Lightweight ML-Based Drone Intrusion Detection System Through Model Compression

Abstract

The adoption of drones in diverse domains (e.g., surveillance, agriculture, and disaster management), together with their integration of advanced technologies and dependence on wireless communication, has significantly increased the need to secure drone networks against cyber threats. Traditional network-based intrusion detection systems (NIDS) can be insufficient against novel or adaptive cyber threats and exceed the computational limits of drones. Thus, we need lightweight and efficient drone-specific NIDS solutions. This dissertation addresses this concern with the goal of achieving an effective balance between security, efficiency, and model accuracy without significantly compromising detection performance. Hence, two complementary main contributions are proposed: First, a lightweight ML-based NIDS optimized for individual drones, utilizing a quantized deep neural network (DNN) through post-training quantization (PTQ), enabling real-time, on-board intrusion detection. Second, a framework for swarm-based deployments that leverage federated learning and knowledge distillation to enable distributed training and lightweight model deployment while preserving data privacy and minimizing communication overhead. Both contributions were evaluated using real-world drone network datasets. The first contribution achieved 95.03% accuracy with significantly reduced model size and inference latency, making it suit- able for real-time and onboard deployment. The second contribution was deployed using Raspberry Pi 4 devices and demonstrated improved accuracy, convergence, and communication efficiency, achieving up to 76% reduction in communication overhead and 29% lower CPU usage. The results demonstrate the practicality and effectiveness of the proposed solutions in meeting the unique demands of both individual and swarm-based drone deployments, while achieving a robust balance between security and efficiency.