In recent years, unmanned aerial vehicles (UAVs) have become increasingly prevalent across various domains, prompting a growing need for study into their security issues [2]. One of the fundamental topics is device authentication. However, traditional authentication methods may not be suitable for UAV classication. Classic techniques based on angle of arrival (AOA) and time dierence of arrival (TDOA) are often ineective due to the mobile nature of even legitimate APs within the operation area [8].
In response to these challenges, radio frequency (RF) ngerprinting has emerged as a promising authentication method in related domains [4,15]. This technique involves classifying the inherent physical imperfections in the analog circuitry of RF emitters that arise during the manufacturing process. These imperfections slightly aect the transmitted signals without compromising device performance, creating unique ngerprints for each device. Compared to traditional authentication methods, RF ngerprinting oers enhanced robustness against tampering and spoong, thereby improving the security of RF devices [11]. By leveraging powerful deep neural networks (DNNs), RF ngerprinting can achieve high performance and be easily deployed. These characteristics make RF ngerprinting particularly well-suited for UAV authentication, addressing the limitations of conventional methods and providing a more reliable solution for this emerging security challenge.
Challenges. Despite the promise of RF ngerprinting for UAV authentication, several signicant challenges persist. First, while DNN-based RF ngerprinting systems can achieve high performance in known domains, they often struggle with the domain shift problem, performing poorly in unknown domains such as dierent times or distances [5]. Second, unlike stationary RF devices like Wi-Fi routers, UAVs are constantly hovering or moving in the air. This mobility introduces complex channel variations and exacerbates the domain shift problem, making it even more challenging to maintain consistent ngerprinting performance across varying conditions [8]. Third, gathering a large dataset of UAV signals from new domains to retrain DNNs is both impractical and resource-intensive. This limitation hinders adapting UAV ngerprinting systems to new environments or conditions. Given these challenges, there is a need for a lightweight solution that can improve cross-domain ngerprinting performance without extensive retraining or unseen domain data collection, thereby signicantly enhancing the viability of RF ngerprinting for UAV authentication.
Our Solution. To address these challenges, we propose a solution based on prototypical networks (PTNs) [7] for crossdomain UAV ngerprinting. First, we redesign the PTN to optimize a feature extractor that can identify stable ngerprint features across dierent domains using similarity metrics. As a few-shot learning (FSL) paradigm, PTN requires only a small amount of labeled data during inference to improve accuracy. Second, we implement a ne-tuning process for the trained feature extractor to accommodate the diverse unseen domains generated by UAVs' hovering nature. This allows the model to adapt to complex new domains. Third, we design and apply a data augmentation technique during the ne-tuning stage to further enhance classication accuracy. Overall, our contributions are as follows:
• To the best of our knowledge, this is the rst work to deploy FSL to mitigate domain shift issues in UAV ngerprinting. Our approach eliminates the need for extensive data collection and cumbersome model training processes, thereby enhancing classication performance and overall system viability. • We carefully design data augmentation and employ ne-tuning with only a few data, which further enhances system accuracy without signicantly increasing overhead. • Our comprehensive experimental evaluation shows that our proposed method can improve the classication accuracy by about 20% in the best case, demonstrating the eectiveness of our approach.
2 Background and Related Work 2.1 RF Fingerprinting RF ngerprinting has emerged as a promising technique for identifying wireless devices based on their unique hardware imperfections. This physical layer identication method offers enhanced resistance to spoong and replay attacks [9].
The development of powerful deep learning techniques has enabled automatic extraction of RF ngerprint features, leading to widespread adoption in various device identication applications [5, 12]. Typically, DNN-based RF ngerprinting systems use raw in-phase/quadrature (I/Q) data as input, leveraging the DNN's ability to eectively process and classify complex signal characteristics. The process of RF ngerprinting generally involves two key components: feature extraction and multi-class identication. Accurate feature extraction is crucial for successfully distinguishing dierent RF ngerprints [6, 16]. In UAV ngerprinting, Soltani et al. propose a multi-classier scheme with a two-step score-based aggregation method and data augmentation to enhance cross-domain performance [8]. However, this method may be time-intensive for training and deployment. Zhao et al. employ auxiliary classier Wasserstein generative adversarial networks (ACW-GANs) for feature identication [14]. Cai et al. develop a lightweight backbone network using lightweight multiscale convolution (LMSC) blocks, reducing model size while improving feature extraction capabilities in a simulation environment [1].
FSL oers a signicant advantage over traditional deep learning approaches by enabling models to generalize to new classes and domains using only a limited number of examples [13]. This capability makes FSL particularly valuable in scenarios where data is scarce or costly to obtain. FSL allows for rapid adaptation to new tasks in data-constrained environments, a feature that has led to its deployment in various related domains [10,15]. Consequently, it is well-suited for cross-domain UAV ngerprinting systems.
In this paper, we use a base dataset E base to train a feature extractor 5 \ . Then, we create a support set E support consisting of a small number of labeled samples and a query set E query containing data that we need to classify. The # -way -shot learning scheme, a common approach in FSL, refers to training the model on # classes with labeled examples per class [13].
Overall, there are some key distinctions between our work and previous studies. First, we address the domain shift issue by considering both time and distance variations. Second, we adapt the PTN structure to better suit UAV ngerprinting, enabling lightweight deployment. Third, we design a data augmentation strategy combined with ne-tuning specically tailored for UAV ngerprinting to enhance performance.
Fig. 1 provides an overview of our proposed cross-domain UAV ngerprinting system, which consists of two main stages. During the training stage, a feature extractor is trained to extract ngerprints using the base set. In the inference stage, the feature extractor is rapidly ne-tuned to generate prototypes for UAV classication. This section will explain each stage of the system in detail.
To address the domain shift challenge, we modify the PTN to train an eective feature extractor for cross-domain UAV ngerprinting. Our goal is to train a robust feature extractor capable of generalizing across dierent domains. We employ ResNet-18 [3] as the feature extractor 5 \ , as convolutional neural networks (CNNs) have demonstrated their ability to extract ngerprints from I/Q data [5]. To adapt ResNet-18 for our specic task, we modify only the rst input layer to accommodate the dimensions of our I/Q data, which has a size of 2 ⇥ 256. To ensure that the embedding vectors lie on a hypersphere with a constant radius, we add an ! 2 -norm layer before the nal output layer as follows:
where 5 0 \ (x 8 ) is the output embedding before the ! 2 -norm k•k 2 , and 5 \ (x 8 ) represents the nal feature embedding.
After feature extraction, a nal classier ⇠ (•) is added on the top and can be adjusted according to the number of UAVs. The feature extractor is trained using a traditional supervised learning scheme, utilizing our base set E base = {x 8 , ~8 } ⌫ 8=1 as the training data. We employ the classic multi-class crossentropy loss function to guide parameter optimization:
(2)
Once a feature extractor is well-trained, we can generate prototypes for inference. The prototype for each UAV is determined by averaging all the feature embedding vectors belonging to that class. The computation of prototypes can be expressed as follows:
where c 8 denotes the prototypes of the UAV ~8 , and = denotes the number of samples in each class in the dataset. These prototypes encapsulate the essential characteristics of specic UAV classes, providing a generalized representation that remains relatively invariant across dierent domains. By computing prototypes for each UAV, we obtain stable representations crucial for the classication process, allowing for accurate and reliable identication across diverse domains.
Once the prototypes of each UAV have been established, the model can use them to generate predictions for new samples. This is accomplished by comparing the feature embedding of a new sample to the prototypes of each class. In our experimental setup, we quantify this comparison using cosine similarity, which is calculated as follows:
where D represents the similarity matrix between the input sample x 8 and the prototypes of all known UAV classes. The function 3 (c, 5 \ (x 8 )) calculates the cosine similarity between a class prototype c and the feature embedding 5 \ (x 8 ) of the input sample. Cosine similarity values range from 1 to 1, with values closer to 1 indicating higher similarity. The model assigns the input sample to the class whose prototype yields the highest similarity score, thus determining the identication result.
The hovering and mobility capabilities of UAVs introduce additional complexities to dierent domains, necessitating ne-tuning to enhance system performance. However, our system relies on a limited number of samples from the support set, which may impede the model's ability to generalize eectively to complex tasks. To address this issue, we incorporate data augmentation during the ne-tuning process. Data augmentation is a widely adopted technique that enhances DNNs' generalization capabilities, enabling more accurate predictions on previously unseen data.
In this paper, we recognize that UAV domains signicantly impact data magnitude. Consequently, we leverage this domain-specic data information to design our augmentation strategy as follows:
where ! denotes the size of the data, x B 9 0 is the augmented data derived from the support set E support , and `(x B 9 ) and f (x B 9 ) represents the mean and standard deviation of the data x B 9 , respectively. By integrating this augmentation strategy with ne-tuning, our well-trained feature extractor can better generalize to new domains.
Alorithm 1 describes the pseudocode for training the feature extractor 5 \ . The process begins with a base set E base for initial training and a support set E support for subsequent ne-tuning and inference. First, we use the base set to train a feature extractor that can generate stable UAV ngerprints. Next, we apply specially designed data augmentation techniques to expand the support set during the inference stage. This augmented dataset is then used to ne-tune the feature Algorithm 1 PTN training and ne-tuning INPUT: Base set E base = {(x 8 , ~8 )}, support set E support = {(x B 9 , ~B 9 )}, feature extractor 5 \ , classier ⇠, learning rate ;A , ne-tuning learning rate ;A 0 , hyperparameter U OUTPUT: ne-tuned feature extractor 5 \ Step 1: Train with base set 1: for number of epoch do 2: for (x 8 , ~8 ) 2 E base do 3: L ⇠A>BB⇢=CA>?~(⇠ (5 \ (x 8 )), ~8 ) 4: end for 5: \ \ ;A • oe \ L 6: end for Step 2: Construct augmented set 7: for (x B 9 , ~B 9 ) 2 E support do 8: x B 9 0 x B 9 + U • # (`(x B 9 ), f (x B 9 ); !) 9: ~B 9 0 ~B 9 10: E augment E support + {(x B 9 0 , ~B 9 0 )} 11: end for Step 3: Fine-tune the feature extractor 12: for number of epoch do 13: for (x 0 : , ~0 : ) 2 E augment do 14: L ⇠A>BB⇢=CA>?~(⇠ (5 \ (x 0 : )), ~0 : ) 15: end for 16:
\ \ ;A 0 • oe \ L 17: end for 18: return 5 \ extractor, adapting it to new domains. The nal predictions are made by comparing the cosine similarities between feature embeddings and each prototype c. The class with the highest similarity to the embedding is assigned to the input.
In all experiments, the learning rate was set to 0.001. B⌘>C , # @D4A ~, ne-tuning epochs and ne-tuning learning rate were set to 5, 15, 10, and 0.0001 respectively. The value of # F0~w as set to the size of all labels for the datasets. The experiments were conducted on a server with an Intel Xeon E5-2650L v4 CPU and 8 NVIDIA GeForce GTX 1080Ti GPU.
This paper conducts experiments on a public UAV ngerprinting dataset [8]. The dataset comprises 7 identical DJI M100 UAVs as transmitters, with an Ettus USRP X310 equipped with a UBX 160 USRP daughterboard serving as the receiver to collect I/Q samples from DJI's non-standard, proprietary waveform. The data collection encompasses 4 non-overlapping bursts at dierent times and 4 distinct distances (6, 9, 12, and 15 feet). To address the domain shift problem, we construct our base set (5112 I/Q samples) using data from distances of 6 and 9 feet, taken from bursts 1, 2, and 3. This conguration allows us to evaluate our UAV ngerprinting system across various domain partitions, considering both time and distance variations.
Table 1 presents the cross-domain UAV ngerprinting results. The high classication accuracy in the source domain demonstrates CNN's strong capability for extracting UAV ngerprints. For unseen domains, results are presented as 'a-b', where 'a' represents distance in feet and 'b' represents burst time. '6-4' and '9-4' indicate tests at distances of 6 and 9 feet during burst 4, which occurs in a dierent time domain than the base set. Classication accuracy decreases by approximately 14% and 6% respectively in these cases. Notably, accuracy drops more signicantly for unseen distances compared to dierent time domains. The most severe case is '15-3', where accuracy falls to only about 32%, which is inadequate for reliable UAV ngerprinting. These results suggest that while the CNN performs well in known domains, its performance degrades in unseen scenarios, particularly with changes in distance. In our FSL experiments, we use 5 examples from the support set to learn about unseen domains and 15 examples from the query set for prediction. While the vanilla PTN method slightly outperformed the CNN approach overall, it performed worse in the '9-4', '12-2', and '12-3' scenarios. These results highlight the limitations of vanilla PTN, despite its ability to incorporate some target domain information from a small number of examples.
To address these limitations, we apply ne-tuning, which increases classication accuracy for all cross-domain cases. The improvement is particularly signicant for the '15-2' and '15-3' cases, where accuracy increased by about 20%. Furthermore, applying our designed data augmentation technique further improves classication accuracy across all cases. The most notable improvement is observed for '15-4', which has both dierent time and distance domains than the source domains, increasing accuracy by about 4% compared to netuning alone and 18% compared to the CNN approach.
Fig. 2 presents the classication accuracy of our proposed FSL and data augmentation-aided UAV ngerprinting method under various k-shot settings. Fine-tuning and augmentation with 1-shot have limited eectiveness due to the constraints of using only a single support set sample. Nevertheless, they still outperform the vanilla PTN in 5-shot settings. Generally, increasing the number of support samples improves classication accuracy. However, this improvement is not linear. For instance, the accuracy gain from 7-shot to 10-shot is minimal, whereas the increase from 1-shot to 3-shot is substantial. This non-linear improvement demonstrates that our method achieves eective results without requiring a large number of domain samples, making it a lightweight solution. We will further validate this eciency in the following subsection.
For UAV ngerprinting, extensive ne-tuning on unseen domains is impractical due to time constraints, highlighting the need for a lightweight cross-domain solution. Fig. 3 shows the time required to ne-tune 5 support samples over 10 epochs and predict 15 query samples across various scenarios. In the case of 1-shot PTN with ne-tuning and data augmentation, the process takes around 0.35 seconds, making it practical for unseen domain adaptation. Notably, vanilla ne-tuning for the 5-shot case consumes roughly the same amount of time as ne-tuning with data augmentation in the 3-shot case. The most time-consuming scenario is the 10-shot PTN, averaging about 1.37 seconds. This overhead is considered acceptable since the process involves 10 epochs of ne-tuning and 15 sample predictions.
Combining these time costs and previous k-shot results, we conclude that deploying 10-shot learning is unnecessary, as it only shows a minimal increase in accuracy compared to 5-shot learning while being about 0.6 seconds slower. Our method thus demonstrates a signicant increase in crossdomain accuracy without incurring large overhead, striking an optimal balance between performance and eciency.
Fig. 4 presents the performance of our proposed method under varying numbers of query samples, ranging from 5 to 25. The results demonstrate a remarkably consistent performance across this range, with only minimal variations observed as the number of query samples increases. This tight performance bound across dierent query settings is a strong indicator of our method's stability and robustness in the context of cross-domain UAV ngerprinting.
This paper proposes a lightweight cross-domain UAV ngerprinting method to mitigate domain shift issues. To achieve this, we rst modify PTN to generate stable ngerprint features. Then, we devise specic data augmentation and apply ne-tuning to further improve performance on the unseen domains. Comprehensive experimental results validate our method's eectiveness in mitigating domain shift, demonstrating robust UAV identication across varied conditions.