status

Completed

year

2026

type

Implementation-Based Technical Paper

Abstract

This implementation-based technical paper presents a deep learning drone detection system for counter-UAV cybersecurity using the YOLOv8 object detection framework. The study begins from the security risks created by the rapid growth of unmanned aerial vehicles, where unauthorized drones may be used for illegal surveillance, smuggling, data interception, or cyber-physical attacks near airports, military zones, critical infrastructure, and government facilities. In this context, detection is treated as the first layer of counter-UAV defense, because a security system cannot classify, track, respond to, or neutralize an aerial threat before detecting its presence. The paper focuses on a vision-based detection approach, which can work with camera input and provide interpretable visual evidence through bounding boxes and class labels. The implemented system uses a custom annotated image dataset containing approximately 1,360 UAV images and two detection classes: multirotor drones and fixed-wing drones. Images were collected and curated to include variation in drone size, background, viewing angle, and illumination, then annotated in YOLO format using normalized bounding box coordinates. The model is built with YOLOv8n, the lightweight nano variant of YOLOv8, selected for its speed and practical suitability for real-time or near-real-time detection. Training was performed using the Ultralytics YOLOv8 framework for 51 epochs, with 640×640 input resolution, batch size of 32, mixed precision support, automatic device selection, and checkpoint-based model selection. The system is not limited to training results only; it also supports deployment through static image detection, recorded video processing, live camera stream detection, and a Streamlit web interface that allows users to upload images and receive annotated detection outputs with class labels and per-class counts. The experimental evaluation uses precision, recall, mAP@0.5, and mAP@0.5:0.95. Results show that the trained model achieved 89.4% precision, indicating that most predicted UAV detections were correct and that the model can reduce false alarms in security-sensitive contexts. However, the recall value of 17.3% shows that many UAV instances were missed, especially small or distant drones, which remains a major challenge in visual UAV detection. The model achieved 30.0% mAP@0.5 and 23.5% mAP@0.5:0.95, while training losses decreased consistently, suggesting stable learning behavior. The paper concludes that YOLOv8-based deep learning can provide a practical foundation for drone detection in counter-UAV cybersecurity, but further work is needed to improve detection sensitivity through larger and more diverse datasets, stronger augmentation, larger model variants, small-object detection techniques, preprocessing support, and real-world video evaluation.