agent decisions, making them audit-able and allowing one to intelligently debug. Creating a threat-sensitive reward structure helping to ensure both the efficiency of communications and the robustness of a system. Test and compare the suggested framework with various scenarios of attacks and mobility options with benchmark FANET networks and realistic UAV mobility models. The article provides an introduction, which gives the background, problems and incentive of securing UAV swarm-based FANETs with an Explainable MARL approach. Then, it provides the review of work related to it such as current approaches to FANET security, reinforcement learning approaches in the context of autonomous networks, and application to XAI in autonomous networks. The section with the proposed framework explains the architecture with agent communication strategies, threat modeling, and explainability techniques integration. Then, there is the experiment setup and evaluation, in which the simulation environments, data sets, and accuracies are specified. The results and discussion section focus on evaluating the Performance of the model in relation to adaptability, security, and interpretability. At last, the final part is the conclusion, which summarizes the contributions and proposes the future research on secure and transparent UAV swarm communication. Literature review The literature review will critically analyses the recent developments in MARL, trust-based communication, and XAI of the UAV swarm systems. It brings up emerging security, interpretability, and scalability issues that drive the proposed EMARL-XAI framework. A deep reinforcement learning-based secure routing protocol (DRL-SRP) in FANETs was suggested by Dahmane et al. (2024) to improve the resilience of communication against jamming attacks 21 . This was done by training UAV agents to use the Deep Q-Network (DQN) method where the routing of each UAV was selected depending on local observation of the state with regard to the current signal-to-noise and number of hops in the path. The simulations were carried out based on generated custom UAV mobility traces under the OMNeT + + platform 22 . Though the method shown to have better PDR and lower latency than traditional protocols, no explainability and the need to train centrally hindered the scalability of the method in real time swarmed applications. Hassan et al. proposed a Q-learning-based trust-aware routing scheme in FANETs 23 . The algorithm focused on the dynamic assessment of the trust of neighboring nodes grounded on the ability to transfer packets and energy consumption. The model designed by the authors was checked against the NS-3 simulator and the synthetic network topologies designed in the NS-3 model with different network densities 24 . The results indicated an increase in resiliency against insider attacks, as well as a sharp increase in throughput, but the system did not provide transparency in decision-making, and did not perform well when generalizing in high mobility environments characteristic of drone swarms. Tong et al. (2025) introduced a blockchain-based authentication and data integrity protocol for UAV swarms 25 . Their approach integrated Ethereum smart contracts with a cluster-based routing mechanism to ensure secure peer verification and transaction logging. Simulations used a synthetic dataset of UAV telemetry and communication logs. The method achieved high resistance to spoofing and Sybil attacks 26 ; however, the added computational overhead and latency due to blockchain consensus made it less suitable for time-critical swarm missions. Ceviz et al. (2025) presented a federated learning (FL)-based intrusion detection system for UAV networks 27 . Their framework trained local lightweight CNN models on each UAV to detect anomalies in real-time, followed by secure aggregation at the central node. The UNSW-NB15 dataset was used to simulate intrusion traffic. While the FL approach-maintained data privacy and achieved high detection accuracy (> 93%), it lacked interpretability and was susceptible to poisoning attacks if one or more nodes became compromised. UAV networks The Deep Deterministic Policy Gradient- (DDPG)-based multi-agent system was designed by Huang et al. (2025) to perform secure and adaptive routing in UAV networks 28 . A continuous-valued action policy was learned so as to maximize link reliability and energy consumption in adversary environment by each agent 29 . The evaluation of the model was performed with the use of UAVNetSim in the dynamic topology changes 30 . The findings demonstrated that there were considerable gains in network lifetime and robustness, however, the method was very time consuming to train and it provided very little in the way of insight into the decision process since it is a black-box optimization. Dubey et al. (2023) suggested a rule-based trust assessment mechanism using fuzzy logic combined with reinforcement learning to detect malicious UAV in a FANET 31 . Its methodology included fuzzy logic to calculate trust-related scores to utilize table-based Q-learning, which drove routing based upon trust paths. MATLAB was used to create 30 UAV nodes simulation scenarios 32 . The hybrid model increased reliability and detection accuracy, however its rule base design was fixed and needed to be manually tuned, constraining its flexibility in adapting to a new type of attack or other unexpected behaviors. Huang et al. (2025) explored the use of attention-based graph neural networks (GNNs) to model UAV swarm communication as a dynamic graph 28 . The model captured inter-UAV dependencies and leveraged attention mechanisms to prioritize critical communication links. Synthetic graph datasets based on flight path simulations were employed. The approach yielded superior link prediction accuracy and adaptive routing performance. However, the absence of explainability tools and high GPU memory demands posed significant deployment challenges. Liao et al. (2024) implemented a game-theoretic anti-jamming algorithm using stochastic learning automata for UAV swarm networks 29 . The UAVs iteratively learned optimal frequency hopping strategies based on adversarial behavior patterns. Evaluations were conducted on a custom simulator with synthetic jamming models. The strategy was better than the static frequency selection in jamming resilience, although it had slow convergence and unstable in multiple agents coordination in a rapidly fluctuating environment. Group A In their work, Zeng et al. (2025) developed an ensemble machine learning structure to identify intrusion in FANETs based on Random Forest (RBEs, XGBoost, and LSTM classes, with the LSTM class to be used in the final stage of the ensemble framework structure 33 . It was a model trained and tested using BoT-IoT dataset 34 . The composite attained a detection rate of 98.2% and cut false positive in very large percentage. However, the solution did not provide model interpretability and needed the centralized training that is detrimental to real-time use in a UAV swarm with distributed architectures. Huang et al. (2025) presented a secure protocol, which is also adaptive by using ant colony optimization (ACO)-reinforcement learning to search paths adaptively 35 . Every UAV was an agent that adapted the pheromone paths to the Q-values feedback. Python was used as custom simulation environments. Although the algorithm showed improved path reliability and reduced energy consumption compared to baseline ACO, it was still having issues in terms of scalability because of complexity of computation and lack of explainable routing behavior. A context-aware reinforcement learning protocol, also made by Zabihi et al. (2023), proposes to implement secureized UAV communication in FANETs 36 . The model capitalized on the environment to provide a Proximal Policy Optimization (PPO) that navigated routing choices by means of node density, delivery strength, and mobility 37 . The NS-3 network simulator was simulated where the flight paths of UAVs were randomized. The system not only had improved the stability of the routes but also had a reduced delay in the adversarial condition. Nevertheless, its inadequacy to be applied in explainable autonomous networks was due to the absence of interpretability layer and transparency of its policy learning. Yang et al. (2025) introduced an abnormal detection system based on Long Short-Term Memory (LSTM) autoencoders to track the UAV telemetry data to reveal abrupt behavior changes 38 . The temporal UAV datasets used to train the model had a mix of BoT-IoT and flight simulation data. The findings had high percentages of true positive as well as low computational overheads. The primary shortcoming was the inability to rely on other defense mechanisms other than detection, and the model lacked aspects of explainability that could provide a clear picture of anomaly classes to human operators. Another recent success is a hierarchical reinforcement learning framework applied by Wang et al. to coordinate Swarm UAVs against the attack scenario in a scalable setting 39 . The objectives of missions were set by the high-level agents and the Deep Q-Learning was used by the low-level agents to choose the best routes and avoid interference. The model was tested using the UAVSim tool with various jamming and eavesdropping profiles. The hierarchical setup improved modularity and response to threats, but coordination between layers suffered delays and lacked interpretability for mission planners. Zeng et al. (2025) introduced an adversarial training method to improve the robustness of neural network-based intrusion detection systems (IDS) in UAV networks 33 . A generative adversarial network (GAN) was used to simulate attack traffic and augment the training of a CNN classifier. The model was evaluated on the NSL-KDD dataset 40 . It achieved improved detection rates against novel threats but remained susceptible to interpretability challenges and required significant training resources. Wang et al. (2024) proposed a multi-agent deep reinforcement learning model for dynamic spectrum access in cognitive UAV networks 34 . The agents learned collaborative access policies using MADDPG, considering interference and primary user activity. Simulations used custom scenarios based on air traffic control spectrum sharing. The system improved spectrum utilization and fairness, but performance degraded under high UAV mobility, and agent coordination became unstable without centralized feedback. Huang et al. (2025) integrated a blockchain-based access control layer with UAV swarms to secure inter-agent communication and prevent unauthorized access 35 . Hyperledger Fabric was used for identity verification, while UAVs employed smart contracts for decentralized authentication. The model was tested using synthetic UAV interaction logs. Although the system enhanced authentication integrity, its heavy consensus requirements and lack of adaptability under real-time conditions hindered suitability for highly mobile FANETs. Zabihi et al. (2023) introduced a fuzzy reinforcement learning model for security-aware path planning in UAV networks 36 . Fuzzy logic was used to interpret environmental inputs (e.g., threat level, energy reserve), and Q-learning guided path selection accordingly. MATLAB-based simulations demonstrated improved survivability and mission completion rates. However, the hand-tuned fuzzy rule sets reduced generalizability, and the absence of XAI techniques made it difficult to analyze agent decision rationale. Li et al. (2021) proposed a reinforcement learning-based anti-jamming strategy where each UAV adapted transmission power and frequency based on Q-table updates 37 . Using a custom-built simulator for FANETs under frequency-hopping attacks, the method showed improved signal integrity and packet delivery. Despite its low complexity, the model struggled in environments with rapidly varying interference patterns and lacked explainable decision auditing. Yang et al. (2025) presented a hybrid Gated Recurrent Unit (GRU) and convolutional model for anomaly detection in UAV swarm telemetry data 38 . The system was trained on both temporal and spatial features using simulated mission logs. It outperformed traditional IDS models in accuracy and precision but lacked defense-triggering capabilities and was deployed as a passive monitoring tool without an explainable output layer for understanding predictions. Roy et al. (2020) explored interpretable AI techniques in UAV swarm control by embedding decision trees into deep reinforcement learning networks 39 . The Tree-CNN hybrid allowed post-hoc analysis of agent actions using rule extraction. Using AirSim for swarm simulation, the model achieved moderate interpretability while maintaining acceptable performance. Nevertheless, it led to the lost interpretability at the price of lower adaptability to complex dynamics in the environment and slower convergence in multi-agent coordination tasks. To recapitulate, although different studies have been conducted about MARL, trust mechanisms, and XAI, the possibility to combine them into one comprehensive framework that will encompass all three elements (security, interpretability, and real-time performance) in UAV swarms has not been fully achieved. This disparity provides the encouragement behind the EMARL-XAI model, which is the fusion of these elements to ensure greater reliability and decision articulation. This study proposes a Digital Twin–enabled UAV multiple-target tracking framework that optimizes local and global swarm topology using a Time–Spatial Graph Learning algorithm 41 . The method significantly reduces tracking latency and improves target-tracking success rates. However, it focuses on topology control and cooperative tracking rather than secure communication. The work does not address trust modeling, adversarial threats, or explainable decision-making in FANETs. This work presents a Versatile Digital Twin (VerDT) framework for UAV-based intelligent logistics, integrating DT-driven resource scheduling and path planning for low-latency and energy-efficient distribution 42 . The system enhances logistics performance but assumes reliable communication and trusted nodes. It does not incorporate secure routing, MARL-based coordination, or XAI for decision transparency in adversarial FANET environments. Recent work on secure optimization for cell-free MIMO STIN systems improves spectral efficiency and physical-layer secrecy using distributed optimization techniques 43 . These methods rely on infrastructure-assisted processing and centralized coordination. In contrast, our work focuses on decentralized UAV swarm communication security, trust modeling, and explainability—capabilities not addressed in cell-free MIMO STIN optimization. Studies on UAV-assisted dual-hop V2V MIMO-NOMA systems analyze BLER performance under finite blocklength constraints to ensure reliable ultra-low-latency communication 44 . These methods focus on PHY-layer reliability for vehicular networks but do not address multi-agent coordination, malicious node detection, or explainable secure routing as required in FANET environments. Power optimization and ARS selection strategies in cell-free multi-ARS communication systems enhance downlink coverage and energy efficiency using centralized resource allocation 45 . While effective in static infrastructure networks, they do not support decentralized trust-aware decision-making or attack resilience required in UAV swarms. Work on NOMA multi-user networks equipped with aerial RIS and aerial relay nodes demonstrates improved signal coverage and multi-user performance 46 . However, these methods primarily target physical-layer performance enhancement and lack mechanisms for MARL-based security, XAI interpretability, or collaborative threat mitigation in highly dynamic FANETs. While these state-of-the-art studies provide significant advances in UAV-enabled resource optimization, MIMO-NOMA performance, and secure cell-free architectures, they focus primarily on PHY/MAC-layer optimization or infrastructure-assisted communication. None of these works address decentralized trust-aware MARL, adversarial resilience, or explainable routing decisions within fully distributed UAV swarms. The proposed EMARL-XAI framework uniquely integrates secure multi-agent learning, dynamic trust modeling, and XAI-based interpretability for robust FANET communication under adversarial conditions. Despite significant progress, prior MARL-based routing frameworks exhibit several critical limitations when applied to highly dynamic FANET environments. First, most existing approaches do not integrate decision interpretability, resulting in opaque routing behaviors that limit operator trust and post-event auditing. Second, trust mechanisms are typically rule-based or heuristic, making them inflexible and prone to misclassification under mobility-induced uncertainty. Third, many models still rely on partially centralized training structures, which reduces scalability and leaves the system vulnerable to single-point failures. Additionally, prior MARL solutions often degrade under adversarial mobility patterns and struggle to maintain performance in large-scale swarms due to communication overhead and unstable convergence. These gaps directly motivate the design choices of the proposed EMARL-XAI framework. Methods In order to succeed in solving the problems of secure and adaptive communication in the UAV swarm-based FANETs, this paper introduces a new EMARL scheme. Methodology aims to combine decentralized learning and interpretability to improve the intelligence of autonomous UAV agents in their operation and increase its trustworthiness. Every UAV that constitutes the swarm is an autonomous learning agent able to make routing, coord and threat response decisions real-time on the basis of local observation and swarm knowledge. This framework includes many important aspects, such as a thoroughly designed system architecture, a multi-agent reinforcement learning (MARL) approach to FANETs in dynamic scenarios, and an explainability layer, which includes suggestions on how to behave in such scenarios by applying and employing interpretable AI approaches and concepts. Also, a strong threat model has been specified to create an adversarial environment that resembles reality and a simulation environment that emulates effects of various attacks and the behavior of networks. The methodology also describes the communication protocols, learning architecture as well as the performance parameters used in assessing the effectiveness of the system in providing secure, resilient and transparent communication of UAV swarms. System architecture overview The given framework is based on the decentralized multi-agent structure designed to operate the use of the UAV swarms in Flying Ad-Hoc Networks (FANETs), assuming that each UAV is an autonomous agent, and it is able to sense its surrounding environment, make a smart choice, and securely coordinate in it with the adjacent nodes. The communication model contained in the swarm takes advantage of the dynamic and mobile characteristics of FANETs and allows UAVs to create an adaptive mesh network where topology changes regularly because of high mobility, environmental barriers, or opposing threats. All UAVs form communication connection with the other group members within its transmission range and they use multi-hop relay schemes to provide an end-to-end connection in the network. The single agent interaction model assumes a decentralized reinforcement learning framework, with each UAV (agent) observing locally available state information, i.e., relative position, signal quality, trust scores and observed anomalies and acting, i.e. route updates, threat alerts or cluster reconfiguration. UAV coordination is performed through a series of lightweight state messages and policy updates. The new scalability, fault tolerance and responsiveness in highly-varying environments also improve using this agent-based approach. In order to equate the trust and transparency in the behavior of the agent, the architecture provides a layer of Explainable Artificial Intelligence (XAI). In this module, human-interpretable explanations of the actions of all the agents are produced based on such methods as SHAP (Shapley Additive explanations), LIME (Local Interpretable Model-Agnostic Explanations), and attention visualization. The security plane, which is embedded at the communication and decision-making plane, is in charge of detecting malicious activity, isolation of compromised nodes and dynamically re-directing data through secure channels based on learned policies. The resulting architecture will make the communication, besides being secure and adaptive, capable of interpretation by the human operators, which will give confidence in automated decision-making procedures. Table 1 presents the key constructs of the suggested EMARL-XAI framework, including the functions of separate UAV agents, the learning engines, and security levels. All modules have to do with making sure that autonomous decision-making, secure communication, and explainable behavior in the UAV swarm network are achieved. Figure 1 provides a high-level diagram of the proposed EMARL framework in FANETs that consists of the communication and observation layers, UAV agents, the MARL engine, and the combined explainability and security modules. The modular architecture enables adaptive learning, real time decisions and interpretable outputs that results in improved trust and resilience. Table 1. Key components of the proposed system Architecture. Component Description UAV Agent Autonomous drone with sensing, computation, and communication capabilities State Observation Unit Captures input data such as signal strength, location, and trust metrics MARL Engine Learns optimal actions using decentralized RL algorithms Communication Layer Facilitates inter-agent message passing and routing decisions XAI Module Applies SHAP, LIME, and visualization to explain decisions Security Module Detects threats, updates trust, and triggers adaptive defense strategies Open in a new tab Fig. 1. Open in a new tab System Architecture of the Proposed Explainable MARL-Based Framework for FANETs. Mathematical formulation of the EMARL system model To enhance reproducibility and provide a precise system representation, the proposed EMARL framework is formally modeled using a Decentralized Partially Observable Markov Decision Process (Dec-POMDP) defined as: Where: is the number of UAV agents in the swarm. is the global environment state space. is the joint action space. denotes local observations for each UAV. is the transition function. represents the rewards functions. is the discount factor. 1. Global and Local State Representation. Each UAV has a local observable state: Where: : UAV’s position and velocity. : measured signal-to-noise ratio. : residual energy. : trust score for neighbor . : recent attack/alert history. : communication metrics (latency, hop count). The global state is: However, due to partial observability, each agent only uses for decision-making. 2. Action Space. Each UAV executes an action: Formally: Where represents neighbors within communication range. 3. Reward Function. The reward for UAV is designed as: Where: are weighting coefficients. : successful packet delivery. end-to-end latency. false positive rate. trust reinforcement for selecting reliable neighbors. 4. Trust Computing Model. Trust between UAV and neighbor is updated based on: Where: forwarding reliability. communication consistency. dropped packet ratio. trust update rate. feature weight. A UAV is flagged malicious if: 5. Policy representation (MADDPG). Each agent has an actor network: and a centralized critic: Actor update: Critic update via Bellman equation: 6. Communication model. UAV-to-UAV link exists if: where is Euclidean distance and is transmission range. Communication success probability: This models jamming and channel noise. Threat model and assumptions To analyze the security and resilience of the proposed EMARL framework in UAV-based FANETs to an intensive threat it is necessary to establish a hybrid threat model in a situation that incorporates diversity of adversarial actions, environmental restrictions, and working assumptions. FANETs are inherently vulnerable due to their dynamic topologies, decentralized structure, open wireless medium, and resource-constrained nodes. Therefore, the system is modeled to simulate realistic threat conditions that may arise during mission-critical swarm operations. The threat model includes various attack types, such as: Jamming attacks, where adversaries emit high-power signals to interfere with legitimate UAV communication. Spoofing attacks, wherein a malicious node impersonates a legitimate UAV to intercept or redirect communication. Sybil attacks, where a single compromised node generates multiple fake identities to manipulate routing decisions or flood the network. Blackhole attacks, where malicious UAVs drop all incoming packets instead of forwarding them. Replay attacks, involving retransmission of previously captured valid packets to confuse the swarm or reroute data. To simulate realistic scenarios, several system assumptions are made: Each UAV is equipped with sufficient computational power to execute lightweight reinforcement learning algorithms and communicate with neighboring agents. UAVs operate within a limited and predefined communication range, with adaptive connectivity depending on altitude, motion speed, and environmental factors. All UAVs are GPS-enabled and can localize their positions relative to the swarm. A subset of UAVs may behave maliciously, but the majority are assumed to be trustworthy and cooperative. The trust and adversarial behavior modeling component plays a critical role in the learning process. Each agent maintains a dynamic trust score for neighboring UAVs based on behavioral metrics such as packet forwarding success rate, consistency in message exchange, and communication reliability. These scores influence decision-making within the MARL framework and help agents adaptively route information away from suspected malicious entities. Trust updates are decentralized and influenced by both direct and indirect observations, improving robustness against stealthy or coordinated attacks. Table 2 outlines the threat landscape and underlying assumptions considered in the FANET security model. It details the types of adversarial attacks addressed, UAV and network capabilities assumed, and the trust-based mechanisms used for mitigating malicious behavior. Figure 2 presents the threat modeling and trust evaluation process within the EMARL framework, showing how UAV agents observe environmental states, compute trust under adversarial conditions (e.g., jamming, spoofing, and Sybil attacks), and make decentralized decisions using the MARL engine. This pipeline enables adaptive response and secure coordination against dynamic security threats. Table 2. Summary of threats and assumptions in the FANET security Model. Aspect Description Attack Types Considered Jamming, Spoofing, Sybil, Blackhole, Replay UAV Capabilities Assumed GPS-enabled, moderate processing power, secure local storage Communication Assumptions Limited range, high mobility, intermittent connectivity Trust Evaluation Based on forwarding reliability, interaction consistency, and message logs Malicious Behavior Handling Isolation, rerouting, and trust-based policy adjustment Open in a new tab Fig. 2. Open in a new tab Threat modeling and trust evaluation flow within the EMARL framework, illustrating UAV interactions, adversarial threats, and decentralized decision-making. Multi-agent reinforcement learning framework The proposed framework uses Multi-Agent Reinforcement Learning (MARL) since it could facilitate safe and adaptive swarm coordination in dynamic FANET. Every swarming UAV can be represented as an autonomous agent which develops the best communication and coordination strategy due to interaction with the environment and the surrounding agents. The learning process is facilitated by a reinforcement signal that helps reward cooperative behavior, secure routing and live time threat avoidance. The state space, the action space and the reward function characterize each UAV agent: The state space contains observations which are the position of local nodes, their velocities, available energy, signal-noise ratio (SNR), trust scores of neighbors, current statistics of communications, and previous detection of anomalies or attacks. The action space includes routing choice (e.g., selection of next-hop), trust management (e.g. penalize or promote a node), evasive actions (e.g. re-clustering or detour path), and dissemination of information (e.g. broadcasting security alerts). The reward model is well constructed to support various goals, namely, achieving maximum levels of communication reliability, minimal exposure to malicious nodes, sparing energy, and the connectivity of the network. Rewards are provided to positive data transmission, safe destination choice and trustful evasion of adversaries and penalty when packets are lost, compromised node destination chosen or loop communication. We formally define the total reward received by UAV agent at time as: Where: : reward for successful packet delivery. : bonus for selecting high-trust neighbors. : penalty for excessive end-to-end delay. : penalty for high transmission or mobility energy. : penalty for interacting with low-trust or malicious nodes. The weight coefficients { } regulate the importance of each behavioral aspect during training. The reward function integrates five components that guide the MARL agents toward secure and efficient routing behavior. Successful packet delivery and selection of high-trust neighbors provide positive rewards, while delay, excessive energy consumption, and interactions with low-trust or malicious nodes result in proportional penalties. This balanced reward structure ensures that agents simultaneously optimize communication reliability, security, latency, and energy efficiency during policy learning. In the case of the learning algorithm, the framework uses MADDPG because it shares a few advantages with the learning algorithm such as touch action space and collaborative training in decentralized environments. By means of MADDPG, each UAV runs an actor-critic structure with personalized policy optimization yet makes use of international information in training. Deep Q-Network (DQN) and Proximal Policy Optimization (PPO) algorithms, too, are put in place to conduct the end-to-end comparison actions, tactics, along with policies, with regard to stability, convergence speed, and policy robustness. The learning architecture facilitates entire decentralized and some kind of centralized training. Since it works in the decentralized configuration, every agent adjusts its policy only on the basis of localistic information and experience, which facilitates scalability and self-sustainability of policy adaptation in high-mobility scenarios. In the partially centralized training mode, the agents exchange some personal experiences, or trust values with some central learner, during the training period, which leads to a convergence after a shorter period and better coordination on a global basis. Yet, the deployment is completely decentralized and ensures the retention of time-critical flexibility and the adaptability to the disruptions of communications. Table 3. Summary of MARL Agent Design and Learning Parameters. Component Description State Space Position, energy, SNR, trust scores, attack history, path quality Action Space Next-hop selection, trust updates, path re-routing, alert broadcasting Reward Function +1 for secure delivery, –1 for selecting malicious paths, –0.5 for unnecessary hops Learning Algorithm MADDPG (primary), with DQN and PPO used for comparison Training Mode Decentralized execution; optionally centralized training for accelerated convergence Policy Architecture Actor-Critic models with neural networks trained per agent Open in a new tab This control system is based on MARL and allows UAVs to acquire a good coordination strategy when they face threat, adjust to the changing topology in FANETs, and guarantees them to exchange information securely besides having explainable decision paths in the incorporated XAI modules. Presents the detailed design of MARL agents within the proposed framework, including their input-output representations, reward shaping strategy, and learning configurations. It also highlights the use of MADDPG as the primary learning algorithm alongside comparative models, supporting scalable and adaptive decision-making in the UAV swarm. Figure 3 illustrates the end-to-end workflow of the proposed MARL-based framework, where multiple UAV agents interact in a decentralized manner, share experience, and optimize their policies via actor-critic learning to enhance coordination, adaptability, and cumulative performance. Fig. 3. Open in a new tab Workflow of the proposed MARL-based framework showing decentralized UAV agent interaction, shared experience, and policy optimization through actor-critic learning. Algorithm 1. Open in a new tab EMARL-XAI Training and Execution using MADDPG in EMARL Framework. Explainability integration (XAI Module) To enhance transparency, trust, and operational oversight within the autonomous UAV swarm communication system, the proposed framework integrates a dedicated Explainable Artificial Intelligence (XAI) module. This module enables human operators and system administrators to interpret the decision-making processes of individual UAV agents by generating comprehensible, instance-level and global explanations for the policies learned through Multi-Agent Reinforcement Learning (MARL). The XAI module employs a hybrid set of explainability techniques: SHAP (SHapley Additive exPlanations): Utilized to assign feature importance scores that quantify the contribution of each input variable (e.g., trust score, SNR, hop count) to an agent’s decision. This technique offers consistency and local accuracy, making it well-suited for critical decision tracing. LIME (Local Interpretable Model-Agnostic Explanations): Applied to create simplified surrogate models around specific decisions, allowing users to understand policy behavior under different environmental conditions or adversarial settings. Policy Visualization: Used to generate visual trajectories of action selection over time, helping mission planners analyze how decisions evolve in response to dynamic threats and mobility patterns. Attention Maps: Integrated into deep reinforcement learning layers (where applicable) to highlight which parts of the input data the model focused on while making a particular routing or defense decision. Each UAV agent is equipped with a local XAI interface that captures its internal policy decisions and outputs interpretable justifications either in real-time or during post-mission audits. These explanations are tied to contextual information such as environment state, trust updates, and reward feedback, forming an explainable decision trace for each autonomous agent. This is particularly critical in adversarial scenarios where verifying the correctness and rationale of decisions (e.g., isolating a peer node or avoiding a communication link) is essential for mission reliability and post-event validation. To assess the effectiveness of the XAI module, several interpretability metrics are adopted, including: Fidelity: Measures how well the explanation model approximates the true model’s behavior. Local Accuracy: Evaluates how precisely the explanation reflects the agent’s policy for a specific instance. Human Trust Score: Derived through expert evaluations, quantifying the perceived reliability and usefulness of the generated explanations. Explanation Time: Measures the computational cost and latency of generating explanations during inference. This explainability layer ensures that the learning and decision processes are not only functional but also interpretable, accountable, and auditable—thus aligning the system with modern standards for trustworthy AI in autonomous systems. Table 4 provides an overview of the explainability techniques integrated into the EMARL framework, detailing their objectives, output formats, evaluation metrics, and application contexts. These techniques—SHAP, LIME, policy visualization, and attention maps—collectively enhance the interpretability of agent behavior and support human trust and system transparency. Figure 4 demonstrates the integration of explainability mechanisms—SHAP, LIME-based local approximation, attention visualization, and policy tracing—into the MARL framework, enabling on-demand interpretation of UAV agent decisions based on input features such as trust and SNR. Table 4. Overview of explainability techniques and evaluation Metrics. Technique Purpose Output Type Suitability Metric Description Usage Context SHAP Feature contribution analysis Feature importance plots High accuracy, model-specific Fidelity Accuracy of explanation model vs. actual policy Global model interpretability LIME Local surrogate modeling Rule-based explanation Fast, model-agnostic Local Accuracy Precision of explanation for single decisions Instance-level trust Policy Visualization Action trajectory insight Time-series plots Interpret policy dynamics Human Trust Score Expert-rated interpretability and usability Field operator confidence Attention Maps Visual focus of the model on input data Heatmaps Deep model interpretability Explanation Time Time cost of generating explanations Real-time feasibility Open in a new tab Fig. 4. Open in a new tab Flowchart showing how SHAP, LIME, attention visualization, and policy tracing are integrated into the MARL pipeline to generate interpretable decision outputs from UAV swarm agents. Communication and coordination protocol The proposed EMARL framework will include the adaptive, effective, and robust communication and coordination protocol appropriate in dynamic FANET environments to permit effective, secure, and adaptive communication between the UAVs composing the UAV swarm. This protocol determines the information sharing, information routing, and network connectivity with UAV agents that must work in different operational and adversarial circumstances. An adaptive routing approach is at the center of the protocol, and according to this algorithm, each UAV chooses the best communication paths based on the observation of the located environment and policies learned. In comparison to a static or greedy routing methodology, this kind of strategy makes use of the current routing decisions within dynamic updates of parameters that include: trust scores, signal strength, remaining energy, amount of hops, and recent behavior history of close nodes. Particular UAV agents, trained through MARL, consider many paths and choose the one which best maximizes multi-objective utility function answering security, latency and reliability. The structure also incorporates a light inter-agent message services to help in enabling decentralized collaboration. UAVs broadcast periodical beacon messages with local state information and trust information as well as alert flags when suspicious activities are detected. In order to minimize on communications overhead only essential and prioritized data is conveyed and message compression is implemented where perfectible. Such collaboration is necessary in order that the agents can sense organized attacks that may be Sybil or blackhole attack and redirect data streams respectively. In the event of link failures or compromised nodes, the system executes a series of fault-tolerant procedures. When a link drops or an agent becomes unresponsive, neighboring UAVs immediately re-evaluate the local network topology and re-initiate the routing selection process based on updated trust and connectivity metrics. If a UAV is detected to be malicious—either through consistent deviation from expected behavior or via consensus among multiple neighboring agents—it is quarantined from the swarm by flagging its trust score below a defined threshold. Subsequent communication is re-routed to bypass the compromised node. The protocol integrates reinforcement-learned behaviors with secure communication primitives to support real-time adaptability, resilience against adversarial disruptions, and decentralized coordination across the UAV swarm. Table 5 outlines the essential features of the proposed communication and coordination protocol, emphasizing adaptive routing, secure inter-agent interactions, and failure resilience. The protocol ensures efficient, trustworthy message dissemination while maintaining robustness against dynamic link conditions and malicious behaviors within the UAV swarm. Figure 5 illustrates the secure communication process within the UAV swarm, incorporating trust evaluation, detection of malicious nodes, and adaptive re-routing based on learned policies to ensure robust and reliable routing decisions. Table 5. Key features of the communication and coordination Protocol. Feature Description Routing Strategy Reinforcement-learned path selection based on trust, latency, and energy Inter-Agent Message Passing Periodic beacons for state sharing, alerts, and trust updates Security Mechanism Trust-score-based isolation of suspicious nodes Failure Recovery On-the-fly re-routing in case of broken links or UAV dropouts Communication Optimization Selective broadcast, compression, and minimal overhead transmission Open in a new tab Fig. 5. Open in a new tab Secure UAV swarm communication with adaptive routing and malicious node handling. Simulation environment and parameter configuration To evaluate the effectiveness, scalability, and robustness of the proposed EMARL framework, a comprehensive simulation environment is developed that integrates network-level communication modeling with intelligent agent-based learning and explainability modules. The simulation setup enables detailed analysis of UAV swarm behavior under various dynamic and adversarial conditions. The simulation framework is constructed using a hybrid environment combining: NS-3: for modeling low-level wireless communication protocols, packet transmission, and network delays; AirSim: for simulating UAV mobility dynamics, trajectory planning, and environmental interactions; Custom Python-based MARL Simulator: built using PyTorch and OpenAI Gym interfaces, this component handles agent training, policy execution, XAI module integration, and trust management logic. The network topology consists of 20 to 50 UAV agents randomly distributed in a 2D/3D bounded airspace, with variable altitude constraints. UAVs communicate over Wi-Fi (IEEE 802.11p) links, with dynamic connectivity influenced by mobility and environmental noise. Mobility models include Random Waypoint and Gauss-Markov mobility to simulate both exploratory and mission-specific patterns. To assess security resilience, several attack scenarios are introduced, including jamming, spoofing, Sybil, and blackhole attacks, where a fraction of UAVs are designated as compromised agents exhibiting malicious behavior patterns. The training configuration for the MARL agents is optimized using the MADDPG algorithm. Agents are trained over 5,000 episodes with each episode consisting of 200 steps. The learning rate is set to 0.001 for the actor network and 0.0005 for the critic network. A discount factor (γ) of 0.95 is used, and soft target update parameter (τ) is fixed at 0.01. Experience replay buffers are employed with a size of 10^5 transitions. Exploration noise follows an Ornstein-Uhlenbeck process to encourage diverse policy exploration during training. For the XAI components, SHAP and LIME models are configured with 100 background samples and a perturbation radius of 5%, ensuring explanation fidelity while maintaining computational efficiency. Attention maps and policy visualization modules are updated every 10 steps during execution to provide near real-time interpretability. Table 6 presents the simulation configuration and parameter settings used to evaluate the proposed EMARL-XAI framework. It details the integration of NS-3 and AirSim for network and mobility simulation, the setup of the MARL engine and learning parameters, as well as the types of attacks and explainability techniques applied during training and evaluation. Figure 6 illustrates the integrated simulation environment used for evaluating the UAV swarm, showcasing the interaction between AirSim for UAV dynamics, NS-3 for network communication, a custom MARL engine for learning, attack injection mechanisms, and real-time explainability modules such as SHAP and LIME. Table 6. Simulation configuration and Parameters. Component Configuration Details Network Simulator NS-3 (Wi-Fi 802.11p, UDP routing, dynamic topology) Mobility Simulator AirSim (Random Waypoint, Gauss-Markov UAV movement) MARL Engine Python-based MADDPG (PyTorch, OpenAI Gym interface) # of UAV Agents 20–50 Attack Types Simulated Jamming, Spoofing, Sybil, Blackhole Learning Rate (Actor/Critic) 0.001/0.0005 Discount Factor (γ) 0.95 Target Update (τ) 0.01 Replay Buffer Size 100,000 transitions Exploration Noise Ornstein-Uhlenbeck process XAI Techniques SHAP, LIME, Attention Maps, Policy Visualization Explanation Frequency Every 10 steps during execution Open in a new tab Fig. 6. Open in a new tab Simulation setup combining AirSim, NS-3, MARL engine, and XAI modules for UAV swarm evaluation. To accurately model FANET behavior in smart-city mission scenarios, we explicitly incorporate UAV mobility dynamics using a hybrid mobility model. The primary mobility pattern is the Gauss–Markov model, chosen for its ability to generate smooth, time-correlated velocity and heading changes that closely resemble real UAV flight characteristics in urban environments. This model reduces abrupt transitions, making it appropriate for low-altitude urban operations, pursuit-tracking, and distributed surveillance missions. For comparison, a Random Waypoint (RWP) dataset is also included to evaluate performance under more stochastic mobility patterns typically encountered in wide-area exploratory missions. UAV velocity is varied across 5 m/s, 10 m/s, 15 m/s, and 20 m/s, reflecting realistic hover, cruise, and rapid-response modes seen in FANET applications. The velocity parameter directly influences link duration, neighborhood stability, and the frequency of routing updates—making it a critical factor for EMARL policy robustness. A dedicated performance evaluation is conducted to analyze the effects of UAV velocity on Packet Delivery Ratio (PDR), end-to-end delay, accuracy, and convergence stability. Performance metrics To comprehensively evaluate the effectiveness, robustness, and transparency of the proposed EMARL framework for UAV swarm-based FANETs, a broad set of performance metrics is employed. These metrics are categorized into three key domains: communication and learning performance, interpretability assessment, and system scalability and energy efficiency. For the communication and learning performance, standard network-level and algorithmic metrics are used: Accuracy: Measures the overall success rate of secure communication path selection and malicious node avoidance. PDR: Calculates the proportion of successfully delivered packets relative to those sent, serving as a primary indicator of communication reliability under dynamic and adversarial conditions. End-to-End Delay: Evaluates the time taken for packets to traverse from source to destination UAVs, considering both routing efficiency and network congestion. False Positive Rate (FPR): Quantifies the rate at which benign UAVs are incorrectly identified as malicious, which directly impacts trust computation and routing behavior. Convergence Time: Denotes the number of training episodes required for MARL agents to reach policy stability across the swarm, reflecting the learning system’s efficiency. In the domain of explainability and interpretability, several specialized metrics are employed to assess the performance of the integrated XAI components: Fidelity: Measures how closely the explanation model (e.g., SHAP or LIME) mimics the original MARL agent’s decision boundaries. Feature Importance Score: Derived from SHAP values, indicating the relative contribution of each input variable (e.g., trust score, SNR) to the agent’s action selection. Human Trustworthiness Score: Collected through user studies and expert evaluations, reflecting how easily system operators can understand and trust the AI-generated explanations. Lastly, the system-level performance is analyzed in terms of: Energy Efficiency: Measured as the average energy consumption per successful data transmission or per episode, capturing the overhead introduced by communication, computation, and coordination. Scalability: Evaluated by testing the framework across varying swarm sizes (e.g., from 10 to 50 UAVs), examining the impact on learning speed, PDR, and delay. These multi-faceted metrics ensure a holistic understanding of the framework’s operational quality, learning behavior, interpretability, and deployment viability. Table 7 summarizes the performance metrics employed to evaluate the EMARL-XAI framework across three key dimensions: communication and learning efficiency, explainability and trustworthiness, and overall system performance. These metrics provide a comprehensive assessment of the framework’s robustness, interpretability, and scalability in dynamic and adversarial FANET environments. Figure 7 shows the classification of the most important performance measures that are used to evaluate the EMARL-based UAV swarm framework, such as communication and learning metrics, interpretability measures, and system-level performance measures, e.g., scalability and energy efficiency. Table 7. Summary of performance Metrics. Category Metric Description Communication & Learning Accuracy Success rate of routing and malicious node avoidance PDR Ratio of successfully delivered packets End-to-End Delay Time taken for data to reach the destination False Positive Rate Incorrect classification of benign UAVs as malicious Convergence Time Episodes needed for policy stabilization Explainability & Trust Fidelity Similarity between explanation and actual model behavior Feature Importance Score Contribution level of input features (e.g., trust, SNR) Human Trust Score Subjective rating of explanation clarity and usefulness System-Level Energy Efficiency Average energy consumed per delivery or episode Scalability Model performance across increasing UAV swarm sizes Open in a new tab Fig. 7. Open in a new tab Key performance metric categories used to evaluate the proposed EMARL-based UAV swarm framework. Experimental results In order to demonstrate the correctness, strength, and interpretability of the introduced EMARL framework, a chain of comprehensive simulations and empirical testing took place. This segment illustrates the findings of the conducted experiment in a variety of operation and adversarial settings, which proved that the framework is capable of supporting high communication performance, flexibility to changing network topologies, and offering human intelligible explanations of decisions taken by autonomous agents. Such performance measures as PDR, end-to-end delay, false positive rate, convergence time, interpretability indicators, and energy effectiveness were used to measure the system behavior in a holistic manner. They are compared with conventional routing protocols and non-explainable MARL baselines to emphasize the benefits of a combination of the trust-based communication, explainability, and reinforcement learning into a single workflow based on a large-scale and secure framework. Evaluation setup A comprehensive simulation was carried out to determine the performance of the proposed EMARL framework system-wide under varying network configurations and attack scenarios as well as the location of the attack and agent behaviors. The evaluation infrastructure intended to explore communication effectiveness, flexibility and comprehensibility needed to the structure, in benign and adversarial environments, of the swarm UAV deployment realistic context. Testing was carried on high-performance computing system with Intel Core i9-12900 K CPU, 64 GB RAM, and NVIDIA RTX 3090 GPU. The simulation environment combined three main steps (i) NS-3 to model the dynamics of network level communication and wireless interference; (ii) Air Sim to simulate UAV mobility and in the real-world flight dynamics; and (iii) custom Python based MARL engine implemented using Porch and OpenAI Gym API to train and execute reinforcement learning policies. Explainability modules including SHAP, LIME, and policy visualization were implemented using the SHAP and ELI5 libraries. Test scenarios included both static and mobile UAV deployments, varying from 20 to 50 agents, configured to perform collaborative surveillance missions over a 1 km² area. UAVs followed a Random Waypoint or Gauss-Markov mobility model, with an average speed of 10–15 m/s and communication ranges between 150 and 200 m. The attack scenarios incorporated in the tests included jamming, Sybil, spoofing, and blackhole attacks. In each trial, 10% to 30% of the UAVs were configured as compromised nodes injecting adversarial behavior. To benchmark the proposed EMARL framework, several baseline models were implemented for comparative analysis: AODV: A classical reactive routing protocol widely used in mobile ad hoc networks. Q-Routing with Trust Evaluation: A lightweight reinforcement learning-based protocol with integrated trust scores. Standard MARL (without XAI and trust): The same MADDPG-based architecture as EMARL, excluding explainability and trust computation layers. All experiments were repeated across 20 independent runs, and the results are reported with standard deviation and 95% confidence intervals to ensure statistical reliability. The results are averaged and visualized in the subsequent sections. Table 8 outlines the experimental evaluation setup, detailing the hardware specifications, simulation tools, UAV swarm parameters, and baseline models used for comparison. It also specifies the adversarial scenarios and the number of trials conducted to ensure consistent and statistically reliable performance assessment of the EMARL-XAI framework. Figure 8 illustrates the complete workflow of the EMARL evaluation system, highlighting the integration of NS-3 for network simulation, AirSim for UAV dynamics, a custom MARL engine for policy learning, and interpretability modules like SHAP and LIME for explanation generation. Table 8. Evaluation setup Parameters. Parameter Value/Description Hardware Intel i9-12900 K, 64 GB RAM, NVIDIA RTX 3090 GPU Network Simulator NS-3 (802.11p protocol, UDP routing) Mobility Simulator AirSim (Random Waypoint, Gauss-Markov) MARL Engine PyTorch-based MADDPG (OpenAI Gym interface) Explainability Tools SHAP (TreeExplainer), LIME, ELI5, Attention Visualization Swarm Size 20 to 50 UAVs Speed/Range 10–15 m/s; 150–200 m communication range Attack Models Jamming, Sybil, Spoofing, Blackhole (10–30% adversarial UAVs) Baseline Models for Comparison AODV, Q-Routing with Trust, Standard MARL (no XAI or trust) Number of Trials 20 runs per scenario Open in a new tab Fig. 8. Open in a new tab Workflow showing the integration of NS-3, AirSim, MARL engine, and XAI modules for EMARL evaluation. To ensure reproducibility, the adversarial attack models were instantiated using fixed and quantifiable parameters. Jamming attacks were simulated by injecting interference signals at 5–10 dBm above the noise floor over a 2–3 MHz bandwidth, targeting the control and data channels. Sybil attacks were instantiated by allowing each compromised UAV to generate 3–6 falsified identities broadcasting counterfeit beacon packets at regular intervals. Spoofing attacks were modeled by altering GPS coordinates with a displacement of 8–15 m, leading to incorrect position reporting. For blackhole attacks, malicious UAVs dropped 80–100% of the packets routed through them while advertising artificially high trust values. These parameterized settings reflect realistic adversarial behavior and enable consistent replication of the experimental scenarios. Communication and routing performance The communication and routing performance of the proposed EMARL framework was evaluated using key metrics such as PDR, accuracy, end-to-end delay, and false positive rate (FPR) under both benign and adversarial conditions. These metrics were assessed to demonstrate the framework’s effectiveness in maintaining reliable and secure communication within the UAV swarm in the presence of dynamic topology changes and targeted network attacks. In a non-adversarial setup, EMARL-based UAV swarm had average PDR of 96.3%, which was much higher than standard baseline protocols or AODV (84.1%) and a standard MARL (89.7%). The reason of this improvement is that in trust-aware routing based on decentralized reinforcement learning, reliable nodes and paths are preferred. Secure route selection also had its accuracy high average at 94.6% portraying the tendency of the system to avoid using compromised or unreliable UAVs. Incorporation of the dynamic trust assessment in the MARL policy provided worthwhile improvement in agent capacity to identify and avoid adversarial operation. As far as the end-to-end delay is concerned, the EMARL framework displayed less latency than baseline models with a mean delay of 61.2 ms even though computation of the real-time policy and trust are updating. This indicates the effectiveness of the system and its scalability in swarm mission. Regarding false positive rate (FPR)—defined as the rate at which benign UAVs were incorrectly classified as malicious—the EMARL framework recorded a low average of 4.3%, owing to the robustness of its multi-agent trust consensus and decision traceability via the XAI layer. Performance under adversarial conditions (10–30% of UAVs behaving maliciously) showed a modest decline across all metrics but remained superior to benchmark models, confirming the resilience of the proposed approach. Table 9 presents the communication and routing performance metrics under both normal and adversarial conditions, comparing the proposed EMARL framework with AODV and standard MARL baselines. The results demonstrate EMARL’s superior packet delivery ratio, accuracy, and reduced latency, while maintaining a significantly lower false positive rate even in hostile environments. Figure 9 presents a comparative analysis of EMARL (under both normal and adversarial conditions), AODV, and Standard MARL across key communication and routing performance metrics, including PDR, accuracy, end-to-end delay, and false positive rate. Table 9. Communication and routing performance metrics (Mean ± SD across 20 runs). Metric EMARL (Normal) EMARL (Adversarial) AODV (Baseline) Standard MARL Packet Delivery Ratio (%) 96.3 ± 1.1 90.7 ± 1.4 84.1 ± 1.8 89.7 ± 1.5 Accuracy (%) 94.6 ± 0.9 88.4 ± 1.3 76.2 ± 2.1 85.5 ± 1.7 End-to-End Delay (ms) 61.2 ± 2.4 70.5 ± 3.1 85.9 ± 3.8 78.6 ± 3.5 False Positive Rate (%) 4.3 ± 0.4 7.8 ± 0.6 13.6 ± 0.9 10.9 ± 0.7 Open in a new tab Fig. 9. Open in a new tab Comparative performance of EMARL, AODV, and Standard MARL across key communication and routing metrics. Learning convergence and agent behavior To evaluate the learning dynamics of the proposed EMARL framework, we analyzed the convergence behavior of the multi-agent reinforcement learning process, the evolution of cumulative rewards over time, and the stability of learned policies across multiple training episodes. These points are important to support the reliability, soundness, and scalability of MARL-based integration of cooperation strategies in dynamic UAV swarm networks. Training took place on 5,000 episodes and each episode had 200 decision steps. Per agent average cumulative reward was monitored and recorded after every set time to check the progress of learning. The EMARL framework converged in a smooth and consistent fashion and after a plateau, cumulative rewards grew to about the 3, 200 episode mark (Fig. 10 ). This means that the agents have been able to establish the lasting policy towards safe communication, threat prevention, and energy-efficient routing. Fig. 10. Open in a new tab Convergence comparison showing faster and more stable learning in EMARL versus standard MARL. However, The EMARL framework converged more rapidly, and its learning curve exhibited significantly lower variance compared to the baseline MARL model that lacked explainability and trust integration, further showing the stabilizing influence of the involvement of the trust-supervised decision-making process and interpretable feedback. Also, a series of validation runs conducted periodically demonstrated that EMARL agents were consistent in their policies even in so highly dynamic and partially observable topologies of the swarms. It was also instrumental to show that based on the behavioral analysis the learned policies favored energy-efficient and secure paths, avoided low-trust nodes and took appropriate countermeasures in reaction to adversarial and environmental conditions to the routing strategies. After convergence, the decision-making process also demonstrated little policy oscillation thus suggestive of stability of policies over a long-term and deployability of the model being realizable in practice. Table 10 synthesizes the performance of the MARL training, with the better convergence rate, average reward, as well as policy stability of the EMARL framework over the standard training of the MARL baseline. The findings reveal that EMARL can learn and adjust to different topology dynamically and exhibit the same or similar learning behavior in cases of changing episodes. Figure 10 shows the convergence of learning behavior of EMARL versus Standard MARL with respect to convergence speed and average cumulative rewards where the faster convergence and greater average cumulative rewards of EMARL shows the ability to learn a policy more effectively. Table 10. MARL training performance Summary. Metric EMARL Framework Standard MARL (Baseline) Episodes to Convergence ~ 3,200 ~ 4,100 Max Average Reward per Agent 182.6 164.3 Convergence Variance (Last 500 eps) 3.8 9.5 Policy Stability (Post-Convergence) High Moderate Response to Dynamic Topology Adaptive Reactive/Delayed Open in a new tab Explainability analysis The introduction of explainable artificial intelligence (XAI) into the EMARL system was tested to gauge the interpretability, confidence and clarity of actions made by UAV agents in routing and responding to threat activities. The SHAP (SHapley Additive exPlanations), LIME (Local Interpretable Model-Agnostic Explanations) and the attention-based visualization approach of the SHAP, LIME were applied in the XAI module to trace, justify and visualize the agent behavior in the normal and adversarial cases. SHAP was employed to measure the impact of every input variable, include trust score, signal strength, node distance, to the ultimate action that was determined by the agent. Mean of SHAP based feature importance analysis indicated the top two most significant features within the swarm were trust score and signal-to-noise ratio (SNR). This aligned well with human expert expectations, enhancing the model’s transparency and justifiability. LIME explanations were generated to locally approximate the policy model for specific decision instances. These surrogate models offered interpretable linear rule sets that explained why a certain UAV was preferred or avoided during communication. LIME’s explanations were particularly effective in real-time alert validation and post-event decision auditing. Attention maps were generated from the policy layers of the deep learning model, visually highlighting the regions of input data that influenced routing and threat detection decisions. These maps were especially useful in time-series input settings, showing how recent patterns of trust degradation or delay spikes shaped policy output. To evaluate the quality of explainability, the system was tested using both quantitative and qualitative metrics. Quantitatively, SHAP and LIME yielded fidelity scores of 93.1% and 89.4%, respectively, indicating high approximation accuracy relative to the underlying MARL policy. Additionally, a user study involving six domain experts was conducted to assign a human trust score, which averaged 8.7 out of 10, reflecting high perceived reliability, clarity, and usefulness of the explanations in operational settings. Table 11 presents the evaluation results of the integrated explainability techniques within the EMARL framework. SHAP and LIME demonstrate high fidelity in aligning with model behavior, with trust score emerging as the most influential feature. Attention visualizations offer distinct, time-aware insights, while human evaluators rated the system highly in terms of interpretability and operational usefulness. Figure 11 provides visual outputs from three interpretability tools applied within EMARL: SHAP highlights the most influential features (e.g., trust score, SNR), LIME clarifies routing choices (e.g., which nodes to avoid or select), and attention maps indicate spatial focus areas of UAV agents during decision-making. Table 11. Explainability evaluation Results. Technique Metric Score/Rating SHAP Fidelity 93.1% Top Feature Trust Score LIME Fidelity 89.4% Local Explanation Clarity High Attention Visual Focus Quality Distinct and Time-Aware Human Evaluation Trust Score (1–10) 8.7 Explanation Usefulness Operationally Effective Open in a new tab Fig. 11. Open in a new tab Visual examples of SHAP, LIME, and attention map outputs for interpreting EMARL agent decisions. Scalability and energy efficiency evaluation To assess the robustness and practicality of the proposed EMARL framework in real-world swarm applications, we conducted experiments focusing on scalability and energy efficiency across varying UAV swarm sizes and mission intensities. The evaluation measured system performance in terms of communication overhead, learning stability, and energy consumption per episode or per successful data delivery. Scalability was tested using swarm sizes of 10, 20, 30, 50, and 75 UAVs, deployed over a 2 km² dynamic operational area. Key performance metrics—PDR, average delay, and convergence time—remained stable as the swarm size increased, confirming the framework’s ability to maintain policy performance under growing agent populations. The average decline in PDR was limited to 4.2% from smallest to largest swarm size, demonstrating resilient communication behavior. Energy efficiency was evaluated using a hybrid simulation setup with UAV power models defined in AirSim, accounting for energy costs of communication, computation, and mobility. The average energy consumption per successful delivery remained within acceptable operational thresholds, increasing linearly with swarm size but benefiting from learned route optimization and avoidance of untrusted or congested nodes. Moreover, the introduction of partial policy sharing and distributed coordination allowed the system to avoid redundant computation, reducing overall network energy load. The EMARL model exhibited 28.6% less energy consumption per delivery compared to standard MARL and 34.9% lower than conventional trust-based models without learning capabilities. Table 12 outlines the scalability and energy efficiency outcomes of the EMARL framework as the UAV swarm size increases. While packet delivery ratio and delay remain within acceptable thresholds, convergence time slightly increases with swarm size, and energy consumption per delivery rises modestly, demonstrating the framework’s resilience and scalability under growing network demands. Figure 12 illustrates how increasing the swarm size affects PDR, average delay, and energy consumption in the EMARL framework. As the number of UAVs rises, PDR slightly declines, average delay increases, and energy per delivery also rises—highlighting the trade-offs between scalability, latency, and efficiency. Table 12. Scalability and energy efficiency Results. Swarm Size PDR (%) Avg Delay (ms) Convergence Episode Energy/Delivery (Wh) 10 UAVs 96.5 58.3 2900 1.72 20 UAVs 95.2 60.1 3050 2.03 30 UAVs 94.1 62.4 3170 2.37 50 UAVs 93.0 65.6 3320 2.84 75 UAVs 92.3 68.9 3470 3.18 Open in a new tab Fig. 12. Open in a new tab Impact of swarm size on PDR, delay, and energy consumption in the EMARL framework. Comparative analysis with baseline models To demonstrate the effectiveness and superiority of the proposed EMARL-XAI framework, a detailed comparative analysis was conducted against three baseline models: (i) AODV, a classical reactive routing protocol; (ii) Trust-Based Q-Routing, a lightweight reinforcement learning model augmented with trust metrics; and (iii) Standard MARL, a multi-agent reinforcement learning approach without integrated explainability or trust evaluation components. Each model was evaluated under identical experimental conditions using a 30-node UAV swarm operating in a dynamic adversarial environment with 20% malicious UAVs. The comparison focused on key performance indicators including PDR, accuracy of secure routing decisions, end-to-end delay, false positive rate (FPR), energy consumption per delivery, and explanation trustworthiness (in the case of models providing interpretability). Results summarized in Table 13 ; Fig. 13 clearly indicate that EMARL-XAI outperforms all baseline models across most metrics. Table 13. Comparative results: EMARL-XAI vs. Baseline Models. Metric EMARL-XAI AODV Trust-Based Q-Routing Standard MARL PDR (%) 94.1 79.5 88.7 89.4 Secure Routing Accuracy (%) 91.5 72.1 83.4 86.0 End-to-End Delay (ms) 62.4 88.2 70.5 68.7 False Positive Rate (%) 5.6 14.3 9.2 8.8 Energy/Delivery (Wh) 2.37 2.89 2.12 2.41 Explanation Trust Score (1–10) 8.7 — — — Open in a new tab Fig. 13. Open in a new tab Radar plot comparing EMARL-XAI with baseline models across key performance metrics, highlighting its superior accuracy, trustworthiness, and communication efficiency. Specifically, EMARL-XAI achieved the highest PDR (94.1%) and secure routing accuracy (91.5%), attributable to its integrated trust assessment and adaptive MARL policies. The false positive rate was also lowest (5.6%), reducing the likelihood of misclassifying benign UAVs as threats. In contrast, AODV exhibited poor performance under adversarial conditions due to its reactive nature and lack of trust awareness, while standard MARL struggled with trust ambiguity and lack of policy transparency. Importantly, EMARL-XAI was the only model to offer high-quality explanations, achieving a human trust score of 8.7/10, which reinforced operator confidence in automated decision-making. Although the Trust-Based Q-Routing approach consumed slightly less energy, its lower accuracy and higher FPR compromised overall mission reliability. Table 13 presents a comprehensive comparative analysis of the proposed EMARL-XAI framework against three baseline models—AODV, Trust-Based Q-Routing, and Standard MARL. The results highlight EMARL-XAI’s superior performance across key metrics, including higher packet delivery and secure routing accuracy, reduced delay and false positives, improved energy efficiency, and enhanced interpretability as indicated by a high explanation trust score. Figure 13 shows a radar plot comparing EMARL-XAI with AODV, Trust-Based Q-Routing, and Standard MARL. EMARL-XAI outperforms the baselines in PDR, accuracy, trust score, and inverse delay, energy, and FPR, demonstrating its overall superiority in communication reliability, efficiency, and explainability. Ablation study To understand the contribution of individual components within the EMARL-XAI framework, an ablation study was conducted to analyze the impact of removing core modules—namely the explainability (XAI) module, the trust computation layer, and specific adversarial conditions—on the system’s overall performance. This analysis helps to validate the architectural choices and highlights the functional importance of each module in ensuring robustness, interpretability, and secure swarm communication. Three ablation variants were defined: EMARL w/o XAI: The framework was trained and deployed without explainability support. Agent decisions were based on the same MARL policy but without interpretability mechanisms. EMARL w/o Trust: Trust score computation and malicious behavior monitoring were disabled. Agents relied solely on environmental inputs (e.g., signal strength, distance). EMARL under No Attack: The model was tested in a purely benign environment, with no jamming, spoofing, or Sybil attacks introduced. To further assess the contribution of the reward-shaping mechanism, an additional ablation experiment was conducted in which the reward weights ​ were flattened to equal values, thereby removing their behavioral prioritization. This modification resulted in a noticeable performance drop, with PDR decreasing by 6.8% , accuracy dropping by 7.3% , and FPR increasing by 4.1% . The results indicate that the carefully designed reward prioritization is critical for guiding agents toward secure, low-latency, and energy-efficient behavior, and that removing this structure leads to diluted learning signals and less stable policy convergence. The results, summarized in Table 14 and visualized in Fig. 14 , show a notable performance degradation when either the XAI or trust components were removed. Specifically, removing the trust mechanism led to a 9.6% drop in accuracy and an increase in the false positive rate by 5.4%, confirming its vital role in identifying and isolating malicious UAVs. Excluding the XAI module did not impact routing accuracy significantly but resulted in a loss of operator trust, with the human trust score declining from 8.7 to 4.1. This underlines the importance of explainability in human-centric and regulated autonomous systems. Table 14. Ablation study Results. Configuration PDR (%) Accuracy (%) FPR (%) Avg Delay (ms) Human Trust Score (1–10) Full EMARL-XAI 94.1 91.5 5.6 62.4 8.7 EMARL w/o XAI 94.1 91.2 5.8 62.3 4.1 EMARL w/o Trust 90.7 81.9 11.0 68.9 6.2 EMARL under No Attack 96.9 93.2 3.9 57.9 8.9 Open in a new tab Fig. 14. Open in a new tab Heatmap showing the impact of removing XAI and trust components on key system performance metrics. Interestingly, performance improved slightly in the absence of adversarial attacks (PDR + 2.8%, delay − 4.5 ms), confirming the EMARL framework’s ability to handle disruptions while maintaining efficient operations. Table 14 presents the ablation study results evaluating different configurations of the EMARL framework. The removal of the XAI module significantly reduces human trust scores despite similar communication performance, while omitting trust mechanisms increases the false positive rate and lowers accuracy. Performance under no attack shows optimal metrics, validating the framework’s robustness, trust effectiveness, and explainability contributions. Figure 14 presents a heatmap visualizing the ablation study results. It highlights the degradation in Accuracy, Trust Score, and False Positive Rate when the XAI or trust modules are removed from EMARL-XAI. The full configuration achieves optimal performance, confirming the critical role of both explainability and trust modeling in secure and efficient UAV swarm operations. Impact of UAV velocity on network performance Given the high mobility inherent to FANET deployments, we conducted an additional set of simulations to examine the relationship between UAV velocity and communication performance. UAV speeds were varied from 5 m/s to 20 m/s, while maintaining identical swarm size, attack conditions, and training settings. Results indicate that EMARL-XAI maintains stable routing accuracy and packet delivery despite increasing velocity, outperforming AODV and standard MARL models. However, as velocity increases beyond 15 m/s, link volatility results in moderate increases in delay and slight drops in PDR—consistent with the mobility constraints of highly dynamic aerial networks. As shown in Table 15 , increasing UAV velocity introduces noticeable variations in PDR, delay, and FPR, highlighting the sensitivity of FANET performance to high-mobility conditions. The Fig. 15 further illustrates these trends, demonstrating a gradual performance degradation as UAV speed increases from 5 to 20 m/s. Table 15. Effect of UAV velocity on key performance Metrics. UAV Velocity (m/s) PDR (%) End-to-End Delay (ms) Accuracy (%) FPR (%) 5 m/s 96.8 58.2 92.7 5.1 10 m/s 95.4 61.3 91.5 5.6 15 m/s 93.9 65.8 89.3 6.2 20 m/s 92.1 69.5 87.8 7.0 Open in a new tab Fig. 15. Open in a new tab Effect of UAV velocity on Packet Delivery Ratio, end-to-end delay, and false positive rate as vvv increases from 5 to 20 m/s. Discussion The EMARL-XAI model is an adequate way of combating the fundamental issues central to the idea of security, adaptability, and interpretability in UAV swarms-based FANETs. This research uses Multi-Agent Reinforcement Learning (MARL), incorporates trust-aware decision-making, and includes Explainable Artificial Intelligence (XAI) to come up with a complete solution that does not only increase the communication robustness but also offers transparency into the agent behaviors, which is an essential necessity in such aerial networks of missions. The achieved results of the simulation done across different performance measures, namely PDR, end-to-end delay, accuracy, energy consumption and false positive rate indicate that EMARL-XAI is far much superior in comparison to their baseline models, AODV, Trust-Based Q-Routing and Standard MARL. The proposed frame work had a 15–20% increase in PDR and accuracy with a lower false positive rate and delay compared to the traditional, even under adversarial conditions (jamming and Sybil attacks). In addition, introduction of a trust assessment mechanism facilitated the ability of the system to actively isolate and reduce the impact of malicious actors in a bid to create network resiliency. Another peculiarity of the presented research is combining the explainability modules, such as SHAP, LIME, and attention-based policy visualization. The modules will offer insights into decision-making process of the UAV agents in a human-understandable way, thus increasing confidence and system visibility to the individual. Fidelity and trust scores obtained based on interpretability scores also supported the values of the usage of the XAI component to provide trustful autonomous decisions. The ablation study also supports the idea that the function of the trust module and the role of XAI were the keys to maintaining high performance of the system. The elimination of one element or the other resulted in a quantifiably significant decline in trust in communication and responsibility of decisions, thus demonstrating that they are invaluable. Also, the scalability analysis showed that the EMARL-XAI framework would experience a stable performance when the size of the UAV swarm changed and few/no additional communication overhead or energy requirements are experienced. Although the framework is heavy-duty, when being used in a virtual setting, some considerations arise. Generation of real-time XAI may impose heavy computational loads that will pose a challenge to the deployment on resource-constrained UAV devices. Also, simulated models of attacks may be varied and may not capture all dynamics of the adversarial tactics in the real world. The restrictions indicate the potential directions of further research that include testing lightweight models of XAI and expanding on threat modeling by means of real-life FANET testbeds or hybrid simulation-emulation systems. To sum up, the study will have provided a unique, modular, and scalable framework of security architecture that incorporates decentralized learning, real-time explainability, and cross-system adaptive trust modeling of FANETs. It forms a solid basis towards making reliable and smart UAV swarm communication systems and thereby paves a way to be deployed in active and hostile aerial surface like disaster recovery, precision agriculture and border surveillance. Conclusion The given study introduces a framework of Explainable Multi-Agent Reinforcement Learning, EMARL-XAI, and attempts to improve the resilience, flexibility, and explainability of communication in UAV swarm-based FANETs. The proposed framework enables distributed MARL trained with the MADDPG algorithm along with the dynamic trust assessment approach and incorporated XAI approaches to overcome the main shortcomings of traditional UAV network protocols, that is, dependence on the adversarial attack, non-transparency of decisions, and low scalability. Through a large number of simulations on a variety of networks, compared to existing traditional methods like AODV, Trust-Based Q-Routing, and noisy regular MARL, EMARL-XAI also shows much better results in such important metrics as PDR, accuracy, false positive rate, communication delay, and energy efficiency. The integration of SHAP, LIME, and attention visualization modules allows for real-time interpretability of agent decisions, significantly improving system transparency and user trust—key considerations in autonomous mission-critical systems. Ablation studies confirm the essential role of both the XAI and trust components in maintaining system robustness under various adversarial scenarios. Furthermore, scalability assessments show that EMARL-XAI maintains consistent performance as the number of UAV agents’ increases, indicating its practical applicability in large-scale aerial deployments. EMARL-XAI represents a significant step toward intelligent, secure, and interpretable UAV swarm communication. Future work will focus on deploying the framework in real-world UAV platforms, optimizing computational efficiency of XAI modules, and extending the system to support collaborative sensing and task allocation in heterogeneous aerial networks. Acknowledgements The authors extend their appreciation to the Deanship of Research and Graduate Studies at King Khalid University for funding this work through Large Research Project under grant number RGP2/283/46. Author contributions H.K.A. and A.Q. conceptualized the study and designed the methodology. A.Q. implemented the software, conducted simulations, and curated the data. B.A. performed the formal analysis and prepared Figs. 1, 2 and 3. H.K.A. and B.A. wrote the main manuscript text. Y.G. supervised the project, handled correspondence, and contributed to manuscript revision and editing.All authors reviewed and approved the final manuscript. Funding The authors extend their appreciation to the Deanship of Research and Graduate Studies at King Khalid University for funding this work through large Research Project under grant number RGP2/283/46. Princess Nourah bint Abdulrehman University Researchers Supporting Project number (PNURSP2026R384), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia. Data availability The data that support the findings of this study are available from the corresponding author upon reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Romero, C. A. T. et al. FANET and MANET, a support and composition relationship. Comput Mater. Contin , 82 , 2, 1699-1732 (2025). 2. Abbas, S., Abu Talib, M., Ahmed, I. & Belal, O. Integration of UAVs and fanets in disaster management: A review on Applications, challenges and future directions. Trans. Emerg. Telecommun Technol. 35 (12), e70023 (2024). [ Google Scholar ] 3. Javed, S. et al. State-of-the-art and future research challenges in uav swarms. IEEE Internet Things J. 11 (11), 19023–19045 (2024). [ Google Scholar ] 4. Priyadarshi, R. & Kumar, R. R. Evolution of swarm intelligence: a systematic review of particle swarm and ant colony optimization approaches in modern research. Arch Comput. Methods Eng , 32 , 1–42 (2025). 5. Mazhar, S. et al. State-of-the-art authentication and verification schemes in vanets: A survey. Veh. Commun. 49 , 100804 (2024). [ Google Scholar ] 6. Sahraoui, S. & Bachir, A. Lightweight consensus mechanisms in the internet of blockchained things: thorough analysis and research directions. Digit Commun. Networks , 11 , 1246-1261 (2025). 7. Saeed, W. & Omlin, C. Explainable AI (XAI): A systematic meta-survey of current challenges and future opportunities. Knowledge-based Syst. 263 , 110273 (2023). [ Google Scholar ] 8. Tian, Y. et al. UAVs Meet llms: overviews and perspectives towards agentic low-altitude mobility. Inf. Fusion . 122 , 103158 (2025). [ Google Scholar ] 9. Qiu, T., Chen, N., Li, K., Qiao, D. & Fu, Z. Heterogeneous ad hoc networks: Architectures, advances and challenges. Ad Hoc Netw. 55 , 143–152 (2017). [ Google Scholar ] 10. Chen, Y. J., Chang, D. K. & Zhang, C. Autonomous tracking using a swarm of uavs: A constrained multi-agent reinforcement learning approach. IEEE Trans. Veh. Technol. 69 (11), 13702–13717 (2020). [ Google Scholar ] 11. Zhou, Z., Liu, G. & Tang, Y. Multiagent reinforcement learning: Methods, trustworthiness, applications in intelligent vehicles, and challenges. IEEE Trans. Intell. Veh , 9 , 8190-8211 (2024). 12. Sifakis, J. & Harel, D. Trustworthy autonomous system development. ACM Trans. Embed Comput. Syst. 22 (3), 1–24 (2023). [ Google Scholar ] 13. Ali, S. et al. Explainable artificial intelligence (XAI): what we know and what is left to attain trustworthy artificial intelligence. Inf. Fusion . 99 , 101805 (2023). [ Google Scholar ] 14. Alharbi, A. M., Alshehri, G. & Elhag, S. Reinforcement learning of emerging swarm technologies: A literature review, in Proceedings of the Future Technologies Conference , pp. 478–494. (2024). 15. Javaid, S. et al. Communication and control in collaborative uavs: recent advances and future trends. IEEE Trans. Intell. Transp. Syst. 24 (6), 5719–5739 (2023). [ Google Scholar ] 16. Xiong, H. et al. Towards explainable artificial intelligence (XAI): A data mining perspective, arXiv Prepr. arXiv2401.04374 , (2024). 17. Kaur, R. Artificial intelligence for cybersecurity: literature review and future research directions, no. January , 97 231-252 (2023). 18. Sapkota, R., Roumeliotis, K. I. & Karkee, M. UAVs Meet Agentic AI: A Multidomain Survey of Autonomous Aerial Intelligence and Agentic UAVs, arXiv Prepr. arXiv2506.08045 , (2025). 19. Hussain, A. et al. Computing challenges of UAV networks: A comprehensive survey. Comput Mater. Contin , 81 (2), 1999-2051 (2024). 20. McMillan, L. & Varga, L. A review of the use of artificial intelligence methods in infrastructure systems. Eng. Appl. Artif. Intell. 116 , 105472 (2022). [ Google Scholar ] 21. Dahmane, W. M., Ouchani, S. & Bouarfa, H. Smart cities services and solutions: A systematic review. Data Inf. Manag . 9 (2), 100087 (2025). [ Google Scholar ] 22. Ghelani, J., Gharia, P. & El-Ocla, H. Gradient monitored reinforcement learning for jamming attack detection in fanets. IEEE Access. 12 , 23081–23095 (2024). [ Google Scholar ] 23. Hassan, S. M., Mohamad, M. M., Muchtar, F. B. & Dawoodi, F. B. Y. P. Enhancing MANET security through federated learning and multiobjective optimization: A Trust-aware routing framework. IEEE Access , 12 , 181149-181178 (2024). 24. Badini, N., Marchese, M. & Patrone, F. NS-3-based 5G Satellite-Terrestrial Integrated Network Simulator, MELECON 2022 - IEEE Mediterr. Electrotech. Conf. Proc. , pp. 154–159, (2022). 25. Tong, X., Hamzei, M. & Jafari, N. Towards secure and efficient data aggregation in Blockchain-Driven IoT environments: A comprehensive and systematic study. Trans. Emerg. Telecommun Technol. 36 (2), e70061 (2025). [ Google Scholar ] 26. Hajiheidari, S., Wakil, K., Badri, M. & Navimipour, N. J. Intrusion detection systems in the internet of things: A comprehensive investigation. Comput. Networks . 160 , 165–191 (2019). [ Google Scholar ] 27. Ceviz, O., Sen, S. & Sadioglu, P. Distributed intrusion detection in dynamic networks of uavs using few-shot federated learning, arXiv Prepr. arXiv2501.13213 , (2025). 28. Huang, S. et al. Graph attentional based agglomerative cluster for UAV swarm networks. IEEE Trans. Netw. Sci. Eng , 12 , 3311 (2025). 29. Liao, C. et al. Game theory and Multi-Agent DRL based Anti-Jamming transmission for integrated Air-Ground network. IEEE Trans. Veh. Technol , 73 , 19565-1981 (2024). 30. AL-Syouf, R. A., Bani-Hani, R. M. & AL-Jarrah, O. Y. Machine learning approaches to intrusion detection in unmanned aerial vehicles (UAVs). Neural Comput. Appl. 36 (29), 18009–18041 (2024). [ Google Scholar ] 31. Dubey, G. P. et al. Optimal path selection using reinforcement learning based ant colony optimization algorithm in IoT-Based wireless sensor networks with 5G technology. Comput. Commun. 212 , 377–389 (2023). [ Google Scholar ] 32. Ben Aissa, S. & Ben Letaifa, A. UAV communications with machine learning: challenges, applications and open issues. Arab. J. Sci. Eng. 47 (2), 1559–1579 (2022). [ Google Scholar ] 33. Zeng, Q. & Nait-Abdesselam, F. Enhancing UAV network security: A Human-in-the-Loop and GAN-Based approach to intrusion detection. IEEE Internet Things J , 12 , 20870-20884 (2025). 34. Wang, D., Wang, J., Wang, J. & Liu, J. Spectrum sharing in cognitive UAV networks based on Multi-Agent reinforcement learning. IEEE J. Miniaturization Air Sp Syst , 6 , 82-91 (2024). 35. Huang, J., Huang, K., Jackson, K. & Hughes, C. AI agent safety and security considerations, in Agentic AI: Theories and Practices, Springer, (Farman Ali) 369–407 (2025). 36. Zabihi, Z., Eftekhari Moghadam, A. M. & Rezvani, M. H. Reinforcement learning methods for computation offloading: a systematic review. ACM Comput. Surv. 56 (1), 1–41 (2023). [ Google Scholar ] 37. Li, Z. et al. UAV networks against multiple maneuvering smart jamming with knowledge-based reinforcement learning. IEEE Internet Things J. 8 (15), 12289–12310 (2021). [ Google Scholar ] 38. Yang, T., Lu, Y., Deng, H. & Tang, C. A hybrid multimodal neural network-based anomaly detection model for UAVs. IEEE Trans. Aerosp. Electron. Syst , 61 , 10273-10291 (2025). 39. Roy, D., Panda, P. & Roy, K. Tree-CNN: a hierarchical deep convolutional neural network for incremental learning. Neural Netw. 121 , 148–160 (2020). [ DOI ] [ PubMed ] [ Google Scholar ] 40. Tian, Q. et al. An intrusion detection approach based on improved deep belief network. Appl. Intell. 50 (10), 3162–3178 (2020). [ Google Scholar ] 41. Zhou, L. et al. Cooperative digital twins enhanced UAV topology optimization for Multi-Target tracking. IEEE Trans. Commun. , 73 , 1 (2025). 42. Zhou, L., Leng, S. & Quek, T. Q. S. VerDT: A Versatile Digital Twins Framework for UAVs-Based Industrial Cyber-Physical Systems, IEEE Trans. Mob. Comput. , vol. 24, no. 10, pp. 10099–10117, Oct. (2025). 43. Hao, W. et al. Robust security energy efficiency optimization for RIS-aided cell-free networks with multiple eavesdroppers. IEEE Trans. Commun. 72 (12), 7401–7416 (2024). [ Google Scholar ] 44. Do, P. N., Khanh, P. X., Le, H. N., Phuong, N. T. & Hiep, P. T. and others, On the BLER Performance of UAV-Assisted Dual-Hop V2V MIMO-NOMA Systems in Finite Blocklength Regime, Digit. Signal Process. , p. 105625, (2025). 45. Duc, B. A., Hoang, T. M., Phuong, N. T., Tran, X. N. & Hiep, P. T. Power optimization and ARSs selection strategies in downlink cell-free multi-ARSs communication systems. Arab. J. Sci. Eng. 50 (8), 5647–5664 (2025). [ Google Scholar ] 46. López-Villegas, I., Mart\’\inez-Rios, E. A., Izquierdo-Reyes, J., Bustamante-Bello, R. & Falcone, F. A systematic literature review of emergency communications assisted by unnamed aerial vehicles. Ad Hoc Networks , 182 , 104063, (2025). Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request. Articles from Scientific Reports are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (3.5 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top