Soft Actor Critic Reinforcement Learning for Decentralized Secured Drone System
Lakshmaiya N, S. Kaliappan, M. Muthukannan, R.S. Priyadharshini, Ramya M
Abstract
Unmanned Aerial Vehicles (UAVs) are gaining more interest in the latest technologies based on their growing features. It is primarily used in emergency response systems, including disaster management, surveillance, and cybersecurity. As the drones operate autonomously, there is a high risk to the security of the data collected by the drones, which can be misused or subject to data spoofing. This occurs when the centralised system is opted for, where the sensed information is transmitted to it. It is also seen that the intrusion detection in the drone system needs to be communicated to the drones to avoid further loss of information. Based on this objective, this paper proposes the Federated Learning approach for the drone system, which uses only the gradients for data transmission. Soft Actor Critic Reinforcement Learning helps in detecting anomalies and determining necessary actions based on the critic's reward update. Thus, the proposed system functions dynamically by detecting intrusion activity and mitigating it by communicating the information to both the onboard and offboard systems of drone control. The proposed work was evaluated using the AIDER dataset from GitHub, which showed 92% trajectory accuracy with 3% collision loss.