AI-Driven Safety and Security for UAVs: From Machine Learning to Large Language Models

Abstract

As unmanned aerial vehicle (UAV) applications expand across logistics, agriculture, and emergency response, safety and security threats are becoming increasingly complex. Addressing these evolving threats, including physical safety and network security threats, requires continued advancement by integrating traditional artificial intelligence (AI) tools such as machine learning (ML) and deep learning (DL), which contribute to significantly enhancing UAV safety and security. Large language models (LLMs), a cutting-edge trend in the AI field, are associated with strong capabilities for learning and adapting across various environments. Their emergence reflects a broader trend toward intelligent systems that may eventually demonstrate behavior comparable to human-level reasoning. This paper summarizes the typical safety and security threats affecting UAVs, reviews the progress of traditional AI technologies, as described in the literature, and identifies strategies for reducing the impact of such threats. It also highlights the limitations of traditional AI technologies and summarizes the current application status of LLMs in UAV safety and security. Finally, this paper discusses the challenges and future research directions for improving UAV safety and security with LLMs. By leveraging their advanced capabilities, LLMs offer potential benefits in critical domains such as urban air traffic management, precision agriculture, and emergency response, fostering transformative progress toward adaptive, reliable, and secure UAV systems that address modern operational complexities.

1. Introduction

1.1. Motivation

Unmanned aerial vehicles (UAVs) have demonstrated significant application value across multiple fields because of their exceptional autonomy, flexibility, and adaptability. Low-altitude flight activities in various scenarios, such as package delivery [1], agriculture [2], fire detection [3], and other operational tasks, have led to the development of comprehensive economic models that integrate and support advancements across related fields.

However, as the number of deployed UAVs increases and their missions diversify, UAV safety and security issues have become increasingly complex. Potential social problems associated with UAVs include flight disruptions, collision or crash incidents, unauthorized flights [4], and no-fly zone designations. In a complex and three-dimensional low-altitude traffic environment, large-scale UAV operations may encounter various dynamic and complex obstacles and threats, including high-risk possibilities of collisions between aircraft and buildings or other obstacles [5]. Such incidents may seriously threaten airspace order, public property, and even personal safety. The complexity of these issues underscores the need for technical solutions that consider physical safety and network security.

To address these challenges, artificial intelligence (AI) technologies have been applied to expand UAV applications, particularly in areas such as image recognition and target tracking [6]. Moreover, AI technology has become the core support for UAV safety and security. For example, deep reinforcement learning (DRL), recurrent neural networks (RNNs), and long short-term memory (LSTM) networks can be employed to optimize UAV resource allocation, latency estimation, and energy management effectiveness [7]. In terms of environmental sensing and decision-making, the combination of convolutional neural networks (CNNs) and distributed reinforcement learning (RL) has supported improvements in obstacle avoidance and the route planning efficiency of UAV swarms in complex scenarios [8]. In threat detection and cybersecurity, LSTM has been used to analyze sensor data for mechanical failure and network intrusion prediction. However, the limitations of AI algorithms, including machine learning (ML) algorithms, constrain their broader deployment and application. For example, traditional AI algorithms still require further improvement in rapidly changing environments to ensure the stability of their output results. The output accuracy of many data-driven ML models is highly dependent on the availability of large amounts of high-quality training data. However, acquiring such data can be difficult due to data privacy restrictions and regional regulatory limitations, which significantly affect the output accuracy of ML models, especially in critical safety and security applications.

To address these limitations, large language models (LLMs) present a promising alternative. LLMs can enhance model stability in dynamic environments by supporting contextual understanding and high-level reasoning, which can help UAVs adapt more effectively to changes. Moreover, by leveraging the vast knowledge and natural language understanding capabilities of LLMs, UAVs can make more informed decisions with less reliance on extensive, high-quality training data, which is particularly crucial for critical UAV safety and security applications where data availability is limited. Moreover, LLMs facilitate natural language interactions between operators and UAVs, allowing real-time adjustments and interventions to ensure safety and security. In light of these potential benefits, researchers have begun exploring the integration of LLMs with UAV systems to increase their performance and capabilities.

Recently, in 2024, Sun et al. explored the integration of LLMs and UAV communication systems [9]. In 2024, Javid et al. evaluated ongoing efforts to integrate LLMs with UAV systems, discussing (in detail) the applicability and potential of LLMs in UAV communication, data processing, decision-making, and spectrum sensing. Their work highlighted the advantages of LLMs in data processing and decision-making [10]. For example, several existing flight control models for UAV swarms have been developed based on GPT-4, such as FlockGPT [11] and Aeroverse [12]. UAV path planning models based on Gemini, such as LEVIOSA [13], have been introduced, and emergency rescue response models have been established based on LLaMA3 and Claude3, such as Say-REAPEx [14]. These models interpret and respond to natural language commands through LLMs, simplifying the control process of UAVs and enhancing their autonomous decision-making capabilities [15,16]. Moreover, Tian et al. thoroughly explored the integration of LLMs and UAVs, indicating that LLMs can enhance the autonomy and intelligence of UAVs in low-altitude mobile systems and proposed a UAV architecture that realizes autonomous sensing, reasoning, memory, and tool use [17]. These papers either focus on applying LLMs to specific UAV scenarios or comprehensively summarize the integration of LLMs with UAV communication or UAV network architectures. However, exploration remains limited in areas involving the combination and application of LLMs for UAV safety and security.

Considering the current research gap regarding the use of LLMs in UAV safety and security, this paper summarizes the present state of AI applications in this domain. Moreover, this paper comprehensively studies and examines the advantages and challenges of LLMs in achieving UAV safety and security. Section 1 outlines the motivation behind this paper, investigates the current body of research, summarizes the existing surveys, and presents the contributions of this paper. Section 2 presents an analysis of the safety and security threats associated with UAVs, reviews the development of AI technology, and discusses the potential applications of LLMs in UAV safety and security. Section 3 discusses the application of traditional AI techniques in UAV safety in detail, including physical safety design, battery safety management, sensor spoofing, collision avoidance, path planning, and flight control. It also highlights the limitations of traditional AI technologies and discusses the potential advantages of LLMs. Section 4 focuses on UAV security, discussing the applications of traditional AI techniques, including sensing security technology, radio interference defense, identity authentication, cryptography technology, and self-organizing network (SON) communication security. In addition, it presents an analysis of the shortcomings of traditional AI methods and highlights the potential application space of LLMs. Section 5 discusses the challenges in deploying LLMs for UAV safety and security and outlines future research directions. Section 6 summarizes and concludes this paper. The structure of our survey is presented in Figure 1.

1.2. State of the Art

UAV safety and security have become core concerns in both academia and industry, each with distinct focuses and approaches. There are surveys that categorize this issue into two categories: safety and security [18,19]. Safety pertains to the physical integrity of UAV hardware and the prevention of risks to human safety from incidents such as collisions or crashes. This aspect emphasizes the reliability of components, such as structural integrity and stability, ensuring that mechanical failures are minimized to avoid operational compromise or threats to human life. Conversely, while safety focuses on maintaining physical reliability and preventing harm, security is concerned with protecting UAV systems from intentional disruptions to data and control. Security involves safeguarding UAV networks and communications against threats such as unauthorized access and cyberattacks. It depends on robust encryption techniques and resilient network management strategies to ensure secure communication and command integrity. Together, these two domains form the foundation of UAV safety and security research, jointly addressing the multifaceted challenges associated with UAV deployment across diverse application scenarios.

UAV safety, a cornerstone of hardware reliability, has undergone significant research advancements aimed at enhancing resilience against physical threats. In 2022, Zhang et al. explored collision safety for lightweight UAVs and recommended integrated active and passive safety systems to mitigate real-world risk [20]. In 2023, Derhab et al. further classified physical attacks on internet of unmanned aerial vehicles (IoUAVs), underscoring the need for robust countermeasures. Current trends show a shift from passive detection to active defenses, leveraging technologies such as multisensor data fusion and AI-driven anomaly detection [21]. In 2024, Wei et al. analyzed vulnerabilities in flight controllers, sensors, and airframes and proposed hardware-specific security measures to counter side-channel attacks [22]. Efforts have been made to increase sensor reliability through multisensor data fusion and encrypted communication links to counteract vulnerabilities such as spoofing attacks [23]. Additionally, innovations in UAV physical safety have been reported, with approaches involving neural network-based systems for predicting battery life and AI-driven self-healing materials to address potential damage [24,25]. More recently, the integration of LLMs and DRL has revolutionized real-time obstacle detection and autonomous navigation, enabling UAVs to adapt dynamically to complex environments [26]. These advancements collectively enhance UAV safety and reliability in diverse scenarios.

UAV security research has also become increasingly critical, with a focus on safeguarding network communication integrity. Traditional approaches focus on cryptographic methods and essential sensor protection to ensure communication integrity and environmental awareness. However, as threats evolve, research has shifted toward integrating AI algorithms like ML techniques to increase robustness. For example, AI-driven multifactor authentication and lightweight cryptographic algorithms were developed to address dynamic environments and resource constraints [27,28]. More recently, emphasis has been placed on leveraging LLMs and deep learning (DL) for real-time threat detection and adaptive defense mechanisms. LLMs are now used to generate lightweight authentication policies, predict interference sources, and optimize frequency-hopping strategies, significantly improving UAV resilience against electromagnetic interference and spoofing attacks [29]. Additionally, multimodal data fusion approaches incorporating LLMs have supported enhanced UAV sensing capabilities, contributing to more accurate detection of anomalies and improved situational awareness [17]. These advancements collectively signify a transition from static, traditional security measures toward dynamic, AI-driven solutions that enhance UAV security in complex and evolving threat landscapes.

Standardization provides a critical framework for UAV safety and security worldwide. Currently, laws and regulations exist to support the healthy development and stable operation of the low-altitude economy. In 2019, the European Aviation Safety Agency (EASA) issued two European Union (EU) general regulations on UAVs, Commission Delegated Regulations 2019/945 [30] and 2019/947 [31]. These regulations categorize UAVs on the basis of operational risk and establish a registration system, a remote identification system, and a geo-awareness system to lay the foundation for U-Space. In 2023, the EU released a directive on the protection against unmanned aircraft systems [32] to address UAV-related threats and ensure that the increasing number of UAVs and rapid technological advancements do not lead to an uncontrolled surge in risk to civil airspace. The Unmanned Aircraft Systems Standardization Collaborative (UASSC) under the American National Standards Institute (ANSI) is responsible for the development of relevant standards, which are intended to promote the safe, reliable, and effective operation of UAV systems and to provide the UAV system industry with unified technical benchmarks and guidance [33]. Additionally, 3GPP is the main organization for mobile communications standardization. Specification 3GPP TS 22.125 defines the requirements for UAV system operations, including remote identification and tracking operational and security requirements [34]. China has released more than 150 policy documents related to the low-altitude economy. In 2024, the Safe Administration of Civil Unmanned Aerial Vehicle Operations and other regulations were enacted. To provide a comprehensive overview of the diverse applications and challenges in UAV safety and security, Figure 2 illustrates a panoramic view of UAV operations across multiple domains.

1.3. Comparison of Related Surveys

In the field of UAV safety and security, the literature has thoroughly explored the potential application areas and classification of related threats. For example, in 2023, Mohsan et al. identified UAVs as promising tools across various fields, such as disaster management, remote sensing, search and rescue, and intelligent navigation [35]. However, operational limitations and safety and security issues remain. In 2024, Ceviz et al. conducted an attack surface analysis to identify potential vulnerabilities in UAVs and flying ad-hoc networks (FANETs), categorized several types of attacks, and discussed preventive measures and detection solutions in detail [36]. They also analyzed the impact of four routing attacks on FANETs through simulation experiments. In 2024, Laghari et al. analyzed the advantages and disadvantages of communication protocols such as Wi-Fi, ZigBee, and LoRaWAN and examined UAV security from three aspects: encryption, authentication, and data protection [37]. Their study highlighted many applications of UAVs in environmental monitoring, disaster response, and industrial safety and security. They noted that integrating UAVs and AI technology can significantly enhance UAV performance and efficiency, improving flight control, navigation, decision-making, sensing capabilities, and human–computer interaction.

In addition to traditional algorithms addressing UAV safety and security issues, AI-assisted strategies have gained prominence as a research hotspot. In 2023, Sai et al. presented the first comprehensive review of AI applications in the UAV field [38]. They classified existing papers on application scenarios, AI algorithms, and AI training paradigms and discussed the use of different AI algorithms in UAV applications. In 2022, Abro et al. provided a detailed discussion on the classification, communication methods, and architecture of UAVs, with a particular focus on UAV regulatory standards as well as potential vulnerabilities and types of attacks [39]. They also proposed solutions such as ML-based intrusion detection systems. In 2022, Pandey et al. highlighted the severe threats that UAVs face due to their open operational environment, such as cyberattacks and eavesdropping on communication links [40]. They also discussed security issues in UAV communication, including attack types such as navigation, data injection, and software installation. In addition, they proposed various mitigation strategies and security performance metrics, such as physical layer security techniques, ML methods, software-defined networking, fog computing, and blockchain. In 2023, Mekdad et al. focused on UAV safety, security, and privacy issues, systematically categorizing them into four aspects: hardware, software, communication, and sensing. For each aspect, they thoroughly explored common vulnerabilities, potential threats, active and passive attack methods, and possible defense technologies, such as ML-based intrusion detection techniques [41]. Sarker et al. focused on UAV autonomy, network resource management and planning, multiple access and routing protocols, power control, and energy efficiency [42]. They reported that AI-based UAV networks are both technically and economically feasible for designing and deploying next-generation networks.

However, current reviews are conducted mainly by more targeted organizations regarding the analysis of how AI techniques can be applied to address threats in UAV safety and security. For example, in 2022, Mcenroe et al. provided a detailed analysis of the impact of edge AI approaches on UAV technologies, such as autonomous navigation, swarm control, energy management, security and privacy, computer vision, and communication [43]. In 2023, Telli et al. used data from the Scopus database to analyze the activity and inter-relationships of different research directions [44]. They provided a detailed analysis of the development and research relationships of technologies such as UAV antennas, obstacle detection, flight control, and the Internet of Things (IoT). Nevertheless, they lacked an analysis focused on UAV security.

Moreover, in 2024, Sarikaya et al. provided a detailed overview of UAV system architectures and explored the security threats targeting UAV systems [45]. They emphasized the need for RL-based UAV security solutions for each type of attack. However, this work focuses more on UAV security and lacks a discussion of UAV safety. In 2024, Zolfaghari et al. reviewed the security challenges of UAVs, including jamming attacks, search attacks, spoofing attacks, electromagnetic attacks, and deauthentication attacks, as well as related security control measures [46]. They also emphasized the role of AI technologies in UAV safety and security. In 2024, Adil et al. explored UAV-assisted IoT, discussing key issues such as the characteristics of UAVs and the IoT, communication complexity, self-healing and self-organizing capabilities, random communication changes, and cybersecurity threats [47]. They highlighted the potential of ML, DL, and RL approaches in enhancing the security of UAV-assisted IoT. However, these papers also lack discussion on UAV safety.

Although AI techniques have demonstrated strong application potential in the field of UAV security, the limitations of traditional AI approaches in UAV safety and security have been explicitly noted [48]. Although advances have been made in recognition capabilities, detection algorithms are still sensitive to adversarial attacks, and model accuracy still requires improvement. In addition, it is necessary to ensure, through other means, that data analysis conducted through ML algorithms can fully protect the privacy and security of UAV data.

The emergence of generative artificial intelligence (GAI) and LLM technologies offers promising avenues to address some of these limitations. In 2023, Telli et al. indicated that the integration of a GAI or LLMs with UAV security will likely become a standard trajectory in future research [44,49]. In 2024, Kaleem et al. reported that GAI methods could contribute to generating UAV communication training data, optimizing communication models, reconstructing environments, and determining the location of UAVs in the absence of global positioning system (GPS) signals by interpreting radio signals [50]. In 2024 and 2025, Javaid et al. and Tian et al. combined LLMs with UAV network design and visual recognition, demonstrating the potential of LLMs in the UAV field [10,17]. However, they did not directly mention UAV security. In 2024, Sun et al. and Zhao et al. thoroughly discussed the classification of GAI methods and summarized the roles and limitations of advanced GAI models, such as generative adversarial networks (GANs), autoencoders, variational autoencoders (VAEs), and diffusion models (DMs), in addressing physical layer security challenges, including communication confidentiality, authentication, availability, resilience, and integrity [9,49]. They also noted the potential of GAI methods in addressing the sparsity and incompleteness of physical layer security data, providing heuristic ideas and novel insights. A comparison of related surveys is shown in

Table 1. Comparison of related surveys.

Ref.	Year	LLM			Countermeasures		Traditional AI
		Taxonomy	Applications	Challenges	UAV Safety	UAV Security	Taxonomy	Applications	Limitations
[43]	2022				✓		✓	✓	✓
[39]	2022				✓	✓		✓
[40]	2022				✓	✓		✓
[35]	2022				✓	✓
[44]	2023		✓		✓			✓
[41]	2023				✓	✓		✓
[38]	2023				✓	✓	✓	✓
[42]	2023					✓		✓
[36]	2024				✓	✓
[45]	2024					✓	✓	✓	✓
[48]	2024				✓	✓	✓	✓	✓
[51]	2024				✓	✓	✓	✓
[50]	2024		✓	✓	✓			✓	✓
[9]	2024		✓			✓	✓	✓	✓
[49]	2024		✓	✓		✓	✓	✓
[10]	2024	✓	✓	✓
[46]	2024					✓		✓
[47]	2024					✓	✓	✓
[37]	2024				✓	✓		✓
[17]	2025	✓	✓		✓
Ours	2025	✓	✓	✓	✓	✓	✓	✓	✓

.

Prior surveys on AI applications in UAV safety and security mainly summarize existing technologies or provide taxonomies of AI methods. Our work introduces a novel framework, classifying UAV safety and security threats into five distinct aspects, each based on a structured taxonomy from [18,19]. Unlike prior surveys, we compare traditional AI (ML and DL) with LLMs for UAV safety and security. We analyze traditional AI applications and their limitations, like poor generalizability and lack of interpretability, and we show how LLMs overcome these through multimodal data fusion and chain-of-thought reasoning. We also identify the challenges posed by the current deployment of LLMs and highlight future research directions.

To ensure scientific rigor, we followed a systematic literature selection process [52], conducting a comprehensive database search with relevant keywords, screening articles by titles and abstracts, and reviewing full texts for quality and relevance. Articles were categorized into directly related, indirectly related, and background knowledge groups, and evaluated for transparency, rigor, credibility, and relevance. This methodological rigor, paired with our novel framework, ensures a robust and targeted review, offering actionable insights into LLM applications for UAV safety and security, distinguishing our work from existing surveys.

The contributions of this paper are as follows:

• We present a comprehensive review of traditional AI methods in the context of UAV safety and security, synthesizing their pivotal role in mitigating operational threats;
• We critically evaluate the limitations of conventional AI models, focus on their shortcomings in dynamic, real-time UAV environments, and propose essential areas for advancement to achieve robust and reliable safety and security frameworks;
• We provide an in-depth analysis of the emerging integration of LLMs into UAV safety and security domains, demonstrating their transformative potential through multimodal data fusion, real-time decision-making, and adaptive threat detection, offering insights that may be useful for future research;
• We outline innovative future research directions, emphasizing novel solutions such as multimodal embodied intelligence systems and satellite networks, to address LLM deployment challenges, paving the way for secure and efficient operations.

2. Background

In UAV applications, diverse threats, including safety issues (such as hardware malfunctions) and security risks (such as communication breaches), and their complex interplay require a thorough understanding of vulnerabilities and practical solutions. This chapter aims to lay the foundation for the subsequent exploration of AI-driven safety and security advancements. It begins by exploring the diverse threats to UAV safety and security, highlighting the complexity and multifaceted nature of these challenges. The chapter subsequently delves into the evolution of AI technologies, tracing their development from traditional ML algorithms to the cutting-edge capabilities of LLMs. This progression is contextualized within the broader landscape of UAV applications, emphasizing how AI technologies have become integral to enhancing UAV functionality and addressing safety and security concerns. The chapter concludes with an overview of the potential applications of LLMs in UAV safety and security. It sets the stage for a detailed examination of their capabilities and limitations in the following chapters.

2.1. UAV Safety and Security Threats

UAV safety and security threats span hardware components, communication infrastructures, and interactive environments, presenting a complex challenge. At the hardware level, physical safety design focuses on material durability and flutter mitigation to prevent in-flight structural instability, while battery safety remains critical, as overheating, overdischarge, and aging can lead to catastrophic failures. Environmental interactions amplify threats, with sensor spoofing targeting GPS, inertial measurement units (IMUs), and vision systems to disrupt navigation, collision avoidance, and path planning, especially in the presence of dynamic obstacles such as birds or other UAVs. Loss of flight control, triggered by environmental interference or system vulnerabilities, underscores the need for robust hardware safeguards and adaptive system-level responses to maintain operational stability in dynamic conditions.

UAV security threats within communication systems present a multifaceted challenge to operational reliability and data integrity. Sensing security adds another dimension, with adversaries potentially exploiting communication vulnerabilities to manipulate sensor data, thereby undermining navigation and control. Radio interference resilience is also essential, as the physical and media access control (MAC) layers are susceptible to eavesdropping on unencrypted transmissions and electromagnetic jamming that disrupts links. In self-organizing network (SON) topologies, threats escalate, as attackers may deploy resource exhaustion strategies to disable nodes, traffic jamming to corrupt data, or fast and wormhole attacks to destabilize network structures, threatening coordination among distributed UAV clusters. Information security hinges on robust authentication to prevent impersonation and cryptography to safeguard against man-in-the-middle or replay attacks, which can alter transmitted content or disrupt timing. At the network level, countermeasures are needed against routing attacks such as black holes and flooding, which require encrypted protocols and adaptive strategies to ensure secure, resilient UAV communications. Figure 3 depicts the typical safety and security threats discussed in this paper.

2.1.1. UAV Safety Threats

• Physical Safety Risks

UAVs may suffer physical damage or be exploited for malicious purposes. For example, an attacker may interfere with the regular operation of a UAV by damaging its hardware components, such as motors and sensors [53]. Additionally, UAVs may exhibit vulnerabilities in their physical design. These vulnerabilities may be exploited by attackers, damaging the UAVs. In addition, UAVs may fail during operation due to environmental factors such as strong winds and extreme temperatures or structural issues such as overloading and insufficient material strength [54]. In addition, poorly designed landing systems for UAVs may lead to accidents.

• Battery management challenges

UAV batteries may be hit, crushed, or dropped during flight or storage. These events can damage their internal structures and trigger short circuits, heat generation, or even explosions. Using or storing UAV batteries at extreme temperatures may result in performance degradation or failure. For example, a significant decrease in battery performance at low temperatures may result in the inability of a UAV to fly properly. In addition, storing fully discharged batteries for extended periods can lead to deep discharge, which can cause irreversible damage to battery cells.

• Sensor Spoofing Hazards

UAV airframes are typically equipped with altitude and position sensors such as gyroscopes, accelerometers, and GPSs, which are essential for flight safety and precise navigation. Current threats against such sensors include ultrasonic interference with gyroscopes and GPS spoofing. In a GPS spoofing attack system, attackers can typically execute spoofing attacks in two ways after accessing the GPS receiver of a UAV. One approach involves leveraging the pseudorange calculation principle in GPS ranging. Attackers can process received satellite signals with high fidelity through a jamming device, subsequently delaying and forwarding the signal to deceive the GPS receiver into calculating an incorrect large pseudorange. This approach is referred to as a forwarding spoofing attack. On the other hand, attackers can create GPS jamming signals tailored to the satellite signal characteristics in the target area. They can then transmit these fabricated signals as if they were genuine satellite signals into the target’s signal reception zone. This deception causes the compromised GPS receiver to lock onto the jamming signal source, resulting in erroneous pseudorange and positioning data [55].

• Collision Dangers

Deploying clusters or multiple UAVs offers significant advantages over deploying single UAVs [56]. In challenging and dynamic environments, UAV flight safety is limited by sensor and battery capacities and power limitations. Additional complications arise from reduced visibility due to inclement weather, such as rain and dust, and the complexity of remote monitoring. In addition, behaviors such as ignoring flight rules, making operational errors, and flying lawless UAVs in the dark can exacerbate the risk of collisions. In complex environments, UAVs must account for dynamic obstacles, such as other flying vehicles and birds, and static obstacles, such as buildings and power lines. These obstacles may interfere with the flight path of a UAV and increase the risk of collision [57].

• Loss of Flight Control Threats

Navigational failures occasionally occur during UAV missions, prompting affected UAVs to autonomously search for and rely on nearby UAVs to act as navigators and guide them back to safety. However, if the pilot aircraft is improperly operated or illegally controlled by an attacker, for example, intentionally deviating from its route, a series of chain reactions may be triggered, including collisions, loss of control, and crashes among the vehicles, resulting in property damage and casualties. On the other hand, UAVs usually rely on electronic fencing technology to define their flight range and ensure that they fly in authorized airspace. However, once an electronic fence is illegally attacked or maliciously tampered with, the aircraft may lose control and enter into sensitive areas such as no-fly zones, hazardous zones, or military zones, which not only significantly increases the risk of collision but also triggers safety accidents and even leads to the exposure of national secrets, posing a significant threat to national and civilian safety.

2.1.2. UAV Security Threats

• Sensing Uncertainties

Owing to the open three-dimensional network environment and the vast number of sensing nodes in the UAV sensing system, the information transmission process is also vulnerable to various security threats, such as eavesdropping, tampering, and manipulation. For example, sensing nodes upload information, such as location, sensing capabilities, sensing willingness, and load conditions, to support mission decision-making. If privacy policies and regulations restrict sensing in certain areas, tampering with the node location information may result in the node turning on the sensing function in restricted areas, leading to privacy violations. In addition, although multisensor data fusion technology improves the accuracy of UAVs, it is still not completely resistant to algorithmic attacks.

• Radio Interference Disruptions

As terrestrial radio technologies such as mobile base stations, TV broadcast towers, and Wi-Fi equipment become more widespread and complex, their frequency bands often overlap with those used by low-altitude aircraft. This overlap increases the risk of interference, degrading vehicle-to-vehicle communication quality, causing data loss or complete disruptions, and ultimately impairing the aircraft’s maneuvering, positioning, and obstacle avoidance capabilities. In addition, owing to the broadcast characteristics of wireless channels, UAVs are susceptible to interference from malicious ground sources. This interference may result in transmitted data loss, elevated block error rates, and disconnection from communication links [58]. Such disruptions can interrupt the transmission of critical control data, thereby increasing the risk of crashes or collisions.

• Authentication Vulnerabilities

To establish communication with vehicles, UAVs activate Wi-Fi hotspots, allowing ground-based devices, such as mobile phones, computers, and remote controls, to achieve data interaction and remote control with UAVs by connecting to hotspots. However, this communication method is not completely secure. Attackers can easily identify and access the Wi-Fi links to steal sensitive data or even take complete control of the UAVs by spoofing the control signal remotely or surveillance-broadcast message attacks [59]. On the other hand, UAVs and ground networks, which rely on communication links for controlling commands and data interactions, are equally exposed to threats regarding information theft, information tampering, DOS attacks, and information injection. For example, UAVs may be impersonated by a malicious third party to perform unauthorized operations or enter no-fly zones. In addition, the lack of comparability of UAV identity information may lead to identity theft or forgery [60].

• Nonencrypted Data Exposure

UAV cryptography security risks stem largely from cryptographic algorithm implementations. The mainstream advanced encryption standard (AES)-128 is vulnerable to quantum computing. Lightweight cryptographic algorithms such as CHAM-128/128 exhibit insufficient resistance to side-channel attacks. Flaws in key distribution mechanisms further induce risks, while homomorphic encryption used for telemetry data protection is limited by arithmetic power. In addition, the security of traditional encryption algorithms may be limited by key management and computational complexity, and is susceptible to tampering.

• UAV Networking Complications

UAV networking relies on various communication protocols with vulnerabilities that attackers can exploit to perform man-in-the-middle attacks, DoS attacks, or service disruptions. Current mainstream protocols, such as ad-hoc on-demand distance vector (AODV), optimized link state routing (OLSR), and greedy perimeter stateless routing (GPSR), are vulnerable to multiple security threats in dynamic topology environments. Attackers can exploit protocol vulnerabilities to implement route spoofing, data theft, or even network paralysis [61]. In addition, UAV networks usually adopt a distributed or SON structure, which has some vulnerabilities despite offering flexibility and fault tolerance [62]. The mobility of UAV nodes and frequent topology changes also increase the complexity of network management and easily lead to security vulnerabilities.

2.2. Overview of Artificial Intelligence Technologies

In recent years, AI technology has developed at an explosive rate, rapidly penetrating various fields and extensively promoting technological innovations. The adoption of AI technology has also affected the field of UAVs. There are already applications of AI-assisted UAVs, such as package delivery, fire detection, and real-time target detection [63], which have greatly improved the efficiency of tasks such as military reconnaissance, logistics and distribution, and environmental monitoring.

Therefore, this section focuses on the basic concepts of AI technology, systematically reviewing the technological development from ML and DL methods to LLMs while clarifying their relationships and the differences and connections in their applications, as shown in Figure 4.

2.2.1. Artificial Intelligence

AI is a comprehensive concept that includes techniques “enabling machines to exhibit intelligent human behavior”, such as ML, natural language processing, computer vision, and robotics. These techniques are intended to support machines in understanding, learning, reasoning, and solving problems.

In the 2010s, integration between UAVs and AI-related technologies began to accelerate. With advancements in AI recognition technology, the images captured by UAVs can be detected in real time through a mounted camera, supporting target detection, target recognition, target positioning, target tracking, and other functions [64]. For example, in environmental monitoring, AI models can assist UAVs in analyzing air quality, water pollution, and forest coverage in real time, thus improving the efficiency of ecological protection and warning of disasters [65]. In infrastructure inspection, AI techniques have been combined with UAVs to automatically identify structural defects in bridges, pipelines, and power lines, achieving maintenance accuracy while reducing manual inspection costs [66]. In addition, the AI model can assist in UAV path planning. For example, Shen et al. proposed an AI-based method for determining the hovering position of a UAV, which reduced the UAV flying distance by 11% and energy consumption by 25% compared to the traditional method [67].

2.2.2. Machine Learning

ML is a subset of AI that refers to methods through which machines can identify patterns in data. Applications include predicting the weather, estimating travel times, recommending songs, automatically completing sentences, summarizing articles, and generating never-before-seen images [68]. ML approaches can be divided into three main categories: supervised learning, unsupervised learning, and RL. Supervised learning performs model training and prediction with labeled data. Unsupervised learning models infer intrinsic patterns and data structures without explicit labels. RL models optimize strategies through iterative interaction with the environment, using feedback in the form of rewards or penalties.

When combined with ML techniques, UAVs can autonomously extract the key target features and realize functions such as target classification, state prediction, and optimal path planning, and can improve UAV positioning accuracy by up to 35% over non-learning techniques [69]. For example, in disaster monitoring, a trained ML model can be employed to assist a UAV in quickly identifying affected areas and predicting the progression of a disaster, providing accurate guidance for rescue operations [70].

2.2.3. Deep Learning

DL is a subset of machine learning that focuses on utilizing neural networks to perform tasks such as classification, regression, and representation learning. Drawing inspiration from the neural structure of the human brain, DL techniques focus on automatically identifying important data features and constructing models that do not require manual data structure descriptions. In contrast to traditional ML methods, DL approaches avoid the need for manually creating complex processes to extract features from data. Instead, they automatically learn multilevel, abstract feature representations from large datasets. Key DL models include CNNs, RNNs, Transformers with attention mechanisms, GANs, autoencoders, and DRL models.

Currently, DL technology has driven major advancements in the fields of image, speech, and natural language. In the field of UAVs, DL models have been employed for detecting and tracking objects in images and videos [71], addressing the underlying task scheduling problem [72] and facilitating large-scale task processing. With the development of DL technology, many advanced object detection and tracking methods have been widely applied in various UAV-related tasks, such as environmental monitoring, precision agriculture, and traffic management [73]. For example, a DRL-based formation flight control algorithm [74] can reduce the collision rate of successfully formed UAVs to 3.4%.

2.2.4. Large Language Models

LLMs, a type of DL model characterized by their parameter size, are pretrained with a large amount of data. These models are capable of performing various natural language tasks, such as text generation, classification, and dialogue. Their architecture is heavily based on transformers, which employ self-attention mechanisms [75]. Additionally, these models support parallel processing of each word in a sentence, enhancing context awareness [76]. Notable LLMs include the generative pre-trained Transformer (GPT) series of models developed by OpenAI and the bidirectional encoder representations from transformers (BERT) and T5 models developed by Google. These models build upon the Transformer architecture to improve performance. Figure 5 presents a timeline that illustrates the primary development process of the LLM field by combining the significant milestones and representative models of the primary language models between 2017, when the Transformer architecture was proposed, and 2025.

Currently, advances in multimodal technologies have led to improved multimodal processing capabilities of LLMs, and their application in the UAV field is becoming increasingly prevalent. Compared with traditional AI methods, LLMs can process multimodal data such as images, radar, and text within the same framework and achieve reasonable task planning through pretraining knowledge, reducing response time by up to 62% [77].

2.3. Typical LLMs and Potential Applications

Currently, LLMs have powerful multimodal data processing and reasoning capabilities that enable them to efficiently parse large amounts of sensor, image, audio, and text data from UAV flights. These data enable UAVs to detect abnormal behaviors and potential threats instantly. For example, in scenarios such as border patrols and emergency disaster relief, UAVs can use these data to achieve target identification, abnormal behavior judgment, and automatic response to safety and security events, thus significantly improving the overall protection and risk control effectiveness [78].

LLMs can also be employed to accurately identify problems such as sensor spoofing and flight anomalies and quickly generate emergency response strategies in critical safety scenarios such as path planning and battery health monitoring, significantly improving the flexibility and safety of mission execution. In addition, owing to their long contextual processing and code generation capabilities, LLMs can integrate flight logs, laws, and regulations, as well as safety standard information, to conduct comprehensive safety audits, provide maintenance guidance, and generate safety training materials for UAVs. Moreover, when combined with cutting-edge technologies such as blockchain data verification, LLMs perform excellently in terms of antijamming, authentication, and cryptography. They can provide intelligent solutions for complex tasks such as UAV cluster collaborative security and cybersecurity monitoring, as shown in Figure 6.

For example, Heuser et al. [79] highlight AI-driven forest surveillance for wildfire management in Germany, aligning with LLM-enabled situational awareness in UAVs. Bücheler et al. [80] explored UAV-AI integration for automating construction workflows, reflecting safety-critical UAV applications. Morisaki et al. [81] introduced ChatDrone, a system controlling UAVs via LLM prompts without traditional vision models. Ikeyama et al. [82] demonstrated LLM-generated flight paths from UAV footage using natural language prompts, validated through indoor tests with a Tello UAV, while addressing both autonomy potential and latency limitations.

As different LLMs have different characteristics, their applicability in the field of UAV safety and security varies. Therefore, we selected eight representative LLMs, outlined their core features, and discussed their applications in the UAV safety and security domain, as shown in

Table 2. Key features of LLMs and their applications in UAV safety and security.

LLM	Key Features	Applications
		Safety	Security
Anthropic: Claude, 2023	Claude 3.5 Sonnet: 175B. Hybrid inference model, closed source, cloud deployment, security and ethical safeguards, long text processing.	Battery health monitoring, collision avoidance path planning, physically secure design, safety training material generation, legal advice and compliance checks.	Authentication system design, communication protocol optimization, cryptographic algorithm co-development, anomaly and intrusion detection, privacy protection plan formulation.
OpenAI: GPT, 2018	GPT-4o: 200B. Strong multimodal ability, closed source, cloud deployment, low hallucination rate, low latency.	Sensor spoofing recognition, collision avoidance path planning, flight control recovery, pilot simulation training, UAV cluster collaborative safety.	Anti-interference strategy optimization, identity recognition and authentication, perception data analysis, secure communication protocol optimization, zero-trust security architecture design.
xAI: Grok, 2024	Grok 1: 314B. Partially closed source, cloud deployment, X-platform access, real-time information updates, large computing power.	Battery health monitoring, compliance checks, environmental awareness, risk assessment, emergency response strategy generation, flight restricted area recognition.	Cryptographic algorithm development, anomaly and intrusion detection, radio spectrum security management, data privacy protection plan formulation, security situation prediction.
Google: Gemini, 2023	Gemini Gemma 3: 27B. Closed source, cloud deployment, native support for multimodal, low hallucination rate, strong inference capability.	Physically secure design, sensor spoofing recognition, collision avoidance path planning, environment perception, risk assessment, pilot simulation training.	Anti-jamming strategy optimization, radio spectrum security management, automatic security vulnerability scanning, encrypted communication simulation testing, security situation prediction.
Meta: LLaMA, 2023	LLaMA 3: 8B/70B/405B. Sparse attention mechanism, open source support, community adaptability, lightweight deployment, flexible customization capabilities.	Physical security design, security training material generation, component failure mode identification, flight path and mission planning, rational resource allocation, maintenance guidance.	Intrusion detection, data verification combined with blockchain, encrypted communication simulation testing, security policy evaluation, zero-trust security architecture design.
DeepSeek: DeepSeek, 2025	DeepSeek-V3: 671B. MoE models, open source, local deployment or via API, low cost and free to use, lightweight deployment.	Sensor spoofing recognition, flight loss recovery, flight parameter optimization, component failure identification, UAV cluster cooperative safety.	Cryptographic algorithms development, digital signature verification, radio spectrum security management, automatic security vulnerability scanning, big data security situation prediction.
Alibaba: Qwen, 2023	Qwen 2.5: 72B. MoE models, open source, local deployment support, multitasking capability, extensive knowledge coverage.	Legal compliance checking, flight parameter optimization, failure mode recognition of components, post-accident safety analysis, safety standard compliance validation.	UAV network security perception, digital signature and authentication, real-time security analysis and response, command fraud detection, edge computing security strategy evaluation.
Moonshot: Kimi, 2023	Kimi k1.5: 200B. Multimodal thinking model, closed source, cloud deployment, long text processing, enhanced content security.	Maintenance guidance, security training material generation, task planning and resource allocation, security audit and log analysis, safety standard compliance validation.	Abnormal behavior detection, automatic security vulnerability scanning, blockchain data validation, real-time security analysis, privacy protection scheme development.

.

2.3.1. Claude

Claude, developed by Anthropic, is a hybrid reasoning model built upon the Transformer architecture. This model is designed to process and generate natural language to achieve meaningful human–computer communication. The Claude 3 family currently includes multiple versions—Haiku, Opus, and Sonnet—with the latest version, Claude 3.7 Sonnet, expected to be released in February 2025. Claude 3.7 Sonnet is regarded as the most advanced version, with reports suggesting that it may support up to 1 trillion tokens. Claude offers both a normal dialogue mode and an extended inference mode, enabling in-depth reasoning before generating responses. This functionality helps users perform various tasks, including answering complex questions in various domains, providing information and explaining concepts, analyzing and summarizing information, and performing programming assistance.

In terms of safety and security, Anthropic has emphasized the importance of installing explicit values and behavioral principles in its AI models through the Constitutional AI training methodology [83]. This approach highlights how the exploration of the model’s safety mechanisms informs the design of UAV safety protocols. In addition, Claude has made significant improvements in multimodal data processing and complex reasoning tasks [84], which help analyze real-time image data captured by UAVs to detect safety and security threats such as unauthorized intrusion or equipment failure.

2.3.2. GPT

The GPT series of models, developed by OpenAI, represents a large-scale approach based on the Transformer architecture. These models are pretrained using large volumes of data and exhibit excellent natural language understanding and generation capabilities. The latest version, GPT-4.5, released on 27 February 2025 and internally referred to as Orion, is OpenAI’s largest model to date, with the number of references predicted to exceed 2 trillion. GPT models have been employed in tasks requiring high-quality text creation, programming assistance, information refinement, and complex question answering. These models also support deep reasoning and adaptive learning in multitask and multidomain scenarios, enhancing their reliability and adaptability in professional domain applications [85].

The technical advantages of GPT models have also supported new ideas for intelligent upgrades of UAV systems. Owing to their powerful multimodal data processing capabilities, GPT models can be seamlessly integrated with image recognition, sensor data analysis, and other modules to comprehensively parse the images and environmental data collected by UAVs in real time. Through intelligent monitoring and judgment of potential threats, GPT models can assist UAVs in reasonably planning flight routes, issuing timely warnings to avoid collisions, and providing reasonable response strategies [9].

2.3.3. Grok

Grok is an advanced AI language model developed by xAI that incorporates the most recent Transformer architecture. The latest version, Grok 3, released in February 2025, was trained using a large-scale data center equipped with approximately 200,000 graphics processing units (GPUs) and 10 times the computational power of its predecessor. Unlike other LLMs, Grok has been designed to reflect a sense of humor and a “rebellious spirit”. It has also been designed to answer more challenging questions that other AI models may refuse to answer.

With respect to UAV safety, Grok has been applied in analyses involving battery discharge patterns and thermal data gathered from sensors to predict potential failures and recommend optimal charging strategies to extend battery life [86]. In addition, its deep reasoning capability enables Grok to assess risks quickly and formulate reasonable safety strategies to ensure the stability and reliability of UAV operations in variable flight environments [77]. With respect to UAV security, Grok can monitor the communication links between UAVs and ground control centers in real time, identifying potential hacking or malware intrusion risks and safeguarding the integrity and confidentiality of data transmission [87].

2.3.4. Gemini

Gemini is a new generation of native multimodal models developed by Google. It builds on the Transformer framework and integrates advanced linguistic, visual, and decision-based capabilities with powerful cross-modal reasoning capabilities. The latest version, Gemini 2.0, released in December 2024, exhibits advanced reasoning and planning capabilities and is widely applied in everyday tasks and complex problem solving. In addition, the Gemma series, an open-source model based on Gemini technology, received an update in March 2025, with the release of Gemma 3. This latest version has been optimized for multimodal and long-context tasks, making it the most advanced and portable open-source model to date.

Vectara published a report introducing the “hallucination chart”, in which Gemini-2.0-Flash-001 ranked highest, achieving a low hallucination rate of 0.7% [88]. This outcome suggests that Gemini can be applied to generate a reasonable safety policy based on real-time environmental data, optimize the flight path and behavior of a UAV during missions, and assist in obstacle avoidance. Through powerful multimodal data fusion technology, Gemini can also process high-definition images, sensor data, and environmental information collected by UAVs in real time. Gemini can also identify sensor spoofing, provide early warnings [89], comprehensively monitor communication links and network security, and effectively enhance the integrity and reliability of data transmission.

2.3.5. LLaMA

LLaMA is a suite of open-source models developed by Meta, with LLaMA 3.3 representing the most recent release in December 2024. Owing to its open-source nature, LLaMA 3.3 supports flexible customization and secondary training by researchers and enterprises on a community-driven basis, resulting in an ecosystem of richly fine-tuned versions, including Vicuna and Alpaca. The multiparameter scale design from 1B to 405B [90] enables LLaMA 3.3 to meet lightweight deployment requirements, enabling deployment across a wide range of performance requirements in both local and cloud-based environments.

In terms of UAV safety and security, LLaMA 3.3 has excellent application potential. By pretraining with large amounts of high-quality multilingual data, this model can provide accurate and efficient support in the generation of safety training materials, the identification of component failure modes, task planning, and resource allocation. Moreover, the flexible customization capability and excellent code generation level of LLaMA 3.3 give it unique advantages in constructing specialized safety and security models. This model is reliable in critical tasks such as edge computing security policy assessment and zero-trust security architecture design, and it promotes the deep integration of the physical and cybersecurity domains.

2.3.6. DeepSeek

DeepSeek is a conversational robot launched by DeepSeek.corp, an AI startup with the same name, located in Hangzhou, China. The latest version of DeepSeek is DeepSeek-R1, which was officially released in January 2025. This version features a fully upgraded Transformer architecture that incorporates a mixture of experts (MoEs). In terms of performance, DeepSeek fully exploits the advantages of advanced algorithms to achieve fast response and accurate reasoning through efficient resource scheduling. Consistent with open-source principles, this version supports both local and cloud deployment, offering improved data processing efficiency while significantly reducing research and development costs. The training cost of DeepSeek-R1 is only $6 million [91], and the price of processing 1 million tokens is only 17% of that of GPT-4 Turbo. Therefore, this is suitable for small and medium-sized enterprises (SMEs) and individual developers to deploy.

Owing to its low cost and high reliability in the field of UAV safety, DeepSeek can be applied in areas such as flight path planning, real-time monitoring, and anomaly detection. Moreover, this model can be applied to intelligently analyze sensor data and image information, provide timely warnings of potential risks, assist UAVs in formulating reasonable risk avoidance measures, and safeguard critical areas. Furthermore, the open-source ecology of DeepSeek facilitates its continual adaptation to new scenarios, contributing to the advancement of UAV technology toward a new era of intelligence and autonomy.

2.3.7. Qwen

Qwen is an LLM developed by the Alibaba Group. The latest version of the series, Qwen2.5-Max, was officially released in January 2025. This model leverages the advantages of multiexpert parallel computing, with 72B activation parameters and a total of up to 320B parameters, to fully demonstrate the power of DL models in large-scale data processing. Qwen2.5-Max relies on mega-scale training, using 20 trillion tokens for pretraining. This approach equips the model with an extensive knowledge base and robust generalizability [92]. Moreover, the model is integrated with the Alibaba ecosystem, supporting local deployment and flexible access to cloud environments through Alibaba Cloud services. This integration meets the customization needs of developers and users.

Qwen’s multitasking capability supports flight loss recovery and parameter optimization, enabling autonomous planning and safe scheduling for UAVs. Moreover, Qwen can provide professional safety guidance in legal consulting and compliance checking, part failure mode identification, and post-accident safety analysis. It can also be applied to establish UAV security situational awareness and false information detection mechanisms to ensure the accurate implementation of edge computing security policy assessments and safety standard compliance verification, enhancing the overall safety of UAVs.

2.3.8. Kimi

Kimi, an intelligent AI assistant launched by Beijing Moonshot AI, is the world’s first intelligent assistant product to support the input of 2,000,000 Chinese characters [93]. Its latest version, Kimi k1.5, was released on 25 January 2025, and introduces a new multimodal reasoning model. This model is built upon the Transformer architecture, incorporates RL training methods, and primarily operates via a cloud-based application programming interface (API). Kimi k1.5 supports context windows of up to 128,000 tokens and can handle massive amounts of data and ultralong text tasks. Its free-to-use model overcomes traditional subscription limitations. This model also has a built-in content filtering and security auditing mechanism to ensure that the output content is compliant and reliable and that it supports URL input to read web content directly, greatly improving the efficiency of information extraction and text summarization.

Kimi’s powerful long text reading and summarizing capabilities can be applied during pretakeoff safety checks, training material generation, accident analysis, and safety audits. Moreover, the model supports anomalous behavior, intrusion detection, and automated security vulnerability scans. Kimi is also designed to aid in the formulation of data privacy protection schemes and verification of compliance with security standards, thus comprehensively improving the safety and security of UAV operations.

2.4. Lessons Learned

Through the analysis in Section 2, several key insights emerge regarding the evolving role of AI in UAV safety and security. First, the transition from traditional ML to DL, and now to LLMs, marks a significant shift not only in algorithmic complexity but also in the breadth of capabilities applicable to UAV systems. While ML and DL have advanced pattern recognition and fault detection, LLMs enable complex reasoning, multimodal fusion, and improved human–AI interactions. These capabilities are vital for dynamic UAV environments. Table 2 reveals that different models exhibit distinct strengths for UAV-specific tasks. Models like GPT and Gemini demonstrate strong performance in scenarios requiring rapid decision-making under uncertainty, thanks to their generative capabilities and low hallucination rates. For resource-constrained UAVs, open-source and lightweight models, such as LLaMA and DeepSeek, offer practical advantages in terms of cost, customization, and on-board deployment. We also observe emerging specialization among LLMs: Grok shows promise in managing energy resources and performing anomaly detection, while Qwen exhibits strong planning and scheduling capabilities. Claude and Kimi, with their emphasis on ethical alignment and security auditing, are well-suited for log analysis and privacy-preserving applications. These findings underscore the importance of scenario-aware model selection, where LLMs are matched to UAV tasks not solely by parameter size, but by architectural strengths and operational requirements. Meanwhile, we also observe that a single LLM cannot fully address all UAV safety and security needs. However, strategically deploying multiple models with diverse capabilities can greatly improve UAV resilience across various mission profiles.

3. AI for UAV Safety

Ensuring UAV safety is a fundamental requirement for their effective operation in complex environments. UAV safety includes a wide range of technologies to protect UAV systems from attacks, enhance UAV anti-interference capabilities, durability, and reliability, and ensure that UAVs can execute tasks safely and reliably. However, the large number of UAVs, their wide distribution, flexible networking methods, and complex parameters significantly increase the difficulty of ensuring their safety and protection.

The key safety technologies include physical safety design, battery safety management, sensor spoofing detection, collision avoidance, path planning, and loss of flight control protection. Physical safety design, which includes optimizing UAV impact resistance and structural stability through the development of high-performance materials, suppressing flutter, and predicting structural stress, is the basis of UAV safety. Battery safety management involves monitoring the condition of batteries, predicting their remaining power and health status, and combining fault diagnosis and safety protection measures to ensure reliability and safety. By securing the physical design and battery, defenses against sensor spoofing can prevent UAVs from deviating from their flight paths or crashing due to false information. When sensors operate normally, collision avoidance and path planning can be conducted to combine real-time obstacle detection with optimization algorithms to achieve dynamic obstacle avoidance and safe trajectory planning. Finally, loss of flight control protection aims to enhance the autonomous navigation capabilities of UAVs to prevent loss of control due to navigation or electronic fence failures. Together, these technologies facilitate the creation of a layered and comprehensive safety protection framework, ensuring the secure flight operations of UAVs and establishing a strong foundation for their extensive adoption.

3.1. UAV Physical Safety Design

Physical safety design-related technologies are fundamental in UAV safety. Material design and flutter design are two key elements of physical safety design. The development and application of high-performance composite materials can reduce aircraft weight and increase structural strength and damage resistance, thereby improving overall structural efficiency. For complex configurations of low-altitude equipment, flutter analysis and control technologies can effectively prevent structural damage or flight accidents caused by aeroelastic problems during flight, ensuring flight safety.

Traditional physical safety design models simplify nonlinear interactions and assume that mechanisms can be modeled using second-order linear ordinary differential equations. This approach may not fully capture all real-world influencing factors, reducing the model’s accuracy and affecting UAV flight safety [94]. In terms of material design, ML methods have been employed to accelerate the manufacturing process of composite materials to improve material performance [95] and strength [96]. Moreover, to address the perennial issue of insufficient datasets, Wang et al. proposed a transfer learning model based on a CGCNN pretrained with a low-precision PBE dataset and fine-tuned with a high-precision HSE06 dataset to predict the high-precision bandgap of crystal structures [97]. This approach overcomes the problems of low accuracy and overfitting caused by small datasets. In aircraft flutter prediction and suppression, DL models can identify flutter parameters in flight test signals, thereby facilitating improvements in analysis accuracy [94]. Additionally, RL techniques can be employed for active flutter suppression of flexible wings, enhancing flight stability [98]. In 2021, Zheng et al. applied neural networks to the active control of the nonlinear elastic response of aircraft, including flutter suppression, predictive control, and inverse model control [99]. These systems have been implemented to successfully suppress flutter in experiments, demonstrating the potential of neural networks in flutter suppression. In terms of transonic flutter management, Boeing [100] conducted Navier–Stokes simulations and limit cycle oscillation analysis. AI algorithms were applied to identify critical flutter points with an error of less than 5%, significantly outperforming traditional linear methods. AI algorithms have also been applied to predict aircraft structural stress and detect damage. For example, multilayer perceptron neural networks have been used to predict the stress field of wing rib structures, facilitating computational efficiency improvements [101]. Moreover, AI and acoustic emission optimization techniques have been integrated to detect impact damage in aircraft composite structures, enhancing the level of automation in damage detection [102].

However, there may be inconsistent data formats and incomplete records of experimental conditions in existing shared databases. For example, different institutions may use different testing standards or data storage formats, making data fusion difficult and limiting the effectiveness of training AI models. Moreover, given the ever-changing types of aircraft and flight conditions [103], existing models may lack detailed data on image resolution, sensor noise, control surface excitation signal amplitude variations [104], and different flight structural components [101], resulting in low adaptability and inaccurate control models over long-term operations. Moreover, neural networks often perform poorly in low-order modes and frequently exhibit oscillations, which can obscure the detection of flutter and result in inaccurate identification of flutter modes [105]. Additionally, the optimization of aircraft impact-resistant structures must simultaneously address constraints such as strength, weight, and cost, while traditional RL converges slowly in nonconvex spaces.

Currently, LLMs have shown versatility across various applications in physical safety design. A brief depiction of LLMs in physical safety design is presented in Figure 7. First, Berenguer et al. leveraged LLMs to collect and retrieve experimental data. These data were then transformed into a format that complies with the findable, accessible, interoperable, and reusable (FAIR) principles, providing word embedding representations [106]. This approach aids in subsequent exploration and semantic comparison. It can potentially be applied to the transformation of unstructured data, addressing the issue of nonuniform data formats in shared databases. On the other hand, optimization problems in physical safety design typically need to satisfy constraints such as strength, weight, and cost, which inevitably lead to nonconvex problems. Guo et al. combined gradient-based optimizers with LLMs for alternating optimization, utilizing the ability of LLMs to infer better solutions from natural language instructions [107]. This method can address the slow convergence of traditional RL in nonconvex spaces. Additionally, MatterChat [108], a multimodal LLM in material design, can integrate material structure data with text input, improving material property prediction and human–computer interaction performance. This approach holds promise for identifying more suitable high-performance composite materials for UAVs.

3.2. UAV Battery Safety Management

UAV battery safety management technology is a key component for ensuring UAV safety. A battery management system (BMS) is responsible for monitoring parameters such as the state of charge (SOC), state of health (SOH), and remaining useful life (RUL). Through charging control strategies, battery balancing strategies, and hybrid battery energy management strategies, BMSs optimize battery performance and predict battery performance and remaining flight time, as shown in Figure 8. Additionally, BMSs include safety measures such as overtemperature, overvoltage, and short-circuit protection to prevent flight accidents caused by battery failures.

Abnormal changes in battery temperature are the main factors leading to UAV battery degradation, performance decline, and even safety accidents. Traditional temperature management methods usually rely on thermal sensors and simple threshold settings, which cannot predict the risk of battery thermal runaway. DL neural networks can be applied to predict the evolution trend of battery temperatures on the basis of historical and real-time sensor data while considering factors such as the discharge rate, ambient temperature, and load changes. This approach leads to improved prediction accuracy and enables early prediction of potential overheating [109]. RL methods can also be applied to dynamically adjust the charging and discharging strategies of UAV batteries to avoid overheating issues, enhance temperature control management capabilities, monitor and identify abnormal battery parameters, and provide early warnings of battery failure [110,111]. Moreover, the limited endurance time of UAV batteries makes the accurate prediction of the remaining charge crucial. Traditional SOC prediction methods based on open-circuit voltage, internal battery resistance, or Coulomb counting suffer from error accumulation during long-term operation [112]. RNNs and Bayesian optimization models can combine multimodal sensor data such as battery temperature, current, voltage, and ambient humidity to predict the remaining charge, reduce error accumulation, and improve SOC prediction accuracy [113,114]. For UAV battery health status assessment, combining features such as temperature, internal resistance, and voltage curves allows DL or RL models to optimize battery management strategies, predict battery health status and remaining life, enhance the precision of battery management [86], and optimize battery charging and discharging strategies [110,115].

However, owing to confidentiality policies, access to actual operational data is strictly limited [109]. Additionally, battery manufacturers are reluctant to share battery health data [116], and generating data in the laboratory is time-consuming and costly. These factors collectively restrict the application of AI models in battery management, constraining AI model training and optimization due to limitations in data quality and quantity [117,118]. The lack of data indirectly leads to limited research on temperature [119] and charging rate [114] ranges in current papers. Models may not be stable under different temperatures, charging rates, battery ages, and battery types. In addition, they may face challenges in the presence of uncertainty and noise [119], making it difficult to generalize to unseen scenarios. Moreover, BMSs involve complex parameters and strong coupling. Existing AI tools still face challenges in predicting battery status and capturing the dynamic complexity of batteries [120], potentially overestimating model prediction accuracy. There are also potential issues with long-range spatial connections across multiple time scales [121]. Therefore, battery failure and in-flight crash risks cannot be avoided entirely [86]. AI models have poor interpretability, making it difficult for humans to understand the analysis and decision-making processes related to complex battery health factors and their interactions.

Currently, LLMs have demonstrated significant potential for applications in battery safety management. A Transformer-based LLM was introduced to estimate the SOH and RUL of lithium-titanate oxide batteries [122]. This model can identify signs of battery degradation through anomaly detection, enabling predictive maintenance to avoid sudden battery failure, improve energy efficiency, and ensure UAV flight safety. Moreover, GPT4Battery and LLM-based electric vehicle battery internet health management methods leverage the strong generalizability of LLMs to estimate the SOH across different types of lithium-ion batteries adaptively [123]. Even with fine-tuning small-scale data, these methods can predict future battery states such as capacity, internal resistance, and temperature [124], reducing the time and resource consumption compared with those of traditional battery degradation experiments. These improvements facilitate early detection of potential failures, such as battery aging or overheating, and adaptability to the changing distribution of batteries over time, ensuring accurate estimation at the end of battery life. These approaches provide efficient and flexible solutions for UAV battery management.

3.3. UAV Sensor Spoofing Detection

Detecting sensor spoofing is a fundamental step in ensuring UAV safety after physical security design and battery safety. UAV sensor systems, such as GPS, IMUs, and vision sensors, are essential for modern applications but are susceptible to spoofing and jamming attacks. Hostile forces or malicious users can fake or disrupt signals, causing UAVs to deviate from their intended course, miss their targets, or crash, jeopardizing mission success and personnel safety.

GPS spoofing is a common sensor spoofing attack, as illustrated in Figure 9. Traditional GPS spoofing detection and identification schemes rely on signal strength detection, frequency analysis, and theoretical relationships between sensors [125]. Nevertheless, in complex environments, a single sensor may not provide a sufficient amount of reliable data and may fail because of inherent sensor defects. DL-based algorithms with CNNs or LSTM can be applied to improve detection accuracy [126]. RL and DRL models can optimize sensor fusion strategies by effectively combining data from multiple sensors such as accelerometers, gyroscopes, magnetometers, GPSs, and barometers [127,128] to improve detection accuracy. Additionally, combining neural networks with Kalman filtering can adaptively enhance the defensive abilities of a UAV positioning system against spoofing signals [129]. These DL-based algorithms effectively detect GPS spoofing by analyzing features such as complex signal patterns and position deviations [22] or by comparing the differences between theoretical path and actual path losses [130]. Moreover, these algorithms can be combined with malicious process detection techniques [130] to enable real-time monitoring of the safety posture and threat scenarios of UAV cluster end devices [131]. By incorporating transfer learning and DL, existing knowledge and data can be effectively leveraged to improve model generalizability and adaptability [132], improve detection and defensive abilities in practical applications, and facilitate the safe operation of UAVs. Moreover, signal features such as jitter [133], Doppler frequency, and carrier-to-noise density ratios have been incorporated into model training [125] to detect spoofed GPS signals more effectively.

However, GPS signals are transmitted unencrypted and unauthenticated, increasing their vulnerability and sensitivity to adversarial attacks [134]. Currently, publicly available datasets of GPS signal spoofing attacks are scarce, costly to acquire [128,132], difficult to collect [133], cover only a few types of sensor data [127] and spoofing scenarios [125], have difficulty adequately representing the relationships among data [135], and still have a large space for generation and completion, which limits the size and quality of datasets [132,136]. In the future, databases will be needed for different types of spoofing devices in multipath environments and various flights. In addition, a database needs to be built for different types of spoofers in multipath environments and various flight conditions to obtain sufficient data and improve model performance [137]. Moreover, the standards and accuracy of UAV flight trajectory classification require enhancement. During the annotation process, existing trajectory validation methods are limited to binary classification, categorizing trajectories as correct and incorrect. Therefore, some incorrect trajectories may be labeled as correct [138]. Algorithm performance will be further degraded in the face of overly complex or nonfixed flight routes [139], and even if the best model can achieve 95% accuracy, false alarms, missed alarms, and omissions may occur, affecting the detection and identification of GPS spoofing [140].

For sensor data spoofing, an Aero-LLM framework was introduced [141] to deploy different types of LLMs in onboard systems, edge servers, and the cloud to achieve dynamic task allocation and cooperative work. In terms of sensor spoofing detection, the TimesNet model has been applied to preprocess sensor data, such as GPS coordinate data, enabling real-time detection of anomalies. Through fine-tuning, it achieves high accuracy, precision, and recall while maintaining robustness against variance changes in the data, effectively strengthening the defense of a UAV against sensor spoofing attacks. Moreover, in the field of anomaly detection, in 2024, Li et al. proposed a serialization method to convert batch-level anomaly detection from a numerical task to a textual task and used models such as GPT, LLaMA 2, and Mistral to conduct experiments with both synthetic and real data [142]. They demonstrated that GPT-3.5 and GPT-4 are capable of detecting hidden anomalies in data batches without additional distribution-specific model fitting. This discovery offers a novel approach for applying LLMs to sensor data anomalies.

3.4. UAV Collision Avoidance and Path Planning

UAV collision avoidance is a challenging problem that becomes even more difficult in dynamic environments with multiple UAV deployments and dynamic obstacles detected by sensors. Collision avoidance systems range from simple warnings provided to UAV controllers for manual intervention to partially or fully autonomous control systems to prevent collisions. Collision avoidance techniques can be categorized into two parts: dynamic obstacle detection and path planning. Global path planning focuses on determining the optimal route while considering the entire environment. In contrast, local path planning is performed locally in response to environmental changes, and collision avoidance algorithms are executed accordingly [143].

In dynamic obstacle detection and tracking, UAVs fly autonomously in complex and changing environments. Therefore, accurately sensing and avoiding dynamic obstacles is the key to ensuring their safe operation, as illustrated in Figure 10. However, traditional techniques have limitations in dealing with dynamic environments and moving obstacles [57,144,145], especially in complex urban environments. In these environments, UAVs have a limited understanding of complex environments, making it challenging to recognize and track dynamic obstacles accurately. Lightweight DL architectures such as the YOLO-Fastest [146] and RGB-D [147] real-time detection frameworks with multidetector fusion can fully leverage the depth information and color image data provided by RGB-D cameras, improving detection accuracy and real-time performance. Moreover, by combining RGB-D cameras with Kalman filtering [146,148] and decentralized gradient optimization [145], the system can more accurately estimate the position and speed of obstacles, thus improving the detection accuracy of small and high-speed obstacles and realizing clustered collision avoidance. Particle maps [149] simultaneously represent static and dynamic obstacles, enabling the construction of a dynamic map model. Thus, UAVs have stronger autonomous navigation and obstacle avoidance capabilities in complex environments, providing strong support for their widespread adoption. In UAV path planning, the MADER framework [150] is adopted to generate dynamic, safe trajectories through polyhedral collision constraints and MINVO basis functions to avoid collisions effectively. B-spline gradient optimization [151] combines obstacle prediction trajectories and significantly improves the real-time performance of path planning by optimizing the smoothing of trajectories and dynamic adaptability. RL approaches have made notable progress in UAV path planning. RRT* [152] accelerates online path searches through offline training and can quickly adapt to dynamic environment changes.

Many algorithms have demonstrated strong performance in rural or indoor environments [119]. Most methods consider motion blur [146] and can avoid fast-moving spheres. Nevertheless, they do not consider the effect of lighting conditions on sensing. Therefore, the direct application of these algorithms in large-scale complex environments such as those with different lighting conditions is difficult, resulting in poor algorithm generalizability [153]. Moreover, tracking may be lost when obstacles obstruct the camera’s line of sight [146], enter blind zones [148], or appear suddenly [154], or when the screen is overexposed. While dynamic maps provide high-precision environmental information, they also encounter challenges in computational efficiency [149]. In addition, most of the current AI processing solutions lack a mechanism to address hardware failures [154], which may lead to complete system paralysis in the event of hardware failure, seriously affecting the flight safety of UAVs.

LLMs have been innovatively applied in UAV collision avoidance and path planning. In terms of dynamic obstacle avoidance, in 2024, Zhong et al. proposed a vision-based autonomous planning system to predict the trajectories of dynamic obstacles and integrate them with LLMs [155]. LLMs can interpret natural language instructions, such as “avoid the construction area”, and recognize semantic labels of dynamic obstacles, such as ambulances being given priority in terms of avoidance, thereby enriching environmental modeling. Moreover, adjusting the planning logic through natural language, such as dynamically modifying collision avoidance rules, not only reduces the cost of algorithm retraining but also reduces the dependence of operators on programming languages, greatly increasing the autonomy of UAVs and user interactivity. In terms of path planning, LEVIOSA converts natural language input into 3D waypoints to generate executable flight paths for UAV swarms. This allows UAVs to follow accurate flight paths in trajectory generation, synchronization, and collision avoidance. This functionality fosters more intuitive human–machine interactions and advanced multi-UAV coordination [13].

3.5. UAV Flight Control

During mission execution, UAVs occasionally experience navigation failures. In such events, they autonomously search for and rely on nearby UAVs as leaders to guide them back to safe areas. However, a leader UAV may operate improperly or be illegally controlled by attackers, such as being manipulated to deviate from its intended flight path, and chain reactions may be triggered. These reactions include mid-air collisions, loss of control, and crashes, resulting in property damage and casualties. On the other hand, UAVs at low altitudes typically depend on electronic fence technology to define their flight areas and ensure that they fly within authorized airspace. However, suppose that an electronic fence is compromised by illegal attacks or malicious tampering. In this case, the aircraft may lose control and enter into no-fly zones, hazardous areas, military regions, and other sensitive locations. This intrusion significantly increases the risk of collisions and may also trigger safety accidents, posing a significant threat to life and property.

In terms of electronic fence defense and anomaly detection, in 2023, He et al. investigated the use of 5G technology for remote UAV control and addressed issues such as loss of control and crashes through aerial electronic fences [156]. An illustration of a UAV electric fence for a prison is presented in Figure 11. In the case of out-of-control failure, a UAV can return automatically and safely via the electronic fence integrated with GPS-based navigation cloud platforms. These systems support functions such as remote real-time control, data transmission, and image transmission, significantly enhancing UAV safety and extending the flight radius [156]. Regarding UAV flight control, Shen et al. explored the use of RL to control UAV quadrotors, training them in AirSim. By incorporating YOLOv7-tiny for object detection, the system unifies simulation and real-world inputs, reaching a 52% traversal rate in tests. This integration provides valuable frameworks for RL reward mechanisms and object detection in UAV navigation [157]. In 2022, autonomous indoor UAV flight control was performed via monocular camera images, enabling UAVs to predict movement directions, correct positioning, and adjust heading simultaneously. This model was successfully deployed on a real UAV to navigate corridors and make decisions at junctions. A custom dataset supports training and validation, demonstrating the potential of multitask learning for UAV navigation [158]. In 2023, Bohn et al. presented a data-efficient DRL approach for the attitude control of fixed-wing UAVs, requiring only three minutes of flight data. The performance of the learned controller is comparable to that of the ArduPlane PID controller [159]. Transferability was improved by incorporating sim-to-real measures such as domain randomization and actuation latency. The study also provides insights into RL controller behavior through linear analysis, highlighting its potential for enhancing UAV capabilities.

However, in complex electromagnetic and physical environments, such as those involving high-frequency signal interference and malicious lead UAV yaw attacks, electronic fence defense systems can be severely challenged [160], and collaborative positioning algorithms may fail. Moreover, when the physical environment is complex, the model size grows exponentially with the number of unknown cells in the environment, greatly limiting the scalability of existing models. To address this, some authors have proposed that future UAV sampling trajectory algorithms should be combined with crowd-sourcing strategies or RL approaches based on graph attention multi-agents or link feedback to improve the efficiency of UAV collaboration.

For traffic incidents such as UAV crashes, in 2024, Ahmed et al. proposed a multimodal AI-based traffic incident response system [161]. This system combines YOLOv11s for real-time traffic incident detection, uses the visual language model Moondream2 to generate scene descriptions, and leverages LLMs to generate brief accident reports and action recommendations. Its multimodal approach and application of LLMs can provide insights into the detection of and response to UAV flight failures. For example, in practical application scenarios, UAVs can detect abnormal situations through image detection and employ LLMs to generate emergency response suggestions, minimize the probability of triggering safety incidents, and rapidly and accurately respond after they occur to protect life and property as much as possible.

3.6. Lessons Learned

Section 3 provides an investigation of the significant progress of AI-driven UAV safety applications and discusses the limitations of traditional methods.

Table 3. Summary of different AI methods in UAV safety.

Field	Ref.	Contributor	Year	Methods	Main Point
Physical Safety Design	[96]	State Key Laboratory of Nonferrous Metals and Processes	2023	ML	An AdaBoost Regressor-based model predicts Al-Li alloys’ specific modulus and achieves a 4% increase in maximum specific modulus over the dataset, with Alloy 3 exceeding widely-used 2195-T8 by 12.6% while maintaining similar specific strength.
	[99]	Northwestern Polytechnical University	2021	DL	A flutter test signal feature extraction method combining EMD and CNN; achieves 100% accuracy on test datasets with fewer training iterations and lower computational complexity.
	[162]	Northwestern Polytechnical University	2025	DL	A BPNN-based system identification technology; achieves less error for various wing models, improving design efficiency and safety in the preliminary stage.
	[108]	Lawrence Berkeley National Laboratory	2025	LLM	A multimodal LLM for materials science; achieves 0.93 accuracy in materials property prediction and 0.09 RMSE in energy above hull tasks.
	[107]	Harbin Institute of Technology	2024	LLM	A collaborative optimization method combining gradient descent and LLMs for prompt tuning; improves performance in NLP and vision-language tasks over traditional methods.
	[106]	University of Alicante	2024	LLM	A sensor data retrieval method using LLMs converts data to FAIR-compliant formats and creates word embeddings; achieves 0.90 precision and 0.94 MRR.
Battery Safety Management	[121]	University of California	2023	ML	An improved LSTM boosts battery state estimation, and a DDPG-based RL method reduces battery failure rate.
	[86]	Qatar University	2023	DL	A UAV battery management system uses DNN and LSTM for SOC prediction with an MSE of 7.6 $\times {10}^{-4}$ and an EV score of 0.98, as well as RF for SOH estimation with 92% accuracy.
	[109]	Jilin University	2024	DL	TinyML-based neural network models estimate battery SOC, maintain high precision in early stages, and reduce computational load.
	[123]	Xiamen University	2024	LLM	An LLM-based framework for cross-battery state estimation; attains a 2.17% MAE in zero-shot settings, and surpasses latest domain adaptation methods on some datasets.
	[124]	Central South University	2025	LLM	An LLM-based EV battery health management method; achieves 0.31 MAE, 0.23 RMSE, and 0.0063 CRPS.
	[122]	Purdue University	2025	LLM	A Transformer-based LLM framework; identifies early degradation via anomaly detection, reduces MAE to 0.87%, and supports predictive maintenance of EVs.
Sensor Spoofing Detection	[127]	Xidian University	2022	ML	A UAV GPS spoofing detection method based on sensor data and ML; achieves a detection rate of 99.69%, outperforming existing methods like JSA in accuracy, precision, recall, and F1 score.
	[125]	Xidian University	2023	ML	A UAV GPS spoofing detection algorithm using signal features and ML; achieves a detection rate of 94.87%, an EER of about 5%, and a detection cycle of only 0.4 s.
	[134]	King Abdulaziz University	2024	DL	An intrusion detection system (IDS) based on adversarial ML model is proposed to detect UAV GPS spoofing attacks. Through adversarial training, the model achieves an average accuracy of 98%.
	[163]	Beirut Arab University	2025	DL	A framework combining a genetic algorithm-optimized LSTM network enhances UAV GPS spoofing detection, improves classification accuracy from 86.0% to 99.0%, and improves the F1 score from 83.0% to 99.0%, boosting UAV adaptability and anti-spoofing in dynamic scenarios.
	[142]	UC Irvine	2024	LLM	A method using LLMs for tabular data anomaly detection converts numerical data to text and fine-tunes end-to-end performance, and improves AUROC by 6.7 and 8.9 with LLaMA and Mistral after fine-tuning.
	[141]	Oakland University	2025	LLM	An LLM uses a distributed architecture across onboard, edge, and cloud for high precision and efficiency; excels in key metrics while maintaining low memory usage, offering robust network threat defense for UAV operations.
Collision Avoidance and Path Planning	[148]	Carnegie Mellon University	2024	ML	A lightweight dynamic obstacle detection and tracking method using RGB-D cameras; integrates detection and feature association tracking, achieves 0.11 m position error and 0.23 m/s velocity error, outperforming benchmarks in dynamic navigation.
	[152]	University of Illinois	2024	DL	NIRRT*, combining point cloud and neural networks, improves path planning efficiency and convergence.
	[146]	The University of Hong Kong	2022	DL	Gradient-based trajectory planning ensures safety and energy efficiency; achieves 98% success in dynamic obstacle avoidance in simulations and real-world tests.
	[13]	Florida Institute of Technology	2024	LLM	LEVIOSA leverages LLMs to convert natural language commands into UAV swarm flight paths, where GPT-4o achieves the highest average success rate of 76.0% in generating complex trajectories.
	[155]	Wuhan University	2024	LLM	A vision-based UAV planning system; explores LLMs to enhance user–UAV interaction; achieves a maximum pedestrian velocity of 1.7 m/s without collisions and a detection success rate of 80% with four pedestrians.
Flight Control	[157]	National Chung Hsing University	2024	ML	A PPO with CRM reward mechanism achieves a 71% UAV target traversal rate in simulations and a 52% target-crossing rate in a physical environment.
	[158]	National Defense Academy of Japan	2022	DL	A multitask DL model with CNN achieves over 0.91 accuracy in position prediction and good performance in orientation prediction with fewer errors in real-flight tests.
	[164]	Northeastern University	2024	ML	A Bayesian RL-based navigation strategy; improves casualty location efficiency and reduces negative entropy.
	[161]	Jeonbuk National University	2024	LLM	A multimodal framework uses YOLOv11s for real-time accident detection, Moondream2 for scene description, and GPT 4-Turbo for emergency suggestions; achieves 94.7% accuracy, 86.8% recall, and 92.8% mAP@0.5.

presents the research progress of traditional AI methods in different safety fields, categorizing different AI methods. ML and DL techniques have been widely applied in UAV safety, particularly in relatively stable and well-defined scenarios. Traditional AI models can optimize material design, structural stability, and battery management strategies, improve GPS spoofing detection accuracy, obstacle avoidance ability, and path planning efficiency, and achieve safe flight control. These papers not only highlight the potential of AI technology to improve the safety performance of UAVs but also highlight the limitations of current methods in terms of dataset size, model generalizability, accuracy, and adaptability.

Table 4. Limitations of traditional AI models in UAV safety.

	Physical Safety Design	Battery Safety Management	Sensor Spoofing Detection	Collision Avoidance and Path Planning	Flight Control
Lack of Datasets	✓	✓	✓	✓
Generalizability Issues	✓	✓	✓	✓	✓
Limited Accuracy	✓	✓	✓		✓
Sensitivity to Adversarial Attacks			✓		✓
Scalability Issues			✓		✓
Low Adaptability			✓
Lack of Hardware Solution				✓
Low Interpretability		✓
Unstandardized and Highly Diverse Data	✓
Low Optimization Search Efficiency	✓			✓

outlines the limitations of the traditional AI models in UAV safety. With respect to physical safety design, traditional models simplify nonlinear interactions and cannot fully capture the key factors influencing UAV flight safety. In the field of battery safety management, the difficulty of data acquisition leads to unstable model performance under different conditions. Scalability is more prominent in sensor spoofing. In the fields of collision avoidance and path planning, traditional AI models have difficulty dealing with sudden obstacles in dynamic environments and lack a hardware fault response mechanism. With respect to flight control, traditional methods lack sufficient adaptability in complex electromagnetic environments or under malicious attacks. These limitations indicate that although traditional AI models have made some progress in the field of UAV safety, they still need to be improved to meet the increasingly complex UAV safety requirements.

To mitigate the limitations of traditional ML and DL models, as shown in Figure 12, LLMs offer more promising application potential in enhancing UAV safety, particularly in more complex and dynamic scenarios, than traditional ML and DL models do. To be more specific, on the software side, LLMs serve as high-level planners, translating natural language instructions into actionable flight tasks, such as generating 3D waypoints for trajectory planning or coordinating multi-UAV formations [13]. They also process sensor data for real-time decision-making, enabling anomaly detection and secure communication in distributed frameworks [141]. On the hardware side, LLMs are deployed across onboard computers, edge servers with high-performance GPUs, or cloud-based servers, depending on computational demands. For instance, complex tasks like generating emergency response reports in traffic incident scenarios leverage cloud resources [161], while onboard systems handle lighter language processing. LLMs, unlike traditional ML models, offer superior contextual understanding by processing sensor data through optimized frameworks to enhance UAV autonomy [155]. This flexible integration of LLMs with UAV hardware and software significantly boosts autonomy, safety, and user interaction, revolutionizing intelligent aerial systems.

First, despite the limited airborne processing capacity and energy of UAVs, applying LLMs to identify complex objects and dynamic paths poses significant challenges, especially in intricate and rapidly shifting environments. For example, in dense urban settings with moving vehicles and fluctuating obstacles, unpredictable wilderness areas with variable terrain and weather, or contested zones with adversarial UAVs and jamming, LLMs must process real-time data to detect objects such as birds or debris and compute adaptive flight paths. However, advancements in model compression technologies such as pruning, quantization, and knowledge distillation have led to high-performance lightweight LLMs such as DeepSeek-R1 and Qwen. When paired with AI accelerators such as NVIDIA Jetson TX2, these models can be deployed locally on UAVs, efficiently handling such dynamic scenarios. Second, while traditional AI models, such as battery health prediction systems, rely on large, high-quality datasets for training, LLMs can extract general feature representations from broad data and fine-tune with minimal specific data, reducing data needs. This adaptability allows them to function effectively in various changing environments, such as fluctuating battery performance at extreme temperatures or altitudes. Moreover, to address the interpretability issues of earlier AI models, modern LLMs use a chain-of-thought approach, breaking down problems into clear, step-by-step reasoning. This approach facilitates the conversion of implicit calculations into explicit, human-like natural language sequences that are verifiable at each step. By transforming black-box processes into transparent logic chains, LLMs support deployment in complex, dynamic environments, enhancing UAV safety.

4. AI for UAV Security

When performing tasks, UAVs must communicate frequently with ground control stations, other UAVs, and external networks, which encounter various security threats, such as sensed data leakage, radio interference, identity authentication failure, non-encryption, and UAV SON vulnerabilities. If these threats are maliciously exploited, they could lead to mission failure and jeopardize personnel safety, compromise privacy, and even pose risks to national security.

This chapter explores various aspects of AI techniques for UAV communication security, including sensing security, radio anti-interference, identity authentication, cryptography, and UAV SON security technologies. Sensing security ensures that the data collected by UAV sensors are accurate and reliable, forming the foundation for all subsequent security measures. Radio anti-interference helps maintain the integrity of communications by preventing eavesdropping and ensuring that the transmitted data reach the destination intact. Once the data transmission is secured, identity authentication protocols verify the identities of the communicating parties, preventing unauthorized access. Cryptographic techniques further enhance security by encrypting the data, ensuring confidentiality and integrity. Finally, SON security addresses the dynamic nature of UAV networks, protecting against advanced threats such as routing and node-based attacks and ensuring overall security. We analyze existing technologies in these areas and their current state of application in UAV communications and explore how AI techniques can empower these technologies to increase the overall security of UAV systems. By systematically elaborating these contents, we discuss the potential and application prospects of combining LLMs with key UAV security technologies.

4.1. UAV Sensing Security Technology

Sensing security technology integrates communication and sensing functions, allowing systems to detect communication signals and continuously track UAVs while collaboratively monitoring their flight status. This approach provides an effective way to regulate UAVs. With the integration of communication and sensing functionalities, air interface resources and core network resources are shared between communication and sensing services, making it essential to enhance the security of sensing data and services alongside existing communication security measures.

In 2016, Fasano et al. detailed UAV sense-and-avoid technology to enhance security by detecting other aircraft and obstacles and automatically performing evasive maneuvers [166]. In 2018, Chen et al. discussed the applications of DL models in UAV visual sensing, especially in obstacle detection and monocular visual navigation in forest environments [167]. In 2020, He et al. proposed a task-driven method of autonomous sensing and fusion based on UAV swarms that uses the distributed information fusion capability of UAV swarms to achieve autonomous sensing and state estimation of the environment [168]. In 2023, Kurunathan et al. discussed ML applications in UAV operations and communications, including key components such as sensing and feature extraction, feature interpretation and regeneration, trajectory and mission planning, and aerodynamic control and operation [69]. In addition, UAV cooperative sensing technology enhances overall sensing capability and task execution efficiency through information sharing and collaboration among multiple UAVs.

However, AI training approaches rely on sensing data, such as from radar, LiDAR, and camera data. However, the accuracy and reliability of these sensors directly affect the sensing results. Additionally, in UAV clustering or air–ground coordination missions, there are difficulties in fusing sensing data between different UAVs or unmanned vehicles, especially in the case of view orthogonality, communication latency, and packet data loss.

LLMs can support the calibration of sensor data and compensate for the shortcomings of sensors in certain dimensions by learning from a large amount of data and mining the correlations of spatial dimensions. In addition, LLMs contribute to enhanced sensing accuracy and robustness through the fusion of multimodal data and simultaneously help UAVs analyze and process multimodal data. LLMs can aid in optimizing cooperative sensing strategies for UAV clusters. In particular, the multisensor cooperation strategy, driven by distributed intelligence and multimachine collaboration, improves UAV sensing security, achieving more efficient environment sensing and mission execution, as shown in Figure 13.

4.2. UAV Radio Anti-Interference

The open nature of the wireless communication channels in UAV systems increases their vulnerability to external interference signals [169]. For example, an illustration of a jamming attack is presented in Figure 14. In low-altitude communication scenarios, on the one hand, it is necessary to ensure the security of data transmission through wireless channels and reduce the risk of interference. On the other hand, ensuring the security and reliability of the data transmitted between the network elements and the platform system is necessary. Smart antenna technology is a revolutionary advancement in wireless communication, significantly improving the anti-interference capabilities of UAVs by adjusting the weighting value of the antenna elements to form a directional beam via AI algorithms.

Dong et al. proposed a 3D robust beamforming method for supporting UAVs via integrated sensing and communication (ISAC) technology and introduced a dynamic beamforming method for UAV-enabled vehicle-mounted networks [170]. In 2024, Pang et al. proposed a collaborative and secure relay communication method for UAV clusters, which uses distributed antenna arrays and beamforming techniques to resist the collusion of eavesdroppers, effectively improving the security of data transmission [171]. Abdallah et al. discussed how to enhance the defense capabilities of UAVs through smart antenna systems and ML algorithms. Smart antenna systems provide directional gain and interference suppression, and ML algorithms identify and mitigate spoofing signals efficiently [163]. Goudos et al. studied the application of smart antennas in UAV communication, optimizing the power model of UAVs by integrating AI algorithms to improve their communication efficiency and security in complex environments [172].

However, antennas designed by traditional AI algorithms may lead to uneven signal propagation, affecting the communication quality of UAVs. UAV communication requires the real-time adjustment of beam direction and transmission strategy, and smart antenna technology may produce ineffective continuous signal transmission, affecting the range time.

LLMs can address complex multiconditional optimization problems and contribute better design schemes by comprehensively considering the frequency, bandwidth, size, material, and other parameters. Through the prediction capabilities of LLMs [173], the sleep and wake-up strategies of UAVs can be optimized to reduce unnecessary communication and energy consumption while ensuring security. In addition, in multi-UAV scenarios, LLMs can coordinate the beam direction and transmission strategies of multiple UAVs through distributed optimization algorithms.

4.3. UAV Identity Authentication Security

UAV authentication technology is an important means of ensuring the legitimate use of UAVs, preventing illegal intrusion, and safeguarding UAV security. A simplified illustration of the authentication procedure is presented in Figure 15. In addition to traditional methods, such as authentication and key agreement (AKA) used in mobile communication networks, the introduction of UAV service provider-based authentication and trust transfer-based authentication can be considered.

In 2019, Goudos et al. proposed an authentication scheme based on a PUF and fuzzy extractor between UAVs and ground stations [172]. This method considers the influence of hardware noise and environmental changes on authentication and enhances the flexibility and security of authentication at the same time. Mrabet et al. explored the application of ML algorithms in UAV communication identity authentication [174]. In 2020, Allahham et al. proposed a method for UAV detection and identification that is based on DL technology to analyze and classify the RF signal characteristics of UAVs for identity authentication [175]. In 2021, Rubina et al. proposed a CNN-based method to achieve real-time detection and authentication of UAVs, improving the efficiency of identity authentication and reducing the false alarm rate [176]. Li et al. analyzed the communication signal features of UAVs by applying a DL model to verify the identity of UAV pilots, thus helping prevent unauthorized user manipulation [177].

However, UAV authentication involves handling a significant volume of sensitive data, such as biometrics and communication records, which could be vulnerable to leakage or misuse. In addition, traditional AI methods have limited accuracy in processing multimodal data such as images, sounds, and communication signals that need to be fused, affecting UAV authentication accuracy. Moreover, UAV communication environments are complex and changing, and it may be difficult for traditional AI models to adapt quickly to new authentication scenarios or attack patterns.

In contrast, LLMs combined with privacy-preserving methods, such as federated learning [178], and multimodal learning methods, such as CLIP and Florence [179], can enable efficient fusion and feature extraction of multimodal data, improving UAV authentication accuracy. By transferring knowledge pretrained with large amounts of general datasets to UAV authentication scenarios, LLMs reduce their reliance on domain-specific data. The models’ online learning capabilities also allow dynamic adjustments to authentication strategies, adapting seamlessly to environmental changes.

4.4. UAV Cryptography

Cryptography is one of the key technologies for securing UAV communication and protecting the confidentiality, integrity, and authenticity of communication data. However, with the rapid development of computing technology, traditional cryptography algorithms face the risk of cracking. To address this challenge, researchers have continued to explore new cryptographic techniques and methods.

Wu et al. reviewed the application of lightweight cryptographic algorithms in IoT systems, including UAV communication [180]. In 2018, Ozmen et al. proposed Dronecrypt, an efficient encryption framework designed for small UAVs [181] that significantly reduces energy consumption while maintaining high security. In 2024, Saini et al. proposed a novel method that employs an autoencoder neural network trained using the MNIST dataset to generate unique cryptographic keys [182]. These keys are integrated with traditional encryption algorithms to enhance data security during transmission. Alshammari et al. proposed a DL-based hybrid arithmetic optimization algorithm with min–max normalization, feature selection, and threat detection for hyperparameter tuning to increase security and performance [183].

However, traditional AI models have limited generalizability, making it difficult for them to handle new cryptoanalysis attacks [184]. Additionally, compatibility issues in integrating traditional AI and asymmetric encryption systems and the lack of unified AI-assisted encryption system evaluation standards [185] are two major problems. The dynamic changes in IoUAV and the impact of equipment resource constraints on algorithmic selection significantly increase the complexity of encryption algorithms, as shown in Figure 16.

LLMs possess powerful data analysis capabilities that allow them to thoroughly evaluate existing asymmetric encryption algorithms, identify potential loopholes and weaknesses, and optimize the selection of encryption algorithm parameters. Additionally, LLMs can develop specialized benchmark tests for cryptographic fundamentals, encrypted traffic detection, and so on. These benchmark tests can provide a unified evaluation standard for integrating AI and cryptography methods. LLMs can also provide a natural language interface to lower the threshold of using complex cryptosystems [182], assisting in the design of lightweight and highly secure cryptographic solutions.

4.5. UAV Self-Organizing Network Communications Security

UAV SON communication treats UAVs as nodes to achieve autonomous networking, dynamic routing, and data transmission between nodes through wireless communication technology. A representative structure of a UAV SON is presented in Figure 17. In networking, UAV communication protocols aim to support secure and reliable communication between UAVs, ground control stations, and other UAVs. However, the open nature of UAV wireless channels makes them susceptible to security threats such as eavesdropping, tampering, and DoS attacks. In addition, UAVs typically operate under constrained resources and require lightweight, real-time, and highly secure communication protocols to meet their special needs.

In 2020, Chen et al. proposed a self-organized mapping technique that enables UAV arrays to dynamically adjust their topology according to users’ demands and form a hierarchical network, improving the efficiency and security of UAV SON [186]. In 2024, Yaacoub et al. proposed a secure communication scheme for UAV FANETs, which includes three phases: UAV registration, UAV and ground station authentication, and inter-UAV authentication [187]. In disaster response and environmental monitoring, Liu et al. deployed a dynamic communication network using AI-driven UAVs to optimize resource allocation, collect and analyze data, and realize dynamic decision-making [188]. With respect to routing protocols, secure strategies have been explored to protect UAV cluster networks. Mohamad et al. [189] proposed a secure routing protocol named PASER to address UAV routing attack challenges in wireless multihop networks (WMNs).

However, traditional methods often encounter issues with limited bandwidth, especially when multiple UAVs transmit data at the same time. UAVs often experience dynamic topology changes, but most existing papers are based on ideal environments [190,191], leading to an increase in the routing interruption rate. For diverse and extreme environments such as urban, mountainous, and marine environments, traditional AI methods exhibit weak generalizability [190,192] and lack cross-scene migration capabilities.

However, LLMs offer potential solutions for optimizing the topological relationships and bandwidth distribution of UAV networks via optimization algorithms such as the improved NSGA-II algorithm. These models can be combined with virtual reality simulation technology for multimodal data fusion [191], achieving a smooth transition from simulation to real scenarios. When combined with the LLM-based migration learning pretraining cross-scenario generic policy [192], the routing disruption rate in extreme environments, such as heavy rain, is reduced. In addition, LLMs can predict link quality by capturing long-distance dependencies [193,194], improving the accuracy of routing prediction and further reducing the routing disruption rate. With adaptive mesh techniques, LLMs can be applied to generate adaptive routing table update strategies [194] on the basis of real-time physical characteristics, such as airflow velocity and pressure changes during flight.

4.6. Lessons Learned

Section 4 presents a comprehensive investigation of the significant advances in AI technologies in the field of UAV security applications while also discussing the serious limitations of traditional AI approaches.

Table 5. Summary of different AI methods in UAV security.

Field	Ref.	Contributor	Year	Methods	Main Point
Sensing Security Technology	[195]	Ministry of Natural Resources of China	2020	DL	A building recognition method; achieves a recognition accuracy of 93.2% and an average processing time of 74 ms per image.
	[196]	Toingji University	2022	DL	A DL-based ISAC approach; achieves a location accuracy of scatterers at approximately 1 m.
	[164]	University of Cambridge	2023	ML	A comprehensive survey on ML-assisted UAV operations and communications, covering a wide range of applications and techniques.
	[10]	Jeonbuk Tongji University	2024	LLM	A comprehensive survey on integrating LLMs with UAVs, achieving enhanced spectral sensing capabilities and improved decision-making efficiency.
Radio Interference	[197]	Space Engineering University	2022	ML	A game-theoretic learning anti-jamming approach; achieves enhanced communication reliability, efficient resource utilization, and effective countermeasures against intelligent jammers.
	[198]	JiLin University	2024	ML	A UAV swarm-enabled collaborative secure relay communication strategy; achieves a higher secure sum rate and lower energy consumption.
	[199]	University of Electronic Science and Technology of China	2025	DL	A RL-based anti-jamming frequency hopping strategy; achieves a higher SINR and faster learning speed compared to those of traditional methods.
	[155]	Wuhan University	2024	LLM	A vision-based autonomous planning system; achieves a planning cycle of approximately 200 milliseconds.
Dentity Authentication	[175]	Qatar University	2020	DL	A UAV detection and identification method; achieves a detection accuracy of 100%, a type identification accuracy of 94.6%, and a state identification accuracy of 87.4%.
	[176]	Kumoh National Institute of Technology	2021	DL	A CNN-based UAV identification algorithm; achieves a detection rate of 94.5%.
	[177]	Northwestern Polytechnical University	2023	DL	A novel UAV pilot authentication scheme; achieves an authentication accuracy of 95.24% and a malicious hijacking detection accuracy of 96.82%.
	[174]	Prince Sattam bin Abdulaziz University	2024	ML	A comprehensive review of ML algorithms for UAV detection and classification, covering radar, acoustic, visual, RF, and multisensor systems.
	[200]	Northwestern Polytechnical University	2024	ML	A multitask learning-based UAV pilot identification and operation evaluation scheme named MTL-PIE; achieves an identification accuracy of 95.36%, an operation evaluation accuracy of 94.47%, and a processing time of only 35 milliseconds.
	[10]	Toingji University	2024	LLM	A comprehensive review on integrating LLMs with UAVs; supports advanced UAV identity recognition and verification processes.
Cryptography	[182]	Guru Gobind Singh Indraprastha University	2024	ML	A UAV data security enhancement framework; achieves a significant reduction in data tampering risks and efficient data validation cycles.
	[183]	King Abdulaziz University	2024	DL	A hybrid DL-based arithmetic optimization algorithm; achieves high accuracy in threat detection and efficient performance in securing communications.
	[201]	Beihang Universityy	2024	LLM	A federated learning framework for LLMs using encryption and model splitting; achieves secure training with comparable performance to that of centralized models and robust defense against gradient-based attacks.
Self-Organizing Network Communications	[202]	Beijing University of Posts and Telecommunications	2022	DL	A topology-aware resilient routing protocol using adaptive Q-learning; achieves a 25.23% lower overhead, a 9.41% higher packet delivery rate (PDR), and a 5.12% lower energy consumption compared to existing methods.
	[190]	WuHan Maritime Communication Research Institute	2024	ML	An optimized AODV-based routing algorithm; achieves a successful packet delivery rate of above 80%.
	[9]	JiLin University	2024	LLM	A UAV networking security framework using generative AI and DMs; achieves a 25% enhancement in classification accuracy for anomaly detection and robust privacy preservation, with over 10% higher communication rates.
	[193]	MSA University	2024	LLM	A LLM-DaaS framework; achieves near-perfect accuracy in converting free-text user requests into structured UAV operation tasks, significantly enhancing operational efficiency and adaptability in uncertain environments.

summarizes key research advances across different security areas and categorizes AI methods. Traditional AI methods, such as ML and DL, are primarily effective in relatively stable and well-defined UAV security scenarios. These methods contribute to enhanced sensing accuracy by analyzing data from physical sensors, improving radio anti-interference ability, and increasing the accuracy of UAV identification. Additionally, they support the protection of UAV confidentiality and communication security within the SON. These studies highlight the potential of AI technology to improve UAV security while also highlighting the limitations of current methods in terms of dataset size, model generalizability, and the difficulty of fusing heterogeneous data.

Table 6. Limitations of traditional AI models in UAV security.

	Sensing Security Technology	Radio Interference	Identity Authentication	Cryptography	Self-Organizing Network Communications
Lack of Datasets	✓		✓		✓
Generalizability Issues	✓	✓	✓	✓	✓
Limited Accuracy	✓	✓	✓
Sensitivitiy to Adversarial Attacks			✓	✓	✓
Lack of Uniform Standards	✓		✓		✓
Physical and Hardware Limitations	✓	✓
Low Interpretability				✓
Unstandardized and Highly Diverse Data	✓
Limited Resources		✓	✓	✓
Cost and Performance Trade-Off	✓	✓		✓
Lack of Encryption Mechanism				✓	✓

summarizes the limitations of traditional AI models in the field of UAV security. Sensing security still relies too much on physical sensors, and multisensor data are heterogeneous and difficult to fuse. In radio anti-interference, traditional methods have poor generalizability and lack security in the case of new types of attacks. In the field of UAV identification, the problem of maintaining accuracy in complex and changing dynamic environments is more prominent. In terms of cryptography, there is a lack of agreed-upon protocols or standards. UAV SON security relies on sensitive datasets, and it is difficult to guarantee that private data will not be leaked. These limitations indicate that although traditional AI models are well-suited for basic UAV security tasks, they fall short in more complex and dynamic UAV security requirements.

To mitigate the limitations of traditional ML and DL models, LLMs offer promising application potential in UAV security. They are designed to process and analyze heterogeneous data from multiple sources, enabling more comprehensive security threat detection and response. By leveraging their ability to learn from large and diverse datasets, LLMs can adapt to new attacks and provide robust solutions. Additionally, LLMs exhibit real-time decision-making capabilities and employ adaptive security measures, making them ideal for advanced UAV security applications where traditional AI falls short. To be more specific, on the software side, edge-–cloud collaboration is implemented using lightweight frameworks such as TensorFlow Lite and PyTorch Mobile for local data processing, minimizing latency and dependence on cloud resources [203], whereas a real-time operating system optimizes resource management, improves reliability, and reduces the risk of interference [204]. For hardware deployment, sensor fusion with integrated cameras, using LIDAR and millimeter wave radar, provides multimodal environment sensing [155], while embedded GPUs accelerate AI inference to meet real-time demands [205]. Secure communications via 5G and satellite modules enable encrypted data transfer and multi-drone coordination, while a hybrid network architecture balances local edge processing with cloud-based LLMs [206]. LLMs process sensor data through optimized edge frameworks or API-driven cloud interactions, leveraging hardware accelerators like NVIDIA Jetson GPUs for real-time performance. Unlike traditional DL models, which are less adaptable to complex scenarios, LLMs enable dynamic decision-making and natural language processing for enhanced security [155]. Additionally, open-source flight control systems, such as ArduPilot and GymFC, are integrated with AI frameworks to enable autonomous functionality, augmented by security measures [155].

In this context, the development of LLMs provides new opportunities in the field of UAV security, as shown in Figure 18. Owing to their powerful data processing and intelligent decision-making abilities, LLMs can effectively respond to complex security problems in UAV communication, particularly in dynamic and unpredictable scenarios. For example, in environments with fluctuating network conditions, such as urban areas with dense signal interference, remote regions with sparse connectivity, or contested airspace with active jamming, LLMs leverage their extraordinary learning capabilities to achieve a more accurate identity authentication mechanism, resisting forgery and counterfeiting attacks by analyzing real-time communication patterns and adapting to evolving threats. Similarly, using their ability to analyze signal characteristics, LLMs can effectively identify and counteract radio interference in scenarios such as high-traffic air corridors or electronic warfare zones, ensuring communication link stability despite rapid signal noise or jamming intensity changes. LLMs can also facilitate the optimization of communication protocols to enhance the security and reliability of UAV SON, adapting to shifting topologies or sudden node failures, such as those caused by weather disruptions or hardware malfunctions. In addition, when combined with cryptography techniques, LLMs can strengthen encryption algorithms to resist potential threats from emerging technologies, ensuring robust protection, even in future-facing, technologically volatile contexts.

5. Challenges and Future Directions

Despite their impressive semantic understanding, text generation, and mathematical reasoning capabilities, the application of LLMs to UAV safety and security is still in its early stages. Currently, LLMs have not yet reached the low-latency requirements for real-time reasoning to achieve real-time decision-making for UAVs. Additionally, there are some adaptation problems in integrating LLMs with UAV systems. Therefore, this chapter aims to explore the challenges associated with applying LLMs to UAV safety and security and suggests some possible future research directions, as shown in Figure 19.

5.1. Challenges

As LLM applications in UAVs continue to advance, several challenges related to safety and security have become increasingly prominent. Communication latency and resource constraints limit timely system responses in complex scenarios. Moreover, a model’s lack of sufficient domain knowledge limits its versatility and poses a certain risk of privacy disclosure. Therefore, this section discusses the main challenges associated with integrating LLMs and UAVs in terms of safety and security.

5.1.1. Real-Time Processing Latency

The onboard memory of UAVs is typically limited. When high-resolution images and multi-layer Transformer models are executed locally, frequent data reads and writes can cause bandwidth bottlenecks, significantly increasing inference latency [207]. This latency is particularly significant in scenarios such as autonomous UAV obstacle avoidance, target tracking, or response to sudden weather changes, making it difficult to meet the real-time demands of navigation, tactical response, and decision-making [10]. Although current edge computing technologies can significantly reduce the inference latency of LLMs by means of model lightweighting, computational offloading, and hardware acceleration, the performance of lightweight models is often insufficient for complex problems, leading to severe inference latency [208]. For example, a lightweight YOLOv8n model on the NVIDIA Jetson Orin Nano achieves only about 10 FPS after optimization, well below the 30 FPS required for real-time navigation [209]. Additionally, offloading inference tasks to the cloud can cause end-to-end latency exceeding 150 ms. This delay allows a UAV to travel several meters before a decision is made, missing the optimal window for obstacle avoidance [210]. Moreover, in a multi-UAV cooperative task, the difference in reasoning delays of different nodes can cause coordination failure and increase the risk of collision, which seriously affects the safety and efficiency of the task.

5.1.2. Computing Resource Limitations

LLMs often contain billions to hundreds of billions of parameters and require extremely high memory capacity for full-precision deployment, e.g., about 28 GB RAM for a 7 B model [211]. However, UAV systems are subject to stringent weight, power, and thermal management constraints and are often unable to carry high-performance computing devices [212]; even with 8-bit quantization, model memory requirements may still exceed the hardware ceiling. Thus, UAVs typically require edge computing architectures to offload tasks and alleviate computing power and energy consumption bottlenecks [86]. In mission-critical applications, if LLMs cannot provide reasonable security policies in a timely manner, this may lead to problems such as processing timeouts, system crashes, or excessive battery depletion [213]. In the future, improvements can be considered through pruning, quantization, model distillation, and edge computing. However, the memory and power constraints of UAVs in edge computing force edge reasoning to still require a compromise between arithmetic power and endurance. When running LLM tasks, UAV endurance is often cut by nearly 50%, and insufficient power will significantly increase flight risk [10].

5.1.3. Privacy Leakage Risk

With the integration of UAVs with LLMs, the risk of data privacy leakage is also a major issue that cannot be ignored [214]. UAVs often need to collect a large amount of sensitive information, such as geographic location, real-time images, and environmental parameters, while performing missions. These data risk being illegally intercepted, stolen, or tampered with during collection, transmission, storage, and processing sessions [215]. This risk becomes more pronounced when these data are used in LLM training or inference because of the complexity of the internal processing of some models and the lack of transparent monitoring. For example, in military reconnaissance, if an attacker steals the critical data collected by UAVs, it may lead to the disclosure of military secrets and even threaten national security.

5.1.4. Poor Extreme Environmental Adaptation

LLMs must be fine-tuned with targeted domain data to meet the requirements of mission execution. However, the available data can hardly cover all the complex environments that UAVs may encounter, leading to models that exhibit low alignment with real-world conditions [216] and perform poorly in practical applications. For example, in harsh conditions such as strong winds, heavy rain, or extreme cold, sensor data can be disrupted [217]. This may prevent LLMs from quickly adapting to environmental noise, electromagnetic interference, or hardware failures, leading to misjudgments of the environment or incorrect commands. In addition, owing to the lack of high-quality training data, LLMs may also experience some hallucinations, such as fictitious safety procedures or nonexistent navigation reference points [218].

5.1.5. Multi-UAV Collaboration Problems

In a multi-UAV system, if each UAV is equipped with LLMs to achieve autonomous decision-making, respecting communication and cluster collaboration among them is still difficult. Each UAV must share its real-time status, environmental data, and threat information over a fast, secure communication network. However, differences in data formats, communication protocols, and interface standards across devices [219] can easily lead to compatibility issues and data distortion. In addition, under sudden security threats, multi-UAV clusters need to respond quickly and coherently to adjust routes and mission assignments to ensure overall system stability [220]. These actions require strong autonomous decision-making skills while addressing hardware variations and communication latency to ensure safety and security.

5.1.6. Network Quality Susceptibility

In real-time missions, UAVs depend heavily on data transmission. However, poor network conditions can disrupt stable system performance [221]. For example, in areas with severe signal blockage or interference, UAVs may experience communication disruption, resulting in critical data not being uploaded or received on time. Moreover, LLMs with networked search functions can falter when network delays or congestion block instant access to external data, slowing the decision-making process [222]. In addition, network instability is prone to data loss or misdelivery, making it difficult to maintain a consistent response to UAV group coordination and, thus, affecting the overall system performance.

5.2. Future Directions

This section presents possible future research directions, considering the challenges and other considerations discussed earlier. It provides some ideas on how to address the current challenges and highlights the cutting-edge research areas that need attention in LLMs and UAV safety and security. Such explorations are essential to overcome current limitations and unlock the full potential of LLM application in UAV safety and security and even other related fields.

5.2.1. Multimodal Embodied Intelligence System Integration

Multimodal data fusion studies aim to integrate visual, speech, text, and other sensor data to achieve collaborative data processing and feature alignment through DL and cross-modal algorithms [223]. Multimodal embodied intelligent systems can perceive and make decisions about complex environments and have a wide range of application prospects in UAVs, autonomous driving, and robotics. For example, Chen et al. proposed that embedded multimodal large-scale models have great potential in bridging the gaps between sensing, cognition, and action in complex real-world environments, with a focus on embodied sensing navigation, interaction, and simulation [224]. In 2024, Hu et al. proposed a multimodal UAV autonomous exploration system that utilizes DRL to synergize spectral and LiDAR data [225]. This approach significantly enhances the navigation ability of UAVs in complex environments and improves the success rate of obstacle avoidance. In addition, integrating multimodal data requires greater computational effort and places higher demands on the edge deployment of LLMs, so research into quantitative acceleration techniques such as model pruning and edge computing is critical to reducing computational requirements [17].

5.2.2. Satellite Network Communication Enhancement

With the increasing demand for global network coverage, integrating satellite networks and UAV communication systems has become an important means of solving the problem of insufficient signals in remote or disaster areas. By utilizing the global coverage capability of satellite networks, UAVs can achieve stable data transmission and real-time command and scheduling, even in areas not supported by terrestrial base stations, significantly improving the robustness of the communication module [226]. Furthermore, by deploying multi-access edge computing nodes, it is possible to perform local preprocessing, intelligent caching, and rapid decision-making regarding sensor data, thereby further reducing round-trip latency. In 2021, Mao et al. used UAVs and satellites to provide edge computing and cloud computing services for wirelessly powered IoT devices, respectively, and employed a DL-based offloading and optimization strategy to address the dynamics of energy harvesting performance [227]. In 2024, Javaid et al. investigated the integration of LLMs into a comprehensive network of satellites, antennas, and terrestrial networks to provide reliable and seamless communication by improving data flow, signal processing, and network management, enhancing 6G communication technology [228]. Dong et al. also proposed an innovative architecture based on the zero-trust principle to support a blockchain-enabled UAV distribution system through multi-agent situational awareness in LLMs, thus realizing a zero-trust air–ground integrated network [229].

5.2.3. 6G Native Intelligent Communication Network

The 6G native intelligent communication network can help LLMs in UAV safety and security with its smart architecture. For safety, the high speed and low latency of this network enable LLMs to process real-time data for adaptive obstacle avoidance and path planning in complex settings. Its smart edge computing ensures instant decisions, whereas predictive analytics enhances fault prevention. In terms of security, the intelligent protocols and dynamic encryption of 6G protect LLM-UAV communications, adapting to threats instantly. Smart resource allocation and cross-domain coordination ensure reliable, secure connectivity across diverse missions. This intelligent 6G framework empowers LLMs to improve UAV safety and security with exceptional adaptability and efficiency. In 2023, Dwivedi et al. provided an overview of the synergistic integration of AI methods, UAVs, and 6G networks, with a focus on predictive network analysis, autonomous fault detection and self-healing, and dynamic resource allocation [230]. Raja et al. proposed an efficient 6G federated learning energy-efficient scheme for UAV deployments, which improves the data transfer rate while protecting data sensitivity and achieving very low latency [231].

5.2.4. ISAC Empowers Digital Low-Altitude Networks

In UAV networks, ISAC fuses high-precision sensor data and high-speed communication links to realize real-time data sharing and dynamic collaboration among UAVs, providing new technical support for their safety and security [232]. Moreover, when integrated with the deep semantic parsing and multimodal data processing capabilities of LLMs, the ISAC system can not only quickly identify environmental threats and abnormal states but also make accurate decisions on the basis of historical data and real-time information and optimize mission scheduling and resource allocation, thus providing solid data and decision-making support for UAVs in the fields of urban management, logistics and distribution, and emergency rescue. For example, in 2024, Yang et al. investigated the application of embedded artificial intelligence (EAI) in the integration of sensing, communication, computation, and control and proposed an EAI-enabled ISC3 framework for the low-altitude economy [233]. In 2025, Li et al. studied a UAV network with an ISAC system, where multiple UAVs simultaneously determine the location of ground users and provide communication services via radar. This approach optimizes UAV deployment and communication with sensing data by optimizing trajectories to avoid collisions, strengthening UAV safety and security [234].

5.2.5. Low-Altitude Semantic Communication

In UAV networks, low-altitude semantic communication is employed to extract the core semantics of information, significantly reducing the transmission load and latency with the help of the deep semantic parsing capabilities of LLMs. By utilizing multimodal data processing capabilities, LLMs can accurately capture environmental changes and update semantic information for collaborative decision-making between UAVs in real time, ensuring more efficient command and control and resource allocation in complex and variable low-altitude airspace. In 2025, Fan et al. proposed a UAV-based SC framework incorporating a DM, balancing between the performance and sampling steps and, thus, ensuring efficient computation and high-quality image enhancement [235]. Hu et al. investigated an intelligent resource allocation method for multi-UAV-assisted semantic sensing in UAV image sensing task scenarios. This approach involves jointly optimizing UAV trajectories, spectrum bandwidth, transmit power, and the number of semantic symbols transmitted to maximize the quality of experience while minimizing transmission costs [236]. By integrating blockchain technology, it is also possible to provide privacy protection for data access within UAV systems, thus reducing the risk of exposure of LLM training data [237].

6. Conclusions

In this paper, we have focused on AI applications in UAV safety and security, highlighting the significant potential of LLMs as a leading advancement in the AI field. These models have paved the way for new developments in UAV hardware and communication security. Specifically, we have provided a detailed evaluation of the traditional AI algorithms currently used to address UAV safety and security issues, analyzing the limitations that still exist in their deployment. Moreover, we have comprehensively analyzed the characteristics of different LLMs, assessed their applicability in solving UAV safety and security problems, and surveyed how LLMs can fully utilize the integration advantages of LLMs-UAVs to be applied and deployed in various fields of UAV safety and security. Thus, this work lays the foundation for ensuring UAV safety and security. Finally, we have outlined the limitations in deploying LLMs and the key areas that require further research to fully exploit the potential of AI. We envision a future where there is full integration and joint action between LLMs and various enabling technologies for UAVs. This integration and advancement will ensure UAV safety and security and lay the foundation for their extensive, stable, and secure deployment, thus transforming fields such as precision agriculture, disaster response, and urban planning in the future.