Data-Centric Challenges, Techniques, and Impacts: A Survey on Image Data Perturbation
Data-driven deep learning models have revolutionized the ability to understand and model intrinsic data patterns. However, in real-world applications, their reliability is often compromised by inherent randomness, uncertainty, and potential adversarial threats, particularly those originating from data. As a critical component of data engineering, data perturbation has emerged as an effective approach for evaluating and enhancing the robustness of deep learning models, encompassing both inadvertent and deliberate modifications to data. This survey offers a comprehensive review of data perturbations, with a particular focus on the extensively studied image perturbations on classification throughout the development and deployment of deep models. It presents a detailed summary and taxonomy of existing perturbations, their generation methods, interrelationships, implications for model robustness, and explores underlying mechanisms. By systematically analyzing recent advances and best practices across both digital and physical domains, our goal is to provide an in-depth understanding of how data engineering, which is particularly through data perturbation, can be leveraged to develop models that are both accurate and resilient. Additionally, we discuss current limitations in current research and suggest promising directions for future study.