Defense of Military Installations from UAV-Borne Attacks using Deep Learning

Consumer-grade Unmanned Aerial Vehicles (UAVs) are becoming more common capabilities on the modern battlefield, finding use by both formal standing armies and non-state sponsored organizations with small budgets. The threats posed by these UAVs are varied, ranging from intelligence gathered from reconnaissance, spotting for indirect fires, or attacking with a payload on the UAV itself. These dangers create new challenges that militaries must adapt to ensure soldiers are protected and mission completion is possible despite the threat of UAV interdiction. In this paper, we propose an AI object detection model that is capable of identifying UAVs in the visible spectrums and distinguishing them from images that contain no UAVs. This model can be used as a targeting system for anti-UAV countermeasures. The dataset used to train the model consists of 2,000 images. 1000 images are of UAVs and 1000 contain no UAVs. The models we tested were ResNet18, ResNet50, and GoogleNet. GoogleNet achieved the best results, yielding a precision of 0.995, recall of 0.995, F1-score of 0.995, and test accuracy of 98.44%. These results and the initial dataset present a good base for researchers to explore and design a practical solution for defense against small drone attacks.