Adversarial Attacks on Aerial Imagery : The State-of-the-Art and Perspective

In recent years, deep model's feature learning skills have become more compelling, resulting in huge advancements in various artificial intelligence (AI) applications. Specifically, depth and breadth of Computer Vision (CV) have expanded rapidly considering the usage of Deep Neural Networks (DNNs). However, it has been shown in the literature that DNNs are vulnerable to adversarial attacks caused by carefully crafted perturbations through solving complex optimization problems. Although the attacks reveal weaknesses in sophisticated DNN algorithms, they might be seen as an opportunity to address issues in real-world security-critical applications. These attacks represent a paradigm change for circumstances in which vulnerable assets must be concealed from autonomous detection systems onboard drones, Unmanned Aerial Vehicles (UAVs), and satellites. Flying AI-models with strong remote detection and classification capabilities may relay exact target-object kinds on the ground, compromising victim security. The employment of conventional tactics to hide huge stationary and movable assets from autonomous aerial detection has become ineffective for larger areas owing to its cost and applicability. Previous works have explained the broader perspective of adversarial attacks in both digital and physical domains. This is the first effort to characterize the multiplicity of adversarial attacks from the viewpoint of autonomous aerial imaging. In addition to providing a thorough literature review of adversarial attacks on aerial imagery in CV tasks, this paper also offers non-specialists succinct descriptions of technical terms and prospects associated with this direction of study.