Research On The Attack Method Of UAV Visual Sensor

With the maturity and development of drone technology,drones are widely used in military and civilian fields.However,drones are also very vulnerable to malicious attacks,leading to serious consequences.The vision system of a drone plays an important role in obstacle avoidance,tracking,positioning,etc.,and is an important guarantee for drone safety,but few researchers have conducted research on its safety.Therefore,this paper designs two attack methods on the hardware and software of drone vision sensors from the perspective of the attacker.The specific work of this paper is as follows:1)A blinding attack method is proposed for the CMOS of the drone vision sensors.This method uses laser as the attack light source,and sets different wavelengths,light intensity,distance,angles and environmental brightness indoors to study the influence of different variables on the drone's vision sensors.The experiment measures the effect of the attack by calculating the similarity of the pictures before and after the attack,and normalizes the similarity results,and finally builds the attack model.In order to verify the effectiveness of the attack,this paper also conducted an attack experiment in an outdoor environment.The results show that the attack method is feasible and can achieve the blinding effect,leading to the failure of obstacle avoidance,target recognition and tracking functions.2)An attack method based on adversarial examples is proposed for the recognition model of the drone vision sensors.This method takes YOLOv2 as the target model,and generates a "patch" in a certain area of the car to deceive YOLOv2's detection of the car.In order to make the generated adversarial examples effective in the real environment,this method considers the non-printability of the image when designing the loss function,and randomly changes the brightness and contrast of the "patch" during each training to enhance the robustness.Experimental results show that the adversarial example can achieve individual attacks and universal attacks in a digital environment,and has a certain transferability to other models.It can also effectively reduce the probability of vehicles being detected in a real environment.