A Survey on Image Data Perturbations: Challenges, Techniques, and Impacts
Peng-Fei Zhang, Guangdong Bai, Xin-Shun Xu, Zi Helen Huang · 2025
Data representation is fundamental to data-intensive applications, where deep learning models play a vital role. However, real-world environments, characterized by inherent randomness, uncertainty, and potential adversarial threats, raise significant concerns about the reliability of deep models in understanding and capturing intrinsic data patterns. Evaluating the robustness of deep learning models is therefore critical for model improvement. In this survey, we specifically review various data perturbations and assess model robustness in response to them. Data perturbation refers to both inadvertent and deliberate modifications of data that occur during the development and deployment of deep learning models. Understanding how these perturbations affect data interpretation across different models, along with the underlying causes, is essential for characterizing deep learning behaviors and building systems that are not only accurate but also resilient to data variability. Among various data modalities, this survey focuses on providing a comprehensive review of the most widely engaged image perturbations in deep learning. It presents a summary and taxonomy of existing perturbations, their generation methods, applications, interrelationships, implications for model robustness, and underlying mechanisms. By systematically analyzing recent advancements and best practices in both digital and physical domains, our goal is to equip researchers with an in-depth understanding necessary for evaluating and enhancing the robustness of deep models against image data perturbations, ultimately supporting downstream data engineering tasks. Finally, we discuss future research directions.