DEEP CONVOLUTIONAL NEURAL NETWORKS FOR EARLY DETECTION OF DRONE ATTACKS IN PUBLIC EVENTS

Abstract

   This paper presents a deep learning-based framework using Convolutional Neural Networks (CNNs) for the early detection of drone attacks in public events, where unauthorized UAV incursions pose significant threats to safety and security. We compare CNN-based models with traditional detection methods such as radar, RF analyzers, and acoustic sensing, highlighting the limitations of these conventional approaches in noisy, crowded environments. The proposed CNN, optimized with skip connections, leaky ReLU activation, and anchor box tuning, demonstrates superior performance in drone recognition tasks. Experimental results show that the model achieves a validation accuracy of 94.58% within 10 epochs, outperforming ResNet-50 (93.75%), ResNet-18 (90.83%), and Darknet-53 (92.22%), while requiring only 5 million parameters compared to 62 million in Darknet-53. On the test dataset, the proposed CNN achieved a detection accuracy of 77%, significantly higher than the 54% of standard YOLOv3. Furthermore, the lightweight architecture enables real-time inference at over 25 frames per second on GPU hardware, making it feasible for live deployment in event monitoring systems. These findings underscore the potential of CNN-based detection as a scalable and efficient solution for safeguarding public gatherings against malicious drone threats.