More Realistic Audio-Based Drone Detection and Identification Approaches With Machine Learning

With advanced mobility and widespread accessibility, drones have attracted considerable attention, offering a wide range of applications in both military and civilian domains. Nonetheless, beyond their beneficial applications and potential, drones have given rise to serious security and privacy concerns as they are increasingly exploited for malicious activities such as drug smuggling or pose threats to individuals and public facilities. The emergence of security threats by drones has stimulated the development of countermeasure solutions including drone detection and identification. These solutions enable timely responses against security threats by facilitating the detection of drone presence and verifying their identities. In this paper, we propose more realistic audio-based drone detection and identification approaches using machine learning, which account for the presence of background environment noises, an often overlooked factor in current methods. By incorporating the interdependence between the drone’s audio and the surrounding environmental sounds, we simulate real-world scenarios where drones operate. We use drone datasets that capture these interactions to evaluate detection and identification performance. Mel-frequency cepstral coefficients (MFCC) features and support vector machine (SVM) classifiers with various kernels are employed to investigate the effectiveness of our approach across different environments. Our results demonstrate that our approach ensures reliable drone detection, achieving an accuracy of approximately 0.99 and an F1-score of 0.98 in closed-set experiments, and an accuracy and F1-score of about 0.94 in open-set experiments. For drone identification, we show the feasibility of our approach with an accuracy and F1-score of about 0.96 for only known drone classes and approximately 0.90 for both known and unknown drone classes.