In this article, we explore the potential of exploiting mainly the EM emissions that are unintentionally generated as a consequence of electrical currents flowing through the electronic circuits of drones. The generation of these magnetic fields is primarily governed by the principles of the Biot–Savart Law. This law states that the magnitude of a magnetic field produced by a long conductor carrying current I is expressed as , where  is the magnetic constant and r is the distance from the conductor [27].

Figure 2 summarizes our proposed solution Drone-Mag. Essentially, Drone-Mag offers a method for authenticating drones based on their unique EM emissions. The solution first starts by training an autoencoder to detect rogue drones and an SVM model to authenticate drones.

In the first phase, we use a magnetic probe to gather the EM emissions from the drone’s flight controller and ICs while the drone is powered ON but not in flight. In the second phase, we calculate relevant Machine Learning (ML) features based on the collected EM emissions samples. These features serve as input to train or test the autoencoder and SVM models.

When a drone authentication is required, the EM emissions are collected, the relevant features are extracted, and then passed as a test sample to the trained autoencoder model. If the drone successfully passes the autoencoder test, it is recognized as an authorized drone. Otherwise, the drone authentication fails. In cases where individual drone authentication is necessary, the test sample is then evaluated by SVM model specifically trained to differentiate between different authorized drones. If the test sample profile passes this additional test, the authentication process successfully identifies the specific drone. Otherwise, the authentication fails.

Overall, Drone-Mag mainly involves the following two entities:

We define two main modes of operation for Drone-Mag framework:

Overall, Drone-Mag framework consists of five different modules, as depicted in Figure 3:

We employ the following performance evaluation metrics to evaluate the performance of Drone-Mag:

Drone-Mag includes two main phases, namely, the Enrolment Phase and the Authentication Phase, detailed in the following.

All of the tests reported in the following were carried out under standard laboratory conditions across 3 days by collecting the unintentional EM emissions emitted by the target drone under test, with no attempt made to decrease background noise. It is worth emphasizing that the magnetic probe we utilize is a near-field probe, which primarily records EM emissions from a close range around the probe, which enables the most effective noise reduction from the surrounding environment. We collect the EM emissions across 11 different 2 MHz frequency bandwidths, i.e., 30–32, 32–34, 34–36, 36–38, 40–42, 42–44, 44–46, 46–48, 49–51, 75–77, and 112–114 MHz. Our spectrum analyzer experimental setup is shown in Figure 5(a).

Note that we just switch the drone ON, then start collecting the EM emissions using the described setup. The drone is idle with the motors and propellers switched OFF, so the unintentional EM emissions collected are related to the powered ON state of the drone’s electronic components without performing any specific function or movement.

In this section, we reduce the form factor of the experimental setup for Drone-Mag, using cheap Commercial off-the-shelf RTL-SDR instead of the spectrum analyzer to collect the EM emissions captured by the magnetic probe, as shown in Figure 5(b). We select four 2 MHz frequency bandwidths: 30–32, 32–34, 34–36, and 36–38 MHz to test the performance of Drone-Mag.

We first evaluate the profile of unintentional EM emissions generated by different drone flight controllers when the drone is switched ON but not flying. We considered six different drone flight controllers as summarized in Table 1.

Figure 6(a) shows the power spectral density of the unintentional EM emissions of the full 200 MHz bandwidth acquired by the spectrum analyzer from the 6 different flight controllers under test, and Figure 6(b) shows the 10 identical Solo drones considered for evaluating the performance of Drone-Mag, respectively. Each trace in the cited figures lasts for around 150 ms. Due to the normalization phase executed during the Features Extraction module, all of the unintentional EM emissions RSS samples recorded in dBm are normalized to values in the range between 0 and 1. Specifically, the blue color maps values in the range , the cyan corresponds to values in the range , the yellow color is related to values in the range , while the red color indicates values in the range .

While we collect the unintentional EM emissions of each drone flight controller using 200 MHz frequency bandwidth to get a clear insight into their differences as shown in Figure 6(a) and (b), we only consider a 2 MHz fingerprinting bandwidth for drone identification and authentication. We chose four random 2 MHz frequency bandwidths from the EM emissions of the six different drones under test collected using the spectrum analyzer setup shown in Figure 5(a), i.e., 30–32, 49–51, 75–77, and 112–114 MHz; and four consecutive frequency bandwidths, i.e., 30–32, 32–34, 34–46, and 36–38 MHz from the drones’ EM emissions traces collected using the RTL-SDR setup depicted in Figure 5(b), in order to show the scalability of Drone-Mag and its ability to authenticate drones across a wide range of frequency bandwidths. Figure 7(a), (b), (c), and (d) shows the power spectral density of the unintentional EM emissions recorded for the following 2 MHz frequency bandwidths: 30–32, 32–34, 34–46, and 36–38 MHz, respectively. In addition, Figure 8(a)–(c) shows the unintentional EM emissions recorded for the 49–51, 75–77, and 112–114 MHz frequency bandwidths, respectively.

In this section, we detail the features computation process for the EM data collected using both setups: Spectrum Analyzer and RTL-SDR.

In this section, we propose an autoencoder-based solution to check if the EM fingerprint extracted from the drone under analysis belongs to the pool of the EM traces collected from the authorized drones. This step is taken before the identification and Authentication Phase, leaving the other two for further analysis should the drone pass this test. We conducted the experiments within the RTL-SDR reduced form setup as it represents a more efficient and cost-effective setup.

An autoencoder is a type of unsupervised neural network that can learn to compress and encode data efficiently. In addition, it can also learn to reconstruct the data from the compressed representation to a version that closely resembles the original input [16]. An autoencoder can be trained on a dataset built recording the normal expected behavior of some devices, and then it can be used to identify any new devices that deviate from the model built over the cited dataset. This pipeline of actions can be implemented by encoding the profiles of the new drones and comparing them to the encoded representation of the authorized drones. If the encoded representation of the new device is significantly different from the authorized drones’ profiles, it can be flagged as a rogue drone.

Our aim is to test the ability of Drone-Mag to reject rogue drones based on the unique EM emissions profiles. Since the task of classifying identical drones is more challenging than classifying different drones that are distinct even to the naked eye, we consider the 10 identical Solo drones to evaluate Drone-Mag rogue drone detection performance. From the dataset of EM emissions data collected from the 10 identical Solo drones using the minimized setup described in Section 4.2, we build an autoencoder framework, and we use it to train a model utilizing the EM emissions data of 8 out of the 10 identical drones as authorized drones, and we consider the data of the remaining 2 as rogue drones. We use the most relevant 50 FFT points as features out of the 16,000 FFT points collected. The autoencoder architecture used for detecting rogue drones is summarized in Table 2.

In Figure 9, we plot the Mean Squared Error (MSE) of both the eight authorized drones and the two rogue drones across four frequency bandwidths, i.e., 30–32 MHz in Figure 9(a), 32–34 MHz in Figure 9(b), 34–36 MHz in Figure 9(c), and 36–38 MHz in Figure 9(d). We set a separating threshold between authenticated and rogue drones as follows: 0.005, 0.004, 0.0045, and 0.0027, respectively. The selected threshold used to accept or reject the samples of EM emissions of the drone under test can be adjusted according to the desired system performance, either to increase security and reduce false accept errors, or reduce false reject errors and enhance accessibility. A comprehensive performance analysis for Drone-Mag rogue drone detection is summarized in Figure 10.

We use the following performance metrics: Accuracy, Precision, Recall, and F1-score. Drone-Mag has a minimum F1-score of approximately 0.99 for the four frequency bandwidths considered. We chose four consecutive frequency bandwidths, i.e., 30–32, 32–34, 34–36, and 36–38 MHz, and four dispersed ones for the spectrum analyzer setup, i.e., 30–32, 49–51, 75–77, and 112–114 MHz to show that Drone-Mag is scalable and can work across a wide range of frequency bandwidths.

This step is performed only if the drone passes the verification test carried out by the rogue drone detection autoencoder framework. The classification results for each of the two main objectives illustrated in Section 2 are detailed in the following.

To further validate the robustness and environmental resilience of Drone-Mag, we conducted outdoor experiments by collecting a new set of EM traces from 18 drones and flight controllers under real-world conditions. Our goal is to examine whether Drone-Mag’s classification accuracy, previously validated under controlled laboratory conditions, remains consistent when deployed in outdoor scenarios. We collected EM traces from 10 identical Solo drones and 8 additional flight controllers that had not been tested in prior experiments, as summarized in Table 3. Each drone’s EM emissions were recorded twice: once in an indoor laboratory environment and again in an outdoor setting. This dataset was used to assess whether environmental variations impact classification performance. For the data collection, we selected four 2 MHz frequency bandwidths within the 40–48 MHz range: 40–42, 42–44, 44–46, and 46–48 MHz. Each sample was recorded for 250 ms and captured using 16,000 FFT points, with an FFT resolution step of 125 Hz, effectively spanning 2 MHz per bandwidth. For classification purposes, we extracted 35 of the most relevant FFT points from each sample, ensuring that we retained the most distinguishing features of each drone and flight controller. Our ML model was trained exclusively on the EM traces collected indoors. We then tested the model using the outdoor-collected samples to evaluate its generalization performance in different environmental conditions. The classification results were assessed using five key metrics: Accuracy, AUC, Precision, Recall, and F1-score. As summarized in Figure 14, our results show an average F1-score of 0.9822, 0.9721, 0.9796, 0.9925 for the 40–42 MHz, 42–44 MHz, 44–46 MHz, and 46–48 MHz frequency bandwidths, respectively. These results confirm that the EM fingerprinting approach remains consistent and reliable across both indoor and outdoor environments, achieving near-identical classification accuracy. This finding aligns with prior research [10], which demonstrated that near-field EM emissions are unaffected by environmental factors due to their localized nature. A critical factor in the consistency of our results is the nature of near-field EM emissions, which originate from a drone’s internal electronics rather than its wireless communication modules. Since these emissions are detectable only within a very short range (typically under 5 cm), they are inherently resilient to external interference from environmental Radio Frequency (RF) sources. Moreover, this near-field characteristic enhances the security of Drone-Mag, making it resistant to eavesdropping and replay attacks. An adversary attempting to capture and replay an EM fingerprint would need to place a recording device within 5 cm of the drone’s components, a highly impractical scenario. This property also aligns with our assumed military and critical infrastructure deployment scenarios, where drones are expected to land in designated zones free from external electronic interference at distances of 5–10 cm or more. Additionally, we ensured that the fingerprinting process leveraged unoccupied frequencies within the 40–48 MHz range (and also in the 0–32, 32–34, 34–36, 36–38, 49–51, 75–77, and 112–114 MHz). This choice minimizes interference from widely used communication services such as Wi-Fi or Bluetooth, further improving classification robustness. Moreover, these new indoor and outdoor tests were conducted on a new frequency range that had not been tested before in the previous indoor lab-only experiments detailed in Section 4, further demonstrating the scalability of Drone-Mag across the entire 200 MHz bandwidth, as shown in Figure 6(a). Our outdoor experimental results confirm that Drone-Mag provides robust, environment-independent authentication of drones and flight controllers. The near-field nature of the collected EM emissions ensures consistent classification performance across different operational settings, further strengthening its applicability for high-security environments such as military bases and critical infrastructure. The ability to maintain over 0.97 F1-score outdoors validates the feasibility of real-world deployment, reinforcing the scalability and reliability of the Drone-Mag authentication framework. In addition to our outdoor experiments, prior studies on EM fingerprinting of laptops and mobile phones [10, 29] have investigated the effects of different locations, times, and temperatures on EM fingerprints. These studies found that the fingerprints remain consistent across different locations and time instances, while temperature variations affect the fingerprints only after crossing a specific threshold. In our case, the primary operational scenario is a military base or similarly controlled environment, where such environmental factors can be maintained within acceptable bounds, ensuring consistent fingerprinting during each authentication instance.

When comparing the classification performance of Drone-Mag using EM emissions samples collected using the spectrum analyzer and RTL-SDR, we can see that they produce comparable results. However, the spectrum analyzer setup requires less sample time length of 150 ms compared to 250 ms for the EM samples collected using the RTL-SDR setup. In addition, the spectrum analyzer can collect the EM emissions from the drones over a wider frequency bandwidth of up to 2 GHz compared to around 3 MHz for the RTL-SDR. However, the RTL-SDR is much cheaper and portable than the spectrum analyzer. The system administrator can decide on which setup to deploy based on the system requirements and available resources.

We acknowledge that the relatively simple models employed in Drone-Mag—such as the SVM and autoencoder—may face limitations when scaled to accommodate thousands or tens of thousands of UAVs. As the dataset size increases, computational overhead, memory requirements, and training times can also increase. However, the simplicity of these models was chosen to fit the current dataset and ensure real-time performance without the need for extensive computational resources.

For larger datasets, more complex models or alternative architectures, such as convolutional neural networks or transformer-based models, may be employed to maintain scalability and accuracy. These architectures are known for their ability to handle larger feature spaces and complex datasets while maintaining robust performance. However, for the size and nature of our current dataset, these more complex models are not necessary, as the SVM and autoencoder have demonstrated excellent performance with minimal computational overhead.

In addition to the previous point of adopting more complex ML models for larger datasets, several strategies can be employed to enhance Drone-Mag’s scalability to accommodate an increasing number of drones:

While the close proximity requirement for Drone-Mag may seem like a limitation, it actually provides a significant security advantage. Since Drone-Mag utilizes a near-field EM probe, the adversary must be in extremely close physical proximity (within a few centimeters) to the drone to successfully record its EM emissions for a potential replay attack. In practice, this is highly unlikely due to the secure and monitored nature of the areas where drones undergo authentication (e.g., military bases, airports, critical infrastructure). Unlike acoustic or vision-based authentication methods, which can be eavesdropped from the far-field using basic recording devices, Drone-Mag’s near-field requirement makes it nearly impossible for an adversary to passively capture the EM emissions without being detected. In addition, the adversary would also need to get very close to the authentication pad to replay the recorded emissions effectively, adding an additional layer of security.

It is well established that EM and RF-based authentication methods may require periodic re-fingerprinting to account for changes over time. This issue is common across physical-layer fingerprinting methods [28], as demonstrated in prior research:

Impersonation or mimicry attacks involve attempting to imitate the EM fingerprint of a target device by manipulating the software or configurations of a similar device. However, the fundamental discrepancies in EM fingerprints originate from the physical hardware (e.g., the flight controller’s electronic components). Manipulating the software or settings of an attack device may alter its EM fingerprint, but it is nearly impossible to make it identical to the target drone’s fingerprint due to the hardware-dependent nature of the emissions—as commonly assumed in the literature [9, 25]. Studies [10] have shown that the hardware-level variations are subtler yet more defining than the variations caused by software configurations, making successful mimicry highly unlikely.

Unlike other physical-layer fingerprinting techniques such as acoustic or vision, Drone-Mag relies on a near-field EM probe that only detects emissions within close proximity. Hence, in order to record/replay the EM emissions of a specific drone, an adversary would need to get extremely close to the drone during the fingerprinting process. In our scenario, this process typically occurs within a restricted and secure area, such as a military base, airport, or critical infrastructure facility. These locations have stringent security protocols, making it difficult for an adversary to access the authentication area and record/replay the EM fingerprint of a legitimate drone. Furthermore, even if an adversary intercepts the drone outside of the secure zone (e.g., military infrastructure, airports, or other critical locations) and manages to record its EM emissions from near-field proximity, they would still need to enter the secure zone and be in close proximity to the authentication system to conduct a replay attack. Given the controlled nature of such environments and the security measures in place, this adds an additional layer of protection against replay attacks.

For most commercial drones, changing the drone’s firmware is fairly simple. However, in the case of military or classified drones considered in the scenario detailed in Section 2, manipulating the firmware might be protected by hardware or software techniques [32, 57]. Nevertheless, as proven in [26], the firmware modifications can be detected leveraging ICs EM emissions.