Along with the dominant and security critical applications of Artificial Intelligence (AI), existence of its perils cannot be undermined. However, such threats become a blessing in disguise in scenarios where vulnerable assets are required to be hidden from autonomous systems including drones and UAVs. The hazards of Adversarial attacks on deep neural network (DNN) models are well known by research community. The intent of these attacks is to cause the DNN model to make inaccurate predictions or judgements. The aforementioned attacks signify a fundamental change in security-sensitive situations within practical contexts, wherein it becomes imperative to conceal vulnerable assets from self-governing detection mechanisms present on drones and unmanned aerial vehicles (UAVs). This paper presents a unique research in utilizing adversarial attacks to deceive autonomous detection systems onboard aerial reconnaissance, detection and surveillance systems. We provide a framework to deceive YOLOv3 object detection model through patch based attack by camouflaging a parked aircraft from overhead imaging. Our study successfully demonstrated patches when overlaid on minimal area of aircraft, was able to blind detection system with much greater accuracy. This work also aims to highlight the use of such methods in security related applications.
Deceiving Airborne Object Detectors Using Adversarial AI
International Conference on Aeronautical Sciences, Engineering and Technology ; 2023 ; Muscat, Oman October 03, 2023 - October 05, 2023
2023-12-26
8 pages
Article/Chapter (Book)
Electronic Resource
English
Springer Verlag | 2016
|CANCloak: Deceiving Two ECUs with One Frame
TIBKAT | 2021
|Deceiving faces: When plastic surgery challenges face recognition
British Library Online Contents | 2016
|GUIDED MISSILES CAN BE DECEIVING MIST DEVICE OF WATER JET
European Patent Office | 2016
|Appearance can be deceiving: using appearance models in color imaging [6494-15]
British Library Conference Proceedings | 2007
|