The safety protocols for drones encompass a variety of layers and categories, contingent on their intended application, dimensions, and control methods. Generally, drones make use of a communication protocol [26] based on IEEE 802.11, as outlined in [27] . By incorporating both the ground stations and WiFi networks, drones lack encryption technologies, rendering them susceptible to intrusions and potential hijacking [7] . The man-in-the-middle (MITM)-based incursions, with an effective area range of up to 2 km, represent a prevalent method for compromising drones [28] . In the realm of military applications, the IoD has gained increasing prominence, introducing challenges related to privacy and security considerations in the design phase [29] . Safeguarding data requires addressing privacy concerns including data accessibility, information disclosure, and ciphering and deciphering techniques [30] . Recent investigations have categorized security threats into four distinct groups encompassing sensor threats, malign interference with signals, and jeopardized modules. A thorough overview of these categories can be found in Table 1 .

Table 1 provides a literature overview primarily focusing on vulnerability detection in drones and scarce remedial discussions. One research avenue explored the implementation of encryption algorithms for a bidirectional secure data transmission of drones and their stations [29] . The popularity of miniature drones has grown due to their compactness, simultaneously eliciting potential concerns of threats to privacy [39] . Additionally, challenges related to drone security have been thrown light upon by valuable research [26] [40] - [43] . As an example, an effective model assessment proposed by Tian et al. for the IoD is supported by edge computing to enhance data security in drone networks [44] . Similarly, a robust system to secure drone data in both commercial and industrial settings has been proposed by [5] . In 2019, [6] proposed a novel concept that involved gas leakage-sensing drones to ensure timely responses in critical scenarios. Drones are predominantly employed for monitoring purposes in fields such as agriculture and security. In the past decade, security threats concerning drones have become a prominent topic in research circles. The privacy concerns in pertinence to smart

city-based applications are thoroughly addressed in [32] , while Table 1 sheds light on other crucial concerns. Researchers in the realm of cyber-security have shown particular interest in exploring network incursions on drones, with their possible limitations [45] . Commercial sectors are faced with similar challenges and opportunities, as demonstrated in related studies [8] [46] [47] , leveraging blockchain and cryptographic technologies, combined with 5G and IoT-based drones, to ensure secure data transmission [47] . The current manual identification of threat intensity and nature has its limitations, emphasizing the pressing need for a secure and intelligent drone system capable of both investigating attacks and implementing measures to safeguard the integrity of drone data. Some studies endeavored to resolve authentication issues by effectively employing both the agreement [48] and key-enabled data [9] yielding a reliable propagation of drone data. The drones in the commercial sector [48] face a common challenge, which is the potential jeopardizing of drones, UAVs, and drones used in agriculture, facilitated by IoT whose effective solutions are weighted in [10] [11] . Similarly, other relevant reservations to UAVs pertain to tracking via GPS [49] , necessitating impenetrable and trustworthy solutions. The interception and unauthorized access also feature prominently in studies [50] .

Many researchers have extensively applied ML models in various domains, including cloud-based computing, networks, and systems with IoT integrations [51] - [54] , to combat cyber attacks. To illustrate this, [55] synergized self-learning models and supervised learning, utilizing a combination of long short-term memory (LSTM), auto-encoder and random forest (RF) classifiers to pinpoint distributed denial of services (DDoS) attacks. Their approach yielded notable results, achieving accuracy up to 94% for scenarios involving varying traffic levels. The most promising outcomes were the results of their proposed LSTM-RF hybrid model, which employed a window size of 100.

Existing research did not exploit the unparalleled potential of ML models in drone networks for the identification of cyber attacks. Nevertheless, an alternative study recommended the probabilistic viewpoint within a circumscribed cyber-physical system for effective management and detection of actuation attacks [56] . The respective research predominantly focused on the PA2 attack, characterized by the interruption of communication between actuators and the controller. A collection of parallel detectors was introduced in this study, oriented around a hypothesis-testing methodology. For coping with uncertainties, certain objective functions were formulated tailored around detection and control, via a probabilistic take. Additionally, a security access control system was put forth for drones and previously applied ML techniques to enhance network anonymity (wi-net).

Existing studies signify the pressing requirement for a comprehensive resolution to confront threats related to cyber-security whilst ensuring the protection of data relayed through drones. While challenges and reservations pertinent to confidentiality have been put to light, only a limited number have put forth viable strategies to reduce these risks [47] . Though ML has displayed potential in identifying intrusive attempts at various networks, the gaps within the research persist within the network-based applications. Moreover, authentication systems presented in studies might not line up sufficiently for drone networks. Hence, bridging the rift is imperative to render drones compliant with industry and commercial standards, all the while upholding their security and safeguarding their privacy.

Securing drones mandates the adoption of an ingenious surveillance infrastructure that autonomously scrutinizes compromised data and executes context-specific corrective measures, all while eliminating the necessity for manual intervention. Although ML models have shown their mettle in fortifying mobile and sensor-based networks against cyber threats, their potential in fortifying drone-based vehicles remains a largely untrodden path. This study takes on the intricate challenge of redefining access control and authentication techniques for drones through the introduction of an innovative ML-driven approach.

Existing blueprints for commonly employed smart [16] , as delineated in Figure 1 , adhere to a structural hierarchy. To bolster its confidentiality based attributes, an additional stage is incorporated, whereas certain modifications are introduced in the data-processing layer by replacing it with machine intelligence-embedded alternatives.

The data layer stands as a pioneering stage within the intricate framework of industrial drones, primarily because of housing a camera-toting miniature drone or nimble quadcopter. This stratum comes to life through an influx of IoT sensor data, where an ensemble of cutting-edge sensors, encompassing cameras, GPS, and radar, takes the helm. Their collective prowess empowers the drone to not only sense and seize data but also orchestrate their seamless transmission to the ensuing layer in the drone’s data journey.

In the subsequent layer, a UAS drone takes on the pivotal roles of data acquisition and drone flight management. This UAS drone is equipped with a remote interaction module alongside a ground controller. A practical example of this design and architecture concept can be found in (DJIS Phantom-3 drones) originating from (Shenzhen, China). Drones like these have a communication interface with a remote controller for controlled maneuvers. The architecture also accommodates the addition of auxiliary sensors as needed.

Unprocessed raw data is sourced from the IoT devices, via the intermediate edge processing drone layer (EPL) and relayed to the privacy layer (PL). PL plays a pivotal role in data authentication and monitoring transmission to the cloud layer (CL). For cloud communication, the Azure IoT gateway is utilized in conjunction with the IoT gateway.

The drone’s confidentiality layer focuses on authenticating devices and controlling access using ML models. This layer is responsible for enforcing data safety and security, both integral components of the framework. Herein, the detection and mitigation of potential privacy threats take place, as they present a substantial risk to the system. Some examples of privacy threats that can manifest at this level include:

1) A physical incursion hazard signifying unsanctioned access or meddling with physical gadgets sensors or drone apparatus.

2) A user-centric behavioral privacy challenge is intricately entwined with the acquisition of personal data, achieved by meticulously observing and documenting user interactions.

3) An unsanctioned geospatial privacy encroachment, centering on the tracking or revelation of an individual’s whereabouts against their wishes.

Addressing security vulnerabilities necessitates the utilization of a versatile array of authentication techniques and protocols malign actors if left unchecked can leverage an assortment of breaches to launch threats encompassing a diverse range of incursions that can yield havoc on transmissions, control, and functionality. Within the proposed framework deep learning-based algorithms come to the fore for device authentication with early detection of potentially unauthorized penetration. This equips users with the capability to receive timely alerts and adopt preemptive measures to intervene in the wake of these looming threats.

Within the connection layer, the IoT gateways play a pivotal role in establishing connections between the base stations and a cloud IoT hub. Enhanced security, orchestration, and automation are guaranteed by introducing an additional module, permitting connectivity exclusively for authenticated devices. This hub serves as a transmission channel, by effectively facilitating communication at both ends i.e., cloud and network. Here, measures are implemented to grant network access solely to authenticated devices. Data from sensing devices originating off of networks and drones is securely transmitted to the client utilizing both robust crypto-graphic and blockchain technology, ensuring data integrity and safe-keeping.

For analysis, the data is routed to the processing layer, where novel modules are introduced inclusive of machine intelligence and hub services. Instead of relying on individual ML algorithms, it employs a weighted ensemble of CNN models, offering unparalleled adaptability across diverse scenarios and data prerequisites. The crux of this research lies in the application of an intelligent ML approach for device authentication. It is the IoT hub layer’s responsibility to authenticate devices by meticulously time-stamping data emanating from drones at predetermined intervals. These datasets, derived from drone flights, form the bedrock for model training and testing. The model undergoes rigorous training, followed by meticulous testing to assess its efficacy in detecting any ill-activity. If detected, the alerts are generated and the services are suspended with the cloud, for abrupt interventions.

A myriad of security threats loom over flight operations, and among these, interception attacks, where an external entity wrests power over the device, are frequent. Similar perils reside in the realm of unauthorized access, wherein malicious actors manipulate drones to disseminate misleading information, underscoring the multifaceted landscape of security risks in drone operations. Herein, a voting of CNN models is employed in fashion a model which is subsequently utilized for validating fresh aircraft trajectories. For effective evaluation against the metrics, the KDD-CUP 99 dataset was used.

In the drone data storage layer, analyzed data is methodically archived within data centers. This repository encompasses a range of results generated by drones, including drone-specific data extracted via sensors and network data in a cloud-based NoSQL database, which resides within the drone layer. Storage of this valuable data takes place.

NoSQL, signified for multi-faceted data storages, allows for quick and easy retrieval of data. Their proficiency in managing substantial volumes cast them as a preferred choice over conventional databases. Given their inherent automatic indexing, common data structures found in non-relational repositories encapsulate the graph-oriented key-value document-based and columnar formats.

The layer responsible for analyzing drone data brings forth an array of resources and utilities for proficient data observation. This platform harnesses MS Azure Services—the outcomes derived from the Data Visualization Language (DVL) are presented through a mobile application showcasing the forecasts generated by the intelligent model. Concerning a drone’s security level (SL), the Naive Bayes model is skillfully utilized for the detection of drone attacks. The business intelligence structure by stream analytics (SA) disseminates results through the MS Power-BI application and is securely contained in dedicated centers.

The experiments utilized accessible, cost-effective peripherals. For sensing, an Arduino Mega 2560 microcontroller sourced from Somerville, USA, connected with a WiFi module ESP8266 was procured for processing.

Drone selection offers a plethora of size and shape choices, and their deployment depends largely on the use cases. In this particular experiment, the Phantom 3 Standard by DJI, took center stage. It was operated with a custom wireless controller, enabling remote functionalities such as tracking, locating, and identifying objects. These tasks were accomplished using radar sensors that use electromagnetic energy transmission with the demonstration of the superior accuracy of radar sensors over optical ones. HC-SR04 proximity sensors are used in this paper as part of the experimentation. Because of the proficient pattern identification, this sensor offers a multifaceted approach to the task.

The GY-GPS6MV2, crafted by UBLOX, headquartered in Zurcherstrasse, Switzerland, is a GPS signal receiver featuring an onboard NEO-6M chip. It activates an LED indicator, powered by a battery when transmitting or receiving GPS data. Notably, this module showcases a remarkable sensitivity down to −161 dBm.

The sensor, BMP180 stands out in terms of energy efficiency delivering pinpoint accuracy in elevation measurements tailored to specific locations. Its exceptional precision defies its compact size and in terms of altitude and pressure sensors, this OEM module outshines its counterparts. The distinctive communication traits offered by ZigBee, notably the specific ability to transmit varying signal types contributed to its broad usage.

As technology develops, airborne systems such as UAVs continue to provide a wide range of advantages. But as other studies have noted, they also have several drawbacks and raise issues with safety, security, and privacy. To reduce these worries and control unauthorized aerial photography, laws and licensing procedures must be put in place. Globally, authorities have enforced strict regulations to tackle the problem of unapproved aerial photography.

When assessing network security and conducting risk analysis, it is important to recognize that UAVs have distinctive characteristics. Distinctions primarily arise from resource constraints and the broader coverage area inherent to UAVs, which necessitates a different approach to security and risk assessment, as highlighted in earlier research [58] .

The set of regulations that govern drone operations within a specific area is commonly known as the authentication, authorization, and accounting (AAA) framework. This framework provides various benefits to drone operators on their administrative rights and enforces rigorous methods of authentication to avoid unauthorized access by unidentified entities. Tracking and identification of drones alongside the operators can also be carried out in situations of uncertainty and illegal activities, thereby reducing the likelihood of being stalked, breached, or intruded. In addressing these security concerns, several mechatronic engineering solutions have been put forward, as documented in prior research [59] .

The market’s immediate access to reasonably priced drones has raised apprehensions concerning their possible exploitation in illicit activities. Their capability to transport external payloads adds to the concern, as it allows them to transport hazardous chemicals or explosives without attracting undue attention [60] . Moreover, their capacity for accessing remote and difficult-to-reach places is a critical risk. Drone operations over heavily populated areas raise safety concerns because of the possibility of collisions or accidents that could result in tragic events. Several suggestions are made to improve public safety in response to these occurrences and worries:

A pragmatic methodology is proposed for improving drone’s robustness to attempted infiltration of their infrastructure, covering both the hardware and software components. Modern technologies are incorporated into the design of the suggested framework in an effort to reduce exclusivity concerns related to drone operations. To guarantee clarity and the capacity to repeat the findings, the attempted procedures and the dataset used for deep learning tasks are meticulously explained and documented.

A seven-layer-based architecture is presented for drone security. Transference of data takes place from edge processing (EP) to SPL, where the incursions are detected using the proposed ensemble scheme. After EP-based authentication of data, it is propagated to a cloud repo where MS Azure-based authentication procedures are practiced. A pre-trained model is implemented in MS Azure cloud storage, utilizing drone data combined with NSL-KDD, KDD CUP, and Edge-IoT datasets. This model is in charge of forecasting possible attacks, and it is this model that initiates a mobile alert upon detection of an attack.

The following parts are used to build the smart secure drone: A wireless module, namely XBee Pro S1, has been used along with a GY-GPS6MV2 GPS module, HC-SR04 ultrasonic sensor, BMP180 barometric pressure sensor, and Mega 2560 microcontroller. These parts were purchased from Amazon. For central processing, the Mega-2560 micro-controller is used first for the initial jump start laced with adequate processing capacity and controlling prowess. It performs flight control algorithms, gathers sensor data, and establishes communication with other components [62] . In order to monitor flight duration and data, the Mega 2560 microcontroller additionally logs the current time. Obstacle avoidance and detection are handled by the HC-SR04 ultrasonic sensor. The drone can modify its flight path to prevent collisions by using the ultrasonic waves it emits, measuring its reflection time, and using this information to identify objects in the vicinity. Based on this data, it makes decisions about moving and changing its position.

An essential component of providing accurate positioning and navigation data is the GY-GPS6MV2 GPS module. It obtains the position ID of the UAV with its real-time coordinates by receiving signals from GPS satellites. Waypoint navigation, flight planning, and tracking all depend on this data. By measuring atmospheric pressure with the BMP180 barometric pressure sensor, the drone can estimate its altitude with a high degree of accuracy. To maintain stable flight and carry out tasks involving altitude changes, these altitude data are crucial. These parts and sensors improve the secure drone system’s ability to avoid obstacles, navigate precisely, and control altitude. They are essential to its operation. The pressure applied to the drone is also detected by this sensor.

For reliable communication of an UAV with its parent station, an information exchange module for wireless transference of real-time data XBee Pro S1 is essential. Aiding remote control, real-time data monitoring, and transmission of ongoing operation status. These are the requisites for a smart, secure, and reliably operational drone, ensuring the system operates safely and effectively by combining secure communication channels, dependable navigational capabilities, and advanced sensor functionalities.

Real-time drone data, including KDD intrusion detection features^1,2, drone OBD data³, and GPS-based attributes like longitude, latitude, and altitude, were used in this experiment for analysis⁴. Drone datasets are included within the provided link along with a few benchmarks related to intrusion detection and cyber-attack prediction⁵.

For a detailed overview of the dataset classes, please refer to Table 2 .

The weighted ensemble CNN voting model plays a pivotal role in elevating prediction accuracy and bolstering robustness. Figure 2 shows the architecture of the weighted voting ensemble [63] .

${\hat{P}}_{\text{ens}}=\underset{i=1}{\overset{3}{{\displaystyle \sum }}}{W}_i{p}_i$(1)

The class with the highest probability can be selected via:

$P_{\text{out}}=\text{argmax}\left({\hat{P}}_{\text{ens}}\right)$(2)

The results of the experiments are provided in this section. By contrasting the suggested model with other cutting-edge ML prototypes using the drone-based dataset, we were able to evaluate its performance. A 70% training and a 30% testing split was applied to the dataset to create training and testing subsets. PowerEdge T430 GPU by Dell, 8 GB GPU, and Intel-Xeon eight-core CPUs x2 with clock speed at 3.2 GHz with 64 GB DDR4-RAM were used for the experiments. Python and Anaconda were used in these experiments, which were carried out in the Jupyter Notebook environment.

Generic ML models and the suggested weighted voting ensemble of 1-D CNN models are among the classifiers that are assessed. These tasks are all carried out with the help of Python libraries like Tensorflow, Scikit-learn, and Keras. Three types of attacks are identified for the dataset: denial of service (DoS), spoofing, and jamming. The accuracy score is over 99%, which is an exceptional performance.

A comparative analysis of the results is provided in Table 6 for the drone dataset concerning different classifiers. Empirical findings demonstrate that conventional ML models and elementary deep learning architectures performed remarkably well in identifying intrusions in the drone dataset. Table 6 clearly underscores the fact that when it comes to recall, accuracy, precision, and F1, Naive Bayes (NB) performed the worst. However, multilayer perceptron (MLP) demonstrated a marginally higher accuracy of 99.75%. Moreover, accuracy rates surpassing 99% are attained by most of the ML classifiers employed within this study across all evaluation metrics. Strong performance is shown by the suggested WVCNN model, with detailed results given in Table 6 all reaching 99.90%. The drone dataset’s attack classification into categories like Prob, DoS, R2L, and U2R demonstrated this success.

The performance of various models on the real-time drone dataset reveals a range of efficiencies in detecting security threats. Random Forest, with an accuracy of 99.26% and an impressive precision (99.93%), recall (99.97%), and F1

score (99.95%), performs well due to its ability to handle large feature spaces and avoid overfitting through bootstrapped aggregation. Similarly, Logistic Regression achieves an accuracy of 99.64%, and its simplicity allows it to excel when the feature space is well-defined, leading to high precision (99.93%) and F1 score (99.97%). However, logistic regression may struggle with more complex, non-linear relationships that other models handle better. In contrast, the Decision Tree model, with 99.22% accuracy, performs slightly lower due to its tendency to overfit the data, which results in less generalization capability compared to Random Forest. The Naive Bayes algorithm has the weakest performance, with an accuracy of 97.43%, due to its assumption of feature independence, which does not hold true in complex, real-time drone datasets. This limitation makes it less effective in capturing intricate relationships between features, as reflected in its lower F1 score (97.98%). On the other hand, the MLP stands out with an accuracy of 99.75% and near-perfect scores across all metrics. This high performance is attributed to the MLP’s ability to model complex, non-linear relationships, making it highly effective for the drone dataset. However, the most outstanding model is the Weighted Voting CNN (WVCNN), which achieves the highest accuracy of 99.90%, precision (99.92%), recall (99.98%), and F1 score (99.97%). The success of WVCNN can be attributed to its ensemble approach, which combines the strengths of multiple CNN models to improve generalization and handle complex attack patterns. This enables WVCNN to capture subtle distinctions in the data that simpler models, like Naive Bayes and Decision Tree, may miss. The ensemble’s weighted voting mechanism also adds robustness, making it more reliable for detecting real-time threats in drone networks.

With an astounding accuracy of 99.90%, the WVCNN model successfully classified attacks into subcategories of the following types: U2R, R2L, Prob, and DoS. These results are examined in Table 6 . Additionally, Figure 3 presents the validation accuracy curve that shows how well the drone dataset performs relative to other datasets using the suggested method. Drone data is used by the suggested ensemble technique to apprise of any possible intrusions.

In this section, we use different datasets to evaluate the suggested WVCNN model. The datasets used for evaluation are diverse and include the NSL-KDD, WUSTL, Edge-IOT, KDD CUP 99, and X-IIOT datasets [66] . This analysis seeks to clarify the efficacy and resilience of this deep learning-based ensemble technique by thoroughly assessing the model’s performance across a range of cyber-attacks and feature sets.

The confusion matrices given in Figure 4 provide class-wise classification scores for each dataset. The CM provides a detailed insight into the model’s capacity and capability to classify each class without any biases. Figure 4 shows the proposed WVCNN model’s performance for 6 distinct intrusion detection datasets, containing dominant subcategories belonging to the four primary categories as mentioned in Table 2 , i.e., Normal, DoS, UoR, Prob, and RtL attacks respectively. The reason for the better performance of the proposed model is that the use of a weighted voting ensemble of three CNN models allows for capturing diverse features from the input data, resulting in improved robustness. CNNs are particularly adept at identifying patterns in complex data, such as that found in IoT-based drone and satellite networks, which enhances the model capacity to detect subtle anomalies and intrusions. The model is specifically designed to address the unique challenges of IoT-based aerial and satellite systems, which often experience high variability and heterogeneity in communication patterns. By focusing on these unique characteristics, the model optimizes intrusion detection and security for IoT and satellite systems, ensuring that even edge cases and uncommon attack patterns are addressed. These factors collectively contribute to the outstanding performance of the proposed model, making it highly effective in detecting security threats in both IoT-based aerial systems and satellite systems. The box plot and interval plot of the 5-fold accuracy of the proposed model are shown in Figure 5 .

In this subsection, we perform an ablation study on the proposed WVCNN model to investigate the contribution of key components to its overall performance. We systematically modify or remove specific components and analyze their impact on the accuracy, precision, recall, and F1 score.

The following components of the WVCNN model are modified for the ablation study:

Table 7 presents the performance of the WVCNN model under different modifications.

The ablation study demonstrates that each component of the WVCNN model-weighted voting mechanism, depth of convolutional layers, data augmentation, feature size, and dataset size-contributes significantly to its overall performance. The proposed WVCNN model performs best with all components intact, as evidenced by the high accuracy, precision, recall, and F1 score of the original model.

A thorough comparison between the WVCNN and the most advanced models covered in the literature can be found in Table 8 . Ensembles of difference approaches, i.e., PCA&MCA, SVM&ANN, and DT&RFE have been studied in the literature in endeavors to enhance intrusion detection model effectiveness. Furthermore, deep hierarchical models have also been used. However, the presented framework proved to be more effective, attaining an amazing 99.89% accuracy in the drone dataset intrusion detection context.

Certain experimental procedures were executed as shown in Table 8 , using the NSL-KDD [67] and KDD Cup 99 [68] datasets to demonstrate the robustness and

applicability of the proposed approach. The WVCNN model demonstrated its superiority in the field of intrusion detection by outperforming all other models from previous research on both datasets.

Unlike previous studies, which mainly concentrated on single-layered drone system architectures, our suggested framework presents a multi-layered strategy that uses cutting-edge ML models to improve security. Deep learning integration improves technology resilience and allows for adaptable responses to changing cyber threats. The promising results in terms of accuracy and other metrics are evidence of the potential of the proposed technique to improve security amidst intrusive and jeopardizing attempts towards the bidirectionally connected systems, encompassing various elements of aerial vehicle systems. It combines multiple 1-D CNN models in a weighted ensemble. The possibility of unsanctioned access and interceptions are addressed by this novel framework. The following study effectively discusses the synergizing of drone and IoT-based designs, offering a structure that considerably raises resilience and confidentiality. We demonstrated the novel contributions and improvements of our study by comparisons with previous works and evaluations with multiple benchmark datasets, opening the door for more secure and effective drone deployment in a variety of domains.