How to identify uncertified drones with machine learning
Our second post on drones and networks takes an in-depth look at how machine learning can be used to separate drones from ground devices. Such identification is needed to optimize service from an interference perspective for both types of users.
As described in our white paper Drones and networks: Ensuring safe and secure operations, mobile networks are well suited to support low-altitude drone communication and to be integrated with drone traffic management systems to enhance the safety and security of drone operations.
Many use cases require drones to transmit video feeds to their flight controllers, imposing heavy uplink traffic load on the networks.
Therefore, it is important that mobile networks can identify if a user equipment (UE) is a drone UE or a regular ground UE, in order to provide the right service optimization for drone UEs while ensuring the performance of ground UEs is not impacted. The cellular network standard, specified by 3GPP, was recently updated to include support for subscription-based Aerial UE identification and authorization, allowing mobile networks to do this.
But uncertified use of mobile devices and subscriptions, like mounting a regular mobile phone on a drone and streaming video from it, may occur. Such usage may cause increased interference to the network, and adversely impact the uplink performance of users on the ground. One important reason is that aerial radio channels have a higher likelihood of line-of-sight propagation due to the absence of obstacles, so an uplink signal transmitted from a drone UE may reach multiple neighboring non-serving base stations and thus cause increased uplink interference.
As soon as an uncertified drone is detected, interference mitigation techniques can then be initiated.
Machine learning for uncertified aerial UE detection
Machine learning, the science (and art) of programming machines enabling the machines to learn from data, can be utilized for this purpose. Here is the logic:
The radio link characteristics and mobility patterns are different for devices in the sky and devices on the ground. This is illustrated in the figure below, where an aerial UE detects all three cells with good signal quality, while, due to the blockage of buildings, the ground UE receives two of the cells with weaker signal quality.
In more detail, a drone UE operating at a high altitude is expected to have a close to line-of-sight propagation environment that leads to low variance of Reference Signals Received Powers (RSRP) of the strongest cells. Similarly, the Received Signal Strength Indicator (RSSI) statistics of drone devices are different from those of regular ground devices because a mobile device mounted on a drone may receive signals from multiple cells with similar strengths. For indoor UEs in high-rise buildings, the penetration loss of the signal entering the building will heavily degrade the signal quality in comparison to an aerial UE at similar height, enabling the classifier to differentiate the two UE types.
As mentioned, different radio characteristics for ground UEs and drone UEs have been validated in field trials. Some results from the field measurements are shown in the figure below, where the drone UE was flying at the height of 100 meters. The figure compares the distribution of the signal strength variations of the four strongest detected neighbor cells of the drone UE to the counterpart associated with the ground UEs. As is clear from the figure, the signal strength variation in the sky is much smaller than on the ground.
Field measurements of signal strength variations
The machine learning task is then to predict UE type (drone UE or ground UE) based on the radio measurement reports sent by the UEs to the network.
As in any typical machine learning problem, data needs to be prepared for machine learning algorithms. To this end, measurement data from known legitimate drone UEs flying in the network should be collected, in addition to the measurement data from ground UEs. The collected measurement data are then divided into training dataset and test dataset. The training dataset is used for training a drone-detection machine-learning model, while the test dataset is used for assessing the performance of the model.
In our first evaluations, we derived features from the RSSI and RSRP fields in the UE handover measurement report. Two important features besides the serving cell RSRP and the RSSI, are the variation in RSRP of the strongest cells shown in above figure, and the RSRP difference between the two strongest cells.
The next step after deriving the features is to select and train a machine-learning model. The selection of a suitable machine-learning model is inherently subjective to formulation of problems, input features, requirements, and resources at hand. There is a trade-off between model complexity, performance, and computational resources (both hardware and time). Complex models can capture more details of the problem but also have high requirements in terms of computational resources and training data. In contrast, too simple models may not well capture the underlying trend of the data, leading to underfitting. For our purpose, we selected Logistic Regression and Decision Tree. We find that with carefully selected features, these standard machine learning algorithms produce satisfactory results.
Our results show that the uncertified drone detection accuracy increases with drone height, and that more measurement reports improve the detection accuracy in our scenario with a mixture of outdoor, indoor and drone UEs.
Options for deploying machine learning algorithms
How to deploy the machine learning algorithms for uncertified drone detection in mobile networks is another challenge. Since the network nodes are often geographically distributed, a careful design is needed to strike a good balance between amount of signaling between the nodes and uncertified drone detection performance. Two options are shown in the figure below.
In option (a), a central entity stores all the data and trains a single machine learning model. Such an entity could be a network node or a device in the cloud that may be further integrated to be part of a UTM system. Alternatively, in option (b), machine learning models can be trained per cell or group of cells. In this case, each cell or group of cells can run a model locally and exchange the output with neighbors to collaboratively carry out the uncertified drone detection task.
Find out more
Want to learn more? In our article, Rogue Drone Detection: A Machine Learning Approach, we take a deeper look at using machine learning to help identify uncertified (rogue) drones in mobile networks.
Our earlier blog post in the series explains the challenges and requirements for mobile networks to support both aerial and terrestrial users.