Unmanned Aerial Vehicles Communication Interference Detection Based on SSA-BP Neural Network

This paper proposes a method for detecting communication interference in unmanned aerial vehicles (UAVs). First, we train the BP neural network with signal data from interference attacks, and then continuously optimize the BP neural network using the sparrow search algorithm (SSA). After iterations, we obtain the final interference detection model. With this detection model, we can detect whether there are malicious interference signals in the drone flight environment and evaluate the detection model using the detection rate. Finally, our detection model has an accuracy of up to 93.64%.

Rule-based intrusion detection relies on pre-defined behavior-based rules, such as creating a set of rules to define the path that a drone must take to detect hijack attacks [5].When the drone deviates from its intended path, the rule-based intrusion detection system will be triggered, this system be used to detect other types of lethal network attacks.Traditional communication intrusion detection methods have various limitations and drawbacks.These methods require a significant amount of manual configuration and rule definition, which limits their detection capabilities and makes them unable to adapt to new types of intrusions and attacks.Additionally, these methods also consume a lot of computational resources and storage space and are difficult to expand and upgrade.
In recent years, many interference detection techniques have been proposed, including traditional interference detection [6,7] and machine learning-based interference detection.Machine learning-based interference detection technology is highly favored because well-trained machine learning models can process large amounts of data in a short time, improving efficiency.In [8], the authors studied several machine learning models and used them to detect the link status between transmitters and receivers to detect attacks.The performance of interference detection using three different methods, namely neural networks, random forests, and neural networks, was compared on the same datasets, and simulation results showed that random forests had the highest accuracy in detecting interference attacks.In [9], an interference attack detection model based on federated learning was proposed.The weight was trained on a local drone and sent to the global model.Each drone shared the weight and then averaged it to obtain a better detection model to detect interference attacks in a drone flight ad hoc network.
In [10], researchers proposed the CNN-LTSM structure for interference attack detection.They developed a structure that can achieve high interference detection accuracy using a simple network model.Through this pre-trained machine learning, we can reduce the time and resources required for model training because they use previously learned knowledge and can avoid the complexity of training models from scratch.However, complex networks require a lot of computation, which may not be suitable for computing and classifying interference information collected by unmanned aircraft.
Currently, there is a lack of research on unmanned aircraft communication interference detection, and most articles focus on interference countermeasures research.Machine learning algorithms have several advantages over traditional programming methods.They can learn from data and improve over time, making them suitable for applications with complex and dynamic patterns.They can handle large amounts of data and extract meaningful insights from it, which can be used for decision-making and predictive modeling.They are also flexible and can be used across different domains, from image recognition and natural language processing to fraud detection and financial forecasting.Additionally, machine learning algorithms can automate tasks and reduce the need for human intervention, leading to increased efficiency and productivity.
The purpose of this paper is to demonstrate that it is feasible to use the SSA-optimized BP neural network algorithm to detect interference signals received by the receiver module of UAVs.

2.1.Sparrow Search Algorithm (SSA)
In the SSA, sparrow flocks imitate the behavior of bird flocks, dividing the search space into multiple regions and continuously moving and searching between these regions.Each sparrow gradually adjusts its direction and speed of movement by communicating with the positions and speeds of the surrounding sparrows in order to find the global optimal solution.
During foraging, sparrows form populations with different roles, including discoverers, joiners, and scouts.The discoverer's task is to search for and provide directions to food, while joiners follow the discoverers to obtain food.When a scout detects a predator, it immediately issues an alarm signal, and the entire sparrow population adopts anti-predator behavior to protect itself.
The initialization of the sparrow population position and fitness, maximum of iterations, population size, discoverers size, number of scouts, initial values for safety value parameter, and the population of n species are described: In (1), d represents the number of parameters to be optimized.Sorting the population and finding the best fitness, it should be noted that for the first generation of sparrows, the initial best value is obtained.The best individual will have priority in obtaining food.The fitness values of all sparrows are shown below: In ( 2), f is the fitness value, and F x is the fitness function: In (3), n is the total number of training data, and y and y are the true value and network output value of the i-th data.
Updating the position of the explorers, the explorers with better fitness values will have a better chance of finding food.The updated formula for the explorer position in each iteration is: In (4), X , represents the position information of the i-th sparrow in the j-th dimension, t represents the current iteration count, N is the maximum iteration count, R is the warning value, ST is the safety value, and Q is a random number following a normal distribution.L is the unit row vector, and  is a random number between 0 and 1.
Updating the follower's position, the followers constantly observe the situation of the discoverer and compete with the discoverer: In ( 5), X represents the position of the sparrow with the lowest fitness, X is the best location before and A is a random row vector consisting of only two elements, 1 and -1.
Updating the discoverer's position: In ( 6), X represents the current global best position, X is the opposite.β is a random number, K ∈ 1,1 , and f represents the fitness value.The subscript "best" indicates the optimal value.ε is a constant.

2.2.SSA-BP algorithm model
BP model can be described as a type of neural network model based on the back-propagation algorithm.It updates the parameters in the network by calculating the back-propagation of error signals in the network, thereby gradually improving the prediction accuracy of the model.Training a BP neural network can easily result in convergence to a local minimum, causing considerable deviations between the network output and actual values.To prevent this scenario, the most crucial stage of the learning process is updating the network's parameters.
To avoid getting stuck in local optima, the weights and thresholds of the BP module.This is a crucial step in the learning process.One way to achieve global optimization is to use the SSA, which is known for its ability to search for the global optimum.In this study, the BP module was optimized using the SSA to obtain the optimal network weights and thresholds.
SSA-BP Algorithm Flow: (1) Determine the structure and parameters of each layer of the BP module; (2) Initialize SSA parameters including sparrow population size, population space dimension, discovery rate, and safety value; (3) Calculate the fitness value and sort them; (4) Use the Equations ( 4)-( 6) to update the sparrow position and get new fitness value; (5) If the condition is not met, return to step (4); (6) When the error reaches the desired value or iterations are reached, the optimization is completed and the optimal value is assigned to the network weights and thresholds of the BP module.

3.1.Experimental Settings
This experiment uses the drone interference datasets from Michaelevol on GitHub.The dataset contains 23565 signal samples, among which 10,071 are non-interference signals, while 3392, 3367, 3378, and 3357 are respectively interference signals of jamming, single-tone, continuous pulse, and protocol-aware and set to 1, 2, 3, and 4, respectively.Non-interference is set to 0. The features of the signals used in this experiment include nine features.

3.2.Experimental Model Setup
In interference signal detection, the normalized training set is input into the SSA-BP model for iteration, and the topology structure of the BP module is set to 5-8-7-1.The four-layer network can better handle complex data.
After initializing the BP module, the SSA is initialized with a discovery rate of 0.6 and a scouting rate of 0.3, and a maximum iteration number of 15.Since the topology structure of the BP module is set to 5-7-8-1, it means that we need to optimize 178 parameters.In (1), d=178 and n=40, and the security value is set to 0.2.In (2), the range of weights and thresholds is (-1, 1).
The evaluation metrics for the model include accuracy (A), precision (P), recall (R), and F-score (F):

3.3Experimental Result
By continuously optimizing BP using the SSA algorithm, the fitness value decreased continuously with the increase of iterations in the optimization process, as shown in Figure 1.This means that as the model was trained, it gradually acquired more precise parameters, which led to better performance on the testing set.After training, we used the test set to verify the optimized BP neural network model, which achieved an accuracy of 93.64% and an F-score of 0.93.The results suggest that the model has a high level of accuracy in detecting UAV interference and can be effectively utilized in real-world scenarios.Finally, we trained and tested some traditional machine learning algorithms using this dataset, and the results showed that the SSA-BP model was better.The accuracy comparison chart is shown in Figure 2. In summary, the experimental outcomes illustrate that the developed UAV interference detection model has significant practical utility and promising applications.

4.CONCLUSIONS
This paper proposes a method for detecting drone interference using SSA-optimized BP neural networks, which can identify four types of interference signals.The model was validated using accuracy and F-score, with a resulting accuracy of 93.64% and an F-score of 0.93, confirming the effectiveness of the method.In the future, we will explore other types of interference signals, expand the applicability of the interference detection algorithm, and investigate specific solutions for detecting drone interference.